aboutsummaryrefslogtreecommitdiff
path: root/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll')
-rw-r--r--llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll10634
1 files changed, 4910 insertions, 5724 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
index 93c11f1..4372f11 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
@@ -3848,30 +3848,58 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; SI: ; %bb.0:
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT: s_mov_b64 exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_writelane_b32 v18, s30, 0
-; SI-NEXT: v_writelane_b32 v18, s31, 1
-; SI-NEXT: v_writelane_b32 v18, s34, 2
-; SI-NEXT: v_writelane_b32 v18, s35, 3
-; SI-NEXT: v_writelane_b32 v18, s36, 4
-; SI-NEXT: v_writelane_b32 v18, s37, 5
-; SI-NEXT: v_writelane_b32 v18, s38, 6
-; SI-NEXT: v_writelane_b32 v18, s39, 7
-; SI-NEXT: v_writelane_b32 v18, s48, 8
-; SI-NEXT: v_writelane_b32 v18, s49, 9
-; SI-NEXT: v_writelane_b32 v18, s50, 10
-; SI-NEXT: v_writelane_b32 v18, s51, 11
-; SI-NEXT: v_writelane_b32 v18, s52, 12
-; SI-NEXT: v_writelane_b32 v18, s53, 13
-; SI-NEXT: v_writelane_b32 v18, s54, 14
+; SI-NEXT: v_writelane_b32 v20, s30, 0
+; SI-NEXT: v_writelane_b32 v20, s31, 1
+; SI-NEXT: v_writelane_b32 v20, s34, 2
+; SI-NEXT: v_writelane_b32 v20, s35, 3
+; SI-NEXT: v_writelane_b32 v20, s36, 4
+; SI-NEXT: v_writelane_b32 v20, s37, 5
+; SI-NEXT: v_writelane_b32 v20, s38, 6
+; SI-NEXT: v_writelane_b32 v20, s39, 7
+; SI-NEXT: v_writelane_b32 v20, s48, 8
+; SI-NEXT: v_mov_b32_e32 v18, s16
+; SI-NEXT: v_mov_b32_e32 v19, s17
+; SI-NEXT: v_writelane_b32 v20, s49, 9
+; SI-NEXT: v_readfirstlane_b32 s46, v18
+; SI-NEXT: v_mov_b32_e32 v18, s18
+; SI-NEXT: v_readfirstlane_b32 s47, v19
+; SI-NEXT: v_mov_b32_e32 v19, s19
+; SI-NEXT: v_writelane_b32 v20, s50, 10
+; SI-NEXT: v_readfirstlane_b32 s44, v18
+; SI-NEXT: v_mov_b32_e32 v18, s20
+; SI-NEXT: v_readfirstlane_b32 s45, v19
+; SI-NEXT: v_mov_b32_e32 v19, s21
+; SI-NEXT: v_writelane_b32 v20, s51, 11
+; SI-NEXT: v_readfirstlane_b32 s42, v18
+; SI-NEXT: v_mov_b32_e32 v18, s22
+; SI-NEXT: v_readfirstlane_b32 s43, v19
+; SI-NEXT: v_mov_b32_e32 v19, s23
+; SI-NEXT: v_writelane_b32 v20, s52, 12
+; SI-NEXT: v_readfirstlane_b32 s40, v18
+; SI-NEXT: v_mov_b32_e32 v18, s24
+; SI-NEXT: v_readfirstlane_b32 s41, v19
+; SI-NEXT: v_mov_b32_e32 v19, s25
+; SI-NEXT: v_writelane_b32 v20, s53, 13
+; SI-NEXT: v_readfirstlane_b32 s24, v18
+; SI-NEXT: v_mov_b32_e32 v18, s26
+; SI-NEXT: v_readfirstlane_b32 s25, v19
+; SI-NEXT: v_mov_b32_e32 v19, s27
+; SI-NEXT: v_writelane_b32 v20, s54, 14
+; SI-NEXT: v_readfirstlane_b32 s22, v18
+; SI-NEXT: v_mov_b32_e32 v18, s28
+; SI-NEXT: v_readfirstlane_b32 s23, v19
+; SI-NEXT: v_mov_b32_e32 v19, s29
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; SI-NEXT: v_writelane_b32 v18, s55, 15
-; SI-NEXT: v_readfirstlane_b32 s42, v1
-; SI-NEXT: v_readfirstlane_b32 s43, v2
-; SI-NEXT: v_readfirstlane_b32 s40, v3
-; SI-NEXT: v_readfirstlane_b32 s41, v4
+; SI-NEXT: v_writelane_b32 v20, s55, 15
+; SI-NEXT: v_readfirstlane_b32 s20, v18
+; SI-NEXT: v_readfirstlane_b32 s21, v19
+; SI-NEXT: v_readfirstlane_b32 s18, v1
+; SI-NEXT: v_readfirstlane_b32 s19, v2
+; SI-NEXT: v_readfirstlane_b32 s16, v3
+; SI-NEXT: v_readfirstlane_b32 s17, v4
; SI-NEXT: v_readfirstlane_b32 s14, v5
; SI-NEXT: v_readfirstlane_b32 s15, v6
; SI-NEXT: v_readfirstlane_b32 s12, v7
@@ -3883,9 +3911,9 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; SI-NEXT: v_readfirstlane_b32 s6, v13
; SI-NEXT: v_readfirstlane_b32 s7, v14
; SI-NEXT: v_readfirstlane_b32 s4, v15
-; SI-NEXT: s_and_b64 s[44:45], vcc, exec
+; SI-NEXT: s_and_b64 s[26:27], vcc, exec
; SI-NEXT: v_readfirstlane_b32 s5, v16
-; SI-NEXT: v_writelane_b32 v18, s64, 16
+; SI-NEXT: v_writelane_b32 v20, s64, 16
; SI-NEXT: s_cbranch_scc0 .LBB13_4
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: s_lshr_b32 s34, s5, 16
@@ -3894,50 +3922,50 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; SI-NEXT: s_lshr_b32 s37, s11, 16
; SI-NEXT: s_lshr_b32 s38, s13, 16
; SI-NEXT: s_lshr_b32 s39, s15, 16
-; SI-NEXT: s_lshr_b32 s48, s41, 16
-; SI-NEXT: s_lshr_b32 s49, s43, 16
-; SI-NEXT: s_lshr_b32 s50, s29, 16
-; SI-NEXT: s_lshr_b32 s51, s27, 16
+; SI-NEXT: s_lshr_b32 s48, s17, 16
+; SI-NEXT: s_lshr_b32 s49, s19, 16
+; SI-NEXT: s_lshr_b32 s50, s21, 16
+; SI-NEXT: s_lshr_b32 s51, s23, 16
; SI-NEXT: s_lshr_b32 s52, s25, 16
-; SI-NEXT: s_lshr_b32 s53, s23, 16
-; SI-NEXT: s_lshr_b32 s54, s21, 16
-; SI-NEXT: s_lshr_b32 s55, s19, 16
-; SI-NEXT: s_lshr_b32 s64, s17, 16
-; SI-NEXT: s_lshr_b64 s[44:45], s[4:5], 16
-; SI-NEXT: s_lshr_b64 s[46:47], s[6:7], 16
+; SI-NEXT: s_lshr_b32 s53, s41, 16
+; SI-NEXT: s_lshr_b32 s54, s43, 16
+; SI-NEXT: s_lshr_b32 s55, s45, 16
+; SI-NEXT: s_lshr_b32 s64, s47, 16
+; SI-NEXT: s_lshr_b64 s[26:27], s[4:5], 16
+; SI-NEXT: s_lshr_b64 s[28:29], s[6:7], 16
; SI-NEXT: s_lshr_b64 s[56:57], s[8:9], 16
; SI-NEXT: s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT: s_lshr_b64 s[60:61], s[12:13], 16
; SI-NEXT: s_lshr_b64 s[62:63], s[14:15], 16
-; SI-NEXT: s_lshr_b64 s[72:73], s[40:41], 16
-; SI-NEXT: s_lshr_b64 s[74:75], s[42:43], 16
-; SI-NEXT: s_lshr_b64 s[76:77], s[28:29], 16
-; SI-NEXT: s_lshr_b64 s[78:79], s[26:27], 16
+; SI-NEXT: s_lshr_b64 s[72:73], s[16:17], 16
+; SI-NEXT: s_lshr_b64 s[74:75], s[18:19], 16
+; SI-NEXT: s_lshr_b64 s[76:77], s[20:21], 16
+; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
; SI-NEXT: s_lshr_b64 s[88:89], s[24:25], 16
-; SI-NEXT: s_lshr_b64 s[90:91], s[22:23], 16
-; SI-NEXT: s_lshr_b64 s[92:93], s[20:21], 16
-; SI-NEXT: s_lshr_b64 s[94:95], s[18:19], 16
-; SI-NEXT: s_lshr_b64 s[30:31], s[16:17], 16
+; SI-NEXT: s_lshr_b64 s[90:91], s[40:41], 16
+; SI-NEXT: s_lshr_b64 s[92:93], s[42:43], 16
+; SI-NEXT: s_lshr_b64 s[94:95], s[44:45], 16
+; SI-NEXT: s_lshr_b64 s[30:31], s[46:47], 16
; SI-NEXT: s_cbranch_execnz .LBB13_3
; SI-NEXT: .LBB13_2: ; %cmp.true
-; SI-NEXT: s_add_i32 s17, s17, 3
-; SI-NEXT: s_add_i32 s16, s16, 3
-; SI-NEXT: s_add_i32 s19, s19, 3
-; SI-NEXT: s_add_i32 s18, s18, 3
-; SI-NEXT: s_add_i32 s21, s21, 3
-; SI-NEXT: s_add_i32 s20, s20, 3
-; SI-NEXT: s_add_i32 s23, s23, 3
-; SI-NEXT: s_add_i32 s22, s22, 3
-; SI-NEXT: s_add_i32 s25, s25, 3
-; SI-NEXT: s_add_i32 s24, s24, 3
-; SI-NEXT: s_add_i32 s27, s27, 3
-; SI-NEXT: s_add_i32 s26, s26, 3
-; SI-NEXT: s_add_i32 s29, s29, 3
-; SI-NEXT: s_add_i32 s28, s28, 3
+; SI-NEXT: s_add_i32 s47, s47, 3
+; SI-NEXT: s_add_i32 s46, s46, 3
+; SI-NEXT: s_add_i32 s45, s45, 3
+; SI-NEXT: s_add_i32 s44, s44, 3
; SI-NEXT: s_add_i32 s43, s43, 3
; SI-NEXT: s_add_i32 s42, s42, 3
; SI-NEXT: s_add_i32 s41, s41, 3
; SI-NEXT: s_add_i32 s40, s40, 3
+; SI-NEXT: s_add_i32 s25, s25, 3
+; SI-NEXT: s_add_i32 s24, s24, 3
+; SI-NEXT: s_add_i32 s23, s23, 3
+; SI-NEXT: s_add_i32 s22, s22, 3
+; SI-NEXT: s_add_i32 s21, s21, 3
+; SI-NEXT: s_add_i32 s20, s20, 3
+; SI-NEXT: s_add_i32 s19, s19, 3
+; SI-NEXT: s_add_i32 s18, s18, 3
+; SI-NEXT: s_add_i32 s17, s17, 3
+; SI-NEXT: s_add_i32 s16, s16, 3
; SI-NEXT: s_add_i32 s15, s15, 3
; SI-NEXT: s_add_i32 s14, s14, 3
; SI-NEXT: s_add_i32 s13, s13, 3
@@ -3950,149 +3978,149 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; SI-NEXT: s_add_i32 s6, s6, 3
; SI-NEXT: s_add_i32 s5, s5, 3
; SI-NEXT: s_add_i32 s4, s4, 3
-; SI-NEXT: s_lshr_b64 s[44:45], s[4:5], 16
-; SI-NEXT: s_lshr_b64 s[46:47], s[6:7], 16
+; SI-NEXT: s_lshr_b64 s[26:27], s[4:5], 16
+; SI-NEXT: s_lshr_b64 s[28:29], s[6:7], 16
; SI-NEXT: s_lshr_b64 s[56:57], s[8:9], 16
; SI-NEXT: s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT: s_lshr_b64 s[60:61], s[12:13], 16
; SI-NEXT: s_lshr_b64 s[62:63], s[14:15], 16
-; SI-NEXT: s_lshr_b64 s[72:73], s[40:41], 16
-; SI-NEXT: s_lshr_b64 s[74:75], s[42:43], 16
+; SI-NEXT: s_lshr_b64 s[72:73], s[16:17], 16
+; SI-NEXT: s_lshr_b64 s[74:75], s[18:19], 16
; SI-NEXT: s_lshr_b32 s34, s5, 16
; SI-NEXT: s_lshr_b32 s35, s7, 16
; SI-NEXT: s_lshr_b32 s36, s9, 16
; SI-NEXT: s_lshr_b32 s37, s11, 16
; SI-NEXT: s_lshr_b32 s38, s13, 16
; SI-NEXT: s_lshr_b32 s39, s15, 16
-; SI-NEXT: s_lshr_b32 s48, s41, 16
-; SI-NEXT: s_lshr_b32 s49, s43, 16
-; SI-NEXT: s_lshr_b32 s50, s29, 16
-; SI-NEXT: s_lshr_b32 s51, s27, 16
+; SI-NEXT: s_lshr_b32 s48, s17, 16
+; SI-NEXT: s_lshr_b32 s49, s19, 16
+; SI-NEXT: s_lshr_b32 s50, s21, 16
+; SI-NEXT: s_lshr_b32 s51, s23, 16
; SI-NEXT: s_lshr_b32 s52, s25, 16
-; SI-NEXT: s_lshr_b32 s53, s23, 16
-; SI-NEXT: s_lshr_b32 s54, s21, 16
-; SI-NEXT: s_lshr_b32 s55, s19, 16
-; SI-NEXT: s_lshr_b32 s64, s17, 16
-; SI-NEXT: s_lshr_b64 s[76:77], s[28:29], 16
-; SI-NEXT: s_lshr_b64 s[78:79], s[26:27], 16
+; SI-NEXT: s_lshr_b32 s53, s41, 16
+; SI-NEXT: s_lshr_b32 s54, s43, 16
+; SI-NEXT: s_lshr_b32 s55, s45, 16
+; SI-NEXT: s_lshr_b32 s64, s47, 16
+; SI-NEXT: s_lshr_b64 s[76:77], s[20:21], 16
+; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
; SI-NEXT: s_lshr_b64 s[88:89], s[24:25], 16
-; SI-NEXT: s_lshr_b64 s[90:91], s[22:23], 16
-; SI-NEXT: s_lshr_b64 s[92:93], s[20:21], 16
-; SI-NEXT: s_lshr_b64 s[94:95], s[18:19], 16
-; SI-NEXT: s_lshr_b64 s[30:31], s[16:17], 16
+; SI-NEXT: s_lshr_b64 s[90:91], s[40:41], 16
+; SI-NEXT: s_lshr_b64 s[92:93], s[42:43], 16
+; SI-NEXT: s_lshr_b64 s[94:95], s[44:45], 16
+; SI-NEXT: s_lshr_b64 s[30:31], s[46:47], 16
; SI-NEXT: .LBB13_3: ; %end
-; SI-NEXT: s_lshl_b32 s45, s30, 16
-; SI-NEXT: s_and_b32 s16, s16, 0xffff
-; SI-NEXT: s_or_b32 s16, s16, s45
-; SI-NEXT: v_mov_b32_e32 v1, s16
-; SI-NEXT: s_and_b32 s16, s17, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s64, 16
-; SI-NEXT: s_or_b32 s16, s16, s17
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_lshl_b32 s16, s94, 16
-; SI-NEXT: s_and_b32 s17, s18, 0xffff
-; SI-NEXT: s_or_b32 s16, s17, s16
-; SI-NEXT: v_mov_b32_e32 v3, s16
-; SI-NEXT: s_and_b32 s16, s19, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s55, 16
-; SI-NEXT: s_or_b32 s16, s16, s17
-; SI-NEXT: v_mov_b32_e32 v4, s16
-; SI-NEXT: s_lshl_b32 s16, s92, 16
-; SI-NEXT: s_and_b32 s17, s20, 0xffff
-; SI-NEXT: s_or_b32 s16, s17, s16
-; SI-NEXT: v_mov_b32_e32 v5, s16
-; SI-NEXT: s_and_b32 s16, s21, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s54, 16
+; SI-NEXT: s_lshl_b32 s27, s30, 16
+; SI-NEXT: s_and_b32 s29, s46, 0xffff
+; SI-NEXT: s_or_b32 s27, s29, s27
+; SI-NEXT: v_mov_b32_e32 v1, s27
+; SI-NEXT: s_and_b32 s27, s47, 0xffff
+; SI-NEXT: s_lshl_b32 s29, s64, 16
+; SI-NEXT: s_or_b32 s27, s27, s29
+; SI-NEXT: v_mov_b32_e32 v2, s27
+; SI-NEXT: s_lshl_b32 s27, s94, 16
+; SI-NEXT: s_and_b32 s29, s44, 0xffff
+; SI-NEXT: s_or_b32 s27, s29, s27
+; SI-NEXT: v_mov_b32_e32 v3, s27
+; SI-NEXT: s_and_b32 s27, s45, 0xffff
+; SI-NEXT: s_lshl_b32 s29, s55, 16
+; SI-NEXT: s_or_b32 s27, s27, s29
+; SI-NEXT: v_mov_b32_e32 v4, s27
+; SI-NEXT: s_lshl_b32 s27, s92, 16
+; SI-NEXT: s_and_b32 s29, s42, 0xffff
+; SI-NEXT: s_or_b32 s27, s29, s27
+; SI-NEXT: v_mov_b32_e32 v5, s27
+; SI-NEXT: s_and_b32 s27, s43, 0xffff
+; SI-NEXT: s_lshl_b32 s29, s54, 16
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s27, s27, s29
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; SI-NEXT: v_mov_b32_e32 v6, s16
+; SI-NEXT: v_mov_b32_e32 v6, s27
; SI-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 12, v0
-; SI-NEXT: s_lshl_b32 s16, s90, 16
-; SI-NEXT: s_and_b32 s17, s22, 0xffff
+; SI-NEXT: s_lshl_b32 s27, s90, 16
+; SI-NEXT: s_and_b32 s29, s40, 0xffff
; SI-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 16, v0
-; SI-NEXT: s_or_b32 s16, s17, s16
+; SI-NEXT: s_or_b32 s27, s29, s27
; SI-NEXT: buffer_store_dword v5, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 20, v0
; SI-NEXT: s_waitcnt expcnt(3)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s23, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s53, 16
+; SI-NEXT: v_mov_b32_e32 v2, s27
+; SI-NEXT: s_and_b32 s27, s41, 0xffff
+; SI-NEXT: s_lshl_b32 s29, s53, 16
; SI-NEXT: buffer_store_dword v6, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 24, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s27, s27, s29
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s24, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s88, 16
+; SI-NEXT: v_mov_b32_e32 v2, s27
+; SI-NEXT: s_and_b32 s24, s24, 0xffff
+; SI-NEXT: s_lshl_b32 s27, s88, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 28, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s24, s24, s27
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s25, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s52, 16
+; SI-NEXT: v_mov_b32_e32 v2, s24
+; SI-NEXT: s_and_b32 s24, s25, 0xffff
+; SI-NEXT: s_lshl_b32 s25, s52, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 32, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s24, s24, s25
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s26, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s78, 16
+; SI-NEXT: v_mov_b32_e32 v2, s24
+; SI-NEXT: s_and_b32 s22, s22, 0xffff
+; SI-NEXT: s_lshl_b32 s24, s78, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 36, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s22, s22, s24
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s27, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s51, 16
+; SI-NEXT: v_mov_b32_e32 v2, s22
+; SI-NEXT: s_and_b32 s22, s23, 0xffff
+; SI-NEXT: s_lshl_b32 s23, s51, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 40, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s22, s22, s23
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s28, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s76, 16
+; SI-NEXT: v_mov_b32_e32 v2, s22
+; SI-NEXT: s_and_b32 s20, s20, 0xffff
+; SI-NEXT: s_lshl_b32 s22, s76, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 44, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s20, s20, s22
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s29, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s50, 16
+; SI-NEXT: v_mov_b32_e32 v2, s20
+; SI-NEXT: s_and_b32 s20, s21, 0xffff
+; SI-NEXT: s_lshl_b32 s21, s50, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 48, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s20, s20, s21
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s42, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s74, 16
+; SI-NEXT: v_mov_b32_e32 v2, s20
+; SI-NEXT: s_and_b32 s18, s18, 0xffff
+; SI-NEXT: s_lshl_b32 s20, s74, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 52, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s18, s18, s20
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s43, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s49, 16
+; SI-NEXT: v_mov_b32_e32 v2, s18
+; SI-NEXT: s_and_b32 s18, s19, 0xffff
+; SI-NEXT: s_lshl_b32 s19, s49, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 56, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s18, s18, s19
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s40, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s72, 16
+; SI-NEXT: v_mov_b32_e32 v2, s18
+; SI-NEXT: s_and_b32 s16, s16, 0xffff
+; SI-NEXT: s_lshl_b32 s18, s72, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 60, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s16, s16, s18
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s41, 0xffff
+; SI-NEXT: s_and_b32 s16, s17, 0xffff
; SI-NEXT: s_lshl_b32 s17, s48, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 64, v0
; SI-NEXT: s_or_b32 s16, s16, s17
@@ -4156,7 +4184,7 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v2, s8
; SI-NEXT: s_and_b32 s6, s6, 0xffff
-; SI-NEXT: s_lshl_b32 s8, s46, 16
+; SI-NEXT: s_lshl_b32 s8, s28, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x64, v0
; SI-NEXT: s_or_b32 s6, s6, s8
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
@@ -4170,7 +4198,7 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v2, s6
; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_lshl_b32 s6, s44, 16
+; SI-NEXT: s_lshl_b32 s6, s26, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x6c, v0
; SI-NEXT: s_or_b32 s4, s4, s6
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
@@ -4184,25 +4212,25 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x74, v0
; SI-NEXT: v_mov_b32_e32 v1, s4
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT: v_readlane_b32 s64, v18, 16
-; SI-NEXT: v_readlane_b32 s55, v18, 15
-; SI-NEXT: v_readlane_b32 s54, v18, 14
-; SI-NEXT: v_readlane_b32 s53, v18, 13
-; SI-NEXT: v_readlane_b32 s52, v18, 12
-; SI-NEXT: v_readlane_b32 s51, v18, 11
-; SI-NEXT: v_readlane_b32 s50, v18, 10
-; SI-NEXT: v_readlane_b32 s49, v18, 9
-; SI-NEXT: v_readlane_b32 s48, v18, 8
-; SI-NEXT: v_readlane_b32 s39, v18, 7
-; SI-NEXT: v_readlane_b32 s38, v18, 6
-; SI-NEXT: v_readlane_b32 s37, v18, 5
-; SI-NEXT: v_readlane_b32 s36, v18, 4
-; SI-NEXT: v_readlane_b32 s35, v18, 3
-; SI-NEXT: v_readlane_b32 s34, v18, 2
-; SI-NEXT: v_readlane_b32 s31, v18, 1
-; SI-NEXT: v_readlane_b32 s30, v18, 0
+; SI-NEXT: v_readlane_b32 s64, v20, 16
+; SI-NEXT: v_readlane_b32 s55, v20, 15
+; SI-NEXT: v_readlane_b32 s54, v20, 14
+; SI-NEXT: v_readlane_b32 s53, v20, 13
+; SI-NEXT: v_readlane_b32 s52, v20, 12
+; SI-NEXT: v_readlane_b32 s51, v20, 11
+; SI-NEXT: v_readlane_b32 s50, v20, 10
+; SI-NEXT: v_readlane_b32 s49, v20, 9
+; SI-NEXT: v_readlane_b32 s48, v20, 8
+; SI-NEXT: v_readlane_b32 s39, v20, 7
+; SI-NEXT: v_readlane_b32 s38, v20, 6
+; SI-NEXT: v_readlane_b32 s37, v20, 5
+; SI-NEXT: v_readlane_b32 s36, v20, 4
+; SI-NEXT: v_readlane_b32 s35, v20, 3
+; SI-NEXT: v_readlane_b32 s34, v20, 2
+; SI-NEXT: v_readlane_b32 s31, v20, 1
+; SI-NEXT: v_readlane_b32 s30, v20, 0
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT: s_mov_b64 exec, s[4:5]
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
@@ -4233,10 +4261,10 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; SI-NEXT: ; implicit-def: $sgpr37
; SI-NEXT: ; implicit-def: $sgpr56
; SI-NEXT: ; implicit-def: $sgpr36
-; SI-NEXT: ; implicit-def: $sgpr46
+; SI-NEXT: ; implicit-def: $sgpr28
; SI-NEXT: ; implicit-def: $sgpr35
; SI-NEXT: ; implicit-def: $sgpr34
-; SI-NEXT: ; implicit-def: $sgpr44
+; SI-NEXT: ; implicit-def: $sgpr26
; SI-NEXT: s_branch .LBB13_2
;
; VI-LABEL: bitcast_v30i32_to_v60i16_scalar:
@@ -4247,18 +4275,46 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; VI-NEXT: s_mov_b64 exec, s[4:5]
; VI-NEXT: v_writelane_b32 v30, s30, 0
; VI-NEXT: v_writelane_b32 v30, s31, 1
+; VI-NEXT: v_mov_b32_e32 v17, s16
+; VI-NEXT: v_mov_b32_e32 v18, s17
; VI-NEXT: v_writelane_b32 v30, s34, 2
+; VI-NEXT: v_mov_b32_e32 v19, s18
+; VI-NEXT: v_readfirstlane_b32 s56, v17
+; VI-NEXT: v_mov_b32_e32 v17, s19
+; VI-NEXT: v_readfirstlane_b32 s47, v18
+; VI-NEXT: v_mov_b32_e32 v18, s20
; VI-NEXT: v_writelane_b32 v30, s35, 3
+; VI-NEXT: v_readfirstlane_b32 s46, v19
+; VI-NEXT: v_mov_b32_e32 v19, s21
+; VI-NEXT: v_readfirstlane_b32 s45, v17
+; VI-NEXT: v_mov_b32_e32 v17, s22
+; VI-NEXT: v_readfirstlane_b32 s44, v18
+; VI-NEXT: v_mov_b32_e32 v18, s23
; VI-NEXT: v_writelane_b32 v30, s36, 4
+; VI-NEXT: v_readfirstlane_b32 s43, v19
+; VI-NEXT: v_mov_b32_e32 v19, s24
+; VI-NEXT: v_readfirstlane_b32 s42, v17
+; VI-NEXT: v_mov_b32_e32 v17, s25
+; VI-NEXT: v_readfirstlane_b32 s41, v18
+; VI-NEXT: v_mov_b32_e32 v18, s26
; VI-NEXT: v_writelane_b32 v30, s37, 5
+; VI-NEXT: v_readfirstlane_b32 s40, v19
+; VI-NEXT: v_mov_b32_e32 v19, s27
+; VI-NEXT: v_readfirstlane_b32 s26, v17
+; VI-NEXT: v_mov_b32_e32 v17, s28
+; VI-NEXT: v_readfirstlane_b32 s25, v18
+; VI-NEXT: v_mov_b32_e32 v18, s29
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; VI-NEXT: v_writelane_b32 v30, s38, 6
-; VI-NEXT: v_readfirstlane_b32 s45, v0
-; VI-NEXT: v_readfirstlane_b32 s44, v1
-; VI-NEXT: v_readfirstlane_b32 s43, v2
-; VI-NEXT: v_readfirstlane_b32 s42, v3
-; VI-NEXT: v_readfirstlane_b32 s41, v4
-; VI-NEXT: v_readfirstlane_b32 s40, v5
+; VI-NEXT: v_readfirstlane_b32 s24, v19
+; VI-NEXT: v_readfirstlane_b32 s23, v17
+; VI-NEXT: v_readfirstlane_b32 s22, v18
+; VI-NEXT: v_readfirstlane_b32 s21, v0
+; VI-NEXT: v_readfirstlane_b32 s20, v1
+; VI-NEXT: v_readfirstlane_b32 s19, v2
+; VI-NEXT: v_readfirstlane_b32 s18, v3
+; VI-NEXT: v_readfirstlane_b32 s17, v4
+; VI-NEXT: v_readfirstlane_b32 s16, v5
; VI-NEXT: v_readfirstlane_b32 s15, v6
; VI-NEXT: v_readfirstlane_b32 s14, v7
; VI-NEXT: v_readfirstlane_b32 s13, v8
@@ -4273,9 +4329,9 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; VI-NEXT: v_writelane_b32 v30, s39, 7
; VI-NEXT: s_cbranch_scc0 .LBB13_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_lshr_b32 s47, s6, 16
-; VI-NEXT: s_lshr_b32 s56, s8, 16
+; VI-NEXT: s_lshr_b32 s27, s7, 16
+; VI-NEXT: s_lshr_b32 s28, s6, 16
+; VI-NEXT: s_lshr_b32 s29, s8, 16
; VI-NEXT: s_lshr_b32 s57, s9, 16
; VI-NEXT: s_lshr_b32 s58, s10, 16
; VI-NEXT: s_lshr_b32 s59, s11, 16
@@ -4283,26 +4339,26 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; VI-NEXT: s_lshr_b32 s61, s13, 16
; VI-NEXT: s_lshr_b32 s62, s14, 16
; VI-NEXT: s_lshr_b32 s63, s15, 16
-; VI-NEXT: s_lshr_b32 s72, s40, 16
-; VI-NEXT: s_lshr_b32 s73, s41, 16
-; VI-NEXT: s_lshr_b32 s74, s42, 16
-; VI-NEXT: s_lshr_b32 s75, s43, 16
-; VI-NEXT: s_lshr_b32 s76, s44, 16
-; VI-NEXT: s_lshr_b32 s77, s45, 16
-; VI-NEXT: s_lshr_b32 s78, s29, 16
-; VI-NEXT: s_lshr_b32 s79, s28, 16
-; VI-NEXT: s_lshr_b32 s88, s27, 16
-; VI-NEXT: s_lshr_b32 s89, s26, 16
-; VI-NEXT: s_lshr_b32 s90, s25, 16
-; VI-NEXT: s_lshr_b32 s91, s24, 16
-; VI-NEXT: s_lshr_b32 s30, s23, 16
-; VI-NEXT: s_lshr_b32 s31, s22, 16
-; VI-NEXT: s_lshr_b32 s34, s21, 16
-; VI-NEXT: s_lshr_b32 s35, s20, 16
-; VI-NEXT: s_lshr_b32 s36, s19, 16
-; VI-NEXT: s_lshr_b32 s37, s18, 16
-; VI-NEXT: s_lshr_b32 s38, s17, 16
-; VI-NEXT: s_lshr_b32 s39, s16, 16
+; VI-NEXT: s_lshr_b32 s72, s16, 16
+; VI-NEXT: s_lshr_b32 s73, s17, 16
+; VI-NEXT: s_lshr_b32 s74, s18, 16
+; VI-NEXT: s_lshr_b32 s75, s19, 16
+; VI-NEXT: s_lshr_b32 s76, s20, 16
+; VI-NEXT: s_lshr_b32 s77, s21, 16
+; VI-NEXT: s_lshr_b32 s78, s22, 16
+; VI-NEXT: s_lshr_b32 s79, s23, 16
+; VI-NEXT: s_lshr_b32 s88, s24, 16
+; VI-NEXT: s_lshr_b32 s89, s25, 16
+; VI-NEXT: s_lshr_b32 s90, s26, 16
+; VI-NEXT: s_lshr_b32 s91, s40, 16
+; VI-NEXT: s_lshr_b32 s30, s41, 16
+; VI-NEXT: s_lshr_b32 s31, s42, 16
+; VI-NEXT: s_lshr_b32 s34, s43, 16
+; VI-NEXT: s_lshr_b32 s35, s44, 16
+; VI-NEXT: s_lshr_b32 s36, s45, 16
+; VI-NEXT: s_lshr_b32 s37, s46, 16
+; VI-NEXT: s_lshr_b32 s38, s47, 16
+; VI-NEXT: s_lshr_b32 s39, s56, 16
; VI-NEXT: s_cbranch_execnz .LBB13_3
; VI-NEXT: .LBB13_2: ; %cmp.true
; VI-NEXT: s_add_i32 s7, s7, 3
@@ -4315,29 +4371,29 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; VI-NEXT: s_add_i32 s13, s13, 3
; VI-NEXT: s_add_i32 s14, s14, 3
; VI-NEXT: s_add_i32 s15, s15, 3
+; VI-NEXT: s_add_i32 s16, s16, 3
+; VI-NEXT: s_add_i32 s17, s17, 3
+; VI-NEXT: s_add_i32 s18, s18, 3
+; VI-NEXT: s_add_i32 s19, s19, 3
+; VI-NEXT: s_add_i32 s20, s20, 3
+; VI-NEXT: s_add_i32 s21, s21, 3
+; VI-NEXT: s_add_i32 s22, s22, 3
+; VI-NEXT: s_add_i32 s23, s23, 3
+; VI-NEXT: s_add_i32 s24, s24, 3
+; VI-NEXT: s_add_i32 s25, s25, 3
+; VI-NEXT: s_add_i32 s26, s26, 3
; VI-NEXT: s_add_i32 s40, s40, 3
; VI-NEXT: s_add_i32 s41, s41, 3
; VI-NEXT: s_add_i32 s42, s42, 3
; VI-NEXT: s_add_i32 s43, s43, 3
; VI-NEXT: s_add_i32 s44, s44, 3
; VI-NEXT: s_add_i32 s45, s45, 3
-; VI-NEXT: s_add_i32 s29, s29, 3
-; VI-NEXT: s_add_i32 s28, s28, 3
-; VI-NEXT: s_add_i32 s27, s27, 3
-; VI-NEXT: s_add_i32 s26, s26, 3
-; VI-NEXT: s_add_i32 s25, s25, 3
-; VI-NEXT: s_add_i32 s24, s24, 3
-; VI-NEXT: s_add_i32 s23, s23, 3
-; VI-NEXT: s_add_i32 s22, s22, 3
-; VI-NEXT: s_add_i32 s21, s21, 3
-; VI-NEXT: s_add_i32 s20, s20, 3
-; VI-NEXT: s_add_i32 s19, s19, 3
-; VI-NEXT: s_add_i32 s18, s18, 3
-; VI-NEXT: s_add_i32 s17, s17, 3
-; VI-NEXT: s_add_i32 s16, s16, 3
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_lshr_b32 s47, s6, 16
-; VI-NEXT: s_lshr_b32 s56, s8, 16
+; VI-NEXT: s_add_i32 s46, s46, 3
+; VI-NEXT: s_add_i32 s47, s47, 3
+; VI-NEXT: s_add_i32 s56, s56, 3
+; VI-NEXT: s_lshr_b32 s27, s7, 16
+; VI-NEXT: s_lshr_b32 s28, s6, 16
+; VI-NEXT: s_lshr_b32 s29, s8, 16
; VI-NEXT: s_lshr_b32 s57, s9, 16
; VI-NEXT: s_lshr_b32 s58, s10, 16
; VI-NEXT: s_lshr_b32 s59, s11, 16
@@ -4345,137 +4401,137 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; VI-NEXT: s_lshr_b32 s61, s13, 16
; VI-NEXT: s_lshr_b32 s62, s14, 16
; VI-NEXT: s_lshr_b32 s63, s15, 16
-; VI-NEXT: s_lshr_b32 s72, s40, 16
-; VI-NEXT: s_lshr_b32 s73, s41, 16
-; VI-NEXT: s_lshr_b32 s74, s42, 16
-; VI-NEXT: s_lshr_b32 s75, s43, 16
-; VI-NEXT: s_lshr_b32 s76, s44, 16
-; VI-NEXT: s_lshr_b32 s77, s45, 16
-; VI-NEXT: s_lshr_b32 s78, s29, 16
-; VI-NEXT: s_lshr_b32 s79, s28, 16
-; VI-NEXT: s_lshr_b32 s88, s27, 16
-; VI-NEXT: s_lshr_b32 s89, s26, 16
-; VI-NEXT: s_lshr_b32 s90, s25, 16
-; VI-NEXT: s_lshr_b32 s91, s24, 16
-; VI-NEXT: s_lshr_b32 s30, s23, 16
-; VI-NEXT: s_lshr_b32 s31, s22, 16
-; VI-NEXT: s_lshr_b32 s34, s21, 16
-; VI-NEXT: s_lshr_b32 s35, s20, 16
-; VI-NEXT: s_lshr_b32 s36, s19, 16
-; VI-NEXT: s_lshr_b32 s37, s18, 16
-; VI-NEXT: s_lshr_b32 s38, s17, 16
-; VI-NEXT: s_lshr_b32 s39, s16, 16
+; VI-NEXT: s_lshr_b32 s72, s16, 16
+; VI-NEXT: s_lshr_b32 s73, s17, 16
+; VI-NEXT: s_lshr_b32 s74, s18, 16
+; VI-NEXT: s_lshr_b32 s75, s19, 16
+; VI-NEXT: s_lshr_b32 s76, s20, 16
+; VI-NEXT: s_lshr_b32 s77, s21, 16
+; VI-NEXT: s_lshr_b32 s78, s22, 16
+; VI-NEXT: s_lshr_b32 s79, s23, 16
+; VI-NEXT: s_lshr_b32 s88, s24, 16
+; VI-NEXT: s_lshr_b32 s89, s25, 16
+; VI-NEXT: s_lshr_b32 s90, s26, 16
+; VI-NEXT: s_lshr_b32 s91, s40, 16
+; VI-NEXT: s_lshr_b32 s30, s41, 16
+; VI-NEXT: s_lshr_b32 s31, s42, 16
+; VI-NEXT: s_lshr_b32 s34, s43, 16
+; VI-NEXT: s_lshr_b32 s35, s44, 16
+; VI-NEXT: s_lshr_b32 s36, s45, 16
+; VI-NEXT: s_lshr_b32 s37, s46, 16
+; VI-NEXT: s_lshr_b32 s38, s47, 16
+; VI-NEXT: s_lshr_b32 s39, s56, 16
; VI-NEXT: .LBB13_3: ; %end
-; VI-NEXT: s_and_b32 s4, 0xffff, s16
+; VI-NEXT: s_and_b32 s4, 0xffff, s56
; VI-NEXT: s_lshl_b32 s5, s39, 16
; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: s_and_b32 s5, 0xffff, s17
-; VI-NEXT: s_lshl_b32 s16, s38, 16
-; VI-NEXT: s_or_b32 s5, s5, s16
-; VI-NEXT: s_and_b32 s16, 0xffff, s18
-; VI-NEXT: s_lshl_b32 s17, s37, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, 0xffff, s19
-; VI-NEXT: s_lshl_b32 s18, s36, 16
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s18, 0xffff, s20
-; VI-NEXT: s_lshl_b32 s19, s35, 16
-; VI-NEXT: s_or_b32 s18, s18, s19
-; VI-NEXT: s_and_b32 s19, 0xffff, s21
-; VI-NEXT: s_lshl_b32 s20, s34, 16
-; VI-NEXT: s_or_b32 s19, s19, s20
-; VI-NEXT: s_and_b32 s20, 0xffff, s22
-; VI-NEXT: s_lshl_b32 s21, s31, 16
-; VI-NEXT: s_or_b32 s20, s20, s21
-; VI-NEXT: s_and_b32 s21, 0xffff, s23
-; VI-NEXT: s_lshl_b32 s22, s30, 16
-; VI-NEXT: s_or_b32 s21, s21, s22
-; VI-NEXT: s_and_b32 s22, 0xffff, s24
-; VI-NEXT: s_lshl_b32 s23, s91, 16
-; VI-NEXT: s_or_b32 s22, s22, s23
-; VI-NEXT: s_and_b32 s23, 0xffff, s25
-; VI-NEXT: s_lshl_b32 s24, s90, 16
-; VI-NEXT: s_or_b32 s23, s23, s24
-; VI-NEXT: s_and_b32 s24, 0xffff, s26
-; VI-NEXT: s_lshl_b32 s25, s89, 16
-; VI-NEXT: s_or_b32 s24, s24, s25
-; VI-NEXT: s_and_b32 s25, 0xffff, s27
-; VI-NEXT: s_lshl_b32 s26, s88, 16
-; VI-NEXT: s_or_b32 s25, s25, s26
-; VI-NEXT: s_and_b32 s26, 0xffff, s28
-; VI-NEXT: s_lshl_b32 s27, s79, 16
-; VI-NEXT: s_or_b32 s26, s26, s27
-; VI-NEXT: s_and_b32 s27, 0xffff, s29
-; VI-NEXT: s_lshl_b32 s28, s78, 16
-; VI-NEXT: s_or_b32 s27, s27, s28
-; VI-NEXT: s_and_b32 s28, 0xffff, s45
-; VI-NEXT: s_lshl_b32 s29, s77, 16
-; VI-NEXT: s_or_b32 s28, s28, s29
-; VI-NEXT: s_and_b32 s29, 0xffff, s44
-; VI-NEXT: s_lshl_b32 s44, s76, 16
-; VI-NEXT: s_or_b32 s29, s29, s44
+; VI-NEXT: s_and_b32 s5, 0xffff, s47
+; VI-NEXT: s_lshl_b32 s47, s38, 16
+; VI-NEXT: s_or_b32 s5, s5, s47
+; VI-NEXT: s_and_b32 s46, 0xffff, s46
+; VI-NEXT: s_lshl_b32 s47, s37, 16
+; VI-NEXT: s_or_b32 s46, s46, s47
+; VI-NEXT: s_and_b32 s45, 0xffff, s45
+; VI-NEXT: s_lshl_b32 s47, s36, 16
+; VI-NEXT: s_or_b32 s45, s45, s47
+; VI-NEXT: s_and_b32 s44, 0xffff, s44
+; VI-NEXT: s_lshl_b32 s47, s35, 16
+; VI-NEXT: s_or_b32 s44, s44, s47
; VI-NEXT: s_and_b32 s43, 0xffff, s43
-; VI-NEXT: s_lshl_b32 s44, s75, 16
-; VI-NEXT: s_or_b32 s43, s43, s44
+; VI-NEXT: s_lshl_b32 s47, s34, 16
+; VI-NEXT: s_or_b32 s43, s43, s47
; VI-NEXT: s_and_b32 s42, 0xffff, s42
-; VI-NEXT: s_lshl_b32 s44, s74, 16
-; VI-NEXT: s_or_b32 s42, s42, s44
+; VI-NEXT: s_lshl_b32 s47, s31, 16
+; VI-NEXT: s_or_b32 s42, s42, s47
; VI-NEXT: s_and_b32 s41, 0xffff, s41
-; VI-NEXT: s_lshl_b32 s44, s73, 16
-; VI-NEXT: s_or_b32 s41, s41, s44
+; VI-NEXT: s_lshl_b32 s47, s30, 16
+; VI-NEXT: s_or_b32 s41, s41, s47
; VI-NEXT: s_and_b32 s40, 0xffff, s40
-; VI-NEXT: s_lshl_b32 s44, s72, 16
-; VI-NEXT: s_or_b32 s40, s40, s44
+; VI-NEXT: s_lshl_b32 s47, s91, 16
+; VI-NEXT: s_or_b32 s40, s40, s47
+; VI-NEXT: s_and_b32 s26, 0xffff, s26
+; VI-NEXT: s_lshl_b32 s47, s90, 16
+; VI-NEXT: s_or_b32 s26, s26, s47
+; VI-NEXT: s_and_b32 s25, 0xffff, s25
+; VI-NEXT: s_lshl_b32 s47, s89, 16
+; VI-NEXT: s_or_b32 s25, s25, s47
+; VI-NEXT: s_and_b32 s24, 0xffff, s24
+; VI-NEXT: s_lshl_b32 s47, s88, 16
+; VI-NEXT: s_or_b32 s24, s24, s47
+; VI-NEXT: s_and_b32 s23, 0xffff, s23
+; VI-NEXT: s_lshl_b32 s47, s79, 16
+; VI-NEXT: s_or_b32 s23, s23, s47
+; VI-NEXT: s_and_b32 s22, 0xffff, s22
+; VI-NEXT: s_lshl_b32 s47, s78, 16
+; VI-NEXT: s_or_b32 s22, s22, s47
+; VI-NEXT: s_and_b32 s21, 0xffff, s21
+; VI-NEXT: s_lshl_b32 s47, s77, 16
+; VI-NEXT: s_or_b32 s21, s21, s47
+; VI-NEXT: s_and_b32 s20, 0xffff, s20
+; VI-NEXT: s_lshl_b32 s47, s76, 16
+; VI-NEXT: s_or_b32 s20, s20, s47
+; VI-NEXT: s_and_b32 s19, 0xffff, s19
+; VI-NEXT: s_lshl_b32 s47, s75, 16
+; VI-NEXT: s_or_b32 s19, s19, s47
+; VI-NEXT: s_and_b32 s18, 0xffff, s18
+; VI-NEXT: s_lshl_b32 s47, s74, 16
+; VI-NEXT: s_or_b32 s18, s18, s47
+; VI-NEXT: s_and_b32 s17, 0xffff, s17
+; VI-NEXT: s_lshl_b32 s47, s73, 16
+; VI-NEXT: s_or_b32 s17, s17, s47
+; VI-NEXT: s_and_b32 s16, 0xffff, s16
+; VI-NEXT: s_lshl_b32 s47, s72, 16
+; VI-NEXT: s_or_b32 s16, s16, s47
; VI-NEXT: s_and_b32 s15, 0xffff, s15
-; VI-NEXT: s_lshl_b32 s44, s63, 16
-; VI-NEXT: s_or_b32 s15, s15, s44
+; VI-NEXT: s_lshl_b32 s47, s63, 16
+; VI-NEXT: s_or_b32 s15, s15, s47
; VI-NEXT: s_and_b32 s14, 0xffff, s14
-; VI-NEXT: s_lshl_b32 s44, s62, 16
-; VI-NEXT: s_or_b32 s14, s14, s44
+; VI-NEXT: s_lshl_b32 s47, s62, 16
+; VI-NEXT: s_or_b32 s14, s14, s47
; VI-NEXT: s_and_b32 s13, 0xffff, s13
-; VI-NEXT: s_lshl_b32 s44, s61, 16
-; VI-NEXT: s_or_b32 s13, s13, s44
+; VI-NEXT: s_lshl_b32 s47, s61, 16
+; VI-NEXT: s_or_b32 s13, s13, s47
; VI-NEXT: s_and_b32 s12, 0xffff, s12
-; VI-NEXT: s_lshl_b32 s44, s60, 16
-; VI-NEXT: s_or_b32 s12, s12, s44
+; VI-NEXT: s_lshl_b32 s47, s60, 16
+; VI-NEXT: s_or_b32 s12, s12, s47
; VI-NEXT: s_and_b32 s11, 0xffff, s11
-; VI-NEXT: s_lshl_b32 s44, s59, 16
-; VI-NEXT: s_or_b32 s11, s11, s44
+; VI-NEXT: s_lshl_b32 s47, s59, 16
+; VI-NEXT: s_or_b32 s11, s11, s47
; VI-NEXT: s_and_b32 s10, 0xffff, s10
-; VI-NEXT: s_lshl_b32 s44, s58, 16
-; VI-NEXT: s_or_b32 s10, s10, s44
+; VI-NEXT: s_lshl_b32 s47, s58, 16
+; VI-NEXT: s_or_b32 s10, s10, s47
; VI-NEXT: s_and_b32 s9, 0xffff, s9
-; VI-NEXT: s_lshl_b32 s44, s57, 16
-; VI-NEXT: s_or_b32 s9, s9, s44
+; VI-NEXT: s_lshl_b32 s47, s57, 16
; VI-NEXT: s_and_b32 s8, 0xffff, s8
-; VI-NEXT: s_lshl_b32 s44, s56, 16
-; VI-NEXT: s_or_b32 s8, s8, s44
+; VI-NEXT: s_lshl_b32 s29, s29, 16
; VI-NEXT: s_and_b32 s6, 0xffff, s6
-; VI-NEXT: s_lshl_b32 s44, s47, 16
-; VI-NEXT: s_or_b32 s6, s6, s44
+; VI-NEXT: s_lshl_b32 s28, s28, 16
; VI-NEXT: s_and_b32 s7, 0xffff, s7
-; VI-NEXT: s_lshl_b32 s44, s46, 16
-; VI-NEXT: s_or_b32 s7, s7, s44
+; VI-NEXT: s_lshl_b32 s27, s27, 16
+; VI-NEXT: s_or_b32 s9, s9, s47
+; VI-NEXT: s_or_b32 s8, s8, s29
+; VI-NEXT: s_or_b32 s6, s6, s28
+; VI-NEXT: s_or_b32 s7, s7, s27
; VI-NEXT: v_mov_b32_e32 v0, s4
; VI-NEXT: v_mov_b32_e32 v1, s5
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: v_mov_b32_e32 v3, s17
-; VI-NEXT: v_mov_b32_e32 v4, s18
-; VI-NEXT: v_mov_b32_e32 v5, s19
-; VI-NEXT: v_mov_b32_e32 v6, s20
-; VI-NEXT: v_mov_b32_e32 v7, s21
-; VI-NEXT: v_mov_b32_e32 v8, s22
-; VI-NEXT: v_mov_b32_e32 v9, s23
-; VI-NEXT: v_mov_b32_e32 v10, s24
-; VI-NEXT: v_mov_b32_e32 v11, s25
-; VI-NEXT: v_mov_b32_e32 v12, s26
-; VI-NEXT: v_mov_b32_e32 v13, s27
-; VI-NEXT: v_mov_b32_e32 v14, s28
-; VI-NEXT: v_mov_b32_e32 v15, s29
-; VI-NEXT: v_mov_b32_e32 v16, s43
-; VI-NEXT: v_mov_b32_e32 v17, s42
-; VI-NEXT: v_mov_b32_e32 v18, s41
-; VI-NEXT: v_mov_b32_e32 v19, s40
+; VI-NEXT: v_mov_b32_e32 v2, s46
+; VI-NEXT: v_mov_b32_e32 v3, s45
+; VI-NEXT: v_mov_b32_e32 v4, s44
+; VI-NEXT: v_mov_b32_e32 v5, s43
+; VI-NEXT: v_mov_b32_e32 v6, s42
+; VI-NEXT: v_mov_b32_e32 v7, s41
+; VI-NEXT: v_mov_b32_e32 v8, s40
+; VI-NEXT: v_mov_b32_e32 v9, s26
+; VI-NEXT: v_mov_b32_e32 v10, s25
+; VI-NEXT: v_mov_b32_e32 v11, s24
+; VI-NEXT: v_mov_b32_e32 v12, s23
+; VI-NEXT: v_mov_b32_e32 v13, s22
+; VI-NEXT: v_mov_b32_e32 v14, s21
+; VI-NEXT: v_mov_b32_e32 v15, s20
+; VI-NEXT: v_mov_b32_e32 v16, s19
+; VI-NEXT: v_mov_b32_e32 v17, s18
+; VI-NEXT: v_mov_b32_e32 v18, s17
+; VI-NEXT: v_mov_b32_e32 v19, s16
; VI-NEXT: v_mov_b32_e32 v20, s15
; VI-NEXT: v_mov_b32_e32 v21, s14
; VI-NEXT: v_mov_b32_e32 v22, s13
@@ -4527,9 +4583,9 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; VI-NEXT: ; implicit-def: $sgpr59
; VI-NEXT: ; implicit-def: $sgpr58
; VI-NEXT: ; implicit-def: $sgpr57
-; VI-NEXT: ; implicit-def: $sgpr56
-; VI-NEXT: ; implicit-def: $sgpr47
-; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr29
+; VI-NEXT: ; implicit-def: $sgpr28
+; VI-NEXT: ; implicit-def: $sgpr27
; VI-NEXT: s_branch .LBB13_2
;
; GFX9-LABEL: bitcast_v30i32_to_v60i16_scalar:
@@ -4538,20 +4594,48 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_mov_b32_e32 v17, s16
+; GFX9-NEXT: v_mov_b32_e32 v18, s17
+; GFX9-NEXT: v_mov_b32_e32 v19, s18
+; GFX9-NEXT: v_readfirstlane_b32 s6, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s19
+; GFX9-NEXT: v_readfirstlane_b32 s7, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s20
+; GFX9-NEXT: v_readfirstlane_b32 s8, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s21
+; GFX9-NEXT: v_readfirstlane_b32 s9, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s22
+; GFX9-NEXT: v_readfirstlane_b32 s10, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s23
; GFX9-NEXT: v_writelane_b32 v30, s30, 0
+; GFX9-NEXT: v_readfirstlane_b32 s11, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s24
+; GFX9-NEXT: v_readfirstlane_b32 s12, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s25
+; GFX9-NEXT: v_readfirstlane_b32 s13, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s26
; GFX9-NEXT: v_writelane_b32 v30, s31, 1
+; GFX9-NEXT: v_readfirstlane_b32 s14, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s27
+; GFX9-NEXT: v_readfirstlane_b32 s15, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s28
+; GFX9-NEXT: v_readfirstlane_b32 s16, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s29
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; GFX9-NEXT: v_writelane_b32 v30, s34, 2
-; GFX9-NEXT: v_readfirstlane_b32 s6, v0
-; GFX9-NEXT: v_readfirstlane_b32 s7, v1
-; GFX9-NEXT: v_readfirstlane_b32 s8, v2
-; GFX9-NEXT: v_readfirstlane_b32 s9, v3
-; GFX9-NEXT: v_readfirstlane_b32 s10, v4
-; GFX9-NEXT: v_readfirstlane_b32 s11, v5
-; GFX9-NEXT: v_readfirstlane_b32 s12, v6
-; GFX9-NEXT: v_readfirstlane_b32 s13, v7
-; GFX9-NEXT: v_readfirstlane_b32 s14, v8
-; GFX9-NEXT: v_readfirstlane_b32 s15, v9
+; GFX9-NEXT: v_readfirstlane_b32 s17, v19
+; GFX9-NEXT: v_readfirstlane_b32 s18, v17
+; GFX9-NEXT: v_readfirstlane_b32 s19, v18
+; GFX9-NEXT: v_readfirstlane_b32 s20, v0
+; GFX9-NEXT: v_readfirstlane_b32 s21, v1
+; GFX9-NEXT: v_readfirstlane_b32 s22, v2
+; GFX9-NEXT: v_readfirstlane_b32 s23, v3
+; GFX9-NEXT: v_readfirstlane_b32 s24, v4
+; GFX9-NEXT: v_readfirstlane_b32 s25, v5
+; GFX9-NEXT: v_readfirstlane_b32 s26, v6
+; GFX9-NEXT: v_readfirstlane_b32 s27, v7
+; GFX9-NEXT: v_readfirstlane_b32 s28, v8
+; GFX9-NEXT: v_readfirstlane_b32 s29, v9
; GFX9-NEXT: v_readfirstlane_b32 s40, v10
; GFX9-NEXT: v_readfirstlane_b32 s41, v11
; GFX9-NEXT: v_readfirstlane_b32 s42, v12
@@ -4568,30 +4652,30 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX9-NEXT: s_lshr_b32 s57, s42, 16
; GFX9-NEXT: s_lshr_b32 s58, s41, 16
; GFX9-NEXT: s_lshr_b32 s59, s40, 16
-; GFX9-NEXT: s_lshr_b32 s60, s15, 16
-; GFX9-NEXT: s_lshr_b32 s61, s14, 16
-; GFX9-NEXT: s_lshr_b32 s62, s13, 16
-; GFX9-NEXT: s_lshr_b32 s63, s12, 16
-; GFX9-NEXT: s_lshr_b32 s72, s11, 16
-; GFX9-NEXT: s_lshr_b32 s73, s10, 16
-; GFX9-NEXT: s_lshr_b32 s74, s9, 16
-; GFX9-NEXT: s_lshr_b32 s75, s8, 16
-; GFX9-NEXT: s_lshr_b32 s76, s7, 16
-; GFX9-NEXT: s_lshr_b32 s77, s6, 16
-; GFX9-NEXT: s_lshr_b32 s78, s29, 16
-; GFX9-NEXT: s_lshr_b32 s79, s28, 16
-; GFX9-NEXT: s_lshr_b32 s88, s27, 16
-; GFX9-NEXT: s_lshr_b32 s89, s26, 16
-; GFX9-NEXT: s_lshr_b32 s90, s25, 16
-; GFX9-NEXT: s_lshr_b32 s91, s24, 16
-; GFX9-NEXT: s_lshr_b32 s92, s23, 16
-; GFX9-NEXT: s_lshr_b32 s93, s22, 16
-; GFX9-NEXT: s_lshr_b32 s94, s21, 16
-; GFX9-NEXT: s_lshr_b32 s95, s20, 16
-; GFX9-NEXT: s_lshr_b32 s30, s19, 16
-; GFX9-NEXT: s_lshr_b32 s31, s18, 16
-; GFX9-NEXT: s_lshr_b32 s34, s17, 16
-; GFX9-NEXT: s_lshr_b32 s35, s16, 16
+; GFX9-NEXT: s_lshr_b32 s60, s29, 16
+; GFX9-NEXT: s_lshr_b32 s61, s28, 16
+; GFX9-NEXT: s_lshr_b32 s62, s27, 16
+; GFX9-NEXT: s_lshr_b32 s63, s26, 16
+; GFX9-NEXT: s_lshr_b32 s72, s25, 16
+; GFX9-NEXT: s_lshr_b32 s73, s24, 16
+; GFX9-NEXT: s_lshr_b32 s74, s23, 16
+; GFX9-NEXT: s_lshr_b32 s75, s22, 16
+; GFX9-NEXT: s_lshr_b32 s76, s21, 16
+; GFX9-NEXT: s_lshr_b32 s77, s20, 16
+; GFX9-NEXT: s_lshr_b32 s78, s19, 16
+; GFX9-NEXT: s_lshr_b32 s79, s18, 16
+; GFX9-NEXT: s_lshr_b32 s88, s17, 16
+; GFX9-NEXT: s_lshr_b32 s89, s16, 16
+; GFX9-NEXT: s_lshr_b32 s90, s15, 16
+; GFX9-NEXT: s_lshr_b32 s91, s14, 16
+; GFX9-NEXT: s_lshr_b32 s92, s13, 16
+; GFX9-NEXT: s_lshr_b32 s93, s12, 16
+; GFX9-NEXT: s_lshr_b32 s94, s11, 16
+; GFX9-NEXT: s_lshr_b32 s95, s10, 16
+; GFX9-NEXT: s_lshr_b32 s30, s9, 16
+; GFX9-NEXT: s_lshr_b32 s31, s8, 16
+; GFX9-NEXT: s_lshr_b32 s34, s7, 16
+; GFX9-NEXT: s_lshr_b32 s35, s6, 16
; GFX9-NEXT: s_cbranch_execnz .LBB13_3
; GFX9-NEXT: .LBB13_2: ; %cmp.true
; GFX9-NEXT: s_add_i32 s45, s45, 3
@@ -4600,16 +4684,6 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX9-NEXT: s_add_i32 s42, s42, 3
; GFX9-NEXT: s_add_i32 s41, s41, 3
; GFX9-NEXT: s_add_i32 s40, s40, 3
-; GFX9-NEXT: s_add_i32 s15, s15, 3
-; GFX9-NEXT: s_add_i32 s14, s14, 3
-; GFX9-NEXT: s_add_i32 s13, s13, 3
-; GFX9-NEXT: s_add_i32 s12, s12, 3
-; GFX9-NEXT: s_add_i32 s11, s11, 3
-; GFX9-NEXT: s_add_i32 s10, s10, 3
-; GFX9-NEXT: s_add_i32 s9, s9, 3
-; GFX9-NEXT: s_add_i32 s8, s8, 3
-; GFX9-NEXT: s_add_i32 s7, s7, 3
-; GFX9-NEXT: s_add_i32 s6, s6, 3
; GFX9-NEXT: s_add_i32 s29, s29, 3
; GFX9-NEXT: s_add_i32 s28, s28, 3
; GFX9-NEXT: s_add_i32 s27, s27, 3
@@ -4624,61 +4698,71 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX9-NEXT: s_add_i32 s18, s18, 3
; GFX9-NEXT: s_add_i32 s17, s17, 3
; GFX9-NEXT: s_add_i32 s16, s16, 3
+; GFX9-NEXT: s_add_i32 s15, s15, 3
+; GFX9-NEXT: s_add_i32 s14, s14, 3
+; GFX9-NEXT: s_add_i32 s13, s13, 3
+; GFX9-NEXT: s_add_i32 s12, s12, 3
+; GFX9-NEXT: s_add_i32 s11, s11, 3
+; GFX9-NEXT: s_add_i32 s10, s10, 3
+; GFX9-NEXT: s_add_i32 s9, s9, 3
+; GFX9-NEXT: s_add_i32 s8, s8, 3
+; GFX9-NEXT: s_add_i32 s7, s7, 3
+; GFX9-NEXT: s_add_i32 s6, s6, 3
; GFX9-NEXT: s_lshr_b32 s46, s45, 16
; GFX9-NEXT: s_lshr_b32 s47, s44, 16
; GFX9-NEXT: s_lshr_b32 s56, s43, 16
; GFX9-NEXT: s_lshr_b32 s57, s42, 16
; GFX9-NEXT: s_lshr_b32 s58, s41, 16
; GFX9-NEXT: s_lshr_b32 s59, s40, 16
-; GFX9-NEXT: s_lshr_b32 s60, s15, 16
-; GFX9-NEXT: s_lshr_b32 s61, s14, 16
-; GFX9-NEXT: s_lshr_b32 s62, s13, 16
-; GFX9-NEXT: s_lshr_b32 s63, s12, 16
-; GFX9-NEXT: s_lshr_b32 s72, s11, 16
-; GFX9-NEXT: s_lshr_b32 s73, s10, 16
-; GFX9-NEXT: s_lshr_b32 s74, s9, 16
-; GFX9-NEXT: s_lshr_b32 s75, s8, 16
-; GFX9-NEXT: s_lshr_b32 s76, s7, 16
-; GFX9-NEXT: s_lshr_b32 s77, s6, 16
-; GFX9-NEXT: s_lshr_b32 s78, s29, 16
-; GFX9-NEXT: s_lshr_b32 s79, s28, 16
-; GFX9-NEXT: s_lshr_b32 s88, s27, 16
-; GFX9-NEXT: s_lshr_b32 s89, s26, 16
-; GFX9-NEXT: s_lshr_b32 s90, s25, 16
-; GFX9-NEXT: s_lshr_b32 s91, s24, 16
-; GFX9-NEXT: s_lshr_b32 s92, s23, 16
-; GFX9-NEXT: s_lshr_b32 s93, s22, 16
-; GFX9-NEXT: s_lshr_b32 s94, s21, 16
-; GFX9-NEXT: s_lshr_b32 s95, s20, 16
-; GFX9-NEXT: s_lshr_b32 s30, s19, 16
-; GFX9-NEXT: s_lshr_b32 s31, s18, 16
-; GFX9-NEXT: s_lshr_b32 s34, s17, 16
-; GFX9-NEXT: s_lshr_b32 s35, s16, 16
+; GFX9-NEXT: s_lshr_b32 s60, s29, 16
+; GFX9-NEXT: s_lshr_b32 s61, s28, 16
+; GFX9-NEXT: s_lshr_b32 s62, s27, 16
+; GFX9-NEXT: s_lshr_b32 s63, s26, 16
+; GFX9-NEXT: s_lshr_b32 s72, s25, 16
+; GFX9-NEXT: s_lshr_b32 s73, s24, 16
+; GFX9-NEXT: s_lshr_b32 s74, s23, 16
+; GFX9-NEXT: s_lshr_b32 s75, s22, 16
+; GFX9-NEXT: s_lshr_b32 s76, s21, 16
+; GFX9-NEXT: s_lshr_b32 s77, s20, 16
+; GFX9-NEXT: s_lshr_b32 s78, s19, 16
+; GFX9-NEXT: s_lshr_b32 s79, s18, 16
+; GFX9-NEXT: s_lshr_b32 s88, s17, 16
+; GFX9-NEXT: s_lshr_b32 s89, s16, 16
+; GFX9-NEXT: s_lshr_b32 s90, s15, 16
+; GFX9-NEXT: s_lshr_b32 s91, s14, 16
+; GFX9-NEXT: s_lshr_b32 s92, s13, 16
+; GFX9-NEXT: s_lshr_b32 s93, s12, 16
+; GFX9-NEXT: s_lshr_b32 s94, s11, 16
+; GFX9-NEXT: s_lshr_b32 s95, s10, 16
+; GFX9-NEXT: s_lshr_b32 s30, s9, 16
+; GFX9-NEXT: s_lshr_b32 s31, s8, 16
+; GFX9-NEXT: s_lshr_b32 s34, s7, 16
+; GFX9-NEXT: s_lshr_b32 s35, s6, 16
; GFX9-NEXT: .LBB13_3: ; %end
-; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s35
-; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s34
-; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s31
-; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s30
-; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s95
-; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s94
-; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s93
-; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s92
-; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s91
-; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s90
-; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s89
-; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s88
-; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s79
-; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s78
-; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s77
-; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s76
-; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s75
-; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s74
-; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s73
-; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s72
-; GFX9-NEXT: s_pack_ll_b32_b16 s12, s12, s63
-; GFX9-NEXT: s_pack_ll_b32_b16 s13, s13, s62
-; GFX9-NEXT: s_pack_ll_b32_b16 s14, s14, s61
-; GFX9-NEXT: s_pack_ll_b32_b16 s15, s15, s60
+; GFX9-NEXT: s_pack_ll_b32_b16 s4, s6, s35
+; GFX9-NEXT: s_pack_ll_b32_b16 s5, s7, s34
+; GFX9-NEXT: s_pack_ll_b32_b16 s6, s8, s31
+; GFX9-NEXT: s_pack_ll_b32_b16 s7, s9, s30
+; GFX9-NEXT: s_pack_ll_b32_b16 s8, s10, s95
+; GFX9-NEXT: s_pack_ll_b32_b16 s9, s11, s94
+; GFX9-NEXT: s_pack_ll_b32_b16 s10, s12, s93
+; GFX9-NEXT: s_pack_ll_b32_b16 s11, s13, s92
+; GFX9-NEXT: s_pack_ll_b32_b16 s12, s14, s91
+; GFX9-NEXT: s_pack_ll_b32_b16 s13, s15, s90
+; GFX9-NEXT: s_pack_ll_b32_b16 s14, s16, s89
+; GFX9-NEXT: s_pack_ll_b32_b16 s15, s17, s88
+; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s79
+; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s78
+; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s77
+; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s76
+; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s75
+; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s74
+; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s73
+; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s72
+; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s63
+; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s62
+; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s61
+; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s60
; GFX9-NEXT: s_pack_ll_b32_b16 s28, s40, s59
; GFX9-NEXT: s_pack_ll_b32_b16 s29, s41, s58
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s42, s57
@@ -4687,28 +4771,28 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s45, s46
; GFX9-NEXT: v_mov_b32_e32 v0, s4
; GFX9-NEXT: v_mov_b32_e32 v1, s5
-; GFX9-NEXT: v_mov_b32_e32 v2, s16
-; GFX9-NEXT: v_mov_b32_e32 v3, s17
-; GFX9-NEXT: v_mov_b32_e32 v4, s18
-; GFX9-NEXT: v_mov_b32_e32 v5, s19
-; GFX9-NEXT: v_mov_b32_e32 v6, s20
-; GFX9-NEXT: v_mov_b32_e32 v7, s21
-; GFX9-NEXT: v_mov_b32_e32 v8, s22
-; GFX9-NEXT: v_mov_b32_e32 v9, s23
-; GFX9-NEXT: v_mov_b32_e32 v10, s24
-; GFX9-NEXT: v_mov_b32_e32 v11, s25
-; GFX9-NEXT: v_mov_b32_e32 v12, s26
-; GFX9-NEXT: v_mov_b32_e32 v13, s27
-; GFX9-NEXT: v_mov_b32_e32 v14, s6
-; GFX9-NEXT: v_mov_b32_e32 v15, s7
-; GFX9-NEXT: v_mov_b32_e32 v16, s8
-; GFX9-NEXT: v_mov_b32_e32 v17, s9
-; GFX9-NEXT: v_mov_b32_e32 v18, s10
-; GFX9-NEXT: v_mov_b32_e32 v19, s11
-; GFX9-NEXT: v_mov_b32_e32 v20, s12
-; GFX9-NEXT: v_mov_b32_e32 v21, s13
-; GFX9-NEXT: v_mov_b32_e32 v22, s14
-; GFX9-NEXT: v_mov_b32_e32 v23, s15
+; GFX9-NEXT: v_mov_b32_e32 v2, s6
+; GFX9-NEXT: v_mov_b32_e32 v3, s7
+; GFX9-NEXT: v_mov_b32_e32 v4, s8
+; GFX9-NEXT: v_mov_b32_e32 v5, s9
+; GFX9-NEXT: v_mov_b32_e32 v6, s10
+; GFX9-NEXT: v_mov_b32_e32 v7, s11
+; GFX9-NEXT: v_mov_b32_e32 v8, s12
+; GFX9-NEXT: v_mov_b32_e32 v9, s13
+; GFX9-NEXT: v_mov_b32_e32 v10, s14
+; GFX9-NEXT: v_mov_b32_e32 v11, s15
+; GFX9-NEXT: v_mov_b32_e32 v12, s16
+; GFX9-NEXT: v_mov_b32_e32 v13, s17
+; GFX9-NEXT: v_mov_b32_e32 v14, s18
+; GFX9-NEXT: v_mov_b32_e32 v15, s19
+; GFX9-NEXT: v_mov_b32_e32 v16, s20
+; GFX9-NEXT: v_mov_b32_e32 v17, s21
+; GFX9-NEXT: v_mov_b32_e32 v18, s22
+; GFX9-NEXT: v_mov_b32_e32 v19, s23
+; GFX9-NEXT: v_mov_b32_e32 v20, s24
+; GFX9-NEXT: v_mov_b32_e32 v21, s25
+; GFX9-NEXT: v_mov_b32_e32 v22, s26
+; GFX9-NEXT: v_mov_b32_e32 v23, s27
; GFX9-NEXT: v_mov_b32_e32 v24, s28
; GFX9-NEXT: v_mov_b32_e32 v25, s29
; GFX9-NEXT: v_mov_b32_e32 v26, s40
@@ -4760,49 +4844,76 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX11-LABEL: bitcast_v30i32_to_v60i16_scalar:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT: v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
+; GFX11-NEXT: v_dual_mov_b32 v17, s16 :: v_dual_mov_b32 v18, s17
+; GFX11-NEXT: v_dual_mov_b32 v19, s18 :: v_dual_mov_b32 v20, s19
+; GFX11-NEXT: v_dual_mov_b32 v21, s20 :: v_dual_mov_b32 v22, s21
+; GFX11-NEXT: v_dual_mov_b32 v23, s22 :: v_dual_mov_b32 v24, s23
+; GFX11-NEXT: v_dual_mov_b32 v25, s24 :: v_dual_mov_b32 v26, s25
+; GFX11-NEXT: v_dual_mov_b32 v27, s26 :: v_dual_mov_b32 v28, s27
+; GFX11-NEXT: v_dual_mov_b32 v29, s28 :: v_dual_mov_b32 v30, s29
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: v_readfirstlane_b32 s4, v0
-; GFX11-NEXT: v_readfirstlane_b32 s5, v1
-; GFX11-NEXT: v_readfirstlane_b32 s6, v2
-; GFX11-NEXT: v_readfirstlane_b32 s7, v3
-; GFX11-NEXT: v_readfirstlane_b32 s8, v4
-; GFX11-NEXT: v_readfirstlane_b32 s9, v5
-; GFX11-NEXT: v_readfirstlane_b32 s10, v6
-; GFX11-NEXT: v_readfirstlane_b32 s11, v7
-; GFX11-NEXT: v_readfirstlane_b32 s12, v8
-; GFX11-NEXT: v_readfirstlane_b32 s13, v9
-; GFX11-NEXT: v_readfirstlane_b32 s15, v10
-; GFX11-NEXT: v_readfirstlane_b32 s14, v11
+; GFX11-NEXT: v_readfirstlane_b32 s0, v13
+; GFX11-NEXT: v_readfirstlane_b32 s1, v14
+; GFX11-NEXT: v_readfirstlane_b32 s2, v15
+; GFX11-NEXT: v_readfirstlane_b32 s3, v16
+; GFX11-NEXT: v_readfirstlane_b32 s4, v17
+; GFX11-NEXT: v_readfirstlane_b32 s5, v18
+; GFX11-NEXT: v_readfirstlane_b32 s6, v19
+; GFX11-NEXT: v_readfirstlane_b32 s7, v20
+; GFX11-NEXT: v_readfirstlane_b32 s8, v21
+; GFX11-NEXT: v_readfirstlane_b32 s9, v22
+; GFX11-NEXT: v_readfirstlane_b32 s10, v23
+; GFX11-NEXT: v_readfirstlane_b32 s11, v24
+; GFX11-NEXT: v_readfirstlane_b32 s12, v25
+; GFX11-NEXT: v_readfirstlane_b32 s13, v26
+; GFX11-NEXT: v_readfirstlane_b32 s14, v27
+; GFX11-NEXT: v_readfirstlane_b32 s15, v28
+; GFX11-NEXT: v_readfirstlane_b32 s16, v29
+; GFX11-NEXT: v_readfirstlane_b32 s17, v30
+; GFX11-NEXT: v_readfirstlane_b32 s18, v0
+; GFX11-NEXT: v_readfirstlane_b32 s19, v1
+; GFX11-NEXT: v_readfirstlane_b32 s20, v2
+; GFX11-NEXT: v_readfirstlane_b32 s21, v3
+; GFX11-NEXT: v_readfirstlane_b32 s22, v4
+; GFX11-NEXT: v_readfirstlane_b32 s23, v5
+; GFX11-NEXT: v_readfirstlane_b32 s24, v6
+; GFX11-NEXT: v_readfirstlane_b32 s25, v7
+; GFX11-NEXT: v_readfirstlane_b32 s26, v8
+; GFX11-NEXT: v_readfirstlane_b32 s27, v9
+; GFX11-NEXT: v_readfirstlane_b32 s29, v10
+; GFX11-NEXT: v_readfirstlane_b32 s28, v11
; GFX11-NEXT: s_mov_b32 s94, 0
; GFX11-NEXT: s_and_b32 s40, vcc_lo, exec_lo
; GFX11-NEXT: s_cbranch_scc0 .LBB13_4
; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: s_lshr_b32 s40, s14, 16
-; GFX11-NEXT: s_lshr_b32 s41, s15, 16
-; GFX11-NEXT: s_lshr_b32 s42, s13, 16
-; GFX11-NEXT: s_lshr_b32 s43, s12, 16
-; GFX11-NEXT: s_lshr_b32 s44, s11, 16
-; GFX11-NEXT: s_lshr_b32 s45, s10, 16
-; GFX11-NEXT: s_lshr_b32 s46, s9, 16
-; GFX11-NEXT: s_lshr_b32 s47, s8, 16
-; GFX11-NEXT: s_lshr_b32 s56, s7, 16
-; GFX11-NEXT: s_lshr_b32 s57, s6, 16
-; GFX11-NEXT: s_lshr_b32 s58, s5, 16
-; GFX11-NEXT: s_lshr_b32 s59, s4, 16
-; GFX11-NEXT: s_lshr_b32 s60, s29, 16
-; GFX11-NEXT: s_lshr_b32 s61, s28, 16
-; GFX11-NEXT: s_lshr_b32 s62, s27, 16
-; GFX11-NEXT: s_lshr_b32 s63, s26, 16
-; GFX11-NEXT: s_lshr_b32 s72, s25, 16
-; GFX11-NEXT: s_lshr_b32 s73, s24, 16
-; GFX11-NEXT: s_lshr_b32 s74, s23, 16
-; GFX11-NEXT: s_lshr_b32 s75, s22, 16
-; GFX11-NEXT: s_lshr_b32 s76, s21, 16
-; GFX11-NEXT: s_lshr_b32 s77, s20, 16
-; GFX11-NEXT: s_lshr_b32 s78, s19, 16
-; GFX11-NEXT: s_lshr_b32 s79, s18, 16
-; GFX11-NEXT: s_lshr_b32 s88, s17, 16
-; GFX11-NEXT: s_lshr_b32 s89, s16, 16
+; GFX11-NEXT: s_lshr_b32 s40, s28, 16
+; GFX11-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-NEXT: s_lshr_b32 s43, s26, 16
+; GFX11-NEXT: s_lshr_b32 s44, s25, 16
+; GFX11-NEXT: s_lshr_b32 s45, s24, 16
+; GFX11-NEXT: s_lshr_b32 s46, s23, 16
+; GFX11-NEXT: s_lshr_b32 s47, s22, 16
+; GFX11-NEXT: s_lshr_b32 s56, s21, 16
+; GFX11-NEXT: s_lshr_b32 s57, s20, 16
+; GFX11-NEXT: s_lshr_b32 s58, s19, 16
+; GFX11-NEXT: s_lshr_b32 s59, s18, 16
+; GFX11-NEXT: s_lshr_b32 s60, s17, 16
+; GFX11-NEXT: s_lshr_b32 s61, s16, 16
+; GFX11-NEXT: s_lshr_b32 s62, s15, 16
+; GFX11-NEXT: s_lshr_b32 s63, s14, 16
+; GFX11-NEXT: s_lshr_b32 s72, s13, 16
+; GFX11-NEXT: s_lshr_b32 s73, s12, 16
+; GFX11-NEXT: s_lshr_b32 s74, s11, 16
+; GFX11-NEXT: s_lshr_b32 s75, s10, 16
+; GFX11-NEXT: s_lshr_b32 s76, s9, 16
+; GFX11-NEXT: s_lshr_b32 s77, s8, 16
+; GFX11-NEXT: s_lshr_b32 s78, s7, 16
+; GFX11-NEXT: s_lshr_b32 s79, s6, 16
+; GFX11-NEXT: s_lshr_b32 s88, s5, 16
+; GFX11-NEXT: s_lshr_b32 s89, s4, 16
; GFX11-NEXT: s_lshr_b32 s90, s3, 16
; GFX11-NEXT: s_lshr_b32 s91, s2, 16
; GFX11-NEXT: s_lshr_b32 s92, s1, 16
@@ -4810,20 +4921,8 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s94
; GFX11-NEXT: s_cbranch_vccnz .LBB13_3
; GFX11-NEXT: .LBB13_2: ; %cmp.true
-; GFX11-NEXT: s_add_i32 s14, s14, 3
-; GFX11-NEXT: s_add_i32 s15, s15, 3
-; GFX11-NEXT: s_add_i32 s13, s13, 3
-; GFX11-NEXT: s_add_i32 s12, s12, 3
-; GFX11-NEXT: s_add_i32 s11, s11, 3
-; GFX11-NEXT: s_add_i32 s10, s10, 3
-; GFX11-NEXT: s_add_i32 s9, s9, 3
-; GFX11-NEXT: s_add_i32 s8, s8, 3
-; GFX11-NEXT: s_add_i32 s7, s7, 3
-; GFX11-NEXT: s_add_i32 s6, s6, 3
-; GFX11-NEXT: s_add_i32 s5, s5, 3
-; GFX11-NEXT: s_add_i32 s4, s4, 3
-; GFX11-NEXT: s_add_i32 s29, s29, 3
; GFX11-NEXT: s_add_i32 s28, s28, 3
+; GFX11-NEXT: s_add_i32 s29, s29, 3
; GFX11-NEXT: s_add_i32 s27, s27, 3
; GFX11-NEXT: s_add_i32 s26, s26, 3
; GFX11-NEXT: s_add_i32 s25, s25, 3
@@ -4836,36 +4935,48 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX11-NEXT: s_add_i32 s18, s18, 3
; GFX11-NEXT: s_add_i32 s17, s17, 3
; GFX11-NEXT: s_add_i32 s16, s16, 3
+; GFX11-NEXT: s_add_i32 s15, s15, 3
+; GFX11-NEXT: s_add_i32 s14, s14, 3
+; GFX11-NEXT: s_add_i32 s13, s13, 3
+; GFX11-NEXT: s_add_i32 s12, s12, 3
+; GFX11-NEXT: s_add_i32 s11, s11, 3
+; GFX11-NEXT: s_add_i32 s10, s10, 3
+; GFX11-NEXT: s_add_i32 s9, s9, 3
+; GFX11-NEXT: s_add_i32 s8, s8, 3
+; GFX11-NEXT: s_add_i32 s7, s7, 3
+; GFX11-NEXT: s_add_i32 s6, s6, 3
+; GFX11-NEXT: s_add_i32 s5, s5, 3
+; GFX11-NEXT: s_add_i32 s4, s4, 3
; GFX11-NEXT: s_add_i32 s3, s3, 3
; GFX11-NEXT: s_add_i32 s2, s2, 3
; GFX11-NEXT: s_add_i32 s1, s1, 3
; GFX11-NEXT: s_add_i32 s0, s0, 3
-; GFX11-NEXT: s_lshr_b32 s40, s14, 16
-; GFX11-NEXT: s_lshr_b32 s41, s15, 16
-; GFX11-NEXT: s_lshr_b32 s42, s13, 16
-; GFX11-NEXT: s_lshr_b32 s43, s12, 16
-; GFX11-NEXT: s_lshr_b32 s44, s11, 16
-; GFX11-NEXT: s_lshr_b32 s45, s10, 16
-; GFX11-NEXT: s_lshr_b32 s46, s9, 16
-; GFX11-NEXT: s_lshr_b32 s47, s8, 16
-; GFX11-NEXT: s_lshr_b32 s56, s7, 16
-; GFX11-NEXT: s_lshr_b32 s57, s6, 16
-; GFX11-NEXT: s_lshr_b32 s58, s5, 16
-; GFX11-NEXT: s_lshr_b32 s59, s4, 16
-; GFX11-NEXT: s_lshr_b32 s60, s29, 16
-; GFX11-NEXT: s_lshr_b32 s61, s28, 16
-; GFX11-NEXT: s_lshr_b32 s62, s27, 16
-; GFX11-NEXT: s_lshr_b32 s63, s26, 16
-; GFX11-NEXT: s_lshr_b32 s72, s25, 16
-; GFX11-NEXT: s_lshr_b32 s73, s24, 16
-; GFX11-NEXT: s_lshr_b32 s74, s23, 16
-; GFX11-NEXT: s_lshr_b32 s75, s22, 16
-; GFX11-NEXT: s_lshr_b32 s76, s21, 16
-; GFX11-NEXT: s_lshr_b32 s77, s20, 16
-; GFX11-NEXT: s_lshr_b32 s78, s19, 16
-; GFX11-NEXT: s_lshr_b32 s79, s18, 16
-; GFX11-NEXT: s_lshr_b32 s88, s17, 16
-; GFX11-NEXT: s_lshr_b32 s89, s16, 16
+; GFX11-NEXT: s_lshr_b32 s40, s28, 16
+; GFX11-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-NEXT: s_lshr_b32 s43, s26, 16
+; GFX11-NEXT: s_lshr_b32 s44, s25, 16
+; GFX11-NEXT: s_lshr_b32 s45, s24, 16
+; GFX11-NEXT: s_lshr_b32 s46, s23, 16
+; GFX11-NEXT: s_lshr_b32 s47, s22, 16
+; GFX11-NEXT: s_lshr_b32 s56, s21, 16
+; GFX11-NEXT: s_lshr_b32 s57, s20, 16
+; GFX11-NEXT: s_lshr_b32 s58, s19, 16
+; GFX11-NEXT: s_lshr_b32 s59, s18, 16
+; GFX11-NEXT: s_lshr_b32 s60, s17, 16
+; GFX11-NEXT: s_lshr_b32 s61, s16, 16
+; GFX11-NEXT: s_lshr_b32 s62, s15, 16
+; GFX11-NEXT: s_lshr_b32 s63, s14, 16
+; GFX11-NEXT: s_lshr_b32 s72, s13, 16
+; GFX11-NEXT: s_lshr_b32 s73, s12, 16
+; GFX11-NEXT: s_lshr_b32 s74, s11, 16
+; GFX11-NEXT: s_lshr_b32 s75, s10, 16
+; GFX11-NEXT: s_lshr_b32 s76, s9, 16
+; GFX11-NEXT: s_lshr_b32 s77, s8, 16
+; GFX11-NEXT: s_lshr_b32 s78, s7, 16
+; GFX11-NEXT: s_lshr_b32 s79, s6, 16
+; GFX11-NEXT: s_lshr_b32 s88, s5, 16
+; GFX11-NEXT: s_lshr_b32 s89, s4, 16
; GFX11-NEXT: s_lshr_b32 s90, s3, 16
; GFX11-NEXT: s_lshr_b32 s91, s2, 16
; GFX11-NEXT: s_lshr_b32 s92, s1, 16
@@ -4876,47 +4987,47 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s92
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s91
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s59
-; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s58
-; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s57
-; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s56
-; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s47
-; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s46
-; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s45
-; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s44
-; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s43
-; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s42
-; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s41
-; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s40
+; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s89
+; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s88
+; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s79
+; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s78
+; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s77
+; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s76
+; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s75
+; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s74
+; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s73
+; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s72
+; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s63
+; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s62
+; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s61
+; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s60
+; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s59
+; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s58
+; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s57
+; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s56
+; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s47
+; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s46
+; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s45
+; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s44
+; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s43
+; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s41
+; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s40
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-NEXT: v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-NEXT: v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-NEXT: v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-NEXT: v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-NEXT: v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
-; GFX11-NEXT: v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT: v_dual_mov_b32 v28, s29 :: v_dual_mov_b32 v29, s28
; GFX11-NEXT: s_setpc_b64 s[30:31]
; GFX11-NEXT: .LBB13_4:
; GFX11-NEXT: ; implicit-def: $sgpr93
@@ -5032,40 +5143,53 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:4
; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v2
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:112
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:108
+; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(4)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v8
-; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:104
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:100
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v10
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v6
+; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v8
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:96
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:92
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v10
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:84
-; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v14
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
@@ -5096,27 +5220,10 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:56
; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:64
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52
-; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v28
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
@@ -5201,7 +5308,6 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr30
; SI-NEXT: ; kill: killed $vgpr30
; SI-NEXT: ; implicit-def: $vgpr30
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v55
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v61
; SI-NEXT: ; kill: killed $vgpr30
@@ -5346,7 +5452,6 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v55
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v61
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
@@ -5473,7 +5578,7 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; SI-NEXT: v_add_i32_e32 v27, vcc, s6, v27
+; SI-NEXT: v_add_i32_e32 v27, vcc, 0x30000, v27
; SI-NEXT: v_add_i32_e32 v28, vcc, 0x30000, v28
; SI-NEXT: v_add_i32_e32 v29, vcc, 0x30000, v29
; SI-NEXT: .LBB14_4: ; %end
@@ -5494,7 +5599,7 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v60i16_to_v30i32:
@@ -5776,7 +5881,6 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: v_mov_b32_e32 v61, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v28
@@ -5855,6 +5959,7 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v60
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v61
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
@@ -6019,6 +6124,9 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT: s_cbranch_execz .LBB14_4
; GFX9-NEXT: ; %bb.3: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
@@ -6033,9 +6141,6 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v38, v61, s6
; GFX9-NEXT: v_perm_b32 v1, v37, v60, s6
; GFX9-NEXT: v_perm_b32 v2, v62, v59, s6
@@ -6054,6 +6159,10 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: s_waitcnt vmcnt(14)
+; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: s_waitcnt vmcnt(15)
; GFX9-NEXT: v_perm_b32 v9, v9, v44, s6
; GFX9-NEXT: s_waitcnt vmcnt(14)
@@ -6082,10 +6191,6 @@ define <30 x i32> @bitcast_v60i16_to_v30i32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_perm_b32 v21, v21, v48, s6
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_perm_b32 v22, v22, v39, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
@@ -7241,304 +7346,115 @@ define inreg <30 x i32> @bitcast_v60i16_to_v30i32_scalar(<60 x i16> inreg %a, i3
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v156, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v157, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v158, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v159, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v168, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v169, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v170, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v171, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v172, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v173, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v174, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v175, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v184, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v185, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v186, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v187, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v188, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v189, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v190, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v191, s32
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v25, v7
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v191, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v190, v6 :: v_dual_mov_b32 v185, v4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v186, v3 :: v_dual_mov_b32 v187, v2
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v188, v1 :: v_dual_mov_b32 v189, v0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s28, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s19, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s17, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s1, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s0, 16
-; GFX11-TRUE16-NEXT: s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s40, s0, s40
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s41, s1, s41
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s2, s46
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s3, s45
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s16, s44
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s17, s43
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s18, s4
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s19, s5
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s20, s6
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s21, s7
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s22, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s23, s9
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s25, s11
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s26, s12
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s27, s13
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s28, s14
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s29, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB15_4
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s40 :: v_dual_mov_b32 v5, s0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s41 :: v_dual_mov_b32 v9, s1
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s2 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s3 :: v_dual_mov_b32 v35, s5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v44, s6 :: v_dual_mov_b32 v65, s8
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s7 :: v_dual_mov_b32 v77, s9
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v90, s10 :: v_dual_mov_b32 v119, s12
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v104, s11 :: v_dual_mov_b32 v135, s13
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v152, s14
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v170, s15
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB15_3
; GFX11-TRUE16-NEXT: .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s40, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s41, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v189, v189, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v188, v188, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v187, v187, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v186, v186, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v185, v185, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v191, v191, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v190, v190, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, s4, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v35, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v44, s6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v54, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v65, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v77, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v90, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v104, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v119, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v135, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v152, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v170, s15, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: .LBB15_3: ; %end
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v20 :: v_dual_mov_b32 v6, v27
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v35 :: v_dual_mov_b32 v8, v44
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v77 :: v_dual_mov_b32 v12, v90
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v104
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, v135 :: v_dual_mov_b32 v16, v152
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v170 :: v_dual_mov_b32 v18, v189
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v188 :: v_dual_mov_b32 v20, v187
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v186 :: v_dual_mov_b32 v22, v185
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v191 :: v_dual_mov_b32 v24, v190
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v191, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v190, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v189, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v188, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v187, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v186, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v185, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v184, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v175, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v174, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v173, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v172, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v171, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v170, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v169, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v168, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v159, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v158, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v157, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v156, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:124
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:252
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:316
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v14
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v54 :: v_dual_mov_b32 v10, v65
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v119 :: v_dual_mov_b32 v27, v30
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB15_4:
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, v29 :: v_dual_mov_b32 v65, v28
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v66, v30
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v26 :: v_dual_mov_b32 v54, v25
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v53
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v65 :: v_dual_mov_b32 v29, v64
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v30, v66
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168_vgpr169_vgpr170_vgpr171_vgpr172_vgpr173_vgpr174_vgpr175_vgpr176_vgpr177_vgpr178_vgpr179_vgpr180_vgpr181_vgpr182_vgpr183_vgpr184
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_branch .LBB15_2
;
; GFX11-FAKE16-LABEL: bitcast_v60i16_to_v30i32_scalar:
@@ -9056,13 +8972,41 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; SI-LABEL: bitcast_v30i32_to_v60f16_scalar:
; SI: ; %bb.0:
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT: v_mov_b32_e32 v18, s16
+; SI-NEXT: v_mov_b32_e32 v19, s17
+; SI-NEXT: v_readfirstlane_b32 s40, v18
+; SI-NEXT: v_mov_b32_e32 v18, s18
+; SI-NEXT: v_readfirstlane_b32 s41, v19
+; SI-NEXT: v_mov_b32_e32 v19, s19
+; SI-NEXT: v_readfirstlane_b32 s42, v18
+; SI-NEXT: v_mov_b32_e32 v18, s20
+; SI-NEXT: v_readfirstlane_b32 s43, v19
+; SI-NEXT: v_mov_b32_e32 v19, s21
+; SI-NEXT: v_readfirstlane_b32 s44, v18
+; SI-NEXT: v_mov_b32_e32 v18, s22
+; SI-NEXT: v_readfirstlane_b32 s45, v19
+; SI-NEXT: v_mov_b32_e32 v19, s23
+; SI-NEXT: v_readfirstlane_b32 s23, v18
+; SI-NEXT: v_mov_b32_e32 v18, s24
+; SI-NEXT: v_readfirstlane_b32 s24, v19
+; SI-NEXT: v_mov_b32_e32 v19, s25
+; SI-NEXT: v_readfirstlane_b32 s25, v18
+; SI-NEXT: v_mov_b32_e32 v18, s26
+; SI-NEXT: v_readfirstlane_b32 s26, v19
+; SI-NEXT: v_mov_b32_e32 v19, s27
+; SI-NEXT: v_readfirstlane_b32 s27, v18
+; SI-NEXT: v_mov_b32_e32 v18, s28
+; SI-NEXT: v_readfirstlane_b32 s28, v19
+; SI-NEXT: v_mov_b32_e32 v19, s29
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; SI-NEXT: v_readfirstlane_b32 s45, v1
-; SI-NEXT: v_readfirstlane_b32 s44, v2
-; SI-NEXT: v_readfirstlane_b32 s43, v3
-; SI-NEXT: v_readfirstlane_b32 s42, v4
-; SI-NEXT: v_readfirstlane_b32 s41, v5
-; SI-NEXT: v_readfirstlane_b32 s40, v6
+; SI-NEXT: v_readfirstlane_b32 s29, v18
+; SI-NEXT: v_readfirstlane_b32 s22, v19
+; SI-NEXT: v_readfirstlane_b32 s21, v1
+; SI-NEXT: v_readfirstlane_b32 s20, v2
+; SI-NEXT: v_readfirstlane_b32 s19, v3
+; SI-NEXT: v_readfirstlane_b32 s18, v4
+; SI-NEXT: v_readfirstlane_b32 s17, v5
+; SI-NEXT: v_readfirstlane_b32 s16, v6
; SI-NEXT: v_readfirstlane_b32 s15, v7
; SI-NEXT: v_readfirstlane_b32 s14, v8
; SI-NEXT: v_readfirstlane_b32 s13, v9
@@ -9109,48 +9053,48 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v15, s4
; SI-NEXT: s_lshr_b32 s4, s15, 16
; SI-NEXT: v_cvt_f32_f16_e32 v17, s4
-; SI-NEXT: s_lshr_b32 s4, s40, 16
+; SI-NEXT: s_lshr_b32 s4, s16, 16
; SI-NEXT: v_cvt_f32_f16_e32 v19, s4
-; SI-NEXT: s_lshr_b32 s4, s41, 16
+; SI-NEXT: s_lshr_b32 s4, s17, 16
; SI-NEXT: v_cvt_f32_f16_e32 v22, s4
-; SI-NEXT: s_lshr_b32 s4, s42, 16
+; SI-NEXT: s_lshr_b32 s4, s18, 16
; SI-NEXT: v_cvt_f32_f16_e32 v24, s4
-; SI-NEXT: s_lshr_b32 s4, s43, 16
+; SI-NEXT: s_lshr_b32 s4, s19, 16
; SI-NEXT: v_cvt_f32_f16_e32 v26, s4
-; SI-NEXT: s_lshr_b32 s4, s44, 16
+; SI-NEXT: s_lshr_b32 s4, s20, 16
; SI-NEXT: v_cvt_f32_f16_e32 v28, s4
-; SI-NEXT: s_lshr_b32 s4, s45, 16
+; SI-NEXT: s_lshr_b32 s4, s21, 16
; SI-NEXT: v_cvt_f32_f16_e32 v30, s4
-; SI-NEXT: s_lshr_b32 s4, s29, 16
+; SI-NEXT: s_lshr_b32 s4, s22, 16
; SI-NEXT: v_cvt_f32_f16_e32 v32, s4
-; SI-NEXT: s_lshr_b32 s4, s28, 16
+; SI-NEXT: s_lshr_b32 s4, s29, 16
; SI-NEXT: v_cvt_f32_f16_e32 v34, s4
-; SI-NEXT: s_lshr_b32 s4, s27, 16
+; SI-NEXT: s_lshr_b32 s4, s28, 16
; SI-NEXT: v_cvt_f32_f16_e32 v36, s4
-; SI-NEXT: s_lshr_b32 s4, s26, 16
+; SI-NEXT: s_lshr_b32 s4, s27, 16
; SI-NEXT: v_cvt_f32_f16_e32 v39, s4
-; SI-NEXT: s_lshr_b32 s4, s25, 16
+; SI-NEXT: s_lshr_b32 s4, s26, 16
; SI-NEXT: v_cvt_f32_f16_e32 v49, s4
-; SI-NEXT: s_lshr_b32 s4, s24, 16
+; SI-NEXT: s_lshr_b32 s4, s25, 16
; SI-NEXT: v_cvt_f32_f16_e32 v51, s4
-; SI-NEXT: s_lshr_b32 s4, s23, 16
+; SI-NEXT: s_lshr_b32 s4, s24, 16
; SI-NEXT: v_cvt_f32_f16_e32 v53, s4
-; SI-NEXT: s_lshr_b32 s4, s22, 16
+; SI-NEXT: s_lshr_b32 s4, s23, 16
; SI-NEXT: v_cvt_f32_f16_e32 v55, s4
-; SI-NEXT: s_lshr_b32 s4, s21, 16
+; SI-NEXT: s_lshr_b32 s4, s45, 16
; SI-NEXT: v_cvt_f32_f16_e32 v41, s4
-; SI-NEXT: s_lshr_b32 s4, s20, 16
+; SI-NEXT: s_lshr_b32 s4, s44, 16
; SI-NEXT: v_cvt_f32_f16_e32 v43, s4
-; SI-NEXT: s_lshr_b32 s4, s19, 16
+; SI-NEXT: s_lshr_b32 s4, s43, 16
; SI-NEXT: s_waitcnt expcnt(6)
; SI-NEXT: v_cvt_f32_f16_e32 v46, s4
-; SI-NEXT: s_lshr_b32 s4, s18, 16
+; SI-NEXT: s_lshr_b32 s4, s42, 16
; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: v_cvt_f32_f16_e32 v56, s4
-; SI-NEXT: s_lshr_b32 s4, s17, 16
+; SI-NEXT: s_lshr_b32 s4, s41, 16
; SI-NEXT: s_waitcnt expcnt(2)
; SI-NEXT: v_cvt_f32_f16_e32 v58, s4
-; SI-NEXT: s_lshr_b32 s4, s16, 16
+; SI-NEXT: s_lshr_b32 s4, s40, 16
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v60, s4
; SI-NEXT: v_cvt_f32_f16_e32 v4, s9
@@ -9163,35 +9107,34 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v18, s13
; SI-NEXT: v_cvt_f32_f16_e32 v20, s14
; SI-NEXT: v_cvt_f32_f16_e32 v21, s15
-; SI-NEXT: v_cvt_f32_f16_e32 v23, s40
-; SI-NEXT: v_cvt_f32_f16_e32 v25, s41
-; SI-NEXT: v_cvt_f32_f16_e32 v27, s42
-; SI-NEXT: v_cvt_f32_f16_e32 v29, s43
-; SI-NEXT: v_cvt_f32_f16_e32 v31, s44
-; SI-NEXT: v_cvt_f32_f16_e32 v33, s45
-; SI-NEXT: v_cvt_f32_f16_e32 v35, s29
-; SI-NEXT: v_cvt_f32_f16_e32 v37, s28
-; SI-NEXT: v_cvt_f32_f16_e32 v38, s27
-; SI-NEXT: v_cvt_f32_f16_e32 v48, s26
-; SI-NEXT: v_cvt_f32_f16_e32 v50, s25
-; SI-NEXT: v_cvt_f32_f16_e32 v52, s24
-; SI-NEXT: v_cvt_f32_f16_e32 v54, s23
-; SI-NEXT: v_cvt_f32_f16_e32 v40, s22
-; SI-NEXT: v_cvt_f32_f16_e32 v42, s21
-; SI-NEXT: v_cvt_f32_f16_e32 v44, s20
-; SI-NEXT: v_cvt_f32_f16_e32 v45, s19
-; SI-NEXT: v_cvt_f32_f16_e32 v47, s18
-; SI-NEXT: v_cvt_f32_f16_e32 v57, s17
-; SI-NEXT: v_cvt_f32_f16_e32 v59, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v23, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v25, s17
+; SI-NEXT: v_cvt_f32_f16_e32 v27, s18
+; SI-NEXT: v_cvt_f32_f16_e32 v29, s19
+; SI-NEXT: v_cvt_f32_f16_e32 v31, s20
+; SI-NEXT: v_cvt_f32_f16_e32 v33, s21
+; SI-NEXT: v_cvt_f32_f16_e32 v35, s22
+; SI-NEXT: v_cvt_f32_f16_e32 v37, s29
+; SI-NEXT: v_cvt_f32_f16_e32 v38, s28
+; SI-NEXT: v_cvt_f32_f16_e32 v48, s27
+; SI-NEXT: v_cvt_f32_f16_e32 v50, s26
+; SI-NEXT: v_cvt_f32_f16_e32 v52, s25
+; SI-NEXT: v_cvt_f32_f16_e32 v54, s24
+; SI-NEXT: v_cvt_f32_f16_e32 v40, s23
+; SI-NEXT: v_cvt_f32_f16_e32 v42, s45
+; SI-NEXT: v_cvt_f32_f16_e32 v44, s44
+; SI-NEXT: v_cvt_f32_f16_e32 v45, s43
+; SI-NEXT: v_cvt_f32_f16_e32 v47, s42
+; SI-NEXT: v_cvt_f32_f16_e32 v57, s41
+; SI-NEXT: v_cvt_f32_f16_e32 v59, s40
; SI-NEXT: s_cbranch_execnz .LBB17_3
; SI-NEXT: .LBB17_2: ; %cmp.true
-; SI-NEXT: s_add_i32 s16, s16, 3
-; SI-NEXT: s_add_i32 s17, s17, 3
-; SI-NEXT: s_add_i32 s18, s18, 3
-; SI-NEXT: s_add_i32 s19, s19, 3
-; SI-NEXT: s_add_i32 s20, s20, 3
-; SI-NEXT: s_add_i32 s21, s21, 3
-; SI-NEXT: s_add_i32 s22, s22, 3
+; SI-NEXT: s_add_i32 s40, s40, 3
+; SI-NEXT: s_add_i32 s41, s41, 3
+; SI-NEXT: s_add_i32 s42, s42, 3
+; SI-NEXT: s_add_i32 s43, s43, 3
+; SI-NEXT: s_add_i32 s44, s44, 3
+; SI-NEXT: s_add_i32 s45, s45, 3
; SI-NEXT: s_add_i32 s23, s23, 3
; SI-NEXT: s_add_i32 s24, s24, 3
; SI-NEXT: s_add_i32 s25, s25, 3
@@ -9199,12 +9142,13 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; SI-NEXT: s_add_i32 s27, s27, 3
; SI-NEXT: s_add_i32 s28, s28, 3
; SI-NEXT: s_add_i32 s29, s29, 3
-; SI-NEXT: s_add_i32 s45, s45, 3
-; SI-NEXT: s_add_i32 s44, s44, 3
-; SI-NEXT: s_add_i32 s43, s43, 3
-; SI-NEXT: s_add_i32 s42, s42, 3
-; SI-NEXT: s_add_i32 s41, s41, 3
-; SI-NEXT: s_add_i32 s40, s40, 3
+; SI-NEXT: s_add_i32 s22, s22, 3
+; SI-NEXT: s_add_i32 s21, s21, 3
+; SI-NEXT: s_add_i32 s20, s20, 3
+; SI-NEXT: s_add_i32 s19, s19, 3
+; SI-NEXT: s_add_i32 s18, s18, 3
+; SI-NEXT: s_add_i32 s17, s17, 3
+; SI-NEXT: s_add_i32 s16, s16, 3
; SI-NEXT: s_add_i32 s15, s15, 3
; SI-NEXT: s_add_i32 s14, s14, 3
; SI-NEXT: s_add_i32 s13, s13, 3
@@ -9215,26 +9159,26 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; SI-NEXT: s_add_i32 s7, s7, 3
; SI-NEXT: s_add_i32 s6, s6, 3
; SI-NEXT: s_add_i32 s9, s9, 3
-; SI-NEXT: s_lshr_b32 s4, s16, 16
-; SI-NEXT: s_lshr_b32 s5, s17, 16
-; SI-NEXT: s_lshr_b32 s46, s18, 16
-; SI-NEXT: s_lshr_b32 s47, s19, 16
-; SI-NEXT: s_lshr_b32 s56, s20, 16
-; SI-NEXT: s_lshr_b32 s57, s21, 16
-; SI-NEXT: s_lshr_b32 s58, s22, 16
-; SI-NEXT: s_lshr_b32 s59, s23, 16
-; SI-NEXT: s_lshr_b32 s60, s24, 16
-; SI-NEXT: s_lshr_b32 s61, s25, 16
-; SI-NEXT: s_lshr_b32 s62, s26, 16
-; SI-NEXT: s_lshr_b32 s63, s27, 16
-; SI-NEXT: s_lshr_b32 s72, s28, 16
-; SI-NEXT: s_lshr_b32 s73, s29, 16
-; SI-NEXT: s_lshr_b32 s74, s45, 16
-; SI-NEXT: s_lshr_b32 s75, s44, 16
-; SI-NEXT: s_lshr_b32 s76, s43, 16
-; SI-NEXT: s_lshr_b32 s77, s42, 16
-; SI-NEXT: s_lshr_b32 s78, s41, 16
-; SI-NEXT: s_lshr_b32 s79, s40, 16
+; SI-NEXT: s_lshr_b32 s4, s40, 16
+; SI-NEXT: s_lshr_b32 s5, s41, 16
+; SI-NEXT: s_lshr_b32 s46, s42, 16
+; SI-NEXT: s_lshr_b32 s47, s43, 16
+; SI-NEXT: s_lshr_b32 s56, s44, 16
+; SI-NEXT: s_lshr_b32 s57, s45, 16
+; SI-NEXT: s_lshr_b32 s58, s23, 16
+; SI-NEXT: s_lshr_b32 s59, s24, 16
+; SI-NEXT: s_lshr_b32 s60, s25, 16
+; SI-NEXT: s_lshr_b32 s61, s26, 16
+; SI-NEXT: s_lshr_b32 s62, s27, 16
+; SI-NEXT: s_lshr_b32 s63, s28, 16
+; SI-NEXT: s_lshr_b32 s72, s29, 16
+; SI-NEXT: s_lshr_b32 s73, s22, 16
+; SI-NEXT: s_lshr_b32 s74, s21, 16
+; SI-NEXT: s_lshr_b32 s75, s20, 16
+; SI-NEXT: s_lshr_b32 s76, s19, 16
+; SI-NEXT: s_lshr_b32 s77, s18, 16
+; SI-NEXT: s_lshr_b32 s78, s17, 16
+; SI-NEXT: s_lshr_b32 s79, s16, 16
; SI-NEXT: s_lshr_b32 s88, s15, 16
; SI-NEXT: s_lshr_b32 s89, s14, 16
; SI-NEXT: s_lshr_b32 s90, s13, 16
@@ -9255,29 +9199,29 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v18, s13
; SI-NEXT: v_cvt_f32_f16_e32 v20, s14
; SI-NEXT: v_cvt_f32_f16_e32 v21, s15
-; SI-NEXT: v_cvt_f32_f16_e32 v23, s40
-; SI-NEXT: v_cvt_f32_f16_e32 v25, s41
-; SI-NEXT: v_cvt_f32_f16_e32 v27, s42
-; SI-NEXT: v_cvt_f32_f16_e32 v29, s43
-; SI-NEXT: v_cvt_f32_f16_e32 v31, s44
-; SI-NEXT: v_cvt_f32_f16_e32 v33, s45
-; SI-NEXT: v_cvt_f32_f16_e32 v35, s29
-; SI-NEXT: v_cvt_f32_f16_e32 v37, s28
-; SI-NEXT: v_cvt_f32_f16_e32 v38, s27
-; SI-NEXT: v_cvt_f32_f16_e32 v48, s26
-; SI-NEXT: v_cvt_f32_f16_e32 v50, s25
-; SI-NEXT: v_cvt_f32_f16_e32 v52, s24
-; SI-NEXT: v_cvt_f32_f16_e32 v54, s23
-; SI-NEXT: v_cvt_f32_f16_e32 v40, s22
-; SI-NEXT: v_cvt_f32_f16_e32 v42, s21
-; SI-NEXT: v_cvt_f32_f16_e32 v44, s20
-; SI-NEXT: v_cvt_f32_f16_e32 v45, s19
+; SI-NEXT: v_cvt_f32_f16_e32 v23, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v25, s17
+; SI-NEXT: v_cvt_f32_f16_e32 v27, s18
+; SI-NEXT: v_cvt_f32_f16_e32 v29, s19
+; SI-NEXT: v_cvt_f32_f16_e32 v31, s20
+; SI-NEXT: v_cvt_f32_f16_e32 v33, s21
+; SI-NEXT: v_cvt_f32_f16_e32 v35, s22
+; SI-NEXT: v_cvt_f32_f16_e32 v37, s29
+; SI-NEXT: v_cvt_f32_f16_e32 v38, s28
+; SI-NEXT: v_cvt_f32_f16_e32 v48, s27
+; SI-NEXT: v_cvt_f32_f16_e32 v50, s26
+; SI-NEXT: v_cvt_f32_f16_e32 v52, s25
+; SI-NEXT: v_cvt_f32_f16_e32 v54, s24
+; SI-NEXT: v_cvt_f32_f16_e32 v40, s23
+; SI-NEXT: v_cvt_f32_f16_e32 v42, s45
+; SI-NEXT: v_cvt_f32_f16_e32 v44, s44
+; SI-NEXT: v_cvt_f32_f16_e32 v45, s43
; SI-NEXT: s_waitcnt expcnt(5)
-; SI-NEXT: v_cvt_f32_f16_e32 v47, s18
+; SI-NEXT: v_cvt_f32_f16_e32 v47, s42
; SI-NEXT: s_waitcnt expcnt(3)
-; SI-NEXT: v_cvt_f32_f16_e32 v57, s17
+; SI-NEXT: v_cvt_f32_f16_e32 v57, s41
; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_cvt_f32_f16_e32 v59, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v59, s40
; SI-NEXT: v_cvt_f32_f16_e32 v1, vcc_hi
; SI-NEXT: v_cvt_f32_f16_e32 v2, vcc_lo
; SI-NEXT: v_cvt_f32_f16_e32 v3, s95
@@ -9605,18 +9549,46 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; VI-NEXT: s_mov_b64 exec, s[4:5]
; VI-NEXT: v_writelane_b32 v30, s30, 0
; VI-NEXT: v_writelane_b32 v30, s31, 1
+; VI-NEXT: v_mov_b32_e32 v17, s16
+; VI-NEXT: v_mov_b32_e32 v18, s17
; VI-NEXT: v_writelane_b32 v30, s34, 2
+; VI-NEXT: v_mov_b32_e32 v19, s18
+; VI-NEXT: v_readfirstlane_b32 s56, v17
+; VI-NEXT: v_mov_b32_e32 v17, s19
+; VI-NEXT: v_readfirstlane_b32 s47, v18
+; VI-NEXT: v_mov_b32_e32 v18, s20
; VI-NEXT: v_writelane_b32 v30, s35, 3
+; VI-NEXT: v_readfirstlane_b32 s46, v19
+; VI-NEXT: v_mov_b32_e32 v19, s21
+; VI-NEXT: v_readfirstlane_b32 s45, v17
+; VI-NEXT: v_mov_b32_e32 v17, s22
+; VI-NEXT: v_readfirstlane_b32 s44, v18
+; VI-NEXT: v_mov_b32_e32 v18, s23
; VI-NEXT: v_writelane_b32 v30, s36, 4
+; VI-NEXT: v_readfirstlane_b32 s43, v19
+; VI-NEXT: v_mov_b32_e32 v19, s24
+; VI-NEXT: v_readfirstlane_b32 s42, v17
+; VI-NEXT: v_mov_b32_e32 v17, s25
+; VI-NEXT: v_readfirstlane_b32 s41, v18
+; VI-NEXT: v_mov_b32_e32 v18, s26
; VI-NEXT: v_writelane_b32 v30, s37, 5
+; VI-NEXT: v_readfirstlane_b32 s40, v19
+; VI-NEXT: v_mov_b32_e32 v19, s27
+; VI-NEXT: v_readfirstlane_b32 s26, v17
+; VI-NEXT: v_mov_b32_e32 v17, s28
+; VI-NEXT: v_readfirstlane_b32 s25, v18
+; VI-NEXT: v_mov_b32_e32 v18, s29
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; VI-NEXT: v_writelane_b32 v30, s38, 6
-; VI-NEXT: v_readfirstlane_b32 s45, v0
-; VI-NEXT: v_readfirstlane_b32 s44, v1
-; VI-NEXT: v_readfirstlane_b32 s43, v2
-; VI-NEXT: v_readfirstlane_b32 s42, v3
-; VI-NEXT: v_readfirstlane_b32 s41, v4
-; VI-NEXT: v_readfirstlane_b32 s40, v5
+; VI-NEXT: v_readfirstlane_b32 s24, v19
+; VI-NEXT: v_readfirstlane_b32 s23, v17
+; VI-NEXT: v_readfirstlane_b32 s22, v18
+; VI-NEXT: v_readfirstlane_b32 s21, v0
+; VI-NEXT: v_readfirstlane_b32 s20, v1
+; VI-NEXT: v_readfirstlane_b32 s19, v2
+; VI-NEXT: v_readfirstlane_b32 s18, v3
+; VI-NEXT: v_readfirstlane_b32 s17, v4
+; VI-NEXT: v_readfirstlane_b32 s16, v5
; VI-NEXT: v_readfirstlane_b32 s15, v6
; VI-NEXT: v_readfirstlane_b32 s14, v7
; VI-NEXT: v_readfirstlane_b32 s13, v8
@@ -9631,9 +9603,9 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; VI-NEXT: v_writelane_b32 v30, s39, 7
; VI-NEXT: s_cbranch_scc0 .LBB17_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_lshr_b32 s47, s6, 16
-; VI-NEXT: s_lshr_b32 s56, s8, 16
+; VI-NEXT: s_lshr_b32 s27, s7, 16
+; VI-NEXT: s_lshr_b32 s28, s6, 16
+; VI-NEXT: s_lshr_b32 s29, s8, 16
; VI-NEXT: s_lshr_b32 s57, s9, 16
; VI-NEXT: s_lshr_b32 s58, s10, 16
; VI-NEXT: s_lshr_b32 s59, s11, 16
@@ -9641,26 +9613,26 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; VI-NEXT: s_lshr_b32 s61, s13, 16
; VI-NEXT: s_lshr_b32 s62, s14, 16
; VI-NEXT: s_lshr_b32 s63, s15, 16
-; VI-NEXT: s_lshr_b32 s72, s40, 16
-; VI-NEXT: s_lshr_b32 s73, s41, 16
-; VI-NEXT: s_lshr_b32 s74, s42, 16
-; VI-NEXT: s_lshr_b32 s75, s43, 16
-; VI-NEXT: s_lshr_b32 s76, s44, 16
-; VI-NEXT: s_lshr_b32 s77, s45, 16
-; VI-NEXT: s_lshr_b32 s78, s29, 16
-; VI-NEXT: s_lshr_b32 s79, s28, 16
-; VI-NEXT: s_lshr_b32 s88, s27, 16
-; VI-NEXT: s_lshr_b32 s89, s26, 16
-; VI-NEXT: s_lshr_b32 s90, s25, 16
-; VI-NEXT: s_lshr_b32 s91, s24, 16
-; VI-NEXT: s_lshr_b32 s30, s23, 16
-; VI-NEXT: s_lshr_b32 s31, s22, 16
-; VI-NEXT: s_lshr_b32 s34, s21, 16
-; VI-NEXT: s_lshr_b32 s35, s20, 16
-; VI-NEXT: s_lshr_b32 s36, s19, 16
-; VI-NEXT: s_lshr_b32 s37, s18, 16
-; VI-NEXT: s_lshr_b32 s38, s17, 16
-; VI-NEXT: s_lshr_b32 s39, s16, 16
+; VI-NEXT: s_lshr_b32 s72, s16, 16
+; VI-NEXT: s_lshr_b32 s73, s17, 16
+; VI-NEXT: s_lshr_b32 s74, s18, 16
+; VI-NEXT: s_lshr_b32 s75, s19, 16
+; VI-NEXT: s_lshr_b32 s76, s20, 16
+; VI-NEXT: s_lshr_b32 s77, s21, 16
+; VI-NEXT: s_lshr_b32 s78, s22, 16
+; VI-NEXT: s_lshr_b32 s79, s23, 16
+; VI-NEXT: s_lshr_b32 s88, s24, 16
+; VI-NEXT: s_lshr_b32 s89, s25, 16
+; VI-NEXT: s_lshr_b32 s90, s26, 16
+; VI-NEXT: s_lshr_b32 s91, s40, 16
+; VI-NEXT: s_lshr_b32 s30, s41, 16
+; VI-NEXT: s_lshr_b32 s31, s42, 16
+; VI-NEXT: s_lshr_b32 s34, s43, 16
+; VI-NEXT: s_lshr_b32 s35, s44, 16
+; VI-NEXT: s_lshr_b32 s36, s45, 16
+; VI-NEXT: s_lshr_b32 s37, s46, 16
+; VI-NEXT: s_lshr_b32 s38, s47, 16
+; VI-NEXT: s_lshr_b32 s39, s56, 16
; VI-NEXT: s_cbranch_execnz .LBB17_3
; VI-NEXT: .LBB17_2: ; %cmp.true
; VI-NEXT: s_add_i32 s7, s7, 3
@@ -9673,29 +9645,29 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; VI-NEXT: s_add_i32 s13, s13, 3
; VI-NEXT: s_add_i32 s14, s14, 3
; VI-NEXT: s_add_i32 s15, s15, 3
+; VI-NEXT: s_add_i32 s16, s16, 3
+; VI-NEXT: s_add_i32 s17, s17, 3
+; VI-NEXT: s_add_i32 s18, s18, 3
+; VI-NEXT: s_add_i32 s19, s19, 3
+; VI-NEXT: s_add_i32 s20, s20, 3
+; VI-NEXT: s_add_i32 s21, s21, 3
+; VI-NEXT: s_add_i32 s22, s22, 3
+; VI-NEXT: s_add_i32 s23, s23, 3
+; VI-NEXT: s_add_i32 s24, s24, 3
+; VI-NEXT: s_add_i32 s25, s25, 3
+; VI-NEXT: s_add_i32 s26, s26, 3
; VI-NEXT: s_add_i32 s40, s40, 3
; VI-NEXT: s_add_i32 s41, s41, 3
; VI-NEXT: s_add_i32 s42, s42, 3
; VI-NEXT: s_add_i32 s43, s43, 3
; VI-NEXT: s_add_i32 s44, s44, 3
; VI-NEXT: s_add_i32 s45, s45, 3
-; VI-NEXT: s_add_i32 s29, s29, 3
-; VI-NEXT: s_add_i32 s28, s28, 3
-; VI-NEXT: s_add_i32 s27, s27, 3
-; VI-NEXT: s_add_i32 s26, s26, 3
-; VI-NEXT: s_add_i32 s25, s25, 3
-; VI-NEXT: s_add_i32 s24, s24, 3
-; VI-NEXT: s_add_i32 s23, s23, 3
-; VI-NEXT: s_add_i32 s22, s22, 3
-; VI-NEXT: s_add_i32 s21, s21, 3
-; VI-NEXT: s_add_i32 s20, s20, 3
-; VI-NEXT: s_add_i32 s19, s19, 3
-; VI-NEXT: s_add_i32 s18, s18, 3
-; VI-NEXT: s_add_i32 s17, s17, 3
-; VI-NEXT: s_add_i32 s16, s16, 3
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_lshr_b32 s47, s6, 16
-; VI-NEXT: s_lshr_b32 s56, s8, 16
+; VI-NEXT: s_add_i32 s46, s46, 3
+; VI-NEXT: s_add_i32 s47, s47, 3
+; VI-NEXT: s_add_i32 s56, s56, 3
+; VI-NEXT: s_lshr_b32 s27, s7, 16
+; VI-NEXT: s_lshr_b32 s28, s6, 16
+; VI-NEXT: s_lshr_b32 s29, s8, 16
; VI-NEXT: s_lshr_b32 s57, s9, 16
; VI-NEXT: s_lshr_b32 s58, s10, 16
; VI-NEXT: s_lshr_b32 s59, s11, 16
@@ -9703,137 +9675,137 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; VI-NEXT: s_lshr_b32 s61, s13, 16
; VI-NEXT: s_lshr_b32 s62, s14, 16
; VI-NEXT: s_lshr_b32 s63, s15, 16
-; VI-NEXT: s_lshr_b32 s72, s40, 16
-; VI-NEXT: s_lshr_b32 s73, s41, 16
-; VI-NEXT: s_lshr_b32 s74, s42, 16
-; VI-NEXT: s_lshr_b32 s75, s43, 16
-; VI-NEXT: s_lshr_b32 s76, s44, 16
-; VI-NEXT: s_lshr_b32 s77, s45, 16
-; VI-NEXT: s_lshr_b32 s78, s29, 16
-; VI-NEXT: s_lshr_b32 s79, s28, 16
-; VI-NEXT: s_lshr_b32 s88, s27, 16
-; VI-NEXT: s_lshr_b32 s89, s26, 16
-; VI-NEXT: s_lshr_b32 s90, s25, 16
-; VI-NEXT: s_lshr_b32 s91, s24, 16
-; VI-NEXT: s_lshr_b32 s30, s23, 16
-; VI-NEXT: s_lshr_b32 s31, s22, 16
-; VI-NEXT: s_lshr_b32 s34, s21, 16
-; VI-NEXT: s_lshr_b32 s35, s20, 16
-; VI-NEXT: s_lshr_b32 s36, s19, 16
-; VI-NEXT: s_lshr_b32 s37, s18, 16
-; VI-NEXT: s_lshr_b32 s38, s17, 16
-; VI-NEXT: s_lshr_b32 s39, s16, 16
+; VI-NEXT: s_lshr_b32 s72, s16, 16
+; VI-NEXT: s_lshr_b32 s73, s17, 16
+; VI-NEXT: s_lshr_b32 s74, s18, 16
+; VI-NEXT: s_lshr_b32 s75, s19, 16
+; VI-NEXT: s_lshr_b32 s76, s20, 16
+; VI-NEXT: s_lshr_b32 s77, s21, 16
+; VI-NEXT: s_lshr_b32 s78, s22, 16
+; VI-NEXT: s_lshr_b32 s79, s23, 16
+; VI-NEXT: s_lshr_b32 s88, s24, 16
+; VI-NEXT: s_lshr_b32 s89, s25, 16
+; VI-NEXT: s_lshr_b32 s90, s26, 16
+; VI-NEXT: s_lshr_b32 s91, s40, 16
+; VI-NEXT: s_lshr_b32 s30, s41, 16
+; VI-NEXT: s_lshr_b32 s31, s42, 16
+; VI-NEXT: s_lshr_b32 s34, s43, 16
+; VI-NEXT: s_lshr_b32 s35, s44, 16
+; VI-NEXT: s_lshr_b32 s36, s45, 16
+; VI-NEXT: s_lshr_b32 s37, s46, 16
+; VI-NEXT: s_lshr_b32 s38, s47, 16
+; VI-NEXT: s_lshr_b32 s39, s56, 16
; VI-NEXT: .LBB17_3: ; %end
-; VI-NEXT: s_and_b32 s4, 0xffff, s16
+; VI-NEXT: s_and_b32 s4, 0xffff, s56
; VI-NEXT: s_lshl_b32 s5, s39, 16
; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: s_and_b32 s5, 0xffff, s17
-; VI-NEXT: s_lshl_b32 s16, s38, 16
-; VI-NEXT: s_or_b32 s5, s5, s16
-; VI-NEXT: s_and_b32 s16, 0xffff, s18
-; VI-NEXT: s_lshl_b32 s17, s37, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, 0xffff, s19
-; VI-NEXT: s_lshl_b32 s18, s36, 16
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s18, 0xffff, s20
-; VI-NEXT: s_lshl_b32 s19, s35, 16
-; VI-NEXT: s_or_b32 s18, s18, s19
-; VI-NEXT: s_and_b32 s19, 0xffff, s21
-; VI-NEXT: s_lshl_b32 s20, s34, 16
-; VI-NEXT: s_or_b32 s19, s19, s20
-; VI-NEXT: s_and_b32 s20, 0xffff, s22
-; VI-NEXT: s_lshl_b32 s21, s31, 16
-; VI-NEXT: s_or_b32 s20, s20, s21
-; VI-NEXT: s_and_b32 s21, 0xffff, s23
-; VI-NEXT: s_lshl_b32 s22, s30, 16
-; VI-NEXT: s_or_b32 s21, s21, s22
-; VI-NEXT: s_and_b32 s22, 0xffff, s24
-; VI-NEXT: s_lshl_b32 s23, s91, 16
-; VI-NEXT: s_or_b32 s22, s22, s23
-; VI-NEXT: s_and_b32 s23, 0xffff, s25
-; VI-NEXT: s_lshl_b32 s24, s90, 16
-; VI-NEXT: s_or_b32 s23, s23, s24
-; VI-NEXT: s_and_b32 s24, 0xffff, s26
-; VI-NEXT: s_lshl_b32 s25, s89, 16
-; VI-NEXT: s_or_b32 s24, s24, s25
-; VI-NEXT: s_and_b32 s25, 0xffff, s27
-; VI-NEXT: s_lshl_b32 s26, s88, 16
-; VI-NEXT: s_or_b32 s25, s25, s26
-; VI-NEXT: s_and_b32 s26, 0xffff, s28
-; VI-NEXT: s_lshl_b32 s27, s79, 16
-; VI-NEXT: s_or_b32 s26, s26, s27
-; VI-NEXT: s_and_b32 s27, 0xffff, s29
-; VI-NEXT: s_lshl_b32 s28, s78, 16
-; VI-NEXT: s_or_b32 s27, s27, s28
-; VI-NEXT: s_and_b32 s28, 0xffff, s45
-; VI-NEXT: s_lshl_b32 s29, s77, 16
-; VI-NEXT: s_or_b32 s28, s28, s29
-; VI-NEXT: s_and_b32 s29, 0xffff, s44
-; VI-NEXT: s_lshl_b32 s44, s76, 16
-; VI-NEXT: s_or_b32 s29, s29, s44
+; VI-NEXT: s_and_b32 s5, 0xffff, s47
+; VI-NEXT: s_lshl_b32 s47, s38, 16
+; VI-NEXT: s_or_b32 s5, s5, s47
+; VI-NEXT: s_and_b32 s46, 0xffff, s46
+; VI-NEXT: s_lshl_b32 s47, s37, 16
+; VI-NEXT: s_or_b32 s46, s46, s47
+; VI-NEXT: s_and_b32 s45, 0xffff, s45
+; VI-NEXT: s_lshl_b32 s47, s36, 16
+; VI-NEXT: s_or_b32 s45, s45, s47
+; VI-NEXT: s_and_b32 s44, 0xffff, s44
+; VI-NEXT: s_lshl_b32 s47, s35, 16
+; VI-NEXT: s_or_b32 s44, s44, s47
; VI-NEXT: s_and_b32 s43, 0xffff, s43
-; VI-NEXT: s_lshl_b32 s44, s75, 16
-; VI-NEXT: s_or_b32 s43, s43, s44
+; VI-NEXT: s_lshl_b32 s47, s34, 16
+; VI-NEXT: s_or_b32 s43, s43, s47
; VI-NEXT: s_and_b32 s42, 0xffff, s42
-; VI-NEXT: s_lshl_b32 s44, s74, 16
-; VI-NEXT: s_or_b32 s42, s42, s44
+; VI-NEXT: s_lshl_b32 s47, s31, 16
+; VI-NEXT: s_or_b32 s42, s42, s47
; VI-NEXT: s_and_b32 s41, 0xffff, s41
-; VI-NEXT: s_lshl_b32 s44, s73, 16
-; VI-NEXT: s_or_b32 s41, s41, s44
+; VI-NEXT: s_lshl_b32 s47, s30, 16
+; VI-NEXT: s_or_b32 s41, s41, s47
; VI-NEXT: s_and_b32 s40, 0xffff, s40
-; VI-NEXT: s_lshl_b32 s44, s72, 16
-; VI-NEXT: s_or_b32 s40, s40, s44
+; VI-NEXT: s_lshl_b32 s47, s91, 16
+; VI-NEXT: s_or_b32 s40, s40, s47
+; VI-NEXT: s_and_b32 s26, 0xffff, s26
+; VI-NEXT: s_lshl_b32 s47, s90, 16
+; VI-NEXT: s_or_b32 s26, s26, s47
+; VI-NEXT: s_and_b32 s25, 0xffff, s25
+; VI-NEXT: s_lshl_b32 s47, s89, 16
+; VI-NEXT: s_or_b32 s25, s25, s47
+; VI-NEXT: s_and_b32 s24, 0xffff, s24
+; VI-NEXT: s_lshl_b32 s47, s88, 16
+; VI-NEXT: s_or_b32 s24, s24, s47
+; VI-NEXT: s_and_b32 s23, 0xffff, s23
+; VI-NEXT: s_lshl_b32 s47, s79, 16
+; VI-NEXT: s_or_b32 s23, s23, s47
+; VI-NEXT: s_and_b32 s22, 0xffff, s22
+; VI-NEXT: s_lshl_b32 s47, s78, 16
+; VI-NEXT: s_or_b32 s22, s22, s47
+; VI-NEXT: s_and_b32 s21, 0xffff, s21
+; VI-NEXT: s_lshl_b32 s47, s77, 16
+; VI-NEXT: s_or_b32 s21, s21, s47
+; VI-NEXT: s_and_b32 s20, 0xffff, s20
+; VI-NEXT: s_lshl_b32 s47, s76, 16
+; VI-NEXT: s_or_b32 s20, s20, s47
+; VI-NEXT: s_and_b32 s19, 0xffff, s19
+; VI-NEXT: s_lshl_b32 s47, s75, 16
+; VI-NEXT: s_or_b32 s19, s19, s47
+; VI-NEXT: s_and_b32 s18, 0xffff, s18
+; VI-NEXT: s_lshl_b32 s47, s74, 16
+; VI-NEXT: s_or_b32 s18, s18, s47
+; VI-NEXT: s_and_b32 s17, 0xffff, s17
+; VI-NEXT: s_lshl_b32 s47, s73, 16
+; VI-NEXT: s_or_b32 s17, s17, s47
+; VI-NEXT: s_and_b32 s16, 0xffff, s16
+; VI-NEXT: s_lshl_b32 s47, s72, 16
+; VI-NEXT: s_or_b32 s16, s16, s47
; VI-NEXT: s_and_b32 s15, 0xffff, s15
-; VI-NEXT: s_lshl_b32 s44, s63, 16
-; VI-NEXT: s_or_b32 s15, s15, s44
+; VI-NEXT: s_lshl_b32 s47, s63, 16
+; VI-NEXT: s_or_b32 s15, s15, s47
; VI-NEXT: s_and_b32 s14, 0xffff, s14
-; VI-NEXT: s_lshl_b32 s44, s62, 16
-; VI-NEXT: s_or_b32 s14, s14, s44
+; VI-NEXT: s_lshl_b32 s47, s62, 16
+; VI-NEXT: s_or_b32 s14, s14, s47
; VI-NEXT: s_and_b32 s13, 0xffff, s13
-; VI-NEXT: s_lshl_b32 s44, s61, 16
-; VI-NEXT: s_or_b32 s13, s13, s44
+; VI-NEXT: s_lshl_b32 s47, s61, 16
+; VI-NEXT: s_or_b32 s13, s13, s47
; VI-NEXT: s_and_b32 s12, 0xffff, s12
-; VI-NEXT: s_lshl_b32 s44, s60, 16
-; VI-NEXT: s_or_b32 s12, s12, s44
+; VI-NEXT: s_lshl_b32 s47, s60, 16
+; VI-NEXT: s_or_b32 s12, s12, s47
; VI-NEXT: s_and_b32 s11, 0xffff, s11
-; VI-NEXT: s_lshl_b32 s44, s59, 16
-; VI-NEXT: s_or_b32 s11, s11, s44
+; VI-NEXT: s_lshl_b32 s47, s59, 16
+; VI-NEXT: s_or_b32 s11, s11, s47
; VI-NEXT: s_and_b32 s10, 0xffff, s10
-; VI-NEXT: s_lshl_b32 s44, s58, 16
-; VI-NEXT: s_or_b32 s10, s10, s44
+; VI-NEXT: s_lshl_b32 s47, s58, 16
+; VI-NEXT: s_or_b32 s10, s10, s47
; VI-NEXT: s_and_b32 s9, 0xffff, s9
-; VI-NEXT: s_lshl_b32 s44, s57, 16
-; VI-NEXT: s_or_b32 s9, s9, s44
+; VI-NEXT: s_lshl_b32 s47, s57, 16
; VI-NEXT: s_and_b32 s8, 0xffff, s8
-; VI-NEXT: s_lshl_b32 s44, s56, 16
-; VI-NEXT: s_or_b32 s8, s8, s44
+; VI-NEXT: s_lshl_b32 s29, s29, 16
; VI-NEXT: s_and_b32 s6, 0xffff, s6
-; VI-NEXT: s_lshl_b32 s44, s47, 16
-; VI-NEXT: s_or_b32 s6, s6, s44
+; VI-NEXT: s_lshl_b32 s28, s28, 16
; VI-NEXT: s_and_b32 s7, 0xffff, s7
-; VI-NEXT: s_lshl_b32 s44, s46, 16
-; VI-NEXT: s_or_b32 s7, s7, s44
+; VI-NEXT: s_lshl_b32 s27, s27, 16
+; VI-NEXT: s_or_b32 s9, s9, s47
+; VI-NEXT: s_or_b32 s8, s8, s29
+; VI-NEXT: s_or_b32 s6, s6, s28
+; VI-NEXT: s_or_b32 s7, s7, s27
; VI-NEXT: v_mov_b32_e32 v0, s4
; VI-NEXT: v_mov_b32_e32 v1, s5
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: v_mov_b32_e32 v3, s17
-; VI-NEXT: v_mov_b32_e32 v4, s18
-; VI-NEXT: v_mov_b32_e32 v5, s19
-; VI-NEXT: v_mov_b32_e32 v6, s20
-; VI-NEXT: v_mov_b32_e32 v7, s21
-; VI-NEXT: v_mov_b32_e32 v8, s22
-; VI-NEXT: v_mov_b32_e32 v9, s23
-; VI-NEXT: v_mov_b32_e32 v10, s24
-; VI-NEXT: v_mov_b32_e32 v11, s25
-; VI-NEXT: v_mov_b32_e32 v12, s26
-; VI-NEXT: v_mov_b32_e32 v13, s27
-; VI-NEXT: v_mov_b32_e32 v14, s28
-; VI-NEXT: v_mov_b32_e32 v15, s29
-; VI-NEXT: v_mov_b32_e32 v16, s43
-; VI-NEXT: v_mov_b32_e32 v17, s42
-; VI-NEXT: v_mov_b32_e32 v18, s41
-; VI-NEXT: v_mov_b32_e32 v19, s40
+; VI-NEXT: v_mov_b32_e32 v2, s46
+; VI-NEXT: v_mov_b32_e32 v3, s45
+; VI-NEXT: v_mov_b32_e32 v4, s44
+; VI-NEXT: v_mov_b32_e32 v5, s43
+; VI-NEXT: v_mov_b32_e32 v6, s42
+; VI-NEXT: v_mov_b32_e32 v7, s41
+; VI-NEXT: v_mov_b32_e32 v8, s40
+; VI-NEXT: v_mov_b32_e32 v9, s26
+; VI-NEXT: v_mov_b32_e32 v10, s25
+; VI-NEXT: v_mov_b32_e32 v11, s24
+; VI-NEXT: v_mov_b32_e32 v12, s23
+; VI-NEXT: v_mov_b32_e32 v13, s22
+; VI-NEXT: v_mov_b32_e32 v14, s21
+; VI-NEXT: v_mov_b32_e32 v15, s20
+; VI-NEXT: v_mov_b32_e32 v16, s19
+; VI-NEXT: v_mov_b32_e32 v17, s18
+; VI-NEXT: v_mov_b32_e32 v18, s17
+; VI-NEXT: v_mov_b32_e32 v19, s16
; VI-NEXT: v_mov_b32_e32 v20, s15
; VI-NEXT: v_mov_b32_e32 v21, s14
; VI-NEXT: v_mov_b32_e32 v22, s13
@@ -9885,9 +9857,9 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; VI-NEXT: ; implicit-def: $sgpr59
; VI-NEXT: ; implicit-def: $sgpr58
; VI-NEXT: ; implicit-def: $sgpr57
-; VI-NEXT: ; implicit-def: $sgpr56
-; VI-NEXT: ; implicit-def: $sgpr47
-; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr29
+; VI-NEXT: ; implicit-def: $sgpr28
+; VI-NEXT: ; implicit-def: $sgpr27
; VI-NEXT: s_branch .LBB17_2
;
; GFX9-LABEL: bitcast_v30i32_to_v60f16_scalar:
@@ -9896,20 +9868,48 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_mov_b32_e32 v17, s16
+; GFX9-NEXT: v_mov_b32_e32 v18, s17
+; GFX9-NEXT: v_mov_b32_e32 v19, s18
+; GFX9-NEXT: v_readfirstlane_b32 s6, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s19
+; GFX9-NEXT: v_readfirstlane_b32 s7, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s20
+; GFX9-NEXT: v_readfirstlane_b32 s8, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s21
+; GFX9-NEXT: v_readfirstlane_b32 s9, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s22
+; GFX9-NEXT: v_readfirstlane_b32 s10, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s23
; GFX9-NEXT: v_writelane_b32 v30, s30, 0
+; GFX9-NEXT: v_readfirstlane_b32 s11, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s24
+; GFX9-NEXT: v_readfirstlane_b32 s12, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s25
+; GFX9-NEXT: v_readfirstlane_b32 s13, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s26
; GFX9-NEXT: v_writelane_b32 v30, s31, 1
+; GFX9-NEXT: v_readfirstlane_b32 s14, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s27
+; GFX9-NEXT: v_readfirstlane_b32 s15, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s28
+; GFX9-NEXT: v_readfirstlane_b32 s16, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s29
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; GFX9-NEXT: v_writelane_b32 v30, s34, 2
-; GFX9-NEXT: v_readfirstlane_b32 s6, v0
-; GFX9-NEXT: v_readfirstlane_b32 s7, v1
-; GFX9-NEXT: v_readfirstlane_b32 s8, v2
-; GFX9-NEXT: v_readfirstlane_b32 s9, v3
-; GFX9-NEXT: v_readfirstlane_b32 s10, v4
-; GFX9-NEXT: v_readfirstlane_b32 s11, v5
-; GFX9-NEXT: v_readfirstlane_b32 s12, v6
-; GFX9-NEXT: v_readfirstlane_b32 s13, v7
-; GFX9-NEXT: v_readfirstlane_b32 s14, v8
-; GFX9-NEXT: v_readfirstlane_b32 s15, v9
+; GFX9-NEXT: v_readfirstlane_b32 s17, v19
+; GFX9-NEXT: v_readfirstlane_b32 s18, v17
+; GFX9-NEXT: v_readfirstlane_b32 s19, v18
+; GFX9-NEXT: v_readfirstlane_b32 s20, v0
+; GFX9-NEXT: v_readfirstlane_b32 s21, v1
+; GFX9-NEXT: v_readfirstlane_b32 s22, v2
+; GFX9-NEXT: v_readfirstlane_b32 s23, v3
+; GFX9-NEXT: v_readfirstlane_b32 s24, v4
+; GFX9-NEXT: v_readfirstlane_b32 s25, v5
+; GFX9-NEXT: v_readfirstlane_b32 s26, v6
+; GFX9-NEXT: v_readfirstlane_b32 s27, v7
+; GFX9-NEXT: v_readfirstlane_b32 s28, v8
+; GFX9-NEXT: v_readfirstlane_b32 s29, v9
; GFX9-NEXT: v_readfirstlane_b32 s40, v10
; GFX9-NEXT: v_readfirstlane_b32 s41, v11
; GFX9-NEXT: v_readfirstlane_b32 s42, v12
@@ -9926,30 +9926,30 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX9-NEXT: s_lshr_b32 s57, s42, 16
; GFX9-NEXT: s_lshr_b32 s58, s41, 16
; GFX9-NEXT: s_lshr_b32 s59, s40, 16
-; GFX9-NEXT: s_lshr_b32 s60, s15, 16
-; GFX9-NEXT: s_lshr_b32 s61, s14, 16
-; GFX9-NEXT: s_lshr_b32 s62, s13, 16
-; GFX9-NEXT: s_lshr_b32 s63, s12, 16
-; GFX9-NEXT: s_lshr_b32 s72, s11, 16
-; GFX9-NEXT: s_lshr_b32 s73, s10, 16
-; GFX9-NEXT: s_lshr_b32 s74, s9, 16
-; GFX9-NEXT: s_lshr_b32 s75, s8, 16
-; GFX9-NEXT: s_lshr_b32 s76, s7, 16
-; GFX9-NEXT: s_lshr_b32 s77, s6, 16
-; GFX9-NEXT: s_lshr_b32 s78, s29, 16
-; GFX9-NEXT: s_lshr_b32 s79, s28, 16
-; GFX9-NEXT: s_lshr_b32 s88, s27, 16
-; GFX9-NEXT: s_lshr_b32 s89, s26, 16
-; GFX9-NEXT: s_lshr_b32 s90, s25, 16
-; GFX9-NEXT: s_lshr_b32 s91, s24, 16
-; GFX9-NEXT: s_lshr_b32 s92, s23, 16
-; GFX9-NEXT: s_lshr_b32 s93, s22, 16
-; GFX9-NEXT: s_lshr_b32 s94, s21, 16
-; GFX9-NEXT: s_lshr_b32 s95, s20, 16
-; GFX9-NEXT: s_lshr_b32 s30, s19, 16
-; GFX9-NEXT: s_lshr_b32 s31, s18, 16
-; GFX9-NEXT: s_lshr_b32 s34, s17, 16
-; GFX9-NEXT: s_lshr_b32 s35, s16, 16
+; GFX9-NEXT: s_lshr_b32 s60, s29, 16
+; GFX9-NEXT: s_lshr_b32 s61, s28, 16
+; GFX9-NEXT: s_lshr_b32 s62, s27, 16
+; GFX9-NEXT: s_lshr_b32 s63, s26, 16
+; GFX9-NEXT: s_lshr_b32 s72, s25, 16
+; GFX9-NEXT: s_lshr_b32 s73, s24, 16
+; GFX9-NEXT: s_lshr_b32 s74, s23, 16
+; GFX9-NEXT: s_lshr_b32 s75, s22, 16
+; GFX9-NEXT: s_lshr_b32 s76, s21, 16
+; GFX9-NEXT: s_lshr_b32 s77, s20, 16
+; GFX9-NEXT: s_lshr_b32 s78, s19, 16
+; GFX9-NEXT: s_lshr_b32 s79, s18, 16
+; GFX9-NEXT: s_lshr_b32 s88, s17, 16
+; GFX9-NEXT: s_lshr_b32 s89, s16, 16
+; GFX9-NEXT: s_lshr_b32 s90, s15, 16
+; GFX9-NEXT: s_lshr_b32 s91, s14, 16
+; GFX9-NEXT: s_lshr_b32 s92, s13, 16
+; GFX9-NEXT: s_lshr_b32 s93, s12, 16
+; GFX9-NEXT: s_lshr_b32 s94, s11, 16
+; GFX9-NEXT: s_lshr_b32 s95, s10, 16
+; GFX9-NEXT: s_lshr_b32 s30, s9, 16
+; GFX9-NEXT: s_lshr_b32 s31, s8, 16
+; GFX9-NEXT: s_lshr_b32 s34, s7, 16
+; GFX9-NEXT: s_lshr_b32 s35, s6, 16
; GFX9-NEXT: s_cbranch_execnz .LBB17_3
; GFX9-NEXT: .LBB17_2: ; %cmp.true
; GFX9-NEXT: s_add_i32 s45, s45, 3
@@ -9958,16 +9958,6 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX9-NEXT: s_add_i32 s42, s42, 3
; GFX9-NEXT: s_add_i32 s41, s41, 3
; GFX9-NEXT: s_add_i32 s40, s40, 3
-; GFX9-NEXT: s_add_i32 s15, s15, 3
-; GFX9-NEXT: s_add_i32 s14, s14, 3
-; GFX9-NEXT: s_add_i32 s13, s13, 3
-; GFX9-NEXT: s_add_i32 s12, s12, 3
-; GFX9-NEXT: s_add_i32 s11, s11, 3
-; GFX9-NEXT: s_add_i32 s10, s10, 3
-; GFX9-NEXT: s_add_i32 s9, s9, 3
-; GFX9-NEXT: s_add_i32 s8, s8, 3
-; GFX9-NEXT: s_add_i32 s7, s7, 3
-; GFX9-NEXT: s_add_i32 s6, s6, 3
; GFX9-NEXT: s_add_i32 s29, s29, 3
; GFX9-NEXT: s_add_i32 s28, s28, 3
; GFX9-NEXT: s_add_i32 s27, s27, 3
@@ -9982,61 +9972,71 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX9-NEXT: s_add_i32 s18, s18, 3
; GFX9-NEXT: s_add_i32 s17, s17, 3
; GFX9-NEXT: s_add_i32 s16, s16, 3
+; GFX9-NEXT: s_add_i32 s15, s15, 3
+; GFX9-NEXT: s_add_i32 s14, s14, 3
+; GFX9-NEXT: s_add_i32 s13, s13, 3
+; GFX9-NEXT: s_add_i32 s12, s12, 3
+; GFX9-NEXT: s_add_i32 s11, s11, 3
+; GFX9-NEXT: s_add_i32 s10, s10, 3
+; GFX9-NEXT: s_add_i32 s9, s9, 3
+; GFX9-NEXT: s_add_i32 s8, s8, 3
+; GFX9-NEXT: s_add_i32 s7, s7, 3
+; GFX9-NEXT: s_add_i32 s6, s6, 3
; GFX9-NEXT: s_lshr_b32 s46, s45, 16
; GFX9-NEXT: s_lshr_b32 s47, s44, 16
; GFX9-NEXT: s_lshr_b32 s56, s43, 16
; GFX9-NEXT: s_lshr_b32 s57, s42, 16
; GFX9-NEXT: s_lshr_b32 s58, s41, 16
; GFX9-NEXT: s_lshr_b32 s59, s40, 16
-; GFX9-NEXT: s_lshr_b32 s60, s15, 16
-; GFX9-NEXT: s_lshr_b32 s61, s14, 16
-; GFX9-NEXT: s_lshr_b32 s62, s13, 16
-; GFX9-NEXT: s_lshr_b32 s63, s12, 16
-; GFX9-NEXT: s_lshr_b32 s72, s11, 16
-; GFX9-NEXT: s_lshr_b32 s73, s10, 16
-; GFX9-NEXT: s_lshr_b32 s74, s9, 16
-; GFX9-NEXT: s_lshr_b32 s75, s8, 16
-; GFX9-NEXT: s_lshr_b32 s76, s7, 16
-; GFX9-NEXT: s_lshr_b32 s77, s6, 16
-; GFX9-NEXT: s_lshr_b32 s78, s29, 16
-; GFX9-NEXT: s_lshr_b32 s79, s28, 16
-; GFX9-NEXT: s_lshr_b32 s88, s27, 16
-; GFX9-NEXT: s_lshr_b32 s89, s26, 16
-; GFX9-NEXT: s_lshr_b32 s90, s25, 16
-; GFX9-NEXT: s_lshr_b32 s91, s24, 16
-; GFX9-NEXT: s_lshr_b32 s92, s23, 16
-; GFX9-NEXT: s_lshr_b32 s93, s22, 16
-; GFX9-NEXT: s_lshr_b32 s94, s21, 16
-; GFX9-NEXT: s_lshr_b32 s95, s20, 16
-; GFX9-NEXT: s_lshr_b32 s30, s19, 16
-; GFX9-NEXT: s_lshr_b32 s31, s18, 16
-; GFX9-NEXT: s_lshr_b32 s34, s17, 16
-; GFX9-NEXT: s_lshr_b32 s35, s16, 16
+; GFX9-NEXT: s_lshr_b32 s60, s29, 16
+; GFX9-NEXT: s_lshr_b32 s61, s28, 16
+; GFX9-NEXT: s_lshr_b32 s62, s27, 16
+; GFX9-NEXT: s_lshr_b32 s63, s26, 16
+; GFX9-NEXT: s_lshr_b32 s72, s25, 16
+; GFX9-NEXT: s_lshr_b32 s73, s24, 16
+; GFX9-NEXT: s_lshr_b32 s74, s23, 16
+; GFX9-NEXT: s_lshr_b32 s75, s22, 16
+; GFX9-NEXT: s_lshr_b32 s76, s21, 16
+; GFX9-NEXT: s_lshr_b32 s77, s20, 16
+; GFX9-NEXT: s_lshr_b32 s78, s19, 16
+; GFX9-NEXT: s_lshr_b32 s79, s18, 16
+; GFX9-NEXT: s_lshr_b32 s88, s17, 16
+; GFX9-NEXT: s_lshr_b32 s89, s16, 16
+; GFX9-NEXT: s_lshr_b32 s90, s15, 16
+; GFX9-NEXT: s_lshr_b32 s91, s14, 16
+; GFX9-NEXT: s_lshr_b32 s92, s13, 16
+; GFX9-NEXT: s_lshr_b32 s93, s12, 16
+; GFX9-NEXT: s_lshr_b32 s94, s11, 16
+; GFX9-NEXT: s_lshr_b32 s95, s10, 16
+; GFX9-NEXT: s_lshr_b32 s30, s9, 16
+; GFX9-NEXT: s_lshr_b32 s31, s8, 16
+; GFX9-NEXT: s_lshr_b32 s34, s7, 16
+; GFX9-NEXT: s_lshr_b32 s35, s6, 16
; GFX9-NEXT: .LBB17_3: ; %end
-; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s35
-; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s34
-; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s31
-; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s30
-; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s95
-; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s94
-; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s93
-; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s92
-; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s91
-; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s90
-; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s89
-; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s88
-; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s79
-; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s78
-; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s77
-; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s76
-; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s75
-; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s74
-; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s73
-; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s72
-; GFX9-NEXT: s_pack_ll_b32_b16 s12, s12, s63
-; GFX9-NEXT: s_pack_ll_b32_b16 s13, s13, s62
-; GFX9-NEXT: s_pack_ll_b32_b16 s14, s14, s61
-; GFX9-NEXT: s_pack_ll_b32_b16 s15, s15, s60
+; GFX9-NEXT: s_pack_ll_b32_b16 s4, s6, s35
+; GFX9-NEXT: s_pack_ll_b32_b16 s5, s7, s34
+; GFX9-NEXT: s_pack_ll_b32_b16 s6, s8, s31
+; GFX9-NEXT: s_pack_ll_b32_b16 s7, s9, s30
+; GFX9-NEXT: s_pack_ll_b32_b16 s8, s10, s95
+; GFX9-NEXT: s_pack_ll_b32_b16 s9, s11, s94
+; GFX9-NEXT: s_pack_ll_b32_b16 s10, s12, s93
+; GFX9-NEXT: s_pack_ll_b32_b16 s11, s13, s92
+; GFX9-NEXT: s_pack_ll_b32_b16 s12, s14, s91
+; GFX9-NEXT: s_pack_ll_b32_b16 s13, s15, s90
+; GFX9-NEXT: s_pack_ll_b32_b16 s14, s16, s89
+; GFX9-NEXT: s_pack_ll_b32_b16 s15, s17, s88
+; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s79
+; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s78
+; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s77
+; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s76
+; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s75
+; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s74
+; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s73
+; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s72
+; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s63
+; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s62
+; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s61
+; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s60
; GFX9-NEXT: s_pack_ll_b32_b16 s28, s40, s59
; GFX9-NEXT: s_pack_ll_b32_b16 s29, s41, s58
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s42, s57
@@ -10045,28 +10045,28 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s45, s46
; GFX9-NEXT: v_mov_b32_e32 v0, s4
; GFX9-NEXT: v_mov_b32_e32 v1, s5
-; GFX9-NEXT: v_mov_b32_e32 v2, s16
-; GFX9-NEXT: v_mov_b32_e32 v3, s17
-; GFX9-NEXT: v_mov_b32_e32 v4, s18
-; GFX9-NEXT: v_mov_b32_e32 v5, s19
-; GFX9-NEXT: v_mov_b32_e32 v6, s20
-; GFX9-NEXT: v_mov_b32_e32 v7, s21
-; GFX9-NEXT: v_mov_b32_e32 v8, s22
-; GFX9-NEXT: v_mov_b32_e32 v9, s23
-; GFX9-NEXT: v_mov_b32_e32 v10, s24
-; GFX9-NEXT: v_mov_b32_e32 v11, s25
-; GFX9-NEXT: v_mov_b32_e32 v12, s26
-; GFX9-NEXT: v_mov_b32_e32 v13, s27
-; GFX9-NEXT: v_mov_b32_e32 v14, s6
-; GFX9-NEXT: v_mov_b32_e32 v15, s7
-; GFX9-NEXT: v_mov_b32_e32 v16, s8
-; GFX9-NEXT: v_mov_b32_e32 v17, s9
-; GFX9-NEXT: v_mov_b32_e32 v18, s10
-; GFX9-NEXT: v_mov_b32_e32 v19, s11
-; GFX9-NEXT: v_mov_b32_e32 v20, s12
-; GFX9-NEXT: v_mov_b32_e32 v21, s13
-; GFX9-NEXT: v_mov_b32_e32 v22, s14
-; GFX9-NEXT: v_mov_b32_e32 v23, s15
+; GFX9-NEXT: v_mov_b32_e32 v2, s6
+; GFX9-NEXT: v_mov_b32_e32 v3, s7
+; GFX9-NEXT: v_mov_b32_e32 v4, s8
+; GFX9-NEXT: v_mov_b32_e32 v5, s9
+; GFX9-NEXT: v_mov_b32_e32 v6, s10
+; GFX9-NEXT: v_mov_b32_e32 v7, s11
+; GFX9-NEXT: v_mov_b32_e32 v8, s12
+; GFX9-NEXT: v_mov_b32_e32 v9, s13
+; GFX9-NEXT: v_mov_b32_e32 v10, s14
+; GFX9-NEXT: v_mov_b32_e32 v11, s15
+; GFX9-NEXT: v_mov_b32_e32 v12, s16
+; GFX9-NEXT: v_mov_b32_e32 v13, s17
+; GFX9-NEXT: v_mov_b32_e32 v14, s18
+; GFX9-NEXT: v_mov_b32_e32 v15, s19
+; GFX9-NEXT: v_mov_b32_e32 v16, s20
+; GFX9-NEXT: v_mov_b32_e32 v17, s21
+; GFX9-NEXT: v_mov_b32_e32 v18, s22
+; GFX9-NEXT: v_mov_b32_e32 v19, s23
+; GFX9-NEXT: v_mov_b32_e32 v20, s24
+; GFX9-NEXT: v_mov_b32_e32 v21, s25
+; GFX9-NEXT: v_mov_b32_e32 v22, s26
+; GFX9-NEXT: v_mov_b32_e32 v23, s27
; GFX9-NEXT: v_mov_b32_e32 v24, s28
; GFX9-NEXT: v_mov_b32_e32 v25, s29
; GFX9-NEXT: v_mov_b32_e32 v26, s40
@@ -10118,49 +10118,76 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX11-LABEL: bitcast_v30i32_to_v60f16_scalar:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT: v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
+; GFX11-NEXT: v_dual_mov_b32 v17, s16 :: v_dual_mov_b32 v18, s17
+; GFX11-NEXT: v_dual_mov_b32 v19, s18 :: v_dual_mov_b32 v20, s19
+; GFX11-NEXT: v_dual_mov_b32 v21, s20 :: v_dual_mov_b32 v22, s21
+; GFX11-NEXT: v_dual_mov_b32 v23, s22 :: v_dual_mov_b32 v24, s23
+; GFX11-NEXT: v_dual_mov_b32 v25, s24 :: v_dual_mov_b32 v26, s25
+; GFX11-NEXT: v_dual_mov_b32 v27, s26 :: v_dual_mov_b32 v28, s27
+; GFX11-NEXT: v_dual_mov_b32 v29, s28 :: v_dual_mov_b32 v30, s29
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: v_readfirstlane_b32 s4, v0
-; GFX11-NEXT: v_readfirstlane_b32 s5, v1
-; GFX11-NEXT: v_readfirstlane_b32 s6, v2
-; GFX11-NEXT: v_readfirstlane_b32 s7, v3
-; GFX11-NEXT: v_readfirstlane_b32 s8, v4
-; GFX11-NEXT: v_readfirstlane_b32 s9, v5
-; GFX11-NEXT: v_readfirstlane_b32 s10, v6
-; GFX11-NEXT: v_readfirstlane_b32 s11, v7
-; GFX11-NEXT: v_readfirstlane_b32 s12, v8
-; GFX11-NEXT: v_readfirstlane_b32 s13, v9
-; GFX11-NEXT: v_readfirstlane_b32 s15, v10
-; GFX11-NEXT: v_readfirstlane_b32 s14, v11
+; GFX11-NEXT: v_readfirstlane_b32 s0, v13
+; GFX11-NEXT: v_readfirstlane_b32 s1, v14
+; GFX11-NEXT: v_readfirstlane_b32 s2, v15
+; GFX11-NEXT: v_readfirstlane_b32 s3, v16
+; GFX11-NEXT: v_readfirstlane_b32 s4, v17
+; GFX11-NEXT: v_readfirstlane_b32 s5, v18
+; GFX11-NEXT: v_readfirstlane_b32 s6, v19
+; GFX11-NEXT: v_readfirstlane_b32 s7, v20
+; GFX11-NEXT: v_readfirstlane_b32 s8, v21
+; GFX11-NEXT: v_readfirstlane_b32 s9, v22
+; GFX11-NEXT: v_readfirstlane_b32 s10, v23
+; GFX11-NEXT: v_readfirstlane_b32 s11, v24
+; GFX11-NEXT: v_readfirstlane_b32 s12, v25
+; GFX11-NEXT: v_readfirstlane_b32 s13, v26
+; GFX11-NEXT: v_readfirstlane_b32 s14, v27
+; GFX11-NEXT: v_readfirstlane_b32 s15, v28
+; GFX11-NEXT: v_readfirstlane_b32 s16, v29
+; GFX11-NEXT: v_readfirstlane_b32 s17, v30
+; GFX11-NEXT: v_readfirstlane_b32 s18, v0
+; GFX11-NEXT: v_readfirstlane_b32 s19, v1
+; GFX11-NEXT: v_readfirstlane_b32 s20, v2
+; GFX11-NEXT: v_readfirstlane_b32 s21, v3
+; GFX11-NEXT: v_readfirstlane_b32 s22, v4
+; GFX11-NEXT: v_readfirstlane_b32 s23, v5
+; GFX11-NEXT: v_readfirstlane_b32 s24, v6
+; GFX11-NEXT: v_readfirstlane_b32 s25, v7
+; GFX11-NEXT: v_readfirstlane_b32 s26, v8
+; GFX11-NEXT: v_readfirstlane_b32 s27, v9
+; GFX11-NEXT: v_readfirstlane_b32 s29, v10
+; GFX11-NEXT: v_readfirstlane_b32 s28, v11
; GFX11-NEXT: s_mov_b32 s94, 0
; GFX11-NEXT: s_and_b32 s40, vcc_lo, exec_lo
; GFX11-NEXT: s_cbranch_scc0 .LBB17_4
; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: s_lshr_b32 s40, s14, 16
-; GFX11-NEXT: s_lshr_b32 s41, s15, 16
-; GFX11-NEXT: s_lshr_b32 s42, s13, 16
-; GFX11-NEXT: s_lshr_b32 s43, s12, 16
-; GFX11-NEXT: s_lshr_b32 s44, s11, 16
-; GFX11-NEXT: s_lshr_b32 s45, s10, 16
-; GFX11-NEXT: s_lshr_b32 s46, s9, 16
-; GFX11-NEXT: s_lshr_b32 s47, s8, 16
-; GFX11-NEXT: s_lshr_b32 s56, s7, 16
-; GFX11-NEXT: s_lshr_b32 s57, s6, 16
-; GFX11-NEXT: s_lshr_b32 s58, s5, 16
-; GFX11-NEXT: s_lshr_b32 s59, s4, 16
-; GFX11-NEXT: s_lshr_b32 s60, s29, 16
-; GFX11-NEXT: s_lshr_b32 s61, s28, 16
-; GFX11-NEXT: s_lshr_b32 s62, s27, 16
-; GFX11-NEXT: s_lshr_b32 s63, s26, 16
-; GFX11-NEXT: s_lshr_b32 s72, s25, 16
-; GFX11-NEXT: s_lshr_b32 s73, s24, 16
-; GFX11-NEXT: s_lshr_b32 s74, s23, 16
-; GFX11-NEXT: s_lshr_b32 s75, s22, 16
-; GFX11-NEXT: s_lshr_b32 s76, s21, 16
-; GFX11-NEXT: s_lshr_b32 s77, s20, 16
-; GFX11-NEXT: s_lshr_b32 s78, s19, 16
-; GFX11-NEXT: s_lshr_b32 s79, s18, 16
-; GFX11-NEXT: s_lshr_b32 s88, s17, 16
-; GFX11-NEXT: s_lshr_b32 s89, s16, 16
+; GFX11-NEXT: s_lshr_b32 s40, s28, 16
+; GFX11-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-NEXT: s_lshr_b32 s43, s26, 16
+; GFX11-NEXT: s_lshr_b32 s44, s25, 16
+; GFX11-NEXT: s_lshr_b32 s45, s24, 16
+; GFX11-NEXT: s_lshr_b32 s46, s23, 16
+; GFX11-NEXT: s_lshr_b32 s47, s22, 16
+; GFX11-NEXT: s_lshr_b32 s56, s21, 16
+; GFX11-NEXT: s_lshr_b32 s57, s20, 16
+; GFX11-NEXT: s_lshr_b32 s58, s19, 16
+; GFX11-NEXT: s_lshr_b32 s59, s18, 16
+; GFX11-NEXT: s_lshr_b32 s60, s17, 16
+; GFX11-NEXT: s_lshr_b32 s61, s16, 16
+; GFX11-NEXT: s_lshr_b32 s62, s15, 16
+; GFX11-NEXT: s_lshr_b32 s63, s14, 16
+; GFX11-NEXT: s_lshr_b32 s72, s13, 16
+; GFX11-NEXT: s_lshr_b32 s73, s12, 16
+; GFX11-NEXT: s_lshr_b32 s74, s11, 16
+; GFX11-NEXT: s_lshr_b32 s75, s10, 16
+; GFX11-NEXT: s_lshr_b32 s76, s9, 16
+; GFX11-NEXT: s_lshr_b32 s77, s8, 16
+; GFX11-NEXT: s_lshr_b32 s78, s7, 16
+; GFX11-NEXT: s_lshr_b32 s79, s6, 16
+; GFX11-NEXT: s_lshr_b32 s88, s5, 16
+; GFX11-NEXT: s_lshr_b32 s89, s4, 16
; GFX11-NEXT: s_lshr_b32 s90, s3, 16
; GFX11-NEXT: s_lshr_b32 s91, s2, 16
; GFX11-NEXT: s_lshr_b32 s92, s1, 16
@@ -10168,20 +10195,8 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s94
; GFX11-NEXT: s_cbranch_vccnz .LBB17_3
; GFX11-NEXT: .LBB17_2: ; %cmp.true
-; GFX11-NEXT: s_add_i32 s14, s14, 3
-; GFX11-NEXT: s_add_i32 s15, s15, 3
-; GFX11-NEXT: s_add_i32 s13, s13, 3
-; GFX11-NEXT: s_add_i32 s12, s12, 3
-; GFX11-NEXT: s_add_i32 s11, s11, 3
-; GFX11-NEXT: s_add_i32 s10, s10, 3
-; GFX11-NEXT: s_add_i32 s9, s9, 3
-; GFX11-NEXT: s_add_i32 s8, s8, 3
-; GFX11-NEXT: s_add_i32 s7, s7, 3
-; GFX11-NEXT: s_add_i32 s6, s6, 3
-; GFX11-NEXT: s_add_i32 s5, s5, 3
-; GFX11-NEXT: s_add_i32 s4, s4, 3
-; GFX11-NEXT: s_add_i32 s29, s29, 3
; GFX11-NEXT: s_add_i32 s28, s28, 3
+; GFX11-NEXT: s_add_i32 s29, s29, 3
; GFX11-NEXT: s_add_i32 s27, s27, 3
; GFX11-NEXT: s_add_i32 s26, s26, 3
; GFX11-NEXT: s_add_i32 s25, s25, 3
@@ -10194,36 +10209,48 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX11-NEXT: s_add_i32 s18, s18, 3
; GFX11-NEXT: s_add_i32 s17, s17, 3
; GFX11-NEXT: s_add_i32 s16, s16, 3
+; GFX11-NEXT: s_add_i32 s15, s15, 3
+; GFX11-NEXT: s_add_i32 s14, s14, 3
+; GFX11-NEXT: s_add_i32 s13, s13, 3
+; GFX11-NEXT: s_add_i32 s12, s12, 3
+; GFX11-NEXT: s_add_i32 s11, s11, 3
+; GFX11-NEXT: s_add_i32 s10, s10, 3
+; GFX11-NEXT: s_add_i32 s9, s9, 3
+; GFX11-NEXT: s_add_i32 s8, s8, 3
+; GFX11-NEXT: s_add_i32 s7, s7, 3
+; GFX11-NEXT: s_add_i32 s6, s6, 3
+; GFX11-NEXT: s_add_i32 s5, s5, 3
+; GFX11-NEXT: s_add_i32 s4, s4, 3
; GFX11-NEXT: s_add_i32 s3, s3, 3
; GFX11-NEXT: s_add_i32 s2, s2, 3
; GFX11-NEXT: s_add_i32 s1, s1, 3
; GFX11-NEXT: s_add_i32 s0, s0, 3
-; GFX11-NEXT: s_lshr_b32 s40, s14, 16
-; GFX11-NEXT: s_lshr_b32 s41, s15, 16
-; GFX11-NEXT: s_lshr_b32 s42, s13, 16
-; GFX11-NEXT: s_lshr_b32 s43, s12, 16
-; GFX11-NEXT: s_lshr_b32 s44, s11, 16
-; GFX11-NEXT: s_lshr_b32 s45, s10, 16
-; GFX11-NEXT: s_lshr_b32 s46, s9, 16
-; GFX11-NEXT: s_lshr_b32 s47, s8, 16
-; GFX11-NEXT: s_lshr_b32 s56, s7, 16
-; GFX11-NEXT: s_lshr_b32 s57, s6, 16
-; GFX11-NEXT: s_lshr_b32 s58, s5, 16
-; GFX11-NEXT: s_lshr_b32 s59, s4, 16
-; GFX11-NEXT: s_lshr_b32 s60, s29, 16
-; GFX11-NEXT: s_lshr_b32 s61, s28, 16
-; GFX11-NEXT: s_lshr_b32 s62, s27, 16
-; GFX11-NEXT: s_lshr_b32 s63, s26, 16
-; GFX11-NEXT: s_lshr_b32 s72, s25, 16
-; GFX11-NEXT: s_lshr_b32 s73, s24, 16
-; GFX11-NEXT: s_lshr_b32 s74, s23, 16
-; GFX11-NEXT: s_lshr_b32 s75, s22, 16
-; GFX11-NEXT: s_lshr_b32 s76, s21, 16
-; GFX11-NEXT: s_lshr_b32 s77, s20, 16
-; GFX11-NEXT: s_lshr_b32 s78, s19, 16
-; GFX11-NEXT: s_lshr_b32 s79, s18, 16
-; GFX11-NEXT: s_lshr_b32 s88, s17, 16
-; GFX11-NEXT: s_lshr_b32 s89, s16, 16
+; GFX11-NEXT: s_lshr_b32 s40, s28, 16
+; GFX11-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-NEXT: s_lshr_b32 s43, s26, 16
+; GFX11-NEXT: s_lshr_b32 s44, s25, 16
+; GFX11-NEXT: s_lshr_b32 s45, s24, 16
+; GFX11-NEXT: s_lshr_b32 s46, s23, 16
+; GFX11-NEXT: s_lshr_b32 s47, s22, 16
+; GFX11-NEXT: s_lshr_b32 s56, s21, 16
+; GFX11-NEXT: s_lshr_b32 s57, s20, 16
+; GFX11-NEXT: s_lshr_b32 s58, s19, 16
+; GFX11-NEXT: s_lshr_b32 s59, s18, 16
+; GFX11-NEXT: s_lshr_b32 s60, s17, 16
+; GFX11-NEXT: s_lshr_b32 s61, s16, 16
+; GFX11-NEXT: s_lshr_b32 s62, s15, 16
+; GFX11-NEXT: s_lshr_b32 s63, s14, 16
+; GFX11-NEXT: s_lshr_b32 s72, s13, 16
+; GFX11-NEXT: s_lshr_b32 s73, s12, 16
+; GFX11-NEXT: s_lshr_b32 s74, s11, 16
+; GFX11-NEXT: s_lshr_b32 s75, s10, 16
+; GFX11-NEXT: s_lshr_b32 s76, s9, 16
+; GFX11-NEXT: s_lshr_b32 s77, s8, 16
+; GFX11-NEXT: s_lshr_b32 s78, s7, 16
+; GFX11-NEXT: s_lshr_b32 s79, s6, 16
+; GFX11-NEXT: s_lshr_b32 s88, s5, 16
+; GFX11-NEXT: s_lshr_b32 s89, s4, 16
; GFX11-NEXT: s_lshr_b32 s90, s3, 16
; GFX11-NEXT: s_lshr_b32 s91, s2, 16
; GFX11-NEXT: s_lshr_b32 s92, s1, 16
@@ -10234,47 +10261,47 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s92
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s91
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s59
-; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s58
-; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s57
-; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s56
-; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s47
-; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s46
-; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s45
-; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s44
-; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s43
-; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s42
-; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s41
-; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s40
+; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s89
+; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s88
+; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s79
+; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s78
+; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s77
+; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s76
+; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s75
+; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s74
+; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s73
+; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s72
+; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s63
+; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s62
+; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s61
+; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s60
+; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s59
+; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s58
+; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s57
+; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s56
+; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s47
+; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s46
+; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s45
+; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s44
+; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s43
+; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s41
+; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s40
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-NEXT: v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-NEXT: v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-NEXT: v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-NEXT: v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-NEXT: v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
-; GFX11-NEXT: v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT: v_dual_mov_b32 v28, s29 :: v_dual_mov_b32 v29, s28
; GFX11-NEXT: s_setpc_b64 s[30:31]
; GFX11-NEXT: .LBB17_4:
; GFX11-NEXT: ; implicit-def: $sgpr93
@@ -10345,6 +10372,9 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(5)
+; SI-NEXT: v_cvt_f16_f32_e32 v58, v0
+; SI-NEXT: v_cvt_f16_f32_e32 v0, v5
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116
; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32
@@ -10373,23 +10403,12 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:88
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:84
-; SI-NEXT: v_cvt_f16_f32_e32 v58, v0
-; SI-NEXT: v_cvt_f16_f32_e32 v0, v5
; SI-NEXT: v_cvt_f16_f32_e32 v59, v1
; SI-NEXT: v_cvt_f16_f32_e32 v57, v3
-; SI-NEXT: v_cvt_f16_f32_e32 v56, v2
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:104
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:100
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v4
-; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
+; SI-NEXT: v_cvt_f16_f32_e32 v56, v2
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v7
@@ -10399,8 +10418,6 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v9
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v8
@@ -10422,9 +10439,18 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v14
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v17
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:104
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:100
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
+; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v16
@@ -10434,6 +10460,7 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v18
+; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v21
@@ -10471,7 +10498,6 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v61
; SI-NEXT: v_cvt_f16_f32_e32 v61, v49
-; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f16_f32_e32 v49, v55
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -10486,6 +10512,7 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v32
+; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f16_f32_e32 v32, v47
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -11357,7 +11384,6 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: v_mov_b32_e32 v61, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v28
@@ -11436,6 +11462,7 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v60
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v61
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
@@ -11600,6 +11627,9 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT: s_cbranch_execz .LBB18_4
; GFX9-NEXT: ; %bb.3: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
@@ -11614,9 +11644,6 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v38, v61, s6
; GFX9-NEXT: s_movk_i32 s7, 0x200
; GFX9-NEXT: v_perm_b32 v1, v37, v60, s6
@@ -11636,6 +11663,10 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
+; GFX9-NEXT: s_waitcnt vmcnt(14)
+; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: s_waitcnt vmcnt(15)
; GFX9-NEXT: v_perm_b32 v9, v9, v44, s6
; GFX9-NEXT: s_waitcnt vmcnt(14)
@@ -11664,10 +11695,6 @@ define <30 x i32> @bitcast_v60f16_to_v30i32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_perm_b32 v21, v21, v48, s6
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_perm_b32 v22, v22, v39, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
@@ -11988,12 +12015,35 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_cvt_f16_f32_e32 v8, s26
; SI-NEXT: v_cvt_f16_f32_e32 v6, s29
; SI-NEXT: v_cvt_f16_f32_e32 v7, s28
+; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT: v_cvt_f16_f32_e32 v50, v54
; SI-NEXT: v_cvt_f16_f32_e32 v48, v48
; SI-NEXT: v_cvt_f16_f32_e32 v31, v40
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v33
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
@@ -12003,7 +12053,7 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v38
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v44
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -12012,7 +12062,7 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v46
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v47
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -12021,7 +12071,7 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v57
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v58
; SI-NEXT: v_cvt_f16_f32_e32 v58, s16
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
@@ -12032,38 +12082,12 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v60
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_cbranch_scc0 .LBB19_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT: v_or_b32_e32 v3, v10, v3
-; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v34
; SI-NEXT: v_mov_b32_e32 v33, v32
; SI-NEXT: v_or_b32_e32 v10, v32, v10
@@ -12088,12 +12112,12 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v59
; SI-NEXT: v_or_b32_e32 v1, v12, v1
; SI-NEXT: v_or_b32_e32 v2, v11, v2
@@ -12202,12 +12226,10 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_mov_b32_e32 v40, v44
; SI-NEXT: s_cbranch_vccnz .LBB19_5
; SI-NEXT: ; %bb.4: ; %cmp.true
-; SI-NEXT: s_waitcnt expcnt(5)
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT: v_cvt_f32_f16_e32 v0, v59
; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v0, v59
; SI-NEXT: v_cvt_f32_f16_e32 v1, v58
; SI-NEXT: s_waitcnt vmcnt(3)
; SI-NEXT: v_cvt_f32_f16_e32 v8, v33
@@ -12993,304 +13015,115 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v156, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v157, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v158, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v159, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v168, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v169, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v170, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v171, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v172, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v173, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v174, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v175, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v184, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v185, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v186, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v187, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v188, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v189, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v190, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v191, s32
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v25, v7
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v191, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v190, v6 :: v_dual_mov_b32 v185, v4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v186, v3 :: v_dual_mov_b32 v187, v2
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v188, v1 :: v_dual_mov_b32 v189, v0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s28, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s19, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s17, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s1, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s0, 16
-; GFX11-TRUE16-NEXT: s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s40, s0, s40
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s41, s1, s41
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s2, s46
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s3, s45
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s16, s44
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s17, s43
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s18, s4
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s19, s5
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s20, s6
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s21, s7
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s22, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s23, s9
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s25, s11
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s26, s12
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s27, s13
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s28, s14
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s29, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB19_4
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s40 :: v_dual_mov_b32 v5, s0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s41 :: v_dual_mov_b32 v9, s1
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s2 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s3 :: v_dual_mov_b32 v35, s5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v44, s6 :: v_dual_mov_b32 v65, s8
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s7 :: v_dual_mov_b32 v77, s9
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v90, s10 :: v_dual_mov_b32 v119, s12
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v104, s11 :: v_dual_mov_b32 v135, s13
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v152, s14
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v170, s15
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB19_3
; GFX11-TRUE16-NEXT: .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s40 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s41 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v189, 0x200, v189 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v188, 0x200, v188 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v187, 0x200, v187 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v186, 0x200, v186 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v185, 0x200, v185 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v191, 0x200, v191 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v190, 0x200, v190 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v35, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v44, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v54, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v65, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v77, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v90, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v104, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v119, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v135, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v152, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v170, 0x200, s15 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: .LBB19_3: ; %end
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v20 :: v_dual_mov_b32 v6, v27
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v35 :: v_dual_mov_b32 v8, v44
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v77 :: v_dual_mov_b32 v12, v90
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v104
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, v135 :: v_dual_mov_b32 v16, v152
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v170 :: v_dual_mov_b32 v18, v189
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v188 :: v_dual_mov_b32 v20, v187
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v186 :: v_dual_mov_b32 v22, v185
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v191 :: v_dual_mov_b32 v24, v190
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v191, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v190, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v189, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v188, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v187, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v186, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v185, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v184, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v175, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v174, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v173, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v172, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v171, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v170, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v169, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v168, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v159, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v158, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v157, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v156, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:124
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:252
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:316
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v14
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v54 :: v_dual_mov_b32 v10, v65
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v119 :: v_dual_mov_b32 v27, v30
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB19_4:
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, v29 :: v_dual_mov_b32 v65, v28
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v66, v30
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v26 :: v_dual_mov_b32 v54, v25
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v53
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v65 :: v_dual_mov_b32 v29, v64
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v30, v66
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168_vgpr169_vgpr170_vgpr171_vgpr172_vgpr173_vgpr174_vgpr175_vgpr176_vgpr177_vgpr178_vgpr179_vgpr180_vgpr181_vgpr182_vgpr183_vgpr184
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_branch .LBB19_2
;
; GFX11-FAKE16-LABEL: bitcast_v60f16_to_v30i32_scalar:
@@ -17570,40 +17403,53 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:4
; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v2
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:112
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:108
+; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(4)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v8
-; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:104
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:100
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v10
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v6
+; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v8
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:96
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:92
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v10
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:84
-; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v14
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
@@ -17634,27 +17480,10 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:56
; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:64
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52
-; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v28
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
@@ -17739,7 +17568,6 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr30
; SI-NEXT: ; kill: killed $vgpr30
; SI-NEXT: ; implicit-def: $vgpr30
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v55
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v61
; SI-NEXT: ; kill: killed $vgpr30
@@ -17884,7 +17712,6 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v55
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v61
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
@@ -18011,7 +17838,7 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; SI-NEXT: v_add_i32_e32 v27, vcc, s6, v27
+; SI-NEXT: v_add_i32_e32 v27, vcc, 0x30000, v27
; SI-NEXT: v_add_i32_e32 v28, vcc, 0x30000, v28
; SI-NEXT: v_add_i32_e32 v29, vcc, 0x30000, v29
; SI-NEXT: .LBB30_4: ; %end
@@ -18032,7 +17859,7 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v60i16_to_v30f32:
@@ -18314,7 +18141,6 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: v_mov_b32_e32 v61, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v28
@@ -18393,6 +18219,7 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v60
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v61
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
@@ -18557,6 +18384,9 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT: s_cbranch_execz .LBB30_4
; GFX9-NEXT: ; %bb.3: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
@@ -18571,9 +18401,6 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v38, v61, s6
; GFX9-NEXT: v_perm_b32 v1, v37, v60, s6
; GFX9-NEXT: v_perm_b32 v2, v62, v59, s6
@@ -18592,6 +18419,10 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: s_waitcnt vmcnt(14)
+; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: s_waitcnt vmcnt(15)
; GFX9-NEXT: v_perm_b32 v9, v9, v44, s6
; GFX9-NEXT: s_waitcnt vmcnt(14)
@@ -18620,10 +18451,6 @@ define <30 x float> @bitcast_v60i16_to_v30f32(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_perm_b32 v21, v21, v48, s6
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_perm_b32 v22, v22, v39, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
@@ -19779,304 +19606,115 @@ define inreg <30 x float> @bitcast_v60i16_to_v30f32_scalar(<60 x i16> inreg %a,
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v156, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v157, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v158, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v159, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v168, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v169, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v170, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v171, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v172, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v173, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v174, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v175, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v184, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v185, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v186, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v187, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v188, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v189, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v190, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v191, s32
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v25, v7
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v191, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v190, v6 :: v_dual_mov_b32 v185, v4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v186, v3 :: v_dual_mov_b32 v187, v2
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v188, v1 :: v_dual_mov_b32 v189, v0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s28, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s19, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s17, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s1, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s0, 16
-; GFX11-TRUE16-NEXT: s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s40, s0, s40
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s41, s1, s41
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s2, s46
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s3, s45
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s16, s44
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s17, s43
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s18, s4
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s19, s5
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s20, s6
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s21, s7
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s22, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s23, s9
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s25, s11
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s26, s12
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s27, s13
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s28, s14
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s29, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB31_4
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s40 :: v_dual_mov_b32 v5, s0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s41 :: v_dual_mov_b32 v9, s1
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s2 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s3 :: v_dual_mov_b32 v35, s5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v44, s6 :: v_dual_mov_b32 v65, s8
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s7 :: v_dual_mov_b32 v77, s9
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v90, s10 :: v_dual_mov_b32 v119, s12
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v104, s11 :: v_dual_mov_b32 v135, s13
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v152, s14
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v170, s15
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB31_3
; GFX11-TRUE16-NEXT: .LBB31_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s40, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s41, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v189, v189, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v188, v188, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v187, v187, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v186, v186, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v185, v185, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v191, v191, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v190, v190, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, s4, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v35, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v44, s6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v54, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v65, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v77, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v90, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v104, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v119, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v135, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v152, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v170, s15, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: .LBB31_3: ; %end
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v20 :: v_dual_mov_b32 v6, v27
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v35 :: v_dual_mov_b32 v8, v44
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v77 :: v_dual_mov_b32 v12, v90
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v104
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, v135 :: v_dual_mov_b32 v16, v152
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v170 :: v_dual_mov_b32 v18, v189
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v188 :: v_dual_mov_b32 v20, v187
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v186 :: v_dual_mov_b32 v22, v185
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v191 :: v_dual_mov_b32 v24, v190
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v191, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v190, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v189, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v188, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v187, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v186, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v185, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v184, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v175, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v174, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v173, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v172, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v171, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v170, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v169, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v168, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v159, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v158, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v157, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v156, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:124
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:252
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:316
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v14
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v54 :: v_dual_mov_b32 v10, v65
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v119 :: v_dual_mov_b32 v27, v30
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB31_4:
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, v29 :: v_dual_mov_b32 v65, v28
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v66, v30
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v26 :: v_dual_mov_b32 v54, v25
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v53
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v65 :: v_dual_mov_b32 v29, v64
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v30, v66
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168_vgpr169_vgpr170_vgpr171_vgpr172_vgpr173_vgpr174_vgpr175_vgpr176_vgpr177_vgpr178_vgpr179_vgpr180_vgpr181_vgpr182_vgpr183_vgpr184
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_branch .LBB31_2
;
; GFX11-FAKE16-LABEL: bitcast_v60i16_to_v30f32_scalar:
@@ -21565,23 +21203,21 @@ define inreg <60 x half> @bitcast_v30f32_to_v60f16_scalar(<30 x float> inreg %a,
; SI: ; %bb.0:
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; SI-NEXT: v_readfirstlane_b32 s45, v1
-; SI-NEXT: v_readfirstlane_b32 s44, v2
-; SI-NEXT: v_readfirstlane_b32 s43, v3
-; SI-NEXT: v_readfirstlane_b32 s42, v4
-; SI-NEXT: v_readfirstlane_b32 s41, v5
-; SI-NEXT: v_readfirstlane_b32 s40, v6
-; SI-NEXT: v_readfirstlane_b32 s15, v7
-; SI-NEXT: v_readfirstlane_b32 s14, v8
-; SI-NEXT: v_readfirstlane_b32 s13, v9
-; SI-NEXT: v_readfirstlane_b32 s12, v10
-; SI-NEXT: v_readfirstlane_b32 s11, v11
-; SI-NEXT: v_readfirstlane_b32 s10, v12
-; SI-NEXT: v_readfirstlane_b32 s8, v13
-; SI-NEXT: v_readfirstlane_b32 s7, v14
-; SI-NEXT: v_readfirstlane_b32 s6, v15
+; SI-NEXT: v_mov_b32_e32 v31, s16
+; SI-NEXT: v_mov_b32_e32 v29, s17
+; SI-NEXT: v_mov_b32_e32 v50, s18
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
-; SI-NEXT: v_readfirstlane_b32 s9, v16
+; SI-NEXT: v_mov_b32_e32 v51, s19
+; SI-NEXT: v_mov_b32_e32 v49, s20
+; SI-NEXT: v_mov_b32_e32 v48, s21
+; SI-NEXT: v_mov_b32_e32 v39, s22
+; SI-NEXT: v_mov_b32_e32 v38, s23
+; SI-NEXT: v_mov_b32_e32 v37, s24
+; SI-NEXT: v_mov_b32_e32 v36, s25
+; SI-NEXT: v_mov_b32_e32 v35, s26
+; SI-NEXT: v_mov_b32_e32 v34, s27
+; SI-NEXT: v_mov_b32_e32 v33, s28
+; SI-NEXT: v_mov_b32_e32 v32, s29
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
@@ -21600,450 +21236,574 @@ define inreg <60 x half> @bitcast_v30f32_to_v60f16_scalar(<30 x float> inreg %a,
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT: s_cbranch_scc0 .LBB33_4
; SI-NEXT: ; %bb.1: ; %cmp.false
-; SI-NEXT: s_lshr_b32 s4, s9, 16
-; SI-NEXT: s_waitcnt expcnt(4)
-; SI-NEXT: v_cvt_f32_f16_e32 v59, s4
-; SI-NEXT: s_lshr_b32 s4, s6, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v42, s4
-; SI-NEXT: s_lshr_b32 s4, s7, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v54, s4
-; SI-NEXT: s_lshr_b32 s4, s8, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v50, s4
-; SI-NEXT: s_lshr_b32 s4, s10, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v57, s4
-; SI-NEXT: s_lshr_b32 s4, s11, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v35, s4
-; SI-NEXT: s_lshr_b32 s4, s12, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v11, s4
-; SI-NEXT: s_lshr_b32 s4, s13, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v12, s4
-; SI-NEXT: s_lshr_b32 s4, s14, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v15, s4
-; SI-NEXT: s_lshr_b32 s4, s15, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v8, s4
-; SI-NEXT: s_lshr_b32 s4, s40, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v19, s4
-; SI-NEXT: s_lshr_b32 s4, s41, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v22, s4
-; SI-NEXT: s_lshr_b32 s4, s42, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v24, s4
-; SI-NEXT: s_lshr_b32 s4, s43, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v26, s4
-; SI-NEXT: s_lshr_b32 s4, s44, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v28, s4
-; SI-NEXT: s_lshr_b32 s4, s45, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v30, s4
-; SI-NEXT: s_lshr_b32 s4, s29, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v32, s4
-; SI-NEXT: s_lshr_b32 s4, s28, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v34, s4
-; SI-NEXT: s_lshr_b32 s4, s27, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v36, s4
-; SI-NEXT: s_lshr_b32 s4, s26, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v39, s4
-; SI-NEXT: s_lshr_b32 s4, s25, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v49, s4
-; SI-NEXT: s_lshr_b32 s4, s24, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v51, s4
-; SI-NEXT: s_lshr_b32 s4, s23, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v53, s4
-; SI-NEXT: s_lshr_b32 s4, s22, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v55, s4
-; SI-NEXT: s_lshr_b32 s4, s21, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v41, s4
-; SI-NEXT: s_lshr_b32 s4, s20, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v43, s4
-; SI-NEXT: s_lshr_b32 s4, s19, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v46, s4
-; SI-NEXT: s_lshr_b32 s4, s18, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v56, s4
-; SI-NEXT: s_lshr_b32 s4, s17, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v58, s4
-; SI-NEXT: s_lshr_b32 s4, s16, 16
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v16
+; SI-NEXT: v_cvt_f32_f16_e32 v56, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v15
+; SI-NEXT: v_cvt_f32_f16_e32 v46, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v14
+; SI-NEXT: v_cvt_f32_f16_e32 v44, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v13
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v12
+; SI-NEXT: v_cvt_f32_f16_e32 v40, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v11
+; SI-NEXT: v_cvt_f32_f16_e32 v55, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v10
+; SI-NEXT: v_cvt_f32_f16_e32 v53, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v9
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: s_waitcnt expcnt(5)
+; SI-NEXT: v_cvt_f32_f16_e32 v58, v34
; SI-NEXT: s_waitcnt expcnt(3)
-; SI-NEXT: v_cvt_f32_f16_e32 v60, s4
-; SI-NEXT: v_cvt_f32_f16_e32 v14, s9
-; SI-NEXT: v_cvt_f32_f16_e32 v38, s6
-; SI-NEXT: v_cvt_f32_f16_e32 v45, s7
-; SI-NEXT: v_cvt_f32_f16_e32 v47, s8
-; SI-NEXT: v_cvt_f32_f16_e32 v1, s10
-; SI-NEXT: s_waitcnt expcnt(2)
-; SI-NEXT: v_cvt_f32_f16_e32 v61, s11
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_cvt_f32_f16_e32 v62, s12
+; SI-NEXT: v_cvt_f32_f16_e32 v60, v35
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v63, s13
-; SI-NEXT: v_cvt_f32_f16_e32 v20, s14
-; SI-NEXT: v_cvt_f32_f16_e32 v21, s15
-; SI-NEXT: v_cvt_f32_f16_e32 v23, s40
-; SI-NEXT: v_cvt_f32_f16_e32 v25, s41
-; SI-NEXT: v_cvt_f32_f16_e32 v27, s42
-; SI-NEXT: v_cvt_f32_f16_e32 v29, s43
-; SI-NEXT: v_cvt_f32_f16_e32 v31, s44
-; SI-NEXT: v_cvt_f32_f16_e32 v33, s45
-; SI-NEXT: v_cvt_f32_f16_e32 v16, s29
-; SI-NEXT: v_cvt_f32_f16_e32 v37, s28
-; SI-NEXT: v_cvt_f32_f16_e32 v18, s27
-; SI-NEXT: v_cvt_f32_f16_e32 v48, s26
-; SI-NEXT: v_cvt_f32_f16_e32 v17, s25
-; SI-NEXT: v_cvt_f32_f16_e32 v52, s24
-; SI-NEXT: v_cvt_f32_f16_e32 v13, s23
-; SI-NEXT: v_cvt_f32_f16_e32 v40, s22
-; SI-NEXT: v_cvt_f32_f16_e32 v9, s21
-; SI-NEXT: v_cvt_f32_f16_e32 v44, s20
-; SI-NEXT: v_cvt_f32_f16_e32 v5, s19
-; SI-NEXT: v_cvt_f32_f16_e32 v3, s18
-; SI-NEXT: v_cvt_f32_f16_e32 v2, s17
-; SI-NEXT: v_cvt_f32_f16_e32 v7, s16
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v8
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: v_cvt_f32_f16_e32 v62, v36
+; SI-NEXT: v_cvt_f32_f16_e32 v18, v38
+; SI-NEXT: v_cvt_f32_f16_e32 v20, v39
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v7
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: v_cvt_f32_f16_e32 v22, v48
+; SI-NEXT: v_cvt_f32_f16_e32 v24, v49
+; SI-NEXT: v_cvt_f32_f16_e32 v26, v51
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v6
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: v_cvt_f32_f16_e32 v27, v50
+; SI-NEXT: v_cvt_f32_f16_e32 v28, v29
+; SI-NEXT: v_cvt_f32_f16_e32 v30, v31
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v5
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v4
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v3
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v2
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v1
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v32
+; SI-NEXT: v_cvt_f32_f16_e32 v52, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v33
+; SI-NEXT: v_cvt_f32_f16_e32 v54, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v34
+; SI-NEXT: v_cvt_f32_f16_e32 v41, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v35
+; SI-NEXT: v_cvt_f32_f16_e32 v43, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v36
+; SI-NEXT: v_cvt_f32_f16_e32 v45, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v37
+; SI-NEXT: v_cvt_f32_f16_e32 v47, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v38
+; SI-NEXT: v_cvt_f32_f16_e32 v57, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v39
+; SI-NEXT: v_cvt_f32_f16_e32 v59, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v48
+; SI-NEXT: v_cvt_f32_f16_e32 v61, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v49
+; SI-NEXT: v_cvt_f32_f16_e32 v63, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v51
+; SI-NEXT: v_cvt_f32_f16_e32 v19, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v50
+; SI-NEXT: v_cvt_f32_f16_e32 v21, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v29
+; SI-NEXT: v_cvt_f32_f16_e32 v23, v17
+; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v31
+; SI-NEXT: v_cvt_f32_f16_e32 v25, v17
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v16
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v15
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v14
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v13
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v12
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v11
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v10
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v9
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v8
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v7
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v6
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v5
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v4
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v3
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v2
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v1
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v32
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v33
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v37
; SI-NEXT: s_cbranch_execnz .LBB33_3
; SI-NEXT: .LBB33_2: ; %cmp.true
-; SI-NEXT: v_add_f32_e64 v6, s16, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v6
-; SI-NEXT: v_add_f32_e64 v2, s17, 1.0
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; SI-NEXT: v_add_f32_e32 v1, 1.0, v1
+; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v1
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT: v_add_f32_e32 v32, 1.0, v32
+; SI-NEXT: v_add_f32_e32 v33, 1.0, v33
+; SI-NEXT: v_add_f32_e32 v9, 1.0, v9
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v2
-; SI-NEXT: v_add_f32_e64 v3, s18, 1.0
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v32
+; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v9
+; SI-NEXT: v_add_f32_e32 v8, 1.0, v8
+; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v8
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v33
+; SI-NEXT: v_add_f32_e32 v7, 1.0, v7
+; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v7
+; SI-NEXT: v_add_f32_e32 v6, 1.0, v6
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v3
-; SI-NEXT: v_add_f32_e64 v5, s19, 1.0
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v63
+; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v6
+; SI-NEXT: v_add_f32_e32 v5, 1.0, v5
+; SI-NEXT: v_lshrrev_b32_e32 v47, 16, v5
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v5
-; SI-NEXT: v_add_f32_e64 v7, s20, 1.0
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v61
+; SI-NEXT: v_add_f32_e32 v4, 1.0, v4
+; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v4
+; SI-NEXT: v_add_f32_e32 v3, 1.0, v3
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v7
-; SI-NEXT: v_add_f32_e64 v16, s12, 1.0
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; SI-NEXT: v_add_f32_e64 v11, s22, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v16
-; SI-NEXT: v_add_f32_e64 v40, s6, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v11
-; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v40
-; SI-NEXT: v_cvt_f32_f16_e32 v38, v40
-; SI-NEXT: v_cvt_f32_f16_e32 v40, v11
-; SI-NEXT: v_cvt_f32_f16_e32 v11, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; SI-NEXT: v_add_f32_e64 v17, s25, 1.0
-; SI-NEXT: v_add_f32_e64 v28, s45, 1.0
-; SI-NEXT: v_add_f32_e64 v23, s40, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v17
-; SI-NEXT: v_add_f32_e64 v19, s26, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v28
-; SI-NEXT: v_lshrrev_b32_e32 v30, 16, v23
-; SI-NEXT: v_add_f32_e64 v48, s8, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v19
-; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v48
-; SI-NEXT: v_cvt_f32_f16_e32 v47, v48
-; SI-NEXT: v_cvt_f32_f16_e32 v48, v19
-; SI-NEXT: v_cvt_f32_f16_e32 v19, v30
-; SI-NEXT: v_cvt_f32_f16_e32 v30, v49
-; SI-NEXT: v_cvt_f32_f16_e32 v49, v43
-; SI-NEXT: v_add_f32_e64 v15, s24, 1.0
-; SI-NEXT: v_add_f32_e64 v26, s29, 1.0
-; SI-NEXT: v_add_f32_e64 v25, s41, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v15
-; SI-NEXT: v_add_f32_e64 v22, s27, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v26
-; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v25
-; SI-NEXT: v_add_f32_e64 v18, s13, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v22
-; SI-NEXT: v_lshrrev_b32_e32 v12, 16, v18
-; SI-NEXT: v_cvt_f32_f16_e32 v63, v18
-; SI-NEXT: v_cvt_f32_f16_e32 v18, v22
-; SI-NEXT: v_cvt_f32_f16_e32 v22, v32
-; SI-NEXT: v_cvt_f32_f16_e32 v32, v51
-; SI-NEXT: v_cvt_f32_f16_e32 v51, v46
-; SI-NEXT: v_add_f32_e64 v13, s23, 1.0
-; SI-NEXT: v_add_f32_e64 v24, s28, 1.0
-; SI-NEXT: v_add_f32_e64 v27, s42, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v13
-; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v24
-; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v27
-; SI-NEXT: v_add_f32_e64 v37, s10, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v57, 16, v37
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v59
+; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v3
+; SI-NEXT: v_add_f32_e32 v2, 1.0, v2
+; SI-NEXT: v_lshrrev_b32_e32 v41, 16, v2
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v1, v37
-; SI-NEXT: v_cvt_f32_f16_e32 v37, v24
-; SI-NEXT: v_cvt_f32_f16_e32 v24, v34
-; SI-NEXT: v_cvt_f32_f16_e32 v34, v53
-; SI-NEXT: v_cvt_f32_f16_e32 v53, v56
-; SI-NEXT: v_add_f32_e64 v29, s43, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v29
-; SI-NEXT: v_cvt_f32_f16_e32 v62, v16
-; SI-NEXT: v_cvt_f32_f16_e32 v16, v26
-; SI-NEXT: v_cvt_f32_f16_e32 v26, v36
-; SI-NEXT: v_cvt_f32_f16_e32 v36, v55
-; SI-NEXT: v_cvt_f32_f16_e32 v55, v58
-; SI-NEXT: v_add_f32_e64 v9, s21, 1.0
-; SI-NEXT: v_add_f32_e64 v31, s44, 1.0
-; SI-NEXT: v_add_f32_e64 v21, s15, 1.0
-; SI-NEXT: v_add_f32_e64 v20, s14, 1.0
-; SI-NEXT: v_add_f32_e64 v33, s11, 1.0
-; SI-NEXT: v_add_f32_e64 v52, s7, 1.0
-; SI-NEXT: v_add_f32_e64 v44, s9, 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
-; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v31
-; SI-NEXT: v_lshrrev_b32_e32 v8, 16, v21
-; SI-NEXT: v_lshrrev_b32_e32 v10, 16, v20
-; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v33
-; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v52
-; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v44
-; SI-NEXT: v_cvt_f32_f16_e32 v14, v44
-; SI-NEXT: v_cvt_f32_f16_e32 v45, v52
-; SI-NEXT: v_cvt_f32_f16_e32 v61, v33
-; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
-; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
-; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
-; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
-; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
-; SI-NEXT: v_cvt_f32_f16_e32 v33, v28
-; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
-; SI-NEXT: v_cvt_f32_f16_e32 v52, v15
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v57
+; SI-NEXT: v_add_f32_e32 v19, 1.0, v31
+; SI-NEXT: v_add_f32_e32 v21, 1.0, v29
+; SI-NEXT: v_add_f32_e32 v27, 1.0, v50
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v47
+; SI-NEXT: v_add_f32_e32 v26, 1.0, v51
+; SI-NEXT: v_add_f32_e32 v24, 1.0, v49
+; SI-NEXT: v_add_f32_e32 v22, 1.0, v48
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v45
+; SI-NEXT: v_add_f32_e32 v20, 1.0, v39
+; SI-NEXT: v_add_f32_e32 v18, 1.0, v38
+; SI-NEXT: v_add_f32_e32 v17, 1.0, v37
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v43
+; SI-NEXT: v_add_f32_e32 v28, 1.0, v36
+; SI-NEXT: v_add_f32_e32 v30, 1.0, v35
+; SI-NEXT: v_add_f32_e32 v34, 1.0, v34
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v41
+; SI-NEXT: v_add_f32_e32 v10, 1.0, v10
+; SI-NEXT: v_add_f32_e32 v11, 1.0, v11
+; SI-NEXT: v_add_f32_e32 v12, 1.0, v12
+; SI-NEXT: v_add_f32_e32 v13, 1.0, v13
+; SI-NEXT: v_add_f32_e32 v14, 1.0, v14
+; SI-NEXT: v_add_f32_e32 v15, 1.0, v15
+; SI-NEXT: v_add_f32_e32 v16, 1.0, v16
+; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v19
+; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v21
+; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v27
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v26
+; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v24
+; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v22
+; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v20
+; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v18
+; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v17
+; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v28
+; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v30
+; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v34
+; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v33
+; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v32
+; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v10
+; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v11
+; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v12
+; SI-NEXT: v_lshrrev_b32_e32 v42, 16, v13
+; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v14
+; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v15
+; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v16
+; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
+; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v43, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; SI-NEXT: v_cvt_f32_f16_e32 v44, v7
+; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
+; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
-; SI-NEXT: v_cvt_f32_f16_e32 v7, v6
-; SI-NEXT: v_cvt_f32_f16_e32 v59, v59
+; SI-NEXT: v_cvt_f32_f16_e32 v58, v34
+; SI-NEXT: v_cvt_f32_f16_e32 v60, v30
+; SI-NEXT: v_cvt_f32_f16_e32 v62, v28
+; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
+; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
+; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
+; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
+; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
+; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
+; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
+; SI-NEXT: v_cvt_f32_f16_e32 v28, v21
+; SI-NEXT: v_cvt_f32_f16_e32 v30, v19
+; SI-NEXT: v_cvt_f32_f16_e32 v56, v56
+; SI-NEXT: v_cvt_f32_f16_e32 v46, v46
+; SI-NEXT: v_cvt_f32_f16_e32 v44, v44
; SI-NEXT: v_cvt_f32_f16_e32 v42, v42
-; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
-; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
-; SI-NEXT: v_cvt_f32_f16_e32 v57, v57
-; SI-NEXT: v_cvt_f32_f16_e32 v35, v35
-; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
-; SI-NEXT: v_cvt_f32_f16_e32 v15, v10
-; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
-; SI-NEXT: v_cvt_f32_f16_e32 v28, v39
-; SI-NEXT: v_cvt_f32_f16_e32 v39, v41
-; SI-NEXT: v_cvt_f32_f16_e32 v41, v60
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v46, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v56, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v58, v4
-; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v60, v4
+; SI-NEXT: v_cvt_f32_f16_e32 v40, v40
+; SI-NEXT: v_cvt_f32_f16_e32 v55, v55
+; SI-NEXT: v_cvt_f32_f16_e32 v53, v53
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v54
+; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
+; SI-NEXT: v_cvt_f32_f16_e32 v54, v51
+; SI-NEXT: v_cvt_f32_f16_e32 v41, v50
+; SI-NEXT: v_cvt_f32_f16_e32 v43, v35
+; SI-NEXT: v_cvt_f32_f16_e32 v45, v36
+; SI-NEXT: v_cvt_f32_f16_e32 v47, v37
+; SI-NEXT: v_cvt_f32_f16_e32 v57, v38
+; SI-NEXT: v_cvt_f32_f16_e32 v59, v39
+; SI-NEXT: v_cvt_f32_f16_e32 v61, v48
+; SI-NEXT: v_cvt_f32_f16_e32 v63, v49
+; SI-NEXT: v_cvt_f32_f16_e32 v19, v31
+; SI-NEXT: v_cvt_f32_f16_e32 v21, v29
+; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
+; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
+; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT: .LBB33_3: ; %end
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v60
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v7
-; SI-NEXT: v_cvt_f16_f32_e32 v10, v2
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v6, v4
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v58
-; SI-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v4, vcc, 4, v0
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT: v_or_b32_e32 v6, v10, v6
-; SI-NEXT: buffer_store_dword v6, v4, s[0:3], 0 offen
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v56
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v25
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v30
+; SI-NEXT: v_cvt_f16_f32_e32 v3, v28
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v23
+; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v3
-; SI-NEXT: v_add_i32_e32 v10, vcc, 8, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v6, v4
-; SI-NEXT: buffer_store_dword v4, v10, s[0:3], 0 offen
+; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v3, v2
+; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v21
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v46
-; SI-NEXT: v_add_i32_e32 v6, vcc, 12, v0
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v54
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v27
+; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v43
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v44
-; SI-NEXT: v_add_i32_e32 v6, vcc, 16, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v19
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v26
+; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v41
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v9
-; SI-NEXT: v_add_i32_e32 v6, vcc, 20, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v63
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v24
+; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v55
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v40
-; SI-NEXT: v_add_i32_e32 v6, vcc, 24, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v61
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v22
+; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v53
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v13
-; SI-NEXT: v_add_i32_e32 v6, vcc, 28, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v20
+; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v51
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v52
-; SI-NEXT: v_add_i32_e32 v6, vcc, 32, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v57
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v18
+; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v49
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v17
-; SI-NEXT: v_add_i32_e32 v6, vcc, 36, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v47
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v17
+; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v39
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v48
-; SI-NEXT: v_add_i32_e32 v6, vcc, 40, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v45
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v62
+; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v36
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v18
-; SI-NEXT: v_add_i32_e32 v6, vcc, 44, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v43
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v60
+; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v34
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v37
-; SI-NEXT: v_add_i32_e32 v6, vcc, 48, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v41
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v58
+; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v32
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v16
-; SI-NEXT: v_add_i32_e32 v6, vcc, 52, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v54
+; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v30
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v33
-; SI-NEXT: v_add_i32_e32 v6, vcc, 56, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
+; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v28
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v31
-; SI-NEXT: v_add_i32_e32 v6, vcc, 60, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v26
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v29
-; SI-NEXT: v_add_i32_e32 v6, vcc, 64, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v24
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v27
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x44, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v22
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v25
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x48, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v19
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v23
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x4c, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v8
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v21
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x50, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v15
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v20
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x54, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v12
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v63
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x58, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v11
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v62
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x5c, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v35
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v61
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x60, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v53
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x5c, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v57
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v1
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x64, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v55
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x60, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v50
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v47
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x68, v0
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; SI-NEXT: v_or_b32_e32 v4, v5, v4
-; SI-NEXT: buffer_store_dword v4, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v40
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x64, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v45
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x6c, v0
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v42
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v42
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x68, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v38
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x70, v0
-; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; SI-NEXT: v_or_b32_e32 v2, v3, v2
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
-; SI-NEXT: buffer_store_dword v2, v4, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v44
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x6c, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v14
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v46
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x70, v0
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x74, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v56
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_or_b32_e32 v1, v2, v1
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT: buffer_load_dword v63, off, s[0:3], s32 ; 4-byte Folded Reload
@@ -22065,66 +21825,93 @@ define inreg <60 x half> @bitcast_v30f32_to_v60f16_scalar(<30 x float> inreg %a,
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
; SI-NEXT: .LBB33_4:
-; SI-NEXT: ; implicit-def: $vgpr7
-; SI-NEXT: ; implicit-def: $vgpr60
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr58
-; SI-NEXT: ; implicit-def: $vgpr3
-; SI-NEXT: ; implicit-def: $vgpr56
-; SI-NEXT: ; implicit-def: $vgpr5
-; SI-NEXT: ; implicit-def: $vgpr46
-; SI-NEXT: ; implicit-def: $vgpr44
-; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: ; implicit-def: $vgpr9
-; SI-NEXT: ; implicit-def: $vgpr41
-; SI-NEXT: ; implicit-def: $vgpr40
-; SI-NEXT: ; implicit-def: $vgpr55
-; SI-NEXT: ; implicit-def: $vgpr13
; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
; SI-NEXT: ; implicit-def: $vgpr52
-; SI-NEXT: ; implicit-def: $vgpr51
-; SI-NEXT: ; implicit-def: $vgpr17
-; SI-NEXT: ; implicit-def: $vgpr49
-; SI-NEXT: ; implicit-def: $vgpr48
-; SI-NEXT: ; implicit-def: $vgpr39
-; SI-NEXT: ; implicit-def: $vgpr18
-; SI-NEXT: ; implicit-def: $vgpr36
-; SI-NEXT: ; implicit-def: $vgpr37
-; SI-NEXT: ; implicit-def: $vgpr34
-; SI-NEXT: ; implicit-def: $vgpr16
-; SI-NEXT: ; implicit-def: $vgpr32
-; SI-NEXT: ; implicit-def: $vgpr33
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; kill: killed $vgpr52
+; SI-NEXT: ; implicit-def: $vgpr52
+; SI-NEXT: ; implicit-def: $vgpr55
+; SI-NEXT: ; implicit-def: $vgpr40
+; SI-NEXT: ; implicit-def: $vgpr42
+; SI-NEXT: ; implicit-def: $vgpr44
+; SI-NEXT: ; implicit-def: $vgpr46
+; SI-NEXT: ; implicit-def: $vgpr56
; SI-NEXT: ; implicit-def: $vgpr30
-; SI-NEXT: ; implicit-def: $vgpr31
-; SI-NEXT: ; implicit-def: $vgpr28
-; SI-NEXT: ; implicit-def: $vgpr29
-; SI-NEXT: ; implicit-def: $vgpr26
-; SI-NEXT: ; implicit-def: $vgpr27
-; SI-NEXT: ; implicit-def: $vgpr24
; SI-NEXT: ; implicit-def: $vgpr25
-; SI-NEXT: ; implicit-def: $vgpr22
+; SI-NEXT: ; implicit-def: $vgpr28
; SI-NEXT: ; implicit-def: $vgpr23
-; SI-NEXT: ; implicit-def: $vgpr19
+; SI-NEXT: ; implicit-def: $vgpr27
; SI-NEXT: ; implicit-def: $vgpr21
-; SI-NEXT: ; implicit-def: $vgpr8
-; SI-NEXT: ; implicit-def: $vgpr20
-; SI-NEXT: ; implicit-def: $vgpr15
+; SI-NEXT: ; implicit-def: $vgpr26
+; SI-NEXT: ; implicit-def: $vgpr19
+; SI-NEXT: ; implicit-def: $vgpr24
; SI-NEXT: ; implicit-def: $vgpr63
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr62
-; SI-NEXT: ; implicit-def: $vgpr11
+; SI-NEXT: ; implicit-def: $vgpr22
; SI-NEXT: ; implicit-def: $vgpr61
-; SI-NEXT: ; implicit-def: $vgpr35
-; SI-NEXT: ; implicit-def: $vgpr1
+; SI-NEXT: ; implicit-def: $vgpr20
+; SI-NEXT: ; implicit-def: $vgpr59
+; SI-NEXT: ; implicit-def: $vgpr18
; SI-NEXT: ; implicit-def: $vgpr57
+; SI-NEXT: ; implicit-def: $vgpr17
; SI-NEXT: ; implicit-def: $vgpr47
-; SI-NEXT: ; implicit-def: $vgpr50
+; SI-NEXT: ; implicit-def: $vgpr62
; SI-NEXT: ; implicit-def: $vgpr45
+; SI-NEXT: ; implicit-def: $vgpr60
+; SI-NEXT: ; implicit-def: $vgpr43
+; SI-NEXT: ; implicit-def: $vgpr58
+; SI-NEXT: ; implicit-def: $vgpr41
; SI-NEXT: ; implicit-def: $vgpr54
-; SI-NEXT: ; implicit-def: $vgpr38
+; SI-NEXT: ; kill: killed $vgpr52
+; SI-NEXT: ; kill: killed $vgpr55
+; SI-NEXT: ; implicit-def: $vgpr55
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr52
+; SI-NEXT: ; kill: killed $vgpr40
+; SI-NEXT: ; implicit-def: $vgpr40
+; SI-NEXT: ; kill: killed $vgpr42
; SI-NEXT: ; implicit-def: $vgpr42
-; SI-NEXT: ; implicit-def: $vgpr14
-; SI-NEXT: ; implicit-def: $vgpr59
+; SI-NEXT: ; kill: killed $vgpr44
+; SI-NEXT: ; implicit-def: $vgpr44
+; SI-NEXT: ; kill: killed $vgpr46
+; SI-NEXT: ; implicit-def: $vgpr46
+; SI-NEXT: ; kill: killed $vgpr56
+; SI-NEXT: ; implicit-def: $vgpr56
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
+; SI-NEXT: ; kill: killed $vgpr53
+; SI-NEXT: ; implicit-def: $vgpr53
; SI-NEXT: s_branch .LBB33_2
;
; VI-LABEL: bitcast_v30f32_to_v60f16_scalar:
@@ -23044,6 +22831,9 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(5)
+; SI-NEXT: v_cvt_f16_f32_e32 v58, v0
+; SI-NEXT: v_cvt_f16_f32_e32 v0, v5
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116
; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32
@@ -23072,23 +22862,12 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:88
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:84
-; SI-NEXT: v_cvt_f16_f32_e32 v58, v0
-; SI-NEXT: v_cvt_f16_f32_e32 v0, v5
; SI-NEXT: v_cvt_f16_f32_e32 v59, v1
; SI-NEXT: v_cvt_f16_f32_e32 v57, v3
-; SI-NEXT: v_cvt_f16_f32_e32 v56, v2
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:104
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:100
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v4
-; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
+; SI-NEXT: v_cvt_f16_f32_e32 v56, v2
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v7
@@ -23098,8 +22877,6 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v9
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v8
@@ -23121,9 +22898,18 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v14
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v17
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:104
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:100
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
+; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v16
@@ -23133,6 +22919,7 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v18
+; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v21
@@ -23170,7 +22957,6 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v61
; SI-NEXT: v_cvt_f16_f32_e32 v61, v49
-; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f16_f32_e32 v49, v55
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -23185,6 +22971,7 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v32
+; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f16_f32_e32 v32, v47
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -24056,7 +23843,6 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: v_mov_b32_e32 v61, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v28
@@ -24135,6 +23921,7 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v60
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v61
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
@@ -24299,6 +24086,9 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT: s_cbranch_execz .LBB34_4
; GFX9-NEXT: ; %bb.3: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
@@ -24313,9 +24103,6 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v38, v61, s6
; GFX9-NEXT: s_movk_i32 s7, 0x200
; GFX9-NEXT: v_perm_b32 v1, v37, v60, s6
@@ -24335,6 +24122,10 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
+; GFX9-NEXT: s_waitcnt vmcnt(14)
+; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: s_waitcnt vmcnt(15)
; GFX9-NEXT: v_perm_b32 v9, v9, v44, s6
; GFX9-NEXT: s_waitcnt vmcnt(14)
@@ -24363,10 +24154,6 @@ define <30 x float> @bitcast_v60f16_to_v30f32(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_perm_b32 v21, v21, v48, s6
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_perm_b32 v22, v22, v39, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
@@ -24687,12 +24474,35 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
; SI-NEXT: v_cvt_f16_f32_e32 v8, s26
; SI-NEXT: v_cvt_f16_f32_e32 v6, s29
; SI-NEXT: v_cvt_f16_f32_e32 v7, s28
+; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT: v_cvt_f16_f32_e32 v50, v54
; SI-NEXT: v_cvt_f16_f32_e32 v48, v48
; SI-NEXT: v_cvt_f16_f32_e32 v31, v40
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v33
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
@@ -24702,7 +24512,7 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v38
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v44
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -24711,7 +24521,7 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v46
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v47
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -24720,7 +24530,7 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v57
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v58
; SI-NEXT: v_cvt_f16_f32_e32 v58, s16
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
@@ -24731,38 +24541,12 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v60
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_cbranch_scc0 .LBB35_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT: v_or_b32_e32 v3, v10, v3
-; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v34
; SI-NEXT: v_mov_b32_e32 v33, v32
; SI-NEXT: v_or_b32_e32 v10, v32, v10
@@ -24787,12 +24571,12 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v59
; SI-NEXT: v_or_b32_e32 v1, v12, v1
; SI-NEXT: v_or_b32_e32 v2, v11, v2
@@ -24901,12 +24685,10 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
; SI-NEXT: v_mov_b32_e32 v40, v44
; SI-NEXT: s_cbranch_vccnz .LBB35_5
; SI-NEXT: ; %bb.4: ; %cmp.true
-; SI-NEXT: s_waitcnt expcnt(5)
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT: v_cvt_f32_f16_e32 v0, v59
; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v0, v59
; SI-NEXT: v_cvt_f32_f16_e32 v1, v58
; SI-NEXT: s_waitcnt vmcnt(3)
; SI-NEXT: v_cvt_f32_f16_e32 v8, v33
@@ -25692,304 +25474,115 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v156, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v157, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v158, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v159, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v168, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v169, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v170, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v171, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v172, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v173, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v174, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v175, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v184, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v185, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v186, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v187, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v188, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v189, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v190, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v191, s32
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v25, v7
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v191, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v190, v6 :: v_dual_mov_b32 v185, v4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v186, v3 :: v_dual_mov_b32 v187, v2
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v188, v1 :: v_dual_mov_b32 v189, v0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s28, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s19, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s17, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s1, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s0, 16
-; GFX11-TRUE16-NEXT: s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s40, s0, s40
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s41, s1, s41
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s2, s46
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s3, s45
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s16, s44
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s17, s43
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s18, s4
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s19, s5
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s20, s6
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s21, s7
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s22, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s23, s9
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s25, s11
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s26, s12
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s27, s13
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s28, s14
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s29, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB35_4
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s40 :: v_dual_mov_b32 v5, s0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s41 :: v_dual_mov_b32 v9, s1
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s2 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s3 :: v_dual_mov_b32 v35, s5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v44, s6 :: v_dual_mov_b32 v65, s8
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s7 :: v_dual_mov_b32 v77, s9
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v90, s10 :: v_dual_mov_b32 v119, s12
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v104, s11 :: v_dual_mov_b32 v135, s13
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v152, s14
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v170, s15
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB35_3
; GFX11-TRUE16-NEXT: .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s40 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s41 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v189, 0x200, v189 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v188, 0x200, v188 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v187, 0x200, v187 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v186, 0x200, v186 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v185, 0x200, v185 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v191, 0x200, v191 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v190, 0x200, v190 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v35, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v44, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v54, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v65, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v77, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v90, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v104, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v119, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v135, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v152, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v170, 0x200, s15 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: .LBB35_3: ; %end
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v20 :: v_dual_mov_b32 v6, v27
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v35 :: v_dual_mov_b32 v8, v44
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v77 :: v_dual_mov_b32 v12, v90
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v104
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, v135 :: v_dual_mov_b32 v16, v152
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v170 :: v_dual_mov_b32 v18, v189
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v188 :: v_dual_mov_b32 v20, v187
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v186 :: v_dual_mov_b32 v22, v185
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v191 :: v_dual_mov_b32 v24, v190
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v191, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v190, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v189, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v188, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v187, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v186, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v185, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v184, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v175, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v174, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v173, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v172, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v171, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v170, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v169, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v168, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v159, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v158, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v157, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v156, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:124
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:252
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:316
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v14
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v54 :: v_dual_mov_b32 v10, v65
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v119 :: v_dual_mov_b32 v27, v30
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB35_4:
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, v29 :: v_dual_mov_b32 v65, v28
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v66, v30
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v26 :: v_dual_mov_b32 v54, v25
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v53
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v65 :: v_dual_mov_b32 v29, v64
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v30, v66
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168_vgpr169_vgpr170_vgpr171_vgpr172_vgpr173_vgpr174_vgpr175_vgpr176_vgpr177_vgpr178_vgpr179_vgpr180_vgpr181_vgpr182_vgpr183_vgpr184
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_branch .LBB35_2
;
; GFX11-FAKE16-LABEL: bitcast_v60f16_to_v30f32_scalar:
@@ -28056,30 +27649,58 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; SI: ; %bb.0:
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT: s_mov_b64 exec, s[4:5]
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_writelane_b32 v18, s30, 0
-; SI-NEXT: v_writelane_b32 v18, s31, 1
-; SI-NEXT: v_writelane_b32 v18, s34, 2
-; SI-NEXT: v_writelane_b32 v18, s35, 3
-; SI-NEXT: v_writelane_b32 v18, s36, 4
-; SI-NEXT: v_writelane_b32 v18, s37, 5
-; SI-NEXT: v_writelane_b32 v18, s38, 6
-; SI-NEXT: v_writelane_b32 v18, s39, 7
-; SI-NEXT: v_writelane_b32 v18, s48, 8
-; SI-NEXT: v_writelane_b32 v18, s49, 9
-; SI-NEXT: v_writelane_b32 v18, s50, 10
-; SI-NEXT: v_writelane_b32 v18, s51, 11
-; SI-NEXT: v_writelane_b32 v18, s52, 12
-; SI-NEXT: v_writelane_b32 v18, s53, 13
-; SI-NEXT: v_writelane_b32 v18, s54, 14
+; SI-NEXT: v_writelane_b32 v20, s30, 0
+; SI-NEXT: v_writelane_b32 v20, s31, 1
+; SI-NEXT: v_writelane_b32 v20, s34, 2
+; SI-NEXT: v_writelane_b32 v20, s35, 3
+; SI-NEXT: v_writelane_b32 v20, s36, 4
+; SI-NEXT: v_writelane_b32 v20, s37, 5
+; SI-NEXT: v_writelane_b32 v20, s38, 6
+; SI-NEXT: v_writelane_b32 v20, s39, 7
+; SI-NEXT: v_writelane_b32 v20, s48, 8
+; SI-NEXT: v_mov_b32_e32 v18, s16
+; SI-NEXT: v_mov_b32_e32 v19, s17
+; SI-NEXT: v_writelane_b32 v20, s49, 9
+; SI-NEXT: v_readfirstlane_b32 s46, v18
+; SI-NEXT: v_mov_b32_e32 v18, s18
+; SI-NEXT: v_readfirstlane_b32 s47, v19
+; SI-NEXT: v_mov_b32_e32 v19, s19
+; SI-NEXT: v_writelane_b32 v20, s50, 10
+; SI-NEXT: v_readfirstlane_b32 s44, v18
+; SI-NEXT: v_mov_b32_e32 v18, s20
+; SI-NEXT: v_readfirstlane_b32 s45, v19
+; SI-NEXT: v_mov_b32_e32 v19, s21
+; SI-NEXT: v_writelane_b32 v20, s51, 11
+; SI-NEXT: v_readfirstlane_b32 s42, v18
+; SI-NEXT: v_mov_b32_e32 v18, s22
+; SI-NEXT: v_readfirstlane_b32 s43, v19
+; SI-NEXT: v_mov_b32_e32 v19, s23
+; SI-NEXT: v_writelane_b32 v20, s52, 12
+; SI-NEXT: v_readfirstlane_b32 s40, v18
+; SI-NEXT: v_mov_b32_e32 v18, s24
+; SI-NEXT: v_readfirstlane_b32 s41, v19
+; SI-NEXT: v_mov_b32_e32 v19, s25
+; SI-NEXT: v_writelane_b32 v20, s53, 13
+; SI-NEXT: v_readfirstlane_b32 s24, v18
+; SI-NEXT: v_mov_b32_e32 v18, s26
+; SI-NEXT: v_readfirstlane_b32 s25, v19
+; SI-NEXT: v_mov_b32_e32 v19, s27
+; SI-NEXT: v_writelane_b32 v20, s54, 14
+; SI-NEXT: v_readfirstlane_b32 s22, v18
+; SI-NEXT: v_mov_b32_e32 v18, s28
+; SI-NEXT: v_readfirstlane_b32 s23, v19
+; SI-NEXT: v_mov_b32_e32 v19, s29
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; SI-NEXT: v_writelane_b32 v18, s55, 15
-; SI-NEXT: v_readfirstlane_b32 s42, v1
-; SI-NEXT: v_readfirstlane_b32 s43, v2
-; SI-NEXT: v_readfirstlane_b32 s40, v3
-; SI-NEXT: v_readfirstlane_b32 s41, v4
+; SI-NEXT: v_writelane_b32 v20, s55, 15
+; SI-NEXT: v_readfirstlane_b32 s20, v18
+; SI-NEXT: v_readfirstlane_b32 s21, v19
+; SI-NEXT: v_readfirstlane_b32 s18, v1
+; SI-NEXT: v_readfirstlane_b32 s19, v2
+; SI-NEXT: v_readfirstlane_b32 s16, v3
+; SI-NEXT: v_readfirstlane_b32 s17, v4
; SI-NEXT: v_readfirstlane_b32 s14, v5
; SI-NEXT: v_readfirstlane_b32 s15, v6
; SI-NEXT: v_readfirstlane_b32 s12, v7
@@ -28091,9 +27712,9 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; SI-NEXT: v_readfirstlane_b32 s6, v13
; SI-NEXT: v_readfirstlane_b32 s7, v14
; SI-NEXT: v_readfirstlane_b32 s4, v15
-; SI-NEXT: s_and_b64 s[44:45], vcc, exec
+; SI-NEXT: s_and_b64 s[26:27], vcc, exec
; SI-NEXT: v_readfirstlane_b32 s5, v16
-; SI-NEXT: v_writelane_b32 v18, s64, 16
+; SI-NEXT: v_writelane_b32 v20, s64, 16
; SI-NEXT: s_cbranch_scc0 .LBB41_4
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: s_lshr_b32 s34, s5, 16
@@ -28102,30 +27723,30 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; SI-NEXT: s_lshr_b32 s37, s11, 16
; SI-NEXT: s_lshr_b32 s38, s13, 16
; SI-NEXT: s_lshr_b32 s39, s15, 16
-; SI-NEXT: s_lshr_b32 s48, s41, 16
-; SI-NEXT: s_lshr_b32 s49, s43, 16
-; SI-NEXT: s_lshr_b32 s50, s29, 16
-; SI-NEXT: s_lshr_b32 s51, s27, 16
+; SI-NEXT: s_lshr_b32 s48, s17, 16
+; SI-NEXT: s_lshr_b32 s49, s19, 16
+; SI-NEXT: s_lshr_b32 s50, s21, 16
+; SI-NEXT: s_lshr_b32 s51, s23, 16
; SI-NEXT: s_lshr_b32 s52, s25, 16
-; SI-NEXT: s_lshr_b32 s53, s23, 16
-; SI-NEXT: s_lshr_b32 s54, s21, 16
-; SI-NEXT: s_lshr_b32 s55, s19, 16
-; SI-NEXT: s_lshr_b32 s64, s17, 16
-; SI-NEXT: s_lshr_b64 s[44:45], s[4:5], 16
-; SI-NEXT: s_lshr_b64 s[46:47], s[6:7], 16
+; SI-NEXT: s_lshr_b32 s53, s41, 16
+; SI-NEXT: s_lshr_b32 s54, s43, 16
+; SI-NEXT: s_lshr_b32 s55, s45, 16
+; SI-NEXT: s_lshr_b32 s64, s47, 16
+; SI-NEXT: s_lshr_b64 s[26:27], s[4:5], 16
+; SI-NEXT: s_lshr_b64 s[28:29], s[6:7], 16
; SI-NEXT: s_lshr_b64 s[56:57], s[8:9], 16
; SI-NEXT: s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT: s_lshr_b64 s[60:61], s[12:13], 16
; SI-NEXT: s_lshr_b64 s[62:63], s[14:15], 16
-; SI-NEXT: s_lshr_b64 s[72:73], s[40:41], 16
-; SI-NEXT: s_lshr_b64 s[74:75], s[42:43], 16
-; SI-NEXT: s_lshr_b64 s[76:77], s[28:29], 16
-; SI-NEXT: s_lshr_b64 s[78:79], s[26:27], 16
+; SI-NEXT: s_lshr_b64 s[72:73], s[16:17], 16
+; SI-NEXT: s_lshr_b64 s[74:75], s[18:19], 16
+; SI-NEXT: s_lshr_b64 s[76:77], s[20:21], 16
+; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
; SI-NEXT: s_lshr_b64 s[88:89], s[24:25], 16
-; SI-NEXT: s_lshr_b64 s[90:91], s[22:23], 16
-; SI-NEXT: s_lshr_b64 s[92:93], s[20:21], 16
-; SI-NEXT: s_lshr_b64 s[94:95], s[18:19], 16
-; SI-NEXT: s_lshr_b64 s[30:31], s[16:17], 16
+; SI-NEXT: s_lshr_b64 s[90:91], s[40:41], 16
+; SI-NEXT: s_lshr_b64 s[92:93], s[42:43], 16
+; SI-NEXT: s_lshr_b64 s[94:95], s[44:45], 16
+; SI-NEXT: s_lshr_b64 s[30:31], s[46:47], 16
; SI-NEXT: s_cbranch_execnz .LBB41_3
; SI-NEXT: .LBB41_2: ; %cmp.true
; SI-NEXT: s_add_u32 s4, s4, 3
@@ -28140,167 +27761,167 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; SI-NEXT: s_addc_u32 s13, s13, 0
; SI-NEXT: s_add_u32 s14, s14, 3
; SI-NEXT: s_addc_u32 s15, s15, 0
+; SI-NEXT: s_add_u32 s16, s16, 3
+; SI-NEXT: s_addc_u32 s17, s17, 0
+; SI-NEXT: s_add_u32 s18, s18, 3
+; SI-NEXT: s_addc_u32 s19, s19, 0
+; SI-NEXT: s_add_u32 s20, s20, 3
+; SI-NEXT: s_addc_u32 s21, s21, 0
+; SI-NEXT: s_add_u32 s22, s22, 3
+; SI-NEXT: s_addc_u32 s23, s23, 0
+; SI-NEXT: s_add_u32 s24, s24, 3
+; SI-NEXT: s_addc_u32 s25, s25, 0
; SI-NEXT: s_add_u32 s40, s40, 3
; SI-NEXT: s_addc_u32 s41, s41, 0
; SI-NEXT: s_add_u32 s42, s42, 3
; SI-NEXT: s_addc_u32 s43, s43, 0
-; SI-NEXT: s_add_u32 s28, s28, 3
-; SI-NEXT: s_addc_u32 s29, s29, 0
-; SI-NEXT: s_add_u32 s26, s26, 3
-; SI-NEXT: s_addc_u32 s27, s27, 0
-; SI-NEXT: s_add_u32 s24, s24, 3
-; SI-NEXT: s_addc_u32 s25, s25, 0
-; SI-NEXT: s_add_u32 s22, s22, 3
-; SI-NEXT: s_addc_u32 s23, s23, 0
-; SI-NEXT: s_add_u32 s20, s20, 3
-; SI-NEXT: s_addc_u32 s21, s21, 0
-; SI-NEXT: s_add_u32 s18, s18, 3
-; SI-NEXT: s_addc_u32 s19, s19, 0
-; SI-NEXT: s_add_u32 s16, s16, 3
-; SI-NEXT: s_addc_u32 s17, s17, 0
+; SI-NEXT: s_add_u32 s44, s44, 3
+; SI-NEXT: s_addc_u32 s45, s45, 0
+; SI-NEXT: s_add_u32 s46, s46, 3
+; SI-NEXT: s_addc_u32 s47, s47, 0
; SI-NEXT: s_lshr_b32 s34, s5, 16
; SI-NEXT: s_lshr_b32 s35, s7, 16
; SI-NEXT: s_lshr_b32 s36, s9, 16
; SI-NEXT: s_lshr_b32 s37, s11, 16
; SI-NEXT: s_lshr_b32 s38, s13, 16
; SI-NEXT: s_lshr_b32 s39, s15, 16
-; SI-NEXT: s_lshr_b32 s48, s41, 16
-; SI-NEXT: s_lshr_b32 s49, s43, 16
-; SI-NEXT: s_lshr_b32 s50, s29, 16
-; SI-NEXT: s_lshr_b32 s51, s27, 16
+; SI-NEXT: s_lshr_b32 s48, s17, 16
+; SI-NEXT: s_lshr_b32 s49, s19, 16
+; SI-NEXT: s_lshr_b32 s50, s21, 16
+; SI-NEXT: s_lshr_b32 s51, s23, 16
; SI-NEXT: s_lshr_b32 s52, s25, 16
-; SI-NEXT: s_lshr_b32 s53, s23, 16
-; SI-NEXT: s_lshr_b32 s54, s21, 16
-; SI-NEXT: s_lshr_b32 s55, s19, 16
-; SI-NEXT: s_lshr_b32 s64, s17, 16
-; SI-NEXT: s_lshr_b64 s[44:45], s[4:5], 16
-; SI-NEXT: s_lshr_b64 s[46:47], s[6:7], 16
+; SI-NEXT: s_lshr_b32 s53, s41, 16
+; SI-NEXT: s_lshr_b32 s54, s43, 16
+; SI-NEXT: s_lshr_b32 s55, s45, 16
+; SI-NEXT: s_lshr_b32 s64, s47, 16
+; SI-NEXT: s_lshr_b64 s[26:27], s[4:5], 16
+; SI-NEXT: s_lshr_b64 s[28:29], s[6:7], 16
; SI-NEXT: s_lshr_b64 s[56:57], s[8:9], 16
; SI-NEXT: s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT: s_lshr_b64 s[60:61], s[12:13], 16
; SI-NEXT: s_lshr_b64 s[62:63], s[14:15], 16
-; SI-NEXT: s_lshr_b64 s[72:73], s[40:41], 16
-; SI-NEXT: s_lshr_b64 s[74:75], s[42:43], 16
-; SI-NEXT: s_lshr_b64 s[76:77], s[28:29], 16
-; SI-NEXT: s_lshr_b64 s[78:79], s[26:27], 16
+; SI-NEXT: s_lshr_b64 s[72:73], s[16:17], 16
+; SI-NEXT: s_lshr_b64 s[74:75], s[18:19], 16
+; SI-NEXT: s_lshr_b64 s[76:77], s[20:21], 16
+; SI-NEXT: s_lshr_b64 s[78:79], s[22:23], 16
; SI-NEXT: s_lshr_b64 s[88:89], s[24:25], 16
-; SI-NEXT: s_lshr_b64 s[90:91], s[22:23], 16
-; SI-NEXT: s_lshr_b64 s[92:93], s[20:21], 16
-; SI-NEXT: s_lshr_b64 s[94:95], s[18:19], 16
-; SI-NEXT: s_lshr_b64 s[30:31], s[16:17], 16
+; SI-NEXT: s_lshr_b64 s[90:91], s[40:41], 16
+; SI-NEXT: s_lshr_b64 s[92:93], s[42:43], 16
+; SI-NEXT: s_lshr_b64 s[94:95], s[44:45], 16
+; SI-NEXT: s_lshr_b64 s[30:31], s[46:47], 16
; SI-NEXT: .LBB41_3: ; %end
-; SI-NEXT: s_lshl_b32 s45, s30, 16
-; SI-NEXT: s_and_b32 s16, s16, 0xffff
-; SI-NEXT: s_or_b32 s16, s16, s45
-; SI-NEXT: v_mov_b32_e32 v1, s16
-; SI-NEXT: s_and_b32 s16, s17, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s64, 16
-; SI-NEXT: s_or_b32 s16, s16, s17
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_lshl_b32 s16, s94, 16
-; SI-NEXT: s_and_b32 s17, s18, 0xffff
-; SI-NEXT: s_or_b32 s16, s17, s16
-; SI-NEXT: v_mov_b32_e32 v3, s16
-; SI-NEXT: s_and_b32 s16, s19, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s55, 16
-; SI-NEXT: s_or_b32 s16, s16, s17
-; SI-NEXT: v_mov_b32_e32 v4, s16
-; SI-NEXT: s_lshl_b32 s16, s92, 16
-; SI-NEXT: s_and_b32 s17, s20, 0xffff
-; SI-NEXT: s_or_b32 s16, s17, s16
-; SI-NEXT: v_mov_b32_e32 v5, s16
-; SI-NEXT: s_and_b32 s16, s21, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s54, 16
+; SI-NEXT: s_lshl_b32 s27, s30, 16
+; SI-NEXT: s_and_b32 s29, s46, 0xffff
+; SI-NEXT: s_or_b32 s27, s29, s27
+; SI-NEXT: v_mov_b32_e32 v1, s27
+; SI-NEXT: s_and_b32 s27, s47, 0xffff
+; SI-NEXT: s_lshl_b32 s29, s64, 16
+; SI-NEXT: s_or_b32 s27, s27, s29
+; SI-NEXT: v_mov_b32_e32 v2, s27
+; SI-NEXT: s_lshl_b32 s27, s94, 16
+; SI-NEXT: s_and_b32 s29, s44, 0xffff
+; SI-NEXT: s_or_b32 s27, s29, s27
+; SI-NEXT: v_mov_b32_e32 v3, s27
+; SI-NEXT: s_and_b32 s27, s45, 0xffff
+; SI-NEXT: s_lshl_b32 s29, s55, 16
+; SI-NEXT: s_or_b32 s27, s27, s29
+; SI-NEXT: v_mov_b32_e32 v4, s27
+; SI-NEXT: s_lshl_b32 s27, s92, 16
+; SI-NEXT: s_and_b32 s29, s42, 0xffff
+; SI-NEXT: s_or_b32 s27, s29, s27
+; SI-NEXT: v_mov_b32_e32 v5, s27
+; SI-NEXT: s_and_b32 s27, s43, 0xffff
+; SI-NEXT: s_lshl_b32 s29, s54, 16
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s27, s27, s29
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 8, v0
-; SI-NEXT: v_mov_b32_e32 v6, s16
+; SI-NEXT: v_mov_b32_e32 v6, s27
; SI-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 12, v0
-; SI-NEXT: s_lshl_b32 s16, s90, 16
-; SI-NEXT: s_and_b32 s17, s22, 0xffff
+; SI-NEXT: s_lshl_b32 s27, s90, 16
+; SI-NEXT: s_and_b32 s29, s40, 0xffff
; SI-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 16, v0
-; SI-NEXT: s_or_b32 s16, s17, s16
+; SI-NEXT: s_or_b32 s27, s29, s27
; SI-NEXT: buffer_store_dword v5, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 20, v0
; SI-NEXT: s_waitcnt expcnt(3)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s23, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s53, 16
+; SI-NEXT: v_mov_b32_e32 v2, s27
+; SI-NEXT: s_and_b32 s27, s41, 0xffff
+; SI-NEXT: s_lshl_b32 s29, s53, 16
; SI-NEXT: buffer_store_dword v6, v1, s[0:3], 0 offen
; SI-NEXT: v_add_i32_e32 v1, vcc, 24, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s27, s27, s29
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s24, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s88, 16
+; SI-NEXT: v_mov_b32_e32 v2, s27
+; SI-NEXT: s_and_b32 s24, s24, 0xffff
+; SI-NEXT: s_lshl_b32 s27, s88, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 28, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s24, s24, s27
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s25, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s52, 16
+; SI-NEXT: v_mov_b32_e32 v2, s24
+; SI-NEXT: s_and_b32 s24, s25, 0xffff
+; SI-NEXT: s_lshl_b32 s25, s52, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 32, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s24, s24, s25
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s26, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s78, 16
+; SI-NEXT: v_mov_b32_e32 v2, s24
+; SI-NEXT: s_and_b32 s22, s22, 0xffff
+; SI-NEXT: s_lshl_b32 s24, s78, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 36, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s22, s22, s24
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s27, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s51, 16
+; SI-NEXT: v_mov_b32_e32 v2, s22
+; SI-NEXT: s_and_b32 s22, s23, 0xffff
+; SI-NEXT: s_lshl_b32 s23, s51, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 40, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s22, s22, s23
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s28, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s76, 16
+; SI-NEXT: v_mov_b32_e32 v2, s22
+; SI-NEXT: s_and_b32 s20, s20, 0xffff
+; SI-NEXT: s_lshl_b32 s22, s76, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 44, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s20, s20, s22
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s29, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s50, 16
+; SI-NEXT: v_mov_b32_e32 v2, s20
+; SI-NEXT: s_and_b32 s20, s21, 0xffff
+; SI-NEXT: s_lshl_b32 s21, s50, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 48, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s20, s20, s21
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s42, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s74, 16
+; SI-NEXT: v_mov_b32_e32 v2, s20
+; SI-NEXT: s_and_b32 s18, s18, 0xffff
+; SI-NEXT: s_lshl_b32 s20, s74, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 52, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s18, s18, s20
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s43, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s49, 16
+; SI-NEXT: v_mov_b32_e32 v2, s18
+; SI-NEXT: s_and_b32 s18, s19, 0xffff
+; SI-NEXT: s_lshl_b32 s19, s49, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 56, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s18, s18, s19
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s40, 0xffff
-; SI-NEXT: s_lshl_b32 s17, s72, 16
+; SI-NEXT: v_mov_b32_e32 v2, s18
+; SI-NEXT: s_and_b32 s16, s16, 0xffff
+; SI-NEXT: s_lshl_b32 s18, s72, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 60, v0
-; SI-NEXT: s_or_b32 s16, s16, s17
+; SI-NEXT: s_or_b32 s16, s16, s18
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v2, s16
-; SI-NEXT: s_and_b32 s16, s41, 0xffff
+; SI-NEXT: s_and_b32 s16, s17, 0xffff
; SI-NEXT: s_lshl_b32 s17, s48, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 64, v0
; SI-NEXT: s_or_b32 s16, s16, s17
@@ -28364,7 +27985,7 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v2, s8
; SI-NEXT: s_and_b32 s6, s6, 0xffff
-; SI-NEXT: s_lshl_b32 s8, s46, 16
+; SI-NEXT: s_lshl_b32 s8, s28, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x64, v0
; SI-NEXT: s_or_b32 s6, s6, s8
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
@@ -28378,7 +27999,7 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_mov_b32_e32 v2, s6
; SI-NEXT: s_and_b32 s4, s4, 0xffff
-; SI-NEXT: s_lshl_b32 s6, s44, 16
+; SI-NEXT: s_lshl_b32 s6, s26, 16
; SI-NEXT: v_add_i32_e32 v1, vcc, 0x6c, v0
; SI-NEXT: s_or_b32 s4, s4, s6
; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
@@ -28392,25 +28013,25 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x74, v0
; SI-NEXT: v_mov_b32_e32 v1, s4
; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
-; SI-NEXT: v_readlane_b32 s64, v18, 16
-; SI-NEXT: v_readlane_b32 s55, v18, 15
-; SI-NEXT: v_readlane_b32 s54, v18, 14
-; SI-NEXT: v_readlane_b32 s53, v18, 13
-; SI-NEXT: v_readlane_b32 s52, v18, 12
-; SI-NEXT: v_readlane_b32 s51, v18, 11
-; SI-NEXT: v_readlane_b32 s50, v18, 10
-; SI-NEXT: v_readlane_b32 s49, v18, 9
-; SI-NEXT: v_readlane_b32 s48, v18, 8
-; SI-NEXT: v_readlane_b32 s39, v18, 7
-; SI-NEXT: v_readlane_b32 s38, v18, 6
-; SI-NEXT: v_readlane_b32 s37, v18, 5
-; SI-NEXT: v_readlane_b32 s36, v18, 4
-; SI-NEXT: v_readlane_b32 s35, v18, 3
-; SI-NEXT: v_readlane_b32 s34, v18, 2
-; SI-NEXT: v_readlane_b32 s31, v18, 1
-; SI-NEXT: v_readlane_b32 s30, v18, 0
+; SI-NEXT: v_readlane_b32 s64, v20, 16
+; SI-NEXT: v_readlane_b32 s55, v20, 15
+; SI-NEXT: v_readlane_b32 s54, v20, 14
+; SI-NEXT: v_readlane_b32 s53, v20, 13
+; SI-NEXT: v_readlane_b32 s52, v20, 12
+; SI-NEXT: v_readlane_b32 s51, v20, 11
+; SI-NEXT: v_readlane_b32 s50, v20, 10
+; SI-NEXT: v_readlane_b32 s49, v20, 9
+; SI-NEXT: v_readlane_b32 s48, v20, 8
+; SI-NEXT: v_readlane_b32 s39, v20, 7
+; SI-NEXT: v_readlane_b32 s38, v20, 6
+; SI-NEXT: v_readlane_b32 s37, v20, 5
+; SI-NEXT: v_readlane_b32 s36, v20, 4
+; SI-NEXT: v_readlane_b32 s35, v20, 3
+; SI-NEXT: v_readlane_b32 s34, v20, 2
+; SI-NEXT: v_readlane_b32 s31, v20, 1
+; SI-NEXT: v_readlane_b32 s30, v20, 0
; SI-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT: s_mov_b64 exec, s[4:5]
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
@@ -28441,10 +28062,10 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; SI-NEXT: ; implicit-def: $sgpr37
; SI-NEXT: ; implicit-def: $sgpr56
; SI-NEXT: ; implicit-def: $sgpr36
-; SI-NEXT: ; implicit-def: $sgpr46
+; SI-NEXT: ; implicit-def: $sgpr28
; SI-NEXT: ; implicit-def: $sgpr35
; SI-NEXT: ; implicit-def: $sgpr34
-; SI-NEXT: ; implicit-def: $sgpr44
+; SI-NEXT: ; implicit-def: $sgpr26
; SI-NEXT: s_branch .LBB41_2
;
; VI-LABEL: bitcast_v15i64_to_v60i16_scalar:
@@ -28455,18 +28076,46 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; VI-NEXT: s_mov_b64 exec, s[4:5]
; VI-NEXT: v_writelane_b32 v30, s30, 0
; VI-NEXT: v_writelane_b32 v30, s31, 1
+; VI-NEXT: v_mov_b32_e32 v17, s16
+; VI-NEXT: v_mov_b32_e32 v18, s17
; VI-NEXT: v_writelane_b32 v30, s34, 2
+; VI-NEXT: v_mov_b32_e32 v19, s18
+; VI-NEXT: v_readfirstlane_b32 s56, v17
+; VI-NEXT: v_mov_b32_e32 v17, s19
+; VI-NEXT: v_readfirstlane_b32 s47, v18
+; VI-NEXT: v_mov_b32_e32 v18, s20
; VI-NEXT: v_writelane_b32 v30, s35, 3
+; VI-NEXT: v_readfirstlane_b32 s46, v19
+; VI-NEXT: v_mov_b32_e32 v19, s21
+; VI-NEXT: v_readfirstlane_b32 s45, v17
+; VI-NEXT: v_mov_b32_e32 v17, s22
+; VI-NEXT: v_readfirstlane_b32 s44, v18
+; VI-NEXT: v_mov_b32_e32 v18, s23
; VI-NEXT: v_writelane_b32 v30, s36, 4
+; VI-NEXT: v_readfirstlane_b32 s43, v19
+; VI-NEXT: v_mov_b32_e32 v19, s24
+; VI-NEXT: v_readfirstlane_b32 s42, v17
+; VI-NEXT: v_mov_b32_e32 v17, s25
+; VI-NEXT: v_readfirstlane_b32 s41, v18
+; VI-NEXT: v_mov_b32_e32 v18, s26
; VI-NEXT: v_writelane_b32 v30, s37, 5
+; VI-NEXT: v_readfirstlane_b32 s40, v19
+; VI-NEXT: v_mov_b32_e32 v19, s27
+; VI-NEXT: v_readfirstlane_b32 s26, v17
+; VI-NEXT: v_mov_b32_e32 v17, s28
+; VI-NEXT: v_readfirstlane_b32 s25, v18
+; VI-NEXT: v_mov_b32_e32 v18, s29
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; VI-NEXT: v_writelane_b32 v30, s38, 6
-; VI-NEXT: v_readfirstlane_b32 s45, v0
-; VI-NEXT: v_readfirstlane_b32 s44, v1
-; VI-NEXT: v_readfirstlane_b32 s43, v2
-; VI-NEXT: v_readfirstlane_b32 s42, v3
-; VI-NEXT: v_readfirstlane_b32 s41, v4
-; VI-NEXT: v_readfirstlane_b32 s40, v5
+; VI-NEXT: v_readfirstlane_b32 s24, v19
+; VI-NEXT: v_readfirstlane_b32 s23, v17
+; VI-NEXT: v_readfirstlane_b32 s22, v18
+; VI-NEXT: v_readfirstlane_b32 s21, v0
+; VI-NEXT: v_readfirstlane_b32 s20, v1
+; VI-NEXT: v_readfirstlane_b32 s19, v2
+; VI-NEXT: v_readfirstlane_b32 s18, v3
+; VI-NEXT: v_readfirstlane_b32 s17, v4
+; VI-NEXT: v_readfirstlane_b32 s16, v5
; VI-NEXT: v_readfirstlane_b32 s15, v6
; VI-NEXT: v_readfirstlane_b32 s14, v7
; VI-NEXT: v_readfirstlane_b32 s13, v8
@@ -28481,9 +28130,9 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; VI-NEXT: v_writelane_b32 v30, s39, 7
; VI-NEXT: s_cbranch_scc0 .LBB41_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_lshr_b32 s47, s6, 16
-; VI-NEXT: s_lshr_b32 s56, s8, 16
+; VI-NEXT: s_lshr_b32 s27, s7, 16
+; VI-NEXT: s_lshr_b32 s28, s6, 16
+; VI-NEXT: s_lshr_b32 s29, s8, 16
; VI-NEXT: s_lshr_b32 s57, s9, 16
; VI-NEXT: s_lshr_b32 s58, s10, 16
; VI-NEXT: s_lshr_b32 s59, s11, 16
@@ -28491,26 +28140,26 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; VI-NEXT: s_lshr_b32 s61, s13, 16
; VI-NEXT: s_lshr_b32 s62, s14, 16
; VI-NEXT: s_lshr_b32 s63, s15, 16
-; VI-NEXT: s_lshr_b32 s72, s40, 16
-; VI-NEXT: s_lshr_b32 s73, s41, 16
-; VI-NEXT: s_lshr_b32 s74, s42, 16
-; VI-NEXT: s_lshr_b32 s75, s43, 16
-; VI-NEXT: s_lshr_b32 s76, s44, 16
-; VI-NEXT: s_lshr_b32 s77, s45, 16
-; VI-NEXT: s_lshr_b32 s78, s29, 16
-; VI-NEXT: s_lshr_b32 s79, s28, 16
-; VI-NEXT: s_lshr_b32 s88, s27, 16
-; VI-NEXT: s_lshr_b32 s89, s26, 16
-; VI-NEXT: s_lshr_b32 s90, s25, 16
-; VI-NEXT: s_lshr_b32 s91, s24, 16
-; VI-NEXT: s_lshr_b32 s30, s23, 16
-; VI-NEXT: s_lshr_b32 s31, s22, 16
-; VI-NEXT: s_lshr_b32 s34, s21, 16
-; VI-NEXT: s_lshr_b32 s35, s20, 16
-; VI-NEXT: s_lshr_b32 s36, s19, 16
-; VI-NEXT: s_lshr_b32 s37, s18, 16
-; VI-NEXT: s_lshr_b32 s38, s17, 16
-; VI-NEXT: s_lshr_b32 s39, s16, 16
+; VI-NEXT: s_lshr_b32 s72, s16, 16
+; VI-NEXT: s_lshr_b32 s73, s17, 16
+; VI-NEXT: s_lshr_b32 s74, s18, 16
+; VI-NEXT: s_lshr_b32 s75, s19, 16
+; VI-NEXT: s_lshr_b32 s76, s20, 16
+; VI-NEXT: s_lshr_b32 s77, s21, 16
+; VI-NEXT: s_lshr_b32 s78, s22, 16
+; VI-NEXT: s_lshr_b32 s79, s23, 16
+; VI-NEXT: s_lshr_b32 s88, s24, 16
+; VI-NEXT: s_lshr_b32 s89, s25, 16
+; VI-NEXT: s_lshr_b32 s90, s26, 16
+; VI-NEXT: s_lshr_b32 s91, s40, 16
+; VI-NEXT: s_lshr_b32 s30, s41, 16
+; VI-NEXT: s_lshr_b32 s31, s42, 16
+; VI-NEXT: s_lshr_b32 s34, s43, 16
+; VI-NEXT: s_lshr_b32 s35, s44, 16
+; VI-NEXT: s_lshr_b32 s36, s45, 16
+; VI-NEXT: s_lshr_b32 s37, s46, 16
+; VI-NEXT: s_lshr_b32 s38, s47, 16
+; VI-NEXT: s_lshr_b32 s39, s56, 16
; VI-NEXT: s_cbranch_execnz .LBB41_3
; VI-NEXT: .LBB41_2: ; %cmp.true
; VI-NEXT: s_add_u32 s6, s6, 3
@@ -28523,29 +28172,29 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; VI-NEXT: s_addc_u32 s12, s12, 0
; VI-NEXT: s_add_u32 s15, s15, 3
; VI-NEXT: s_addc_u32 s14, s14, 0
-; VI-NEXT: s_add_u32 s41, s41, 3
-; VI-NEXT: s_addc_u32 s40, s40, 0
-; VI-NEXT: s_add_u32 s43, s43, 3
-; VI-NEXT: s_addc_u32 s42, s42, 0
-; VI-NEXT: s_add_u32 s45, s45, 3
-; VI-NEXT: s_addc_u32 s44, s44, 0
-; VI-NEXT: s_add_u32 s28, s28, 3
-; VI-NEXT: s_addc_u32 s29, s29, 0
-; VI-NEXT: s_add_u32 s26, s26, 3
-; VI-NEXT: s_addc_u32 s27, s27, 0
-; VI-NEXT: s_add_u32 s24, s24, 3
-; VI-NEXT: s_addc_u32 s25, s25, 0
-; VI-NEXT: s_add_u32 s22, s22, 3
-; VI-NEXT: s_addc_u32 s23, s23, 0
-; VI-NEXT: s_add_u32 s20, s20, 3
-; VI-NEXT: s_addc_u32 s21, s21, 0
-; VI-NEXT: s_add_u32 s18, s18, 3
-; VI-NEXT: s_addc_u32 s19, s19, 0
-; VI-NEXT: s_add_u32 s16, s16, 3
-; VI-NEXT: s_addc_u32 s17, s17, 0
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_lshr_b32 s47, s6, 16
-; VI-NEXT: s_lshr_b32 s56, s8, 16
+; VI-NEXT: s_add_u32 s17, s17, 3
+; VI-NEXT: s_addc_u32 s16, s16, 0
+; VI-NEXT: s_add_u32 s19, s19, 3
+; VI-NEXT: s_addc_u32 s18, s18, 0
+; VI-NEXT: s_add_u32 s21, s21, 3
+; VI-NEXT: s_addc_u32 s20, s20, 0
+; VI-NEXT: s_add_u32 s23, s23, 3
+; VI-NEXT: s_addc_u32 s22, s22, 0
+; VI-NEXT: s_add_u32 s25, s25, 3
+; VI-NEXT: s_addc_u32 s24, s24, 0
+; VI-NEXT: s_add_u32 s40, s40, 3
+; VI-NEXT: s_addc_u32 s26, s26, 0
+; VI-NEXT: s_add_u32 s42, s42, 3
+; VI-NEXT: s_addc_u32 s41, s41, 0
+; VI-NEXT: s_add_u32 s44, s44, 3
+; VI-NEXT: s_addc_u32 s43, s43, 0
+; VI-NEXT: s_add_u32 s46, s46, 3
+; VI-NEXT: s_addc_u32 s45, s45, 0
+; VI-NEXT: s_add_u32 s56, s56, 3
+; VI-NEXT: s_addc_u32 s47, s47, 0
+; VI-NEXT: s_lshr_b32 s27, s7, 16
+; VI-NEXT: s_lshr_b32 s28, s6, 16
+; VI-NEXT: s_lshr_b32 s29, s8, 16
; VI-NEXT: s_lshr_b32 s57, s9, 16
; VI-NEXT: s_lshr_b32 s58, s10, 16
; VI-NEXT: s_lshr_b32 s59, s11, 16
@@ -28553,137 +28202,137 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; VI-NEXT: s_lshr_b32 s61, s13, 16
; VI-NEXT: s_lshr_b32 s62, s14, 16
; VI-NEXT: s_lshr_b32 s63, s15, 16
-; VI-NEXT: s_lshr_b32 s72, s40, 16
-; VI-NEXT: s_lshr_b32 s73, s41, 16
-; VI-NEXT: s_lshr_b32 s74, s42, 16
-; VI-NEXT: s_lshr_b32 s75, s43, 16
-; VI-NEXT: s_lshr_b32 s76, s44, 16
-; VI-NEXT: s_lshr_b32 s77, s45, 16
-; VI-NEXT: s_lshr_b32 s78, s29, 16
-; VI-NEXT: s_lshr_b32 s79, s28, 16
-; VI-NEXT: s_lshr_b32 s88, s27, 16
-; VI-NEXT: s_lshr_b32 s89, s26, 16
-; VI-NEXT: s_lshr_b32 s90, s25, 16
-; VI-NEXT: s_lshr_b32 s91, s24, 16
-; VI-NEXT: s_lshr_b32 s30, s23, 16
-; VI-NEXT: s_lshr_b32 s31, s22, 16
-; VI-NEXT: s_lshr_b32 s34, s21, 16
-; VI-NEXT: s_lshr_b32 s35, s20, 16
-; VI-NEXT: s_lshr_b32 s36, s19, 16
-; VI-NEXT: s_lshr_b32 s37, s18, 16
-; VI-NEXT: s_lshr_b32 s38, s17, 16
-; VI-NEXT: s_lshr_b32 s39, s16, 16
+; VI-NEXT: s_lshr_b32 s72, s16, 16
+; VI-NEXT: s_lshr_b32 s73, s17, 16
+; VI-NEXT: s_lshr_b32 s74, s18, 16
+; VI-NEXT: s_lshr_b32 s75, s19, 16
+; VI-NEXT: s_lshr_b32 s76, s20, 16
+; VI-NEXT: s_lshr_b32 s77, s21, 16
+; VI-NEXT: s_lshr_b32 s78, s22, 16
+; VI-NEXT: s_lshr_b32 s79, s23, 16
+; VI-NEXT: s_lshr_b32 s88, s24, 16
+; VI-NEXT: s_lshr_b32 s89, s25, 16
+; VI-NEXT: s_lshr_b32 s90, s26, 16
+; VI-NEXT: s_lshr_b32 s91, s40, 16
+; VI-NEXT: s_lshr_b32 s30, s41, 16
+; VI-NEXT: s_lshr_b32 s31, s42, 16
+; VI-NEXT: s_lshr_b32 s34, s43, 16
+; VI-NEXT: s_lshr_b32 s35, s44, 16
+; VI-NEXT: s_lshr_b32 s36, s45, 16
+; VI-NEXT: s_lshr_b32 s37, s46, 16
+; VI-NEXT: s_lshr_b32 s38, s47, 16
+; VI-NEXT: s_lshr_b32 s39, s56, 16
; VI-NEXT: .LBB41_3: ; %end
-; VI-NEXT: s_and_b32 s4, 0xffff, s16
+; VI-NEXT: s_and_b32 s4, 0xffff, s56
; VI-NEXT: s_lshl_b32 s5, s39, 16
; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: s_and_b32 s5, 0xffff, s17
-; VI-NEXT: s_lshl_b32 s16, s38, 16
-; VI-NEXT: s_or_b32 s5, s5, s16
-; VI-NEXT: s_and_b32 s16, 0xffff, s18
-; VI-NEXT: s_lshl_b32 s17, s37, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, 0xffff, s19
-; VI-NEXT: s_lshl_b32 s18, s36, 16
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s18, 0xffff, s20
-; VI-NEXT: s_lshl_b32 s19, s35, 16
-; VI-NEXT: s_or_b32 s18, s18, s19
-; VI-NEXT: s_and_b32 s19, 0xffff, s21
-; VI-NEXT: s_lshl_b32 s20, s34, 16
-; VI-NEXT: s_or_b32 s19, s19, s20
-; VI-NEXT: s_and_b32 s20, 0xffff, s22
-; VI-NEXT: s_lshl_b32 s21, s31, 16
-; VI-NEXT: s_or_b32 s20, s20, s21
-; VI-NEXT: s_and_b32 s21, 0xffff, s23
-; VI-NEXT: s_lshl_b32 s22, s30, 16
-; VI-NEXT: s_or_b32 s21, s21, s22
-; VI-NEXT: s_and_b32 s22, 0xffff, s24
-; VI-NEXT: s_lshl_b32 s23, s91, 16
-; VI-NEXT: s_or_b32 s22, s22, s23
-; VI-NEXT: s_and_b32 s23, 0xffff, s25
-; VI-NEXT: s_lshl_b32 s24, s90, 16
-; VI-NEXT: s_or_b32 s23, s23, s24
-; VI-NEXT: s_and_b32 s24, 0xffff, s26
-; VI-NEXT: s_lshl_b32 s25, s89, 16
-; VI-NEXT: s_or_b32 s24, s24, s25
-; VI-NEXT: s_and_b32 s25, 0xffff, s27
-; VI-NEXT: s_lshl_b32 s26, s88, 16
-; VI-NEXT: s_or_b32 s25, s25, s26
-; VI-NEXT: s_and_b32 s26, 0xffff, s28
-; VI-NEXT: s_lshl_b32 s27, s79, 16
-; VI-NEXT: s_or_b32 s26, s26, s27
-; VI-NEXT: s_and_b32 s27, 0xffff, s29
-; VI-NEXT: s_lshl_b32 s28, s78, 16
-; VI-NEXT: s_or_b32 s27, s27, s28
-; VI-NEXT: s_and_b32 s28, 0xffff, s45
-; VI-NEXT: s_lshl_b32 s29, s77, 16
-; VI-NEXT: s_or_b32 s28, s28, s29
-; VI-NEXT: s_and_b32 s29, 0xffff, s44
-; VI-NEXT: s_lshl_b32 s44, s76, 16
-; VI-NEXT: s_or_b32 s29, s29, s44
+; VI-NEXT: s_and_b32 s5, 0xffff, s47
+; VI-NEXT: s_lshl_b32 s47, s38, 16
+; VI-NEXT: s_or_b32 s5, s5, s47
+; VI-NEXT: s_and_b32 s46, 0xffff, s46
+; VI-NEXT: s_lshl_b32 s47, s37, 16
+; VI-NEXT: s_or_b32 s46, s46, s47
+; VI-NEXT: s_and_b32 s45, 0xffff, s45
+; VI-NEXT: s_lshl_b32 s47, s36, 16
+; VI-NEXT: s_or_b32 s45, s45, s47
+; VI-NEXT: s_and_b32 s44, 0xffff, s44
+; VI-NEXT: s_lshl_b32 s47, s35, 16
+; VI-NEXT: s_or_b32 s44, s44, s47
; VI-NEXT: s_and_b32 s43, 0xffff, s43
-; VI-NEXT: s_lshl_b32 s44, s75, 16
-; VI-NEXT: s_or_b32 s43, s43, s44
+; VI-NEXT: s_lshl_b32 s47, s34, 16
+; VI-NEXT: s_or_b32 s43, s43, s47
; VI-NEXT: s_and_b32 s42, 0xffff, s42
-; VI-NEXT: s_lshl_b32 s44, s74, 16
-; VI-NEXT: s_or_b32 s42, s42, s44
+; VI-NEXT: s_lshl_b32 s47, s31, 16
+; VI-NEXT: s_or_b32 s42, s42, s47
; VI-NEXT: s_and_b32 s41, 0xffff, s41
-; VI-NEXT: s_lshl_b32 s44, s73, 16
-; VI-NEXT: s_or_b32 s41, s41, s44
+; VI-NEXT: s_lshl_b32 s47, s30, 16
+; VI-NEXT: s_or_b32 s41, s41, s47
; VI-NEXT: s_and_b32 s40, 0xffff, s40
-; VI-NEXT: s_lshl_b32 s44, s72, 16
-; VI-NEXT: s_or_b32 s40, s40, s44
+; VI-NEXT: s_lshl_b32 s47, s91, 16
+; VI-NEXT: s_or_b32 s40, s40, s47
+; VI-NEXT: s_and_b32 s26, 0xffff, s26
+; VI-NEXT: s_lshl_b32 s47, s90, 16
+; VI-NEXT: s_or_b32 s26, s26, s47
+; VI-NEXT: s_and_b32 s25, 0xffff, s25
+; VI-NEXT: s_lshl_b32 s47, s89, 16
+; VI-NEXT: s_or_b32 s25, s25, s47
+; VI-NEXT: s_and_b32 s24, 0xffff, s24
+; VI-NEXT: s_lshl_b32 s47, s88, 16
+; VI-NEXT: s_or_b32 s24, s24, s47
+; VI-NEXT: s_and_b32 s23, 0xffff, s23
+; VI-NEXT: s_lshl_b32 s47, s79, 16
+; VI-NEXT: s_or_b32 s23, s23, s47
+; VI-NEXT: s_and_b32 s22, 0xffff, s22
+; VI-NEXT: s_lshl_b32 s47, s78, 16
+; VI-NEXT: s_or_b32 s22, s22, s47
+; VI-NEXT: s_and_b32 s21, 0xffff, s21
+; VI-NEXT: s_lshl_b32 s47, s77, 16
+; VI-NEXT: s_or_b32 s21, s21, s47
+; VI-NEXT: s_and_b32 s20, 0xffff, s20
+; VI-NEXT: s_lshl_b32 s47, s76, 16
+; VI-NEXT: s_or_b32 s20, s20, s47
+; VI-NEXT: s_and_b32 s19, 0xffff, s19
+; VI-NEXT: s_lshl_b32 s47, s75, 16
+; VI-NEXT: s_or_b32 s19, s19, s47
+; VI-NEXT: s_and_b32 s18, 0xffff, s18
+; VI-NEXT: s_lshl_b32 s47, s74, 16
+; VI-NEXT: s_or_b32 s18, s18, s47
+; VI-NEXT: s_and_b32 s17, 0xffff, s17
+; VI-NEXT: s_lshl_b32 s47, s73, 16
+; VI-NEXT: s_or_b32 s17, s17, s47
+; VI-NEXT: s_and_b32 s16, 0xffff, s16
+; VI-NEXT: s_lshl_b32 s47, s72, 16
+; VI-NEXT: s_or_b32 s16, s16, s47
; VI-NEXT: s_and_b32 s15, 0xffff, s15
-; VI-NEXT: s_lshl_b32 s44, s63, 16
-; VI-NEXT: s_or_b32 s15, s15, s44
+; VI-NEXT: s_lshl_b32 s47, s63, 16
+; VI-NEXT: s_or_b32 s15, s15, s47
; VI-NEXT: s_and_b32 s14, 0xffff, s14
-; VI-NEXT: s_lshl_b32 s44, s62, 16
-; VI-NEXT: s_or_b32 s14, s14, s44
+; VI-NEXT: s_lshl_b32 s47, s62, 16
+; VI-NEXT: s_or_b32 s14, s14, s47
; VI-NEXT: s_and_b32 s13, 0xffff, s13
-; VI-NEXT: s_lshl_b32 s44, s61, 16
-; VI-NEXT: s_or_b32 s13, s13, s44
+; VI-NEXT: s_lshl_b32 s47, s61, 16
+; VI-NEXT: s_or_b32 s13, s13, s47
; VI-NEXT: s_and_b32 s12, 0xffff, s12
-; VI-NEXT: s_lshl_b32 s44, s60, 16
-; VI-NEXT: s_or_b32 s12, s12, s44
+; VI-NEXT: s_lshl_b32 s47, s60, 16
+; VI-NEXT: s_or_b32 s12, s12, s47
; VI-NEXT: s_and_b32 s11, 0xffff, s11
-; VI-NEXT: s_lshl_b32 s44, s59, 16
-; VI-NEXT: s_or_b32 s11, s11, s44
+; VI-NEXT: s_lshl_b32 s47, s59, 16
+; VI-NEXT: s_or_b32 s11, s11, s47
; VI-NEXT: s_and_b32 s10, 0xffff, s10
-; VI-NEXT: s_lshl_b32 s44, s58, 16
-; VI-NEXT: s_or_b32 s10, s10, s44
+; VI-NEXT: s_lshl_b32 s47, s58, 16
+; VI-NEXT: s_or_b32 s10, s10, s47
; VI-NEXT: s_and_b32 s9, 0xffff, s9
-; VI-NEXT: s_lshl_b32 s44, s57, 16
-; VI-NEXT: s_or_b32 s9, s9, s44
+; VI-NEXT: s_lshl_b32 s47, s57, 16
; VI-NEXT: s_and_b32 s8, 0xffff, s8
-; VI-NEXT: s_lshl_b32 s44, s56, 16
-; VI-NEXT: s_or_b32 s8, s8, s44
+; VI-NEXT: s_lshl_b32 s29, s29, 16
; VI-NEXT: s_and_b32 s6, 0xffff, s6
-; VI-NEXT: s_lshl_b32 s44, s47, 16
-; VI-NEXT: s_or_b32 s6, s6, s44
+; VI-NEXT: s_lshl_b32 s28, s28, 16
; VI-NEXT: s_and_b32 s7, 0xffff, s7
-; VI-NEXT: s_lshl_b32 s44, s46, 16
-; VI-NEXT: s_or_b32 s7, s7, s44
+; VI-NEXT: s_lshl_b32 s27, s27, 16
+; VI-NEXT: s_or_b32 s9, s9, s47
+; VI-NEXT: s_or_b32 s8, s8, s29
+; VI-NEXT: s_or_b32 s6, s6, s28
+; VI-NEXT: s_or_b32 s7, s7, s27
; VI-NEXT: v_mov_b32_e32 v0, s4
; VI-NEXT: v_mov_b32_e32 v1, s5
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: v_mov_b32_e32 v3, s17
-; VI-NEXT: v_mov_b32_e32 v4, s18
-; VI-NEXT: v_mov_b32_e32 v5, s19
-; VI-NEXT: v_mov_b32_e32 v6, s20
-; VI-NEXT: v_mov_b32_e32 v7, s21
-; VI-NEXT: v_mov_b32_e32 v8, s22
-; VI-NEXT: v_mov_b32_e32 v9, s23
-; VI-NEXT: v_mov_b32_e32 v10, s24
-; VI-NEXT: v_mov_b32_e32 v11, s25
-; VI-NEXT: v_mov_b32_e32 v12, s26
-; VI-NEXT: v_mov_b32_e32 v13, s27
-; VI-NEXT: v_mov_b32_e32 v14, s28
-; VI-NEXT: v_mov_b32_e32 v15, s29
-; VI-NEXT: v_mov_b32_e32 v16, s43
-; VI-NEXT: v_mov_b32_e32 v17, s42
-; VI-NEXT: v_mov_b32_e32 v18, s41
-; VI-NEXT: v_mov_b32_e32 v19, s40
+; VI-NEXT: v_mov_b32_e32 v2, s46
+; VI-NEXT: v_mov_b32_e32 v3, s45
+; VI-NEXT: v_mov_b32_e32 v4, s44
+; VI-NEXT: v_mov_b32_e32 v5, s43
+; VI-NEXT: v_mov_b32_e32 v6, s42
+; VI-NEXT: v_mov_b32_e32 v7, s41
+; VI-NEXT: v_mov_b32_e32 v8, s40
+; VI-NEXT: v_mov_b32_e32 v9, s26
+; VI-NEXT: v_mov_b32_e32 v10, s25
+; VI-NEXT: v_mov_b32_e32 v11, s24
+; VI-NEXT: v_mov_b32_e32 v12, s23
+; VI-NEXT: v_mov_b32_e32 v13, s22
+; VI-NEXT: v_mov_b32_e32 v14, s21
+; VI-NEXT: v_mov_b32_e32 v15, s20
+; VI-NEXT: v_mov_b32_e32 v16, s19
+; VI-NEXT: v_mov_b32_e32 v17, s18
+; VI-NEXT: v_mov_b32_e32 v18, s17
+; VI-NEXT: v_mov_b32_e32 v19, s16
; VI-NEXT: v_mov_b32_e32 v20, s15
; VI-NEXT: v_mov_b32_e32 v21, s14
; VI-NEXT: v_mov_b32_e32 v22, s13
@@ -28735,9 +28384,9 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; VI-NEXT: ; implicit-def: $sgpr59
; VI-NEXT: ; implicit-def: $sgpr58
; VI-NEXT: ; implicit-def: $sgpr57
-; VI-NEXT: ; implicit-def: $sgpr56
-; VI-NEXT: ; implicit-def: $sgpr47
-; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr29
+; VI-NEXT: ; implicit-def: $sgpr28
+; VI-NEXT: ; implicit-def: $sgpr27
; VI-NEXT: s_branch .LBB41_2
;
; GFX9-LABEL: bitcast_v15i64_to_v60i16_scalar:
@@ -28746,20 +28395,48 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_mov_b32_e32 v17, s16
+; GFX9-NEXT: v_mov_b32_e32 v18, s17
+; GFX9-NEXT: v_mov_b32_e32 v19, s18
+; GFX9-NEXT: v_readfirstlane_b32 s6, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s19
+; GFX9-NEXT: v_readfirstlane_b32 s7, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s20
+; GFX9-NEXT: v_readfirstlane_b32 s8, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s21
+; GFX9-NEXT: v_readfirstlane_b32 s9, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s22
+; GFX9-NEXT: v_readfirstlane_b32 s10, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s23
; GFX9-NEXT: v_writelane_b32 v30, s30, 0
+; GFX9-NEXT: v_readfirstlane_b32 s11, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s24
+; GFX9-NEXT: v_readfirstlane_b32 s12, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s25
+; GFX9-NEXT: v_readfirstlane_b32 s13, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s26
; GFX9-NEXT: v_writelane_b32 v30, s31, 1
+; GFX9-NEXT: v_readfirstlane_b32 s14, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s27
+; GFX9-NEXT: v_readfirstlane_b32 s15, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s28
+; GFX9-NEXT: v_readfirstlane_b32 s16, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s29
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; GFX9-NEXT: v_writelane_b32 v30, s34, 2
-; GFX9-NEXT: v_readfirstlane_b32 s6, v0
-; GFX9-NEXT: v_readfirstlane_b32 s7, v1
-; GFX9-NEXT: v_readfirstlane_b32 s8, v2
-; GFX9-NEXT: v_readfirstlane_b32 s9, v3
-; GFX9-NEXT: v_readfirstlane_b32 s10, v4
-; GFX9-NEXT: v_readfirstlane_b32 s11, v5
-; GFX9-NEXT: v_readfirstlane_b32 s12, v6
-; GFX9-NEXT: v_readfirstlane_b32 s13, v7
-; GFX9-NEXT: v_readfirstlane_b32 s14, v8
-; GFX9-NEXT: v_readfirstlane_b32 s15, v9
+; GFX9-NEXT: v_readfirstlane_b32 s17, v19
+; GFX9-NEXT: v_readfirstlane_b32 s18, v17
+; GFX9-NEXT: v_readfirstlane_b32 s19, v18
+; GFX9-NEXT: v_readfirstlane_b32 s20, v0
+; GFX9-NEXT: v_readfirstlane_b32 s21, v1
+; GFX9-NEXT: v_readfirstlane_b32 s22, v2
+; GFX9-NEXT: v_readfirstlane_b32 s23, v3
+; GFX9-NEXT: v_readfirstlane_b32 s24, v4
+; GFX9-NEXT: v_readfirstlane_b32 s25, v5
+; GFX9-NEXT: v_readfirstlane_b32 s26, v6
+; GFX9-NEXT: v_readfirstlane_b32 s27, v7
+; GFX9-NEXT: v_readfirstlane_b32 s28, v8
+; GFX9-NEXT: v_readfirstlane_b32 s29, v9
; GFX9-NEXT: v_readfirstlane_b32 s40, v10
; GFX9-NEXT: v_readfirstlane_b32 s41, v11
; GFX9-NEXT: v_readfirstlane_b32 s42, v12
@@ -28776,30 +28453,30 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX9-NEXT: s_lshr_b32 s57, s42, 16
; GFX9-NEXT: s_lshr_b32 s58, s41, 16
; GFX9-NEXT: s_lshr_b32 s59, s40, 16
-; GFX9-NEXT: s_lshr_b32 s60, s15, 16
-; GFX9-NEXT: s_lshr_b32 s61, s14, 16
-; GFX9-NEXT: s_lshr_b32 s62, s13, 16
-; GFX9-NEXT: s_lshr_b32 s63, s12, 16
-; GFX9-NEXT: s_lshr_b32 s72, s11, 16
-; GFX9-NEXT: s_lshr_b32 s73, s10, 16
-; GFX9-NEXT: s_lshr_b32 s74, s9, 16
-; GFX9-NEXT: s_lshr_b32 s75, s8, 16
-; GFX9-NEXT: s_lshr_b32 s76, s7, 16
-; GFX9-NEXT: s_lshr_b32 s77, s6, 16
-; GFX9-NEXT: s_lshr_b32 s78, s29, 16
-; GFX9-NEXT: s_lshr_b32 s79, s28, 16
-; GFX9-NEXT: s_lshr_b32 s88, s27, 16
-; GFX9-NEXT: s_lshr_b32 s89, s26, 16
-; GFX9-NEXT: s_lshr_b32 s90, s25, 16
-; GFX9-NEXT: s_lshr_b32 s91, s24, 16
-; GFX9-NEXT: s_lshr_b32 s92, s23, 16
-; GFX9-NEXT: s_lshr_b32 s93, s22, 16
-; GFX9-NEXT: s_lshr_b32 s94, s21, 16
-; GFX9-NEXT: s_lshr_b32 s95, s20, 16
-; GFX9-NEXT: s_lshr_b32 s30, s19, 16
-; GFX9-NEXT: s_lshr_b32 s31, s18, 16
-; GFX9-NEXT: s_lshr_b32 s34, s17, 16
-; GFX9-NEXT: s_lshr_b32 s35, s16, 16
+; GFX9-NEXT: s_lshr_b32 s60, s29, 16
+; GFX9-NEXT: s_lshr_b32 s61, s28, 16
+; GFX9-NEXT: s_lshr_b32 s62, s27, 16
+; GFX9-NEXT: s_lshr_b32 s63, s26, 16
+; GFX9-NEXT: s_lshr_b32 s72, s25, 16
+; GFX9-NEXT: s_lshr_b32 s73, s24, 16
+; GFX9-NEXT: s_lshr_b32 s74, s23, 16
+; GFX9-NEXT: s_lshr_b32 s75, s22, 16
+; GFX9-NEXT: s_lshr_b32 s76, s21, 16
+; GFX9-NEXT: s_lshr_b32 s77, s20, 16
+; GFX9-NEXT: s_lshr_b32 s78, s19, 16
+; GFX9-NEXT: s_lshr_b32 s79, s18, 16
+; GFX9-NEXT: s_lshr_b32 s88, s17, 16
+; GFX9-NEXT: s_lshr_b32 s89, s16, 16
+; GFX9-NEXT: s_lshr_b32 s90, s15, 16
+; GFX9-NEXT: s_lshr_b32 s91, s14, 16
+; GFX9-NEXT: s_lshr_b32 s92, s13, 16
+; GFX9-NEXT: s_lshr_b32 s93, s12, 16
+; GFX9-NEXT: s_lshr_b32 s94, s11, 16
+; GFX9-NEXT: s_lshr_b32 s95, s10, 16
+; GFX9-NEXT: s_lshr_b32 s30, s9, 16
+; GFX9-NEXT: s_lshr_b32 s31, s8, 16
+; GFX9-NEXT: s_lshr_b32 s34, s7, 16
+; GFX9-NEXT: s_lshr_b32 s35, s6, 16
; GFX9-NEXT: s_cbranch_execnz .LBB41_3
; GFX9-NEXT: .LBB41_2: ; %cmp.true
; GFX9-NEXT: s_add_u32 s44, s44, 3
@@ -28808,16 +28485,6 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX9-NEXT: s_addc_u32 s43, s43, 0
; GFX9-NEXT: s_add_u32 s40, s40, 3
; GFX9-NEXT: s_addc_u32 s41, s41, 0
-; GFX9-NEXT: s_add_u32 s14, s14, 3
-; GFX9-NEXT: s_addc_u32 s15, s15, 0
-; GFX9-NEXT: s_add_u32 s12, s12, 3
-; GFX9-NEXT: s_addc_u32 s13, s13, 0
-; GFX9-NEXT: s_add_u32 s10, s10, 3
-; GFX9-NEXT: s_addc_u32 s11, s11, 0
-; GFX9-NEXT: s_add_u32 s8, s8, 3
-; GFX9-NEXT: s_addc_u32 s9, s9, 0
-; GFX9-NEXT: s_add_u32 s6, s6, 3
-; GFX9-NEXT: s_addc_u32 s7, s7, 0
; GFX9-NEXT: s_add_u32 s28, s28, 3
; GFX9-NEXT: s_addc_u32 s29, s29, 0
; GFX9-NEXT: s_add_u32 s26, s26, 3
@@ -28832,61 +28499,71 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX9-NEXT: s_addc_u32 s19, s19, 0
; GFX9-NEXT: s_add_u32 s16, s16, 3
; GFX9-NEXT: s_addc_u32 s17, s17, 0
+; GFX9-NEXT: s_add_u32 s14, s14, 3
+; GFX9-NEXT: s_addc_u32 s15, s15, 0
+; GFX9-NEXT: s_add_u32 s12, s12, 3
+; GFX9-NEXT: s_addc_u32 s13, s13, 0
+; GFX9-NEXT: s_add_u32 s10, s10, 3
+; GFX9-NEXT: s_addc_u32 s11, s11, 0
+; GFX9-NEXT: s_add_u32 s8, s8, 3
+; GFX9-NEXT: s_addc_u32 s9, s9, 0
+; GFX9-NEXT: s_add_u32 s6, s6, 3
+; GFX9-NEXT: s_addc_u32 s7, s7, 0
; GFX9-NEXT: s_lshr_b32 s46, s45, 16
; GFX9-NEXT: s_lshr_b32 s47, s44, 16
; GFX9-NEXT: s_lshr_b32 s56, s43, 16
; GFX9-NEXT: s_lshr_b32 s57, s42, 16
; GFX9-NEXT: s_lshr_b32 s58, s41, 16
; GFX9-NEXT: s_lshr_b32 s59, s40, 16
-; GFX9-NEXT: s_lshr_b32 s60, s15, 16
-; GFX9-NEXT: s_lshr_b32 s61, s14, 16
-; GFX9-NEXT: s_lshr_b32 s62, s13, 16
-; GFX9-NEXT: s_lshr_b32 s63, s12, 16
-; GFX9-NEXT: s_lshr_b32 s72, s11, 16
-; GFX9-NEXT: s_lshr_b32 s73, s10, 16
-; GFX9-NEXT: s_lshr_b32 s74, s9, 16
-; GFX9-NEXT: s_lshr_b32 s75, s8, 16
-; GFX9-NEXT: s_lshr_b32 s76, s7, 16
-; GFX9-NEXT: s_lshr_b32 s77, s6, 16
-; GFX9-NEXT: s_lshr_b32 s78, s29, 16
-; GFX9-NEXT: s_lshr_b32 s79, s28, 16
-; GFX9-NEXT: s_lshr_b32 s88, s27, 16
-; GFX9-NEXT: s_lshr_b32 s89, s26, 16
-; GFX9-NEXT: s_lshr_b32 s90, s25, 16
-; GFX9-NEXT: s_lshr_b32 s91, s24, 16
-; GFX9-NEXT: s_lshr_b32 s92, s23, 16
-; GFX9-NEXT: s_lshr_b32 s93, s22, 16
-; GFX9-NEXT: s_lshr_b32 s94, s21, 16
-; GFX9-NEXT: s_lshr_b32 s95, s20, 16
-; GFX9-NEXT: s_lshr_b32 s30, s19, 16
-; GFX9-NEXT: s_lshr_b32 s31, s18, 16
-; GFX9-NEXT: s_lshr_b32 s34, s17, 16
-; GFX9-NEXT: s_lshr_b32 s35, s16, 16
+; GFX9-NEXT: s_lshr_b32 s60, s29, 16
+; GFX9-NEXT: s_lshr_b32 s61, s28, 16
+; GFX9-NEXT: s_lshr_b32 s62, s27, 16
+; GFX9-NEXT: s_lshr_b32 s63, s26, 16
+; GFX9-NEXT: s_lshr_b32 s72, s25, 16
+; GFX9-NEXT: s_lshr_b32 s73, s24, 16
+; GFX9-NEXT: s_lshr_b32 s74, s23, 16
+; GFX9-NEXT: s_lshr_b32 s75, s22, 16
+; GFX9-NEXT: s_lshr_b32 s76, s21, 16
+; GFX9-NEXT: s_lshr_b32 s77, s20, 16
+; GFX9-NEXT: s_lshr_b32 s78, s19, 16
+; GFX9-NEXT: s_lshr_b32 s79, s18, 16
+; GFX9-NEXT: s_lshr_b32 s88, s17, 16
+; GFX9-NEXT: s_lshr_b32 s89, s16, 16
+; GFX9-NEXT: s_lshr_b32 s90, s15, 16
+; GFX9-NEXT: s_lshr_b32 s91, s14, 16
+; GFX9-NEXT: s_lshr_b32 s92, s13, 16
+; GFX9-NEXT: s_lshr_b32 s93, s12, 16
+; GFX9-NEXT: s_lshr_b32 s94, s11, 16
+; GFX9-NEXT: s_lshr_b32 s95, s10, 16
+; GFX9-NEXT: s_lshr_b32 s30, s9, 16
+; GFX9-NEXT: s_lshr_b32 s31, s8, 16
+; GFX9-NEXT: s_lshr_b32 s34, s7, 16
+; GFX9-NEXT: s_lshr_b32 s35, s6, 16
; GFX9-NEXT: .LBB41_3: ; %end
-; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s35
-; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s34
-; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s31
-; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s30
-; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s95
-; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s94
-; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s93
-; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s92
-; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s91
-; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s90
-; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s89
-; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s88
-; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s79
-; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s78
-; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s77
-; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s76
-; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s75
-; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s74
-; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s73
-; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s72
-; GFX9-NEXT: s_pack_ll_b32_b16 s12, s12, s63
-; GFX9-NEXT: s_pack_ll_b32_b16 s13, s13, s62
-; GFX9-NEXT: s_pack_ll_b32_b16 s14, s14, s61
-; GFX9-NEXT: s_pack_ll_b32_b16 s15, s15, s60
+; GFX9-NEXT: s_pack_ll_b32_b16 s4, s6, s35
+; GFX9-NEXT: s_pack_ll_b32_b16 s5, s7, s34
+; GFX9-NEXT: s_pack_ll_b32_b16 s6, s8, s31
+; GFX9-NEXT: s_pack_ll_b32_b16 s7, s9, s30
+; GFX9-NEXT: s_pack_ll_b32_b16 s8, s10, s95
+; GFX9-NEXT: s_pack_ll_b32_b16 s9, s11, s94
+; GFX9-NEXT: s_pack_ll_b32_b16 s10, s12, s93
+; GFX9-NEXT: s_pack_ll_b32_b16 s11, s13, s92
+; GFX9-NEXT: s_pack_ll_b32_b16 s12, s14, s91
+; GFX9-NEXT: s_pack_ll_b32_b16 s13, s15, s90
+; GFX9-NEXT: s_pack_ll_b32_b16 s14, s16, s89
+; GFX9-NEXT: s_pack_ll_b32_b16 s15, s17, s88
+; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s79
+; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s78
+; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s77
+; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s76
+; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s75
+; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s74
+; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s73
+; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s72
+; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s63
+; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s62
+; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s61
+; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s60
; GFX9-NEXT: s_pack_ll_b32_b16 s28, s40, s59
; GFX9-NEXT: s_pack_ll_b32_b16 s29, s41, s58
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s42, s57
@@ -28895,28 +28572,28 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s45, s46
; GFX9-NEXT: v_mov_b32_e32 v0, s4
; GFX9-NEXT: v_mov_b32_e32 v1, s5
-; GFX9-NEXT: v_mov_b32_e32 v2, s16
-; GFX9-NEXT: v_mov_b32_e32 v3, s17
-; GFX9-NEXT: v_mov_b32_e32 v4, s18
-; GFX9-NEXT: v_mov_b32_e32 v5, s19
-; GFX9-NEXT: v_mov_b32_e32 v6, s20
-; GFX9-NEXT: v_mov_b32_e32 v7, s21
-; GFX9-NEXT: v_mov_b32_e32 v8, s22
-; GFX9-NEXT: v_mov_b32_e32 v9, s23
-; GFX9-NEXT: v_mov_b32_e32 v10, s24
-; GFX9-NEXT: v_mov_b32_e32 v11, s25
-; GFX9-NEXT: v_mov_b32_e32 v12, s26
-; GFX9-NEXT: v_mov_b32_e32 v13, s27
-; GFX9-NEXT: v_mov_b32_e32 v14, s6
-; GFX9-NEXT: v_mov_b32_e32 v15, s7
-; GFX9-NEXT: v_mov_b32_e32 v16, s8
-; GFX9-NEXT: v_mov_b32_e32 v17, s9
-; GFX9-NEXT: v_mov_b32_e32 v18, s10
-; GFX9-NEXT: v_mov_b32_e32 v19, s11
-; GFX9-NEXT: v_mov_b32_e32 v20, s12
-; GFX9-NEXT: v_mov_b32_e32 v21, s13
-; GFX9-NEXT: v_mov_b32_e32 v22, s14
-; GFX9-NEXT: v_mov_b32_e32 v23, s15
+; GFX9-NEXT: v_mov_b32_e32 v2, s6
+; GFX9-NEXT: v_mov_b32_e32 v3, s7
+; GFX9-NEXT: v_mov_b32_e32 v4, s8
+; GFX9-NEXT: v_mov_b32_e32 v5, s9
+; GFX9-NEXT: v_mov_b32_e32 v6, s10
+; GFX9-NEXT: v_mov_b32_e32 v7, s11
+; GFX9-NEXT: v_mov_b32_e32 v8, s12
+; GFX9-NEXT: v_mov_b32_e32 v9, s13
+; GFX9-NEXT: v_mov_b32_e32 v10, s14
+; GFX9-NEXT: v_mov_b32_e32 v11, s15
+; GFX9-NEXT: v_mov_b32_e32 v12, s16
+; GFX9-NEXT: v_mov_b32_e32 v13, s17
+; GFX9-NEXT: v_mov_b32_e32 v14, s18
+; GFX9-NEXT: v_mov_b32_e32 v15, s19
+; GFX9-NEXT: v_mov_b32_e32 v16, s20
+; GFX9-NEXT: v_mov_b32_e32 v17, s21
+; GFX9-NEXT: v_mov_b32_e32 v18, s22
+; GFX9-NEXT: v_mov_b32_e32 v19, s23
+; GFX9-NEXT: v_mov_b32_e32 v20, s24
+; GFX9-NEXT: v_mov_b32_e32 v21, s25
+; GFX9-NEXT: v_mov_b32_e32 v22, s26
+; GFX9-NEXT: v_mov_b32_e32 v23, s27
; GFX9-NEXT: v_mov_b32_e32 v24, s28
; GFX9-NEXT: v_mov_b32_e32 v25, s29
; GFX9-NEXT: v_mov_b32_e32 v26, s40
@@ -28968,49 +28645,76 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX11-LABEL: bitcast_v15i64_to_v60i16_scalar:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT: v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
+; GFX11-NEXT: v_dual_mov_b32 v17, s16 :: v_dual_mov_b32 v18, s17
+; GFX11-NEXT: v_dual_mov_b32 v19, s18 :: v_dual_mov_b32 v20, s19
+; GFX11-NEXT: v_dual_mov_b32 v21, s20 :: v_dual_mov_b32 v22, s21
+; GFX11-NEXT: v_dual_mov_b32 v23, s22 :: v_dual_mov_b32 v24, s23
+; GFX11-NEXT: v_dual_mov_b32 v25, s24 :: v_dual_mov_b32 v26, s25
+; GFX11-NEXT: v_dual_mov_b32 v27, s26 :: v_dual_mov_b32 v28, s27
+; GFX11-NEXT: v_dual_mov_b32 v29, s28 :: v_dual_mov_b32 v30, s29
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: v_readfirstlane_b32 s4, v0
-; GFX11-NEXT: v_readfirstlane_b32 s5, v1
-; GFX11-NEXT: v_readfirstlane_b32 s6, v2
-; GFX11-NEXT: v_readfirstlane_b32 s7, v3
-; GFX11-NEXT: v_readfirstlane_b32 s8, v4
-; GFX11-NEXT: v_readfirstlane_b32 s9, v5
-; GFX11-NEXT: v_readfirstlane_b32 s10, v6
-; GFX11-NEXT: v_readfirstlane_b32 s11, v7
-; GFX11-NEXT: v_readfirstlane_b32 s12, v8
-; GFX11-NEXT: v_readfirstlane_b32 s13, v9
-; GFX11-NEXT: v_readfirstlane_b32 s15, v10
-; GFX11-NEXT: v_readfirstlane_b32 s14, v11
+; GFX11-NEXT: v_readfirstlane_b32 s0, v13
+; GFX11-NEXT: v_readfirstlane_b32 s1, v14
+; GFX11-NEXT: v_readfirstlane_b32 s2, v15
+; GFX11-NEXT: v_readfirstlane_b32 s3, v16
+; GFX11-NEXT: v_readfirstlane_b32 s4, v17
+; GFX11-NEXT: v_readfirstlane_b32 s5, v18
+; GFX11-NEXT: v_readfirstlane_b32 s6, v19
+; GFX11-NEXT: v_readfirstlane_b32 s7, v20
+; GFX11-NEXT: v_readfirstlane_b32 s8, v21
+; GFX11-NEXT: v_readfirstlane_b32 s9, v22
+; GFX11-NEXT: v_readfirstlane_b32 s10, v23
+; GFX11-NEXT: v_readfirstlane_b32 s11, v24
+; GFX11-NEXT: v_readfirstlane_b32 s12, v25
+; GFX11-NEXT: v_readfirstlane_b32 s13, v26
+; GFX11-NEXT: v_readfirstlane_b32 s14, v27
+; GFX11-NEXT: v_readfirstlane_b32 s15, v28
+; GFX11-NEXT: v_readfirstlane_b32 s16, v29
+; GFX11-NEXT: v_readfirstlane_b32 s17, v30
+; GFX11-NEXT: v_readfirstlane_b32 s18, v0
+; GFX11-NEXT: v_readfirstlane_b32 s19, v1
+; GFX11-NEXT: v_readfirstlane_b32 s20, v2
+; GFX11-NEXT: v_readfirstlane_b32 s21, v3
+; GFX11-NEXT: v_readfirstlane_b32 s22, v4
+; GFX11-NEXT: v_readfirstlane_b32 s23, v5
+; GFX11-NEXT: v_readfirstlane_b32 s24, v6
+; GFX11-NEXT: v_readfirstlane_b32 s25, v7
+; GFX11-NEXT: v_readfirstlane_b32 s26, v8
+; GFX11-NEXT: v_readfirstlane_b32 s27, v9
+; GFX11-NEXT: v_readfirstlane_b32 s29, v10
+; GFX11-NEXT: v_readfirstlane_b32 s28, v11
; GFX11-NEXT: s_mov_b32 s94, 0
; GFX11-NEXT: s_and_b32 s40, vcc_lo, exec_lo
; GFX11-NEXT: s_cbranch_scc0 .LBB41_4
; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: s_lshr_b32 s40, s14, 16
-; GFX11-NEXT: s_lshr_b32 s41, s15, 16
-; GFX11-NEXT: s_lshr_b32 s42, s13, 16
-; GFX11-NEXT: s_lshr_b32 s43, s12, 16
-; GFX11-NEXT: s_lshr_b32 s44, s11, 16
-; GFX11-NEXT: s_lshr_b32 s45, s10, 16
-; GFX11-NEXT: s_lshr_b32 s46, s9, 16
-; GFX11-NEXT: s_lshr_b32 s47, s8, 16
-; GFX11-NEXT: s_lshr_b32 s56, s7, 16
-; GFX11-NEXT: s_lshr_b32 s57, s6, 16
-; GFX11-NEXT: s_lshr_b32 s58, s5, 16
-; GFX11-NEXT: s_lshr_b32 s59, s4, 16
-; GFX11-NEXT: s_lshr_b32 s60, s29, 16
-; GFX11-NEXT: s_lshr_b32 s61, s28, 16
-; GFX11-NEXT: s_lshr_b32 s62, s27, 16
-; GFX11-NEXT: s_lshr_b32 s63, s26, 16
-; GFX11-NEXT: s_lshr_b32 s72, s25, 16
-; GFX11-NEXT: s_lshr_b32 s73, s24, 16
-; GFX11-NEXT: s_lshr_b32 s74, s23, 16
-; GFX11-NEXT: s_lshr_b32 s75, s22, 16
-; GFX11-NEXT: s_lshr_b32 s76, s21, 16
-; GFX11-NEXT: s_lshr_b32 s77, s20, 16
-; GFX11-NEXT: s_lshr_b32 s78, s19, 16
-; GFX11-NEXT: s_lshr_b32 s79, s18, 16
-; GFX11-NEXT: s_lshr_b32 s88, s17, 16
-; GFX11-NEXT: s_lshr_b32 s89, s16, 16
+; GFX11-NEXT: s_lshr_b32 s40, s28, 16
+; GFX11-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-NEXT: s_lshr_b32 s43, s26, 16
+; GFX11-NEXT: s_lshr_b32 s44, s25, 16
+; GFX11-NEXT: s_lshr_b32 s45, s24, 16
+; GFX11-NEXT: s_lshr_b32 s46, s23, 16
+; GFX11-NEXT: s_lshr_b32 s47, s22, 16
+; GFX11-NEXT: s_lshr_b32 s56, s21, 16
+; GFX11-NEXT: s_lshr_b32 s57, s20, 16
+; GFX11-NEXT: s_lshr_b32 s58, s19, 16
+; GFX11-NEXT: s_lshr_b32 s59, s18, 16
+; GFX11-NEXT: s_lshr_b32 s60, s17, 16
+; GFX11-NEXT: s_lshr_b32 s61, s16, 16
+; GFX11-NEXT: s_lshr_b32 s62, s15, 16
+; GFX11-NEXT: s_lshr_b32 s63, s14, 16
+; GFX11-NEXT: s_lshr_b32 s72, s13, 16
+; GFX11-NEXT: s_lshr_b32 s73, s12, 16
+; GFX11-NEXT: s_lshr_b32 s74, s11, 16
+; GFX11-NEXT: s_lshr_b32 s75, s10, 16
+; GFX11-NEXT: s_lshr_b32 s76, s9, 16
+; GFX11-NEXT: s_lshr_b32 s77, s8, 16
+; GFX11-NEXT: s_lshr_b32 s78, s7, 16
+; GFX11-NEXT: s_lshr_b32 s79, s6, 16
+; GFX11-NEXT: s_lshr_b32 s88, s5, 16
+; GFX11-NEXT: s_lshr_b32 s89, s4, 16
; GFX11-NEXT: s_lshr_b32 s90, s3, 16
; GFX11-NEXT: s_lshr_b32 s91, s2, 16
; GFX11-NEXT: s_lshr_b32 s92, s1, 16
@@ -29018,20 +28722,8 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s94
; GFX11-NEXT: s_cbranch_vccnz .LBB41_3
; GFX11-NEXT: .LBB41_2: ; %cmp.true
-; GFX11-NEXT: s_add_u32 s15, s15, 3
-; GFX11-NEXT: s_addc_u32 s14, s14, 0
-; GFX11-NEXT: s_add_u32 s12, s12, 3
-; GFX11-NEXT: s_addc_u32 s13, s13, 0
-; GFX11-NEXT: s_add_u32 s10, s10, 3
-; GFX11-NEXT: s_addc_u32 s11, s11, 0
-; GFX11-NEXT: s_add_u32 s8, s8, 3
-; GFX11-NEXT: s_addc_u32 s9, s9, 0
-; GFX11-NEXT: s_add_u32 s6, s6, 3
-; GFX11-NEXT: s_addc_u32 s7, s7, 0
-; GFX11-NEXT: s_add_u32 s4, s4, 3
-; GFX11-NEXT: s_addc_u32 s5, s5, 0
-; GFX11-NEXT: s_add_u32 s28, s28, 3
-; GFX11-NEXT: s_addc_u32 s29, s29, 0
+; GFX11-NEXT: s_add_u32 s29, s29, 3
+; GFX11-NEXT: s_addc_u32 s28, s28, 0
; GFX11-NEXT: s_add_u32 s26, s26, 3
; GFX11-NEXT: s_addc_u32 s27, s27, 0
; GFX11-NEXT: s_add_u32 s24, s24, 3
@@ -29044,36 +28736,48 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX11-NEXT: s_addc_u32 s19, s19, 0
; GFX11-NEXT: s_add_u32 s16, s16, 3
; GFX11-NEXT: s_addc_u32 s17, s17, 0
+; GFX11-NEXT: s_add_u32 s14, s14, 3
+; GFX11-NEXT: s_addc_u32 s15, s15, 0
+; GFX11-NEXT: s_add_u32 s12, s12, 3
+; GFX11-NEXT: s_addc_u32 s13, s13, 0
+; GFX11-NEXT: s_add_u32 s10, s10, 3
+; GFX11-NEXT: s_addc_u32 s11, s11, 0
+; GFX11-NEXT: s_add_u32 s8, s8, 3
+; GFX11-NEXT: s_addc_u32 s9, s9, 0
+; GFX11-NEXT: s_add_u32 s6, s6, 3
+; GFX11-NEXT: s_addc_u32 s7, s7, 0
+; GFX11-NEXT: s_add_u32 s4, s4, 3
+; GFX11-NEXT: s_addc_u32 s5, s5, 0
; GFX11-NEXT: s_add_u32 s2, s2, 3
; GFX11-NEXT: s_addc_u32 s3, s3, 0
; GFX11-NEXT: s_add_u32 s0, s0, 3
; GFX11-NEXT: s_addc_u32 s1, s1, 0
-; GFX11-NEXT: s_lshr_b32 s40, s14, 16
-; GFX11-NEXT: s_lshr_b32 s41, s15, 16
-; GFX11-NEXT: s_lshr_b32 s42, s13, 16
-; GFX11-NEXT: s_lshr_b32 s43, s12, 16
-; GFX11-NEXT: s_lshr_b32 s44, s11, 16
-; GFX11-NEXT: s_lshr_b32 s45, s10, 16
-; GFX11-NEXT: s_lshr_b32 s46, s9, 16
-; GFX11-NEXT: s_lshr_b32 s47, s8, 16
-; GFX11-NEXT: s_lshr_b32 s56, s7, 16
-; GFX11-NEXT: s_lshr_b32 s57, s6, 16
-; GFX11-NEXT: s_lshr_b32 s58, s5, 16
-; GFX11-NEXT: s_lshr_b32 s59, s4, 16
-; GFX11-NEXT: s_lshr_b32 s60, s29, 16
-; GFX11-NEXT: s_lshr_b32 s61, s28, 16
-; GFX11-NEXT: s_lshr_b32 s62, s27, 16
-; GFX11-NEXT: s_lshr_b32 s63, s26, 16
-; GFX11-NEXT: s_lshr_b32 s72, s25, 16
-; GFX11-NEXT: s_lshr_b32 s73, s24, 16
-; GFX11-NEXT: s_lshr_b32 s74, s23, 16
-; GFX11-NEXT: s_lshr_b32 s75, s22, 16
-; GFX11-NEXT: s_lshr_b32 s76, s21, 16
-; GFX11-NEXT: s_lshr_b32 s77, s20, 16
-; GFX11-NEXT: s_lshr_b32 s78, s19, 16
-; GFX11-NEXT: s_lshr_b32 s79, s18, 16
-; GFX11-NEXT: s_lshr_b32 s88, s17, 16
-; GFX11-NEXT: s_lshr_b32 s89, s16, 16
+; GFX11-NEXT: s_lshr_b32 s40, s28, 16
+; GFX11-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-NEXT: s_lshr_b32 s43, s26, 16
+; GFX11-NEXT: s_lshr_b32 s44, s25, 16
+; GFX11-NEXT: s_lshr_b32 s45, s24, 16
+; GFX11-NEXT: s_lshr_b32 s46, s23, 16
+; GFX11-NEXT: s_lshr_b32 s47, s22, 16
+; GFX11-NEXT: s_lshr_b32 s56, s21, 16
+; GFX11-NEXT: s_lshr_b32 s57, s20, 16
+; GFX11-NEXT: s_lshr_b32 s58, s19, 16
+; GFX11-NEXT: s_lshr_b32 s59, s18, 16
+; GFX11-NEXT: s_lshr_b32 s60, s17, 16
+; GFX11-NEXT: s_lshr_b32 s61, s16, 16
+; GFX11-NEXT: s_lshr_b32 s62, s15, 16
+; GFX11-NEXT: s_lshr_b32 s63, s14, 16
+; GFX11-NEXT: s_lshr_b32 s72, s13, 16
+; GFX11-NEXT: s_lshr_b32 s73, s12, 16
+; GFX11-NEXT: s_lshr_b32 s74, s11, 16
+; GFX11-NEXT: s_lshr_b32 s75, s10, 16
+; GFX11-NEXT: s_lshr_b32 s76, s9, 16
+; GFX11-NEXT: s_lshr_b32 s77, s8, 16
+; GFX11-NEXT: s_lshr_b32 s78, s7, 16
+; GFX11-NEXT: s_lshr_b32 s79, s6, 16
+; GFX11-NEXT: s_lshr_b32 s88, s5, 16
+; GFX11-NEXT: s_lshr_b32 s89, s4, 16
; GFX11-NEXT: s_lshr_b32 s90, s3, 16
; GFX11-NEXT: s_lshr_b32 s91, s2, 16
; GFX11-NEXT: s_lshr_b32 s92, s1, 16
@@ -29084,47 +28788,47 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s92
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s91
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s59
-; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s58
-; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s57
-; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s56
-; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s47
-; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s46
-; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s45
-; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s44
-; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s43
-; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s42
-; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s41
-; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s40
+; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s89
+; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s88
+; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s79
+; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s78
+; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s77
+; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s76
+; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s75
+; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s74
+; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s73
+; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s72
+; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s63
+; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s62
+; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s61
+; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s60
+; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s59
+; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s58
+; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s57
+; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s56
+; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s47
+; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s46
+; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s45
+; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s44
+; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s43
+; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s41
+; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s40
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-NEXT: v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-NEXT: v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-NEXT: v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-NEXT: v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-NEXT: v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
-; GFX11-NEXT: v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT: v_dual_mov_b32 v28, s29 :: v_dual_mov_b32 v29, s28
; GFX11-NEXT: s_setpc_b64 s[30:31]
; GFX11-NEXT: .LBB41_4:
; GFX11-NEXT: ; implicit-def: $sgpr93
@@ -29240,40 +28944,53 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:4
; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v2
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:112
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:108
+; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(4)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v8
-; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:104
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:100
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v10
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v6
+; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v8
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:96
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:92
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v10
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:84
-; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v14
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
@@ -29304,27 +29021,10 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:56
; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:64
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52
-; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v28
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
@@ -29409,7 +29109,6 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr30
; SI-NEXT: ; kill: killed $vgpr30
; SI-NEXT: ; implicit-def: $vgpr30
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v55
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v61
; SI-NEXT: ; kill: killed $vgpr30
@@ -29554,7 +29253,6 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v55
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v61
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
@@ -29681,7 +29379,7 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; SI-NEXT: v_add_i32_e32 v27, vcc, s6, v27
+; SI-NEXT: v_add_i32_e32 v27, vcc, 0x30000, v27
; SI-NEXT: v_add_i32_e32 v28, vcc, 0x30000, v28
; SI-NEXT: v_add_i32_e32 v29, vcc, 0x30000, v29
; SI-NEXT: .LBB42_4: ; %end
@@ -29702,7 +29400,7 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v60i16_to_v15i64:
@@ -29984,7 +29682,6 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: v_mov_b32_e32 v61, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v28
@@ -30063,6 +29760,7 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v60
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v61
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
@@ -30227,6 +29925,9 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT: s_cbranch_execz .LBB42_4
; GFX9-NEXT: ; %bb.3: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
@@ -30241,9 +29942,6 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v38, v61, s6
; GFX9-NEXT: v_perm_b32 v1, v37, v60, s6
; GFX9-NEXT: v_perm_b32 v2, v62, v59, s6
@@ -30262,6 +29960,10 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: s_waitcnt vmcnt(14)
+; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: s_waitcnt vmcnt(15)
; GFX9-NEXT: v_perm_b32 v9, v9, v44, s6
; GFX9-NEXT: s_waitcnt vmcnt(14)
@@ -30290,10 +29992,6 @@ define <15 x i64> @bitcast_v60i16_to_v15i64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_perm_b32 v21, v21, v48, s6
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_perm_b32 v22, v22, v39, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
@@ -31449,304 +31147,115 @@ define inreg <15 x i64> @bitcast_v60i16_to_v15i64_scalar(<60 x i16> inreg %a, i3
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v156, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v157, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v158, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v159, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v168, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v169, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v170, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v171, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v172, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v173, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v174, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v175, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v184, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v185, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v186, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v187, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v188, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v189, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v190, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v191, s32
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v25, v7
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v191, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v190, v6 :: v_dual_mov_b32 v185, v4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v186, v3 :: v_dual_mov_b32 v187, v2
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v188, v1 :: v_dual_mov_b32 v189, v0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s28, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s19, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s17, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s1, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s0, 16
-; GFX11-TRUE16-NEXT: s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s40, s0, s40
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s41, s1, s41
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s2, s46
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s3, s45
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s16, s44
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s17, s43
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s18, s4
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s19, s5
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s20, s6
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s21, s7
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s22, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s23, s9
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s25, s11
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s26, s12
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s27, s13
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s28, s14
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s29, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB43_4
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s40 :: v_dual_mov_b32 v5, s0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s41 :: v_dual_mov_b32 v9, s1
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s2 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s3 :: v_dual_mov_b32 v35, s5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v44, s6 :: v_dual_mov_b32 v65, s8
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s7 :: v_dual_mov_b32 v77, s9
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v90, s10 :: v_dual_mov_b32 v119, s12
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v104, s11 :: v_dual_mov_b32 v135, s13
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v152, s14
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v170, s15
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB43_3
; GFX11-TRUE16-NEXT: .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s40, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s41, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v189, v189, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v188, v188, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v187, v187, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v186, v186, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v185, v185, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v191, v191, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v190, v190, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, s4, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v35, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v44, s6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v54, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v65, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v77, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v90, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v104, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v119, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v135, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v152, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v170, s15, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: .LBB43_3: ; %end
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v20 :: v_dual_mov_b32 v6, v27
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v35 :: v_dual_mov_b32 v8, v44
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v77 :: v_dual_mov_b32 v12, v90
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v104
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, v135 :: v_dual_mov_b32 v16, v152
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v170 :: v_dual_mov_b32 v18, v189
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v188 :: v_dual_mov_b32 v20, v187
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v186 :: v_dual_mov_b32 v22, v185
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v191 :: v_dual_mov_b32 v24, v190
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v191, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v190, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v189, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v188, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v187, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v186, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v185, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v184, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v175, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v174, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v173, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v172, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v171, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v170, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v169, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v168, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v159, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v158, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v157, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v156, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:124
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:252
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:316
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v14
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v54 :: v_dual_mov_b32 v10, v65
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v119 :: v_dual_mov_b32 v27, v30
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB43_4:
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, v29 :: v_dual_mov_b32 v65, v28
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v66, v30
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v26 :: v_dual_mov_b32 v54, v25
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v53
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v65 :: v_dual_mov_b32 v29, v64
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v30, v66
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168_vgpr169_vgpr170_vgpr171_vgpr172_vgpr173_vgpr174_vgpr175_vgpr176_vgpr177_vgpr178_vgpr179_vgpr180_vgpr181_vgpr182_vgpr183_vgpr184
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_branch .LBB43_2
;
; GFX11-FAKE16-LABEL: bitcast_v60i16_to_v15i64_scalar:
@@ -33281,13 +32790,41 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; SI-LABEL: bitcast_v15i64_to_v60f16_scalar:
; SI: ; %bb.0:
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT: v_mov_b32_e32 v18, s16
+; SI-NEXT: v_mov_b32_e32 v19, s17
+; SI-NEXT: v_readfirstlane_b32 s40, v18
+; SI-NEXT: v_mov_b32_e32 v18, s18
+; SI-NEXT: v_readfirstlane_b32 s43, v19
+; SI-NEXT: v_mov_b32_e32 v19, s19
+; SI-NEXT: v_readfirstlane_b32 s41, v18
+; SI-NEXT: v_mov_b32_e32 v18, s20
+; SI-NEXT: v_readfirstlane_b32 s44, v19
+; SI-NEXT: v_mov_b32_e32 v19, s21
+; SI-NEXT: v_readfirstlane_b32 s42, v18
+; SI-NEXT: v_mov_b32_e32 v18, s22
+; SI-NEXT: v_readfirstlane_b32 s45, v19
+; SI-NEXT: v_mov_b32_e32 v19, s23
+; SI-NEXT: v_readfirstlane_b32 s22, v18
+; SI-NEXT: v_mov_b32_e32 v18, s24
+; SI-NEXT: v_readfirstlane_b32 s46, v19
+; SI-NEXT: v_mov_b32_e32 v19, s25
+; SI-NEXT: v_readfirstlane_b32 s23, v18
+; SI-NEXT: v_mov_b32_e32 v18, s26
+; SI-NEXT: v_readfirstlane_b32 s47, v19
+; SI-NEXT: v_mov_b32_e32 v19, s27
+; SI-NEXT: v_readfirstlane_b32 s24, v18
+; SI-NEXT: v_mov_b32_e32 v18, s28
+; SI-NEXT: v_readfirstlane_b32 s27, v19
+; SI-NEXT: v_mov_b32_e32 v19, s29
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; SI-NEXT: v_readfirstlane_b32 s44, v1
-; SI-NEXT: v_readfirstlane_b32 s45, v2
-; SI-NEXT: v_readfirstlane_b32 s42, v3
-; SI-NEXT: v_readfirstlane_b32 s43, v4
-; SI-NEXT: v_readfirstlane_b32 s40, v5
-; SI-NEXT: v_readfirstlane_b32 s41, v6
+; SI-NEXT: v_readfirstlane_b32 s25, v18
+; SI-NEXT: v_readfirstlane_b32 s26, v19
+; SI-NEXT: v_readfirstlane_b32 s20, v1
+; SI-NEXT: v_readfirstlane_b32 s21, v2
+; SI-NEXT: v_readfirstlane_b32 s18, v3
+; SI-NEXT: v_readfirstlane_b32 s19, v4
+; SI-NEXT: v_readfirstlane_b32 s16, v5
+; SI-NEXT: v_readfirstlane_b32 s17, v6
; SI-NEXT: v_readfirstlane_b32 s14, v7
; SI-NEXT: v_readfirstlane_b32 s15, v8
; SI-NEXT: v_readfirstlane_b32 s12, v9
@@ -33334,48 +32871,48 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v15, s4
; SI-NEXT: s_lshr_b32 s4, s14, 16
; SI-NEXT: v_cvt_f32_f16_e32 v17, s4
-; SI-NEXT: s_lshr_b32 s4, s41, 16
+; SI-NEXT: s_lshr_b32 s4, s17, 16
; SI-NEXT: v_cvt_f32_f16_e32 v19, s4
-; SI-NEXT: s_lshr_b32 s4, s40, 16
+; SI-NEXT: s_lshr_b32 s4, s16, 16
; SI-NEXT: v_cvt_f32_f16_e32 v22, s4
-; SI-NEXT: s_lshr_b32 s4, s43, 16
+; SI-NEXT: s_lshr_b32 s4, s19, 16
; SI-NEXT: v_cvt_f32_f16_e32 v24, s4
-; SI-NEXT: s_lshr_b32 s4, s42, 16
+; SI-NEXT: s_lshr_b32 s4, s18, 16
; SI-NEXT: v_cvt_f32_f16_e32 v26, s4
-; SI-NEXT: s_lshr_b32 s4, s45, 16
+; SI-NEXT: s_lshr_b32 s4, s21, 16
; SI-NEXT: v_cvt_f32_f16_e32 v28, s4
-; SI-NEXT: s_lshr_b32 s4, s44, 16
+; SI-NEXT: s_lshr_b32 s4, s20, 16
; SI-NEXT: v_cvt_f32_f16_e32 v30, s4
-; SI-NEXT: s_lshr_b32 s4, s29, 16
+; SI-NEXT: s_lshr_b32 s4, s26, 16
; SI-NEXT: v_cvt_f32_f16_e32 v32, s4
-; SI-NEXT: s_lshr_b32 s4, s28, 16
+; SI-NEXT: s_lshr_b32 s4, s25, 16
; SI-NEXT: v_cvt_f32_f16_e32 v34, s4
; SI-NEXT: s_lshr_b32 s4, s27, 16
; SI-NEXT: v_cvt_f32_f16_e32 v36, s4
-; SI-NEXT: s_lshr_b32 s4, s26, 16
+; SI-NEXT: s_lshr_b32 s4, s24, 16
; SI-NEXT: v_cvt_f32_f16_e32 v39, s4
-; SI-NEXT: s_lshr_b32 s4, s25, 16
+; SI-NEXT: s_lshr_b32 s4, s47, 16
; SI-NEXT: v_cvt_f32_f16_e32 v49, s4
-; SI-NEXT: s_lshr_b32 s4, s24, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v51, s4
; SI-NEXT: s_lshr_b32 s4, s23, 16
+; SI-NEXT: v_cvt_f32_f16_e32 v51, s4
+; SI-NEXT: s_lshr_b32 s4, s46, 16
; SI-NEXT: v_cvt_f32_f16_e32 v53, s4
; SI-NEXT: s_lshr_b32 s4, s22, 16
; SI-NEXT: v_cvt_f32_f16_e32 v55, s4
-; SI-NEXT: s_lshr_b32 s4, s21, 16
+; SI-NEXT: s_lshr_b32 s4, s45, 16
; SI-NEXT: v_cvt_f32_f16_e32 v41, s4
-; SI-NEXT: s_lshr_b32 s4, s20, 16
+; SI-NEXT: s_lshr_b32 s4, s42, 16
; SI-NEXT: v_cvt_f32_f16_e32 v43, s4
-; SI-NEXT: s_lshr_b32 s4, s19, 16
+; SI-NEXT: s_lshr_b32 s4, s44, 16
; SI-NEXT: s_waitcnt expcnt(6)
; SI-NEXT: v_cvt_f32_f16_e32 v46, s4
-; SI-NEXT: s_lshr_b32 s4, s18, 16
+; SI-NEXT: s_lshr_b32 s4, s41, 16
; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: v_cvt_f32_f16_e32 v56, s4
-; SI-NEXT: s_lshr_b32 s4, s17, 16
+; SI-NEXT: s_lshr_b32 s4, s43, 16
; SI-NEXT: s_waitcnt expcnt(2)
; SI-NEXT: v_cvt_f32_f16_e32 v58, s4
-; SI-NEXT: s_lshr_b32 s4, s16, 16
+; SI-NEXT: s_lshr_b32 s4, s40, 16
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v60, s4
; SI-NEXT: v_cvt_f32_f16_e32 v4, s9
@@ -33388,68 +32925,68 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v18, s12
; SI-NEXT: v_cvt_f32_f16_e32 v20, s15
; SI-NEXT: v_cvt_f32_f16_e32 v21, s14
-; SI-NEXT: v_cvt_f32_f16_e32 v23, s41
-; SI-NEXT: v_cvt_f32_f16_e32 v25, s40
-; SI-NEXT: v_cvt_f32_f16_e32 v27, s43
-; SI-NEXT: v_cvt_f32_f16_e32 v29, s42
-; SI-NEXT: v_cvt_f32_f16_e32 v31, s45
-; SI-NEXT: v_cvt_f32_f16_e32 v33, s44
-; SI-NEXT: v_cvt_f32_f16_e32 v35, s29
-; SI-NEXT: v_cvt_f32_f16_e32 v37, s28
+; SI-NEXT: v_cvt_f32_f16_e32 v23, s17
+; SI-NEXT: v_cvt_f32_f16_e32 v25, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v27, s19
+; SI-NEXT: v_cvt_f32_f16_e32 v29, s18
+; SI-NEXT: v_cvt_f32_f16_e32 v31, s21
+; SI-NEXT: v_cvt_f32_f16_e32 v33, s20
+; SI-NEXT: v_cvt_f32_f16_e32 v35, s26
+; SI-NEXT: v_cvt_f32_f16_e32 v37, s25
; SI-NEXT: v_cvt_f32_f16_e32 v38, s27
-; SI-NEXT: v_cvt_f32_f16_e32 v48, s26
-; SI-NEXT: v_cvt_f32_f16_e32 v50, s25
-; SI-NEXT: v_cvt_f32_f16_e32 v52, s24
-; SI-NEXT: v_cvt_f32_f16_e32 v54, s23
+; SI-NEXT: v_cvt_f32_f16_e32 v48, s24
+; SI-NEXT: v_cvt_f32_f16_e32 v50, s47
+; SI-NEXT: v_cvt_f32_f16_e32 v52, s23
+; SI-NEXT: v_cvt_f32_f16_e32 v54, s46
; SI-NEXT: v_cvt_f32_f16_e32 v40, s22
-; SI-NEXT: v_cvt_f32_f16_e32 v42, s21
-; SI-NEXT: v_cvt_f32_f16_e32 v44, s20
-; SI-NEXT: v_cvt_f32_f16_e32 v45, s19
-; SI-NEXT: v_cvt_f32_f16_e32 v47, s18
-; SI-NEXT: v_cvt_f32_f16_e32 v57, s17
-; SI-NEXT: v_cvt_f32_f16_e32 v59, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v42, s45
+; SI-NEXT: v_cvt_f32_f16_e32 v44, s42
+; SI-NEXT: v_cvt_f32_f16_e32 v45, s44
+; SI-NEXT: v_cvt_f32_f16_e32 v47, s41
+; SI-NEXT: v_cvt_f32_f16_e32 v57, s43
+; SI-NEXT: v_cvt_f32_f16_e32 v59, s40
; SI-NEXT: s_cbranch_execnz .LBB45_3
; SI-NEXT: .LBB45_2: ; %cmp.true
-; SI-NEXT: s_add_u32 s4, s16, 3
-; SI-NEXT: s_addc_u32 s5, s17, 0
-; SI-NEXT: s_lshr_b32 s16, s4, 16
-; SI-NEXT: s_lshr_b32 s17, s5, 16
-; SI-NEXT: s_add_u32 s18, s18, 3
-; SI-NEXT: s_addc_u32 s19, s19, 0
-; SI-NEXT: s_lshr_b32 s46, s18, 16
-; SI-NEXT: s_lshr_b32 s47, s19, 16
-; SI-NEXT: s_add_u32 s20, s20, 3
-; SI-NEXT: s_addc_u32 s21, s21, 0
-; SI-NEXT: s_lshr_b32 s56, s20, 16
-; SI-NEXT: s_lshr_b32 s57, s21, 16
+; SI-NEXT: s_add_u32 s4, s40, 3
+; SI-NEXT: s_addc_u32 s5, s43, 0
+; SI-NEXT: s_lshr_b32 s28, s4, 16
+; SI-NEXT: s_lshr_b32 s29, s5, 16
+; SI-NEXT: s_add_u32 s40, s41, 3
+; SI-NEXT: s_addc_u32 s41, s44, 0
+; SI-NEXT: s_lshr_b32 s43, s40, 16
+; SI-NEXT: s_lshr_b32 s44, s41, 16
+; SI-NEXT: s_add_u32 s42, s42, 3
+; SI-NEXT: s_addc_u32 s45, s45, 0
+; SI-NEXT: s_lshr_b32 s56, s42, 16
+; SI-NEXT: s_lshr_b32 s57, s45, 16
; SI-NEXT: s_add_u32 s22, s22, 3
-; SI-NEXT: s_addc_u32 s23, s23, 0
+; SI-NEXT: s_addc_u32 s46, s46, 0
; SI-NEXT: s_lshr_b32 s58, s22, 16
-; SI-NEXT: s_lshr_b32 s59, s23, 16
+; SI-NEXT: s_lshr_b32 s59, s46, 16
+; SI-NEXT: s_add_u32 s23, s23, 3
+; SI-NEXT: s_addc_u32 s47, s47, 0
+; SI-NEXT: s_lshr_b32 s60, s23, 16
+; SI-NEXT: s_lshr_b32 s61, s47, 16
; SI-NEXT: s_add_u32 s24, s24, 3
-; SI-NEXT: s_addc_u32 s25, s25, 0
-; SI-NEXT: s_lshr_b32 s60, s24, 16
-; SI-NEXT: s_lshr_b32 s61, s25, 16
-; SI-NEXT: s_add_u32 s26, s26, 3
; SI-NEXT: s_addc_u32 s27, s27, 0
-; SI-NEXT: s_lshr_b32 s62, s26, 16
+; SI-NEXT: s_lshr_b32 s62, s24, 16
; SI-NEXT: s_lshr_b32 s63, s27, 16
-; SI-NEXT: s_add_u32 s28, s28, 3
-; SI-NEXT: s_addc_u32 s29, s29, 0
-; SI-NEXT: s_lshr_b32 s72, s28, 16
-; SI-NEXT: s_lshr_b32 s73, s29, 16
-; SI-NEXT: s_add_u32 s44, s44, 3
-; SI-NEXT: s_addc_u32 s45, s45, 0
-; SI-NEXT: s_lshr_b32 s74, s44, 16
-; SI-NEXT: s_lshr_b32 s75, s45, 16
-; SI-NEXT: s_add_u32 s42, s42, 3
-; SI-NEXT: s_addc_u32 s43, s43, 0
-; SI-NEXT: s_lshr_b32 s76, s42, 16
-; SI-NEXT: s_lshr_b32 s77, s43, 16
-; SI-NEXT: s_add_u32 s40, s40, 3
-; SI-NEXT: s_addc_u32 s41, s41, 0
-; SI-NEXT: s_lshr_b32 s78, s40, 16
-; SI-NEXT: s_lshr_b32 s79, s41, 16
+; SI-NEXT: s_add_u32 s25, s25, 3
+; SI-NEXT: s_addc_u32 s26, s26, 0
+; SI-NEXT: s_lshr_b32 s72, s25, 16
+; SI-NEXT: s_lshr_b32 s73, s26, 16
+; SI-NEXT: s_add_u32 s20, s20, 3
+; SI-NEXT: s_addc_u32 s21, s21, 0
+; SI-NEXT: s_lshr_b32 s74, s20, 16
+; SI-NEXT: s_lshr_b32 s75, s21, 16
+; SI-NEXT: s_add_u32 s18, s18, 3
+; SI-NEXT: s_addc_u32 s19, s19, 0
+; SI-NEXT: s_lshr_b32 s76, s18, 16
+; SI-NEXT: s_lshr_b32 s77, s19, 16
+; SI-NEXT: s_add_u32 s16, s16, 3
+; SI-NEXT: s_addc_u32 s17, s17, 0
+; SI-NEXT: s_lshr_b32 s78, s16, 16
+; SI-NEXT: s_lshr_b32 s79, s17, 16
; SI-NEXT: s_add_u32 s14, s14, 3
; SI-NEXT: s_addc_u32 s15, s15, 0
; SI-NEXT: s_lshr_b32 s88, s14, 16
@@ -33480,25 +33017,25 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v18, s12
; SI-NEXT: v_cvt_f32_f16_e32 v20, s15
; SI-NEXT: v_cvt_f32_f16_e32 v21, s14
-; SI-NEXT: v_cvt_f32_f16_e32 v23, s41
-; SI-NEXT: v_cvt_f32_f16_e32 v25, s40
-; SI-NEXT: v_cvt_f32_f16_e32 v27, s43
-; SI-NEXT: v_cvt_f32_f16_e32 v29, s42
-; SI-NEXT: v_cvt_f32_f16_e32 v31, s45
-; SI-NEXT: v_cvt_f32_f16_e32 v33, s44
-; SI-NEXT: v_cvt_f32_f16_e32 v35, s29
-; SI-NEXT: v_cvt_f32_f16_e32 v37, s28
+; SI-NEXT: v_cvt_f32_f16_e32 v23, s17
+; SI-NEXT: v_cvt_f32_f16_e32 v25, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v27, s19
+; SI-NEXT: v_cvt_f32_f16_e32 v29, s18
+; SI-NEXT: v_cvt_f32_f16_e32 v31, s21
+; SI-NEXT: v_cvt_f32_f16_e32 v33, s20
+; SI-NEXT: v_cvt_f32_f16_e32 v35, s26
+; SI-NEXT: v_cvt_f32_f16_e32 v37, s25
; SI-NEXT: v_cvt_f32_f16_e32 v38, s27
-; SI-NEXT: v_cvt_f32_f16_e32 v48, s26
-; SI-NEXT: v_cvt_f32_f16_e32 v50, s25
-; SI-NEXT: v_cvt_f32_f16_e32 v52, s24
-; SI-NEXT: v_cvt_f32_f16_e32 v54, s23
+; SI-NEXT: v_cvt_f32_f16_e32 v48, s24
+; SI-NEXT: v_cvt_f32_f16_e32 v50, s47
+; SI-NEXT: v_cvt_f32_f16_e32 v52, s23
+; SI-NEXT: v_cvt_f32_f16_e32 v54, s46
; SI-NEXT: v_cvt_f32_f16_e32 v40, s22
-; SI-NEXT: v_cvt_f32_f16_e32 v42, s21
-; SI-NEXT: v_cvt_f32_f16_e32 v44, s20
-; SI-NEXT: v_cvt_f32_f16_e32 v45, s19
+; SI-NEXT: v_cvt_f32_f16_e32 v42, s45
+; SI-NEXT: v_cvt_f32_f16_e32 v44, s42
+; SI-NEXT: v_cvt_f32_f16_e32 v45, s41
; SI-NEXT: s_waitcnt expcnt(5)
-; SI-NEXT: v_cvt_f32_f16_e32 v47, s18
+; SI-NEXT: v_cvt_f32_f16_e32 v47, s40
; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: v_cvt_f32_f16_e32 v57, s5
; SI-NEXT: s_waitcnt expcnt(1)
@@ -33529,11 +33066,11 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v55, s58
; SI-NEXT: v_cvt_f32_f16_e32 v41, s57
; SI-NEXT: v_cvt_f32_f16_e32 v43, s56
-; SI-NEXT: v_cvt_f32_f16_e32 v46, s47
-; SI-NEXT: v_cvt_f32_f16_e32 v56, s46
-; SI-NEXT: v_cvt_f32_f16_e32 v58, s17
+; SI-NEXT: v_cvt_f32_f16_e32 v46, s44
+; SI-NEXT: v_cvt_f32_f16_e32 v56, s43
+; SI-NEXT: v_cvt_f32_f16_e32 v58, s29
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v60, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v60, s28
; SI-NEXT: .LBB45_3: ; %end
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
@@ -33830,18 +33367,46 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; VI-NEXT: s_mov_b64 exec, s[4:5]
; VI-NEXT: v_writelane_b32 v30, s30, 0
; VI-NEXT: v_writelane_b32 v30, s31, 1
+; VI-NEXT: v_mov_b32_e32 v17, s16
+; VI-NEXT: v_mov_b32_e32 v18, s17
; VI-NEXT: v_writelane_b32 v30, s34, 2
+; VI-NEXT: v_mov_b32_e32 v19, s18
+; VI-NEXT: v_readfirstlane_b32 s56, v17
+; VI-NEXT: v_mov_b32_e32 v17, s19
+; VI-NEXT: v_readfirstlane_b32 s47, v18
+; VI-NEXT: v_mov_b32_e32 v18, s20
; VI-NEXT: v_writelane_b32 v30, s35, 3
+; VI-NEXT: v_readfirstlane_b32 s46, v19
+; VI-NEXT: v_mov_b32_e32 v19, s21
+; VI-NEXT: v_readfirstlane_b32 s45, v17
+; VI-NEXT: v_mov_b32_e32 v17, s22
+; VI-NEXT: v_readfirstlane_b32 s44, v18
+; VI-NEXT: v_mov_b32_e32 v18, s23
; VI-NEXT: v_writelane_b32 v30, s36, 4
+; VI-NEXT: v_readfirstlane_b32 s43, v19
+; VI-NEXT: v_mov_b32_e32 v19, s24
+; VI-NEXT: v_readfirstlane_b32 s42, v17
+; VI-NEXT: v_mov_b32_e32 v17, s25
+; VI-NEXT: v_readfirstlane_b32 s41, v18
+; VI-NEXT: v_mov_b32_e32 v18, s26
; VI-NEXT: v_writelane_b32 v30, s37, 5
+; VI-NEXT: v_readfirstlane_b32 s40, v19
+; VI-NEXT: v_mov_b32_e32 v19, s27
+; VI-NEXT: v_readfirstlane_b32 s26, v17
+; VI-NEXT: v_mov_b32_e32 v17, s28
+; VI-NEXT: v_readfirstlane_b32 s25, v18
+; VI-NEXT: v_mov_b32_e32 v18, s29
; VI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; VI-NEXT: v_writelane_b32 v30, s38, 6
-; VI-NEXT: v_readfirstlane_b32 s45, v0
-; VI-NEXT: v_readfirstlane_b32 s44, v1
-; VI-NEXT: v_readfirstlane_b32 s43, v2
-; VI-NEXT: v_readfirstlane_b32 s42, v3
-; VI-NEXT: v_readfirstlane_b32 s41, v4
-; VI-NEXT: v_readfirstlane_b32 s40, v5
+; VI-NEXT: v_readfirstlane_b32 s24, v19
+; VI-NEXT: v_readfirstlane_b32 s23, v17
+; VI-NEXT: v_readfirstlane_b32 s22, v18
+; VI-NEXT: v_readfirstlane_b32 s21, v0
+; VI-NEXT: v_readfirstlane_b32 s20, v1
+; VI-NEXT: v_readfirstlane_b32 s19, v2
+; VI-NEXT: v_readfirstlane_b32 s18, v3
+; VI-NEXT: v_readfirstlane_b32 s17, v4
+; VI-NEXT: v_readfirstlane_b32 s16, v5
; VI-NEXT: v_readfirstlane_b32 s15, v6
; VI-NEXT: v_readfirstlane_b32 s14, v7
; VI-NEXT: v_readfirstlane_b32 s13, v8
@@ -33856,9 +33421,9 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; VI-NEXT: v_writelane_b32 v30, s39, 7
; VI-NEXT: s_cbranch_scc0 .LBB45_4
; VI-NEXT: ; %bb.1: ; %cmp.false
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_lshr_b32 s47, s6, 16
-; VI-NEXT: s_lshr_b32 s56, s8, 16
+; VI-NEXT: s_lshr_b32 s27, s7, 16
+; VI-NEXT: s_lshr_b32 s28, s6, 16
+; VI-NEXT: s_lshr_b32 s29, s8, 16
; VI-NEXT: s_lshr_b32 s57, s9, 16
; VI-NEXT: s_lshr_b32 s58, s10, 16
; VI-NEXT: s_lshr_b32 s59, s11, 16
@@ -33866,26 +33431,26 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; VI-NEXT: s_lshr_b32 s61, s13, 16
; VI-NEXT: s_lshr_b32 s62, s14, 16
; VI-NEXT: s_lshr_b32 s63, s15, 16
-; VI-NEXT: s_lshr_b32 s72, s40, 16
-; VI-NEXT: s_lshr_b32 s73, s41, 16
-; VI-NEXT: s_lshr_b32 s74, s42, 16
-; VI-NEXT: s_lshr_b32 s75, s43, 16
-; VI-NEXT: s_lshr_b32 s76, s44, 16
-; VI-NEXT: s_lshr_b32 s77, s45, 16
-; VI-NEXT: s_lshr_b32 s78, s29, 16
-; VI-NEXT: s_lshr_b32 s79, s28, 16
-; VI-NEXT: s_lshr_b32 s88, s27, 16
-; VI-NEXT: s_lshr_b32 s89, s26, 16
-; VI-NEXT: s_lshr_b32 s90, s25, 16
-; VI-NEXT: s_lshr_b32 s91, s24, 16
-; VI-NEXT: s_lshr_b32 s30, s23, 16
-; VI-NEXT: s_lshr_b32 s31, s22, 16
-; VI-NEXT: s_lshr_b32 s34, s21, 16
-; VI-NEXT: s_lshr_b32 s35, s20, 16
-; VI-NEXT: s_lshr_b32 s36, s19, 16
-; VI-NEXT: s_lshr_b32 s37, s18, 16
-; VI-NEXT: s_lshr_b32 s38, s17, 16
-; VI-NEXT: s_lshr_b32 s39, s16, 16
+; VI-NEXT: s_lshr_b32 s72, s16, 16
+; VI-NEXT: s_lshr_b32 s73, s17, 16
+; VI-NEXT: s_lshr_b32 s74, s18, 16
+; VI-NEXT: s_lshr_b32 s75, s19, 16
+; VI-NEXT: s_lshr_b32 s76, s20, 16
+; VI-NEXT: s_lshr_b32 s77, s21, 16
+; VI-NEXT: s_lshr_b32 s78, s22, 16
+; VI-NEXT: s_lshr_b32 s79, s23, 16
+; VI-NEXT: s_lshr_b32 s88, s24, 16
+; VI-NEXT: s_lshr_b32 s89, s25, 16
+; VI-NEXT: s_lshr_b32 s90, s26, 16
+; VI-NEXT: s_lshr_b32 s91, s40, 16
+; VI-NEXT: s_lshr_b32 s30, s41, 16
+; VI-NEXT: s_lshr_b32 s31, s42, 16
+; VI-NEXT: s_lshr_b32 s34, s43, 16
+; VI-NEXT: s_lshr_b32 s35, s44, 16
+; VI-NEXT: s_lshr_b32 s36, s45, 16
+; VI-NEXT: s_lshr_b32 s37, s46, 16
+; VI-NEXT: s_lshr_b32 s38, s47, 16
+; VI-NEXT: s_lshr_b32 s39, s56, 16
; VI-NEXT: s_cbranch_execnz .LBB45_3
; VI-NEXT: .LBB45_2: ; %cmp.true
; VI-NEXT: s_add_u32 s6, s6, 3
@@ -33898,29 +33463,29 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; VI-NEXT: s_addc_u32 s12, s12, 0
; VI-NEXT: s_add_u32 s15, s15, 3
; VI-NEXT: s_addc_u32 s14, s14, 0
-; VI-NEXT: s_add_u32 s41, s41, 3
-; VI-NEXT: s_addc_u32 s40, s40, 0
-; VI-NEXT: s_add_u32 s43, s43, 3
-; VI-NEXT: s_addc_u32 s42, s42, 0
-; VI-NEXT: s_add_u32 s45, s45, 3
-; VI-NEXT: s_addc_u32 s44, s44, 0
-; VI-NEXT: s_add_u32 s28, s28, 3
-; VI-NEXT: s_addc_u32 s29, s29, 0
-; VI-NEXT: s_add_u32 s26, s26, 3
-; VI-NEXT: s_addc_u32 s27, s27, 0
-; VI-NEXT: s_add_u32 s24, s24, 3
-; VI-NEXT: s_addc_u32 s25, s25, 0
-; VI-NEXT: s_add_u32 s22, s22, 3
-; VI-NEXT: s_addc_u32 s23, s23, 0
-; VI-NEXT: s_add_u32 s20, s20, 3
-; VI-NEXT: s_addc_u32 s21, s21, 0
-; VI-NEXT: s_add_u32 s18, s18, 3
-; VI-NEXT: s_addc_u32 s19, s19, 0
-; VI-NEXT: s_add_u32 s16, s16, 3
-; VI-NEXT: s_addc_u32 s17, s17, 0
-; VI-NEXT: s_lshr_b32 s46, s7, 16
-; VI-NEXT: s_lshr_b32 s47, s6, 16
-; VI-NEXT: s_lshr_b32 s56, s8, 16
+; VI-NEXT: s_add_u32 s17, s17, 3
+; VI-NEXT: s_addc_u32 s16, s16, 0
+; VI-NEXT: s_add_u32 s19, s19, 3
+; VI-NEXT: s_addc_u32 s18, s18, 0
+; VI-NEXT: s_add_u32 s21, s21, 3
+; VI-NEXT: s_addc_u32 s20, s20, 0
+; VI-NEXT: s_add_u32 s23, s23, 3
+; VI-NEXT: s_addc_u32 s22, s22, 0
+; VI-NEXT: s_add_u32 s25, s25, 3
+; VI-NEXT: s_addc_u32 s24, s24, 0
+; VI-NEXT: s_add_u32 s40, s40, 3
+; VI-NEXT: s_addc_u32 s26, s26, 0
+; VI-NEXT: s_add_u32 s42, s42, 3
+; VI-NEXT: s_addc_u32 s41, s41, 0
+; VI-NEXT: s_add_u32 s44, s44, 3
+; VI-NEXT: s_addc_u32 s43, s43, 0
+; VI-NEXT: s_add_u32 s46, s46, 3
+; VI-NEXT: s_addc_u32 s45, s45, 0
+; VI-NEXT: s_add_u32 s56, s56, 3
+; VI-NEXT: s_addc_u32 s47, s47, 0
+; VI-NEXT: s_lshr_b32 s27, s7, 16
+; VI-NEXT: s_lshr_b32 s28, s6, 16
+; VI-NEXT: s_lshr_b32 s29, s8, 16
; VI-NEXT: s_lshr_b32 s57, s9, 16
; VI-NEXT: s_lshr_b32 s58, s10, 16
; VI-NEXT: s_lshr_b32 s59, s11, 16
@@ -33928,137 +33493,137 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; VI-NEXT: s_lshr_b32 s61, s13, 16
; VI-NEXT: s_lshr_b32 s62, s14, 16
; VI-NEXT: s_lshr_b32 s63, s15, 16
-; VI-NEXT: s_lshr_b32 s72, s40, 16
-; VI-NEXT: s_lshr_b32 s73, s41, 16
-; VI-NEXT: s_lshr_b32 s74, s42, 16
-; VI-NEXT: s_lshr_b32 s75, s43, 16
-; VI-NEXT: s_lshr_b32 s76, s44, 16
-; VI-NEXT: s_lshr_b32 s77, s45, 16
-; VI-NEXT: s_lshr_b32 s78, s29, 16
-; VI-NEXT: s_lshr_b32 s79, s28, 16
-; VI-NEXT: s_lshr_b32 s88, s27, 16
-; VI-NEXT: s_lshr_b32 s89, s26, 16
-; VI-NEXT: s_lshr_b32 s90, s25, 16
-; VI-NEXT: s_lshr_b32 s91, s24, 16
-; VI-NEXT: s_lshr_b32 s30, s23, 16
-; VI-NEXT: s_lshr_b32 s31, s22, 16
-; VI-NEXT: s_lshr_b32 s34, s21, 16
-; VI-NEXT: s_lshr_b32 s35, s20, 16
-; VI-NEXT: s_lshr_b32 s36, s19, 16
-; VI-NEXT: s_lshr_b32 s37, s18, 16
-; VI-NEXT: s_lshr_b32 s38, s17, 16
-; VI-NEXT: s_lshr_b32 s39, s16, 16
+; VI-NEXT: s_lshr_b32 s72, s16, 16
+; VI-NEXT: s_lshr_b32 s73, s17, 16
+; VI-NEXT: s_lshr_b32 s74, s18, 16
+; VI-NEXT: s_lshr_b32 s75, s19, 16
+; VI-NEXT: s_lshr_b32 s76, s20, 16
+; VI-NEXT: s_lshr_b32 s77, s21, 16
+; VI-NEXT: s_lshr_b32 s78, s22, 16
+; VI-NEXT: s_lshr_b32 s79, s23, 16
+; VI-NEXT: s_lshr_b32 s88, s24, 16
+; VI-NEXT: s_lshr_b32 s89, s25, 16
+; VI-NEXT: s_lshr_b32 s90, s26, 16
+; VI-NEXT: s_lshr_b32 s91, s40, 16
+; VI-NEXT: s_lshr_b32 s30, s41, 16
+; VI-NEXT: s_lshr_b32 s31, s42, 16
+; VI-NEXT: s_lshr_b32 s34, s43, 16
+; VI-NEXT: s_lshr_b32 s35, s44, 16
+; VI-NEXT: s_lshr_b32 s36, s45, 16
+; VI-NEXT: s_lshr_b32 s37, s46, 16
+; VI-NEXT: s_lshr_b32 s38, s47, 16
+; VI-NEXT: s_lshr_b32 s39, s56, 16
; VI-NEXT: .LBB45_3: ; %end
-; VI-NEXT: s_and_b32 s4, 0xffff, s16
+; VI-NEXT: s_and_b32 s4, 0xffff, s56
; VI-NEXT: s_lshl_b32 s5, s39, 16
; VI-NEXT: s_or_b32 s4, s4, s5
-; VI-NEXT: s_and_b32 s5, 0xffff, s17
-; VI-NEXT: s_lshl_b32 s16, s38, 16
-; VI-NEXT: s_or_b32 s5, s5, s16
-; VI-NEXT: s_and_b32 s16, 0xffff, s18
-; VI-NEXT: s_lshl_b32 s17, s37, 16
-; VI-NEXT: s_or_b32 s16, s16, s17
-; VI-NEXT: s_and_b32 s17, 0xffff, s19
-; VI-NEXT: s_lshl_b32 s18, s36, 16
-; VI-NEXT: s_or_b32 s17, s17, s18
-; VI-NEXT: s_and_b32 s18, 0xffff, s20
-; VI-NEXT: s_lshl_b32 s19, s35, 16
-; VI-NEXT: s_or_b32 s18, s18, s19
-; VI-NEXT: s_and_b32 s19, 0xffff, s21
-; VI-NEXT: s_lshl_b32 s20, s34, 16
-; VI-NEXT: s_or_b32 s19, s19, s20
-; VI-NEXT: s_and_b32 s20, 0xffff, s22
-; VI-NEXT: s_lshl_b32 s21, s31, 16
-; VI-NEXT: s_or_b32 s20, s20, s21
-; VI-NEXT: s_and_b32 s21, 0xffff, s23
-; VI-NEXT: s_lshl_b32 s22, s30, 16
-; VI-NEXT: s_or_b32 s21, s21, s22
-; VI-NEXT: s_and_b32 s22, 0xffff, s24
-; VI-NEXT: s_lshl_b32 s23, s91, 16
-; VI-NEXT: s_or_b32 s22, s22, s23
-; VI-NEXT: s_and_b32 s23, 0xffff, s25
-; VI-NEXT: s_lshl_b32 s24, s90, 16
-; VI-NEXT: s_or_b32 s23, s23, s24
-; VI-NEXT: s_and_b32 s24, 0xffff, s26
-; VI-NEXT: s_lshl_b32 s25, s89, 16
-; VI-NEXT: s_or_b32 s24, s24, s25
-; VI-NEXT: s_and_b32 s25, 0xffff, s27
-; VI-NEXT: s_lshl_b32 s26, s88, 16
-; VI-NEXT: s_or_b32 s25, s25, s26
-; VI-NEXT: s_and_b32 s26, 0xffff, s28
-; VI-NEXT: s_lshl_b32 s27, s79, 16
-; VI-NEXT: s_or_b32 s26, s26, s27
-; VI-NEXT: s_and_b32 s27, 0xffff, s29
-; VI-NEXT: s_lshl_b32 s28, s78, 16
-; VI-NEXT: s_or_b32 s27, s27, s28
-; VI-NEXT: s_and_b32 s28, 0xffff, s45
-; VI-NEXT: s_lshl_b32 s29, s77, 16
-; VI-NEXT: s_or_b32 s28, s28, s29
-; VI-NEXT: s_and_b32 s29, 0xffff, s44
-; VI-NEXT: s_lshl_b32 s44, s76, 16
-; VI-NEXT: s_or_b32 s29, s29, s44
+; VI-NEXT: s_and_b32 s5, 0xffff, s47
+; VI-NEXT: s_lshl_b32 s47, s38, 16
+; VI-NEXT: s_or_b32 s5, s5, s47
+; VI-NEXT: s_and_b32 s46, 0xffff, s46
+; VI-NEXT: s_lshl_b32 s47, s37, 16
+; VI-NEXT: s_or_b32 s46, s46, s47
+; VI-NEXT: s_and_b32 s45, 0xffff, s45
+; VI-NEXT: s_lshl_b32 s47, s36, 16
+; VI-NEXT: s_or_b32 s45, s45, s47
+; VI-NEXT: s_and_b32 s44, 0xffff, s44
+; VI-NEXT: s_lshl_b32 s47, s35, 16
+; VI-NEXT: s_or_b32 s44, s44, s47
; VI-NEXT: s_and_b32 s43, 0xffff, s43
-; VI-NEXT: s_lshl_b32 s44, s75, 16
-; VI-NEXT: s_or_b32 s43, s43, s44
+; VI-NEXT: s_lshl_b32 s47, s34, 16
+; VI-NEXT: s_or_b32 s43, s43, s47
; VI-NEXT: s_and_b32 s42, 0xffff, s42
-; VI-NEXT: s_lshl_b32 s44, s74, 16
-; VI-NEXT: s_or_b32 s42, s42, s44
+; VI-NEXT: s_lshl_b32 s47, s31, 16
+; VI-NEXT: s_or_b32 s42, s42, s47
; VI-NEXT: s_and_b32 s41, 0xffff, s41
-; VI-NEXT: s_lshl_b32 s44, s73, 16
-; VI-NEXT: s_or_b32 s41, s41, s44
+; VI-NEXT: s_lshl_b32 s47, s30, 16
+; VI-NEXT: s_or_b32 s41, s41, s47
; VI-NEXT: s_and_b32 s40, 0xffff, s40
-; VI-NEXT: s_lshl_b32 s44, s72, 16
-; VI-NEXT: s_or_b32 s40, s40, s44
+; VI-NEXT: s_lshl_b32 s47, s91, 16
+; VI-NEXT: s_or_b32 s40, s40, s47
+; VI-NEXT: s_and_b32 s26, 0xffff, s26
+; VI-NEXT: s_lshl_b32 s47, s90, 16
+; VI-NEXT: s_or_b32 s26, s26, s47
+; VI-NEXT: s_and_b32 s25, 0xffff, s25
+; VI-NEXT: s_lshl_b32 s47, s89, 16
+; VI-NEXT: s_or_b32 s25, s25, s47
+; VI-NEXT: s_and_b32 s24, 0xffff, s24
+; VI-NEXT: s_lshl_b32 s47, s88, 16
+; VI-NEXT: s_or_b32 s24, s24, s47
+; VI-NEXT: s_and_b32 s23, 0xffff, s23
+; VI-NEXT: s_lshl_b32 s47, s79, 16
+; VI-NEXT: s_or_b32 s23, s23, s47
+; VI-NEXT: s_and_b32 s22, 0xffff, s22
+; VI-NEXT: s_lshl_b32 s47, s78, 16
+; VI-NEXT: s_or_b32 s22, s22, s47
+; VI-NEXT: s_and_b32 s21, 0xffff, s21
+; VI-NEXT: s_lshl_b32 s47, s77, 16
+; VI-NEXT: s_or_b32 s21, s21, s47
+; VI-NEXT: s_and_b32 s20, 0xffff, s20
+; VI-NEXT: s_lshl_b32 s47, s76, 16
+; VI-NEXT: s_or_b32 s20, s20, s47
+; VI-NEXT: s_and_b32 s19, 0xffff, s19
+; VI-NEXT: s_lshl_b32 s47, s75, 16
+; VI-NEXT: s_or_b32 s19, s19, s47
+; VI-NEXT: s_and_b32 s18, 0xffff, s18
+; VI-NEXT: s_lshl_b32 s47, s74, 16
+; VI-NEXT: s_or_b32 s18, s18, s47
+; VI-NEXT: s_and_b32 s17, 0xffff, s17
+; VI-NEXT: s_lshl_b32 s47, s73, 16
+; VI-NEXT: s_or_b32 s17, s17, s47
+; VI-NEXT: s_and_b32 s16, 0xffff, s16
+; VI-NEXT: s_lshl_b32 s47, s72, 16
+; VI-NEXT: s_or_b32 s16, s16, s47
; VI-NEXT: s_and_b32 s15, 0xffff, s15
-; VI-NEXT: s_lshl_b32 s44, s63, 16
-; VI-NEXT: s_or_b32 s15, s15, s44
+; VI-NEXT: s_lshl_b32 s47, s63, 16
+; VI-NEXT: s_or_b32 s15, s15, s47
; VI-NEXT: s_and_b32 s14, 0xffff, s14
-; VI-NEXT: s_lshl_b32 s44, s62, 16
-; VI-NEXT: s_or_b32 s14, s14, s44
+; VI-NEXT: s_lshl_b32 s47, s62, 16
+; VI-NEXT: s_or_b32 s14, s14, s47
; VI-NEXT: s_and_b32 s13, 0xffff, s13
-; VI-NEXT: s_lshl_b32 s44, s61, 16
-; VI-NEXT: s_or_b32 s13, s13, s44
+; VI-NEXT: s_lshl_b32 s47, s61, 16
+; VI-NEXT: s_or_b32 s13, s13, s47
; VI-NEXT: s_and_b32 s12, 0xffff, s12
-; VI-NEXT: s_lshl_b32 s44, s60, 16
-; VI-NEXT: s_or_b32 s12, s12, s44
+; VI-NEXT: s_lshl_b32 s47, s60, 16
+; VI-NEXT: s_or_b32 s12, s12, s47
; VI-NEXT: s_and_b32 s11, 0xffff, s11
-; VI-NEXT: s_lshl_b32 s44, s59, 16
-; VI-NEXT: s_or_b32 s11, s11, s44
+; VI-NEXT: s_lshl_b32 s47, s59, 16
+; VI-NEXT: s_or_b32 s11, s11, s47
; VI-NEXT: s_and_b32 s10, 0xffff, s10
-; VI-NEXT: s_lshl_b32 s44, s58, 16
-; VI-NEXT: s_or_b32 s10, s10, s44
+; VI-NEXT: s_lshl_b32 s47, s58, 16
+; VI-NEXT: s_or_b32 s10, s10, s47
; VI-NEXT: s_and_b32 s9, 0xffff, s9
-; VI-NEXT: s_lshl_b32 s44, s57, 16
-; VI-NEXT: s_or_b32 s9, s9, s44
+; VI-NEXT: s_lshl_b32 s47, s57, 16
; VI-NEXT: s_and_b32 s8, 0xffff, s8
-; VI-NEXT: s_lshl_b32 s44, s56, 16
-; VI-NEXT: s_or_b32 s8, s8, s44
+; VI-NEXT: s_lshl_b32 s29, s29, 16
; VI-NEXT: s_and_b32 s6, 0xffff, s6
-; VI-NEXT: s_lshl_b32 s44, s47, 16
-; VI-NEXT: s_or_b32 s6, s6, s44
+; VI-NEXT: s_lshl_b32 s28, s28, 16
; VI-NEXT: s_and_b32 s7, 0xffff, s7
-; VI-NEXT: s_lshl_b32 s44, s46, 16
-; VI-NEXT: s_or_b32 s7, s7, s44
+; VI-NEXT: s_lshl_b32 s27, s27, 16
+; VI-NEXT: s_or_b32 s9, s9, s47
+; VI-NEXT: s_or_b32 s8, s8, s29
+; VI-NEXT: s_or_b32 s6, s6, s28
+; VI-NEXT: s_or_b32 s7, s7, s27
; VI-NEXT: v_mov_b32_e32 v0, s4
; VI-NEXT: v_mov_b32_e32 v1, s5
-; VI-NEXT: v_mov_b32_e32 v2, s16
-; VI-NEXT: v_mov_b32_e32 v3, s17
-; VI-NEXT: v_mov_b32_e32 v4, s18
-; VI-NEXT: v_mov_b32_e32 v5, s19
-; VI-NEXT: v_mov_b32_e32 v6, s20
-; VI-NEXT: v_mov_b32_e32 v7, s21
-; VI-NEXT: v_mov_b32_e32 v8, s22
-; VI-NEXT: v_mov_b32_e32 v9, s23
-; VI-NEXT: v_mov_b32_e32 v10, s24
-; VI-NEXT: v_mov_b32_e32 v11, s25
-; VI-NEXT: v_mov_b32_e32 v12, s26
-; VI-NEXT: v_mov_b32_e32 v13, s27
-; VI-NEXT: v_mov_b32_e32 v14, s28
-; VI-NEXT: v_mov_b32_e32 v15, s29
-; VI-NEXT: v_mov_b32_e32 v16, s43
-; VI-NEXT: v_mov_b32_e32 v17, s42
-; VI-NEXT: v_mov_b32_e32 v18, s41
-; VI-NEXT: v_mov_b32_e32 v19, s40
+; VI-NEXT: v_mov_b32_e32 v2, s46
+; VI-NEXT: v_mov_b32_e32 v3, s45
+; VI-NEXT: v_mov_b32_e32 v4, s44
+; VI-NEXT: v_mov_b32_e32 v5, s43
+; VI-NEXT: v_mov_b32_e32 v6, s42
+; VI-NEXT: v_mov_b32_e32 v7, s41
+; VI-NEXT: v_mov_b32_e32 v8, s40
+; VI-NEXT: v_mov_b32_e32 v9, s26
+; VI-NEXT: v_mov_b32_e32 v10, s25
+; VI-NEXT: v_mov_b32_e32 v11, s24
+; VI-NEXT: v_mov_b32_e32 v12, s23
+; VI-NEXT: v_mov_b32_e32 v13, s22
+; VI-NEXT: v_mov_b32_e32 v14, s21
+; VI-NEXT: v_mov_b32_e32 v15, s20
+; VI-NEXT: v_mov_b32_e32 v16, s19
+; VI-NEXT: v_mov_b32_e32 v17, s18
+; VI-NEXT: v_mov_b32_e32 v18, s17
+; VI-NEXT: v_mov_b32_e32 v19, s16
; VI-NEXT: v_mov_b32_e32 v20, s15
; VI-NEXT: v_mov_b32_e32 v21, s14
; VI-NEXT: v_mov_b32_e32 v22, s13
@@ -34110,9 +33675,9 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; VI-NEXT: ; implicit-def: $sgpr59
; VI-NEXT: ; implicit-def: $sgpr58
; VI-NEXT: ; implicit-def: $sgpr57
-; VI-NEXT: ; implicit-def: $sgpr56
-; VI-NEXT: ; implicit-def: $sgpr47
-; VI-NEXT: ; implicit-def: $sgpr46
+; VI-NEXT: ; implicit-def: $sgpr29
+; VI-NEXT: ; implicit-def: $sgpr28
+; VI-NEXT: ; implicit-def: $sgpr27
; VI-NEXT: s_branch .LBB45_2
;
; GFX9-LABEL: bitcast_v15i64_to_v60f16_scalar:
@@ -34121,20 +33686,48 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v30, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: v_mov_b32_e32 v17, s16
+; GFX9-NEXT: v_mov_b32_e32 v18, s17
+; GFX9-NEXT: v_mov_b32_e32 v19, s18
+; GFX9-NEXT: v_readfirstlane_b32 s6, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s19
+; GFX9-NEXT: v_readfirstlane_b32 s7, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s20
+; GFX9-NEXT: v_readfirstlane_b32 s8, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s21
+; GFX9-NEXT: v_readfirstlane_b32 s9, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s22
+; GFX9-NEXT: v_readfirstlane_b32 s10, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s23
; GFX9-NEXT: v_writelane_b32 v30, s30, 0
+; GFX9-NEXT: v_readfirstlane_b32 s11, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s24
+; GFX9-NEXT: v_readfirstlane_b32 s12, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s25
+; GFX9-NEXT: v_readfirstlane_b32 s13, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s26
; GFX9-NEXT: v_writelane_b32 v30, s31, 1
+; GFX9-NEXT: v_readfirstlane_b32 s14, v19
+; GFX9-NEXT: v_mov_b32_e32 v19, s27
+; GFX9-NEXT: v_readfirstlane_b32 s15, v17
+; GFX9-NEXT: v_mov_b32_e32 v17, s28
+; GFX9-NEXT: v_readfirstlane_b32 s16, v18
+; GFX9-NEXT: v_mov_b32_e32 v18, s29
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v16
; GFX9-NEXT: v_writelane_b32 v30, s34, 2
-; GFX9-NEXT: v_readfirstlane_b32 s6, v0
-; GFX9-NEXT: v_readfirstlane_b32 s7, v1
-; GFX9-NEXT: v_readfirstlane_b32 s8, v2
-; GFX9-NEXT: v_readfirstlane_b32 s9, v3
-; GFX9-NEXT: v_readfirstlane_b32 s10, v4
-; GFX9-NEXT: v_readfirstlane_b32 s11, v5
-; GFX9-NEXT: v_readfirstlane_b32 s12, v6
-; GFX9-NEXT: v_readfirstlane_b32 s13, v7
-; GFX9-NEXT: v_readfirstlane_b32 s14, v8
-; GFX9-NEXT: v_readfirstlane_b32 s15, v9
+; GFX9-NEXT: v_readfirstlane_b32 s17, v19
+; GFX9-NEXT: v_readfirstlane_b32 s18, v17
+; GFX9-NEXT: v_readfirstlane_b32 s19, v18
+; GFX9-NEXT: v_readfirstlane_b32 s20, v0
+; GFX9-NEXT: v_readfirstlane_b32 s21, v1
+; GFX9-NEXT: v_readfirstlane_b32 s22, v2
+; GFX9-NEXT: v_readfirstlane_b32 s23, v3
+; GFX9-NEXT: v_readfirstlane_b32 s24, v4
+; GFX9-NEXT: v_readfirstlane_b32 s25, v5
+; GFX9-NEXT: v_readfirstlane_b32 s26, v6
+; GFX9-NEXT: v_readfirstlane_b32 s27, v7
+; GFX9-NEXT: v_readfirstlane_b32 s28, v8
+; GFX9-NEXT: v_readfirstlane_b32 s29, v9
; GFX9-NEXT: v_readfirstlane_b32 s40, v10
; GFX9-NEXT: v_readfirstlane_b32 s41, v11
; GFX9-NEXT: v_readfirstlane_b32 s42, v12
@@ -34151,30 +33744,30 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX9-NEXT: s_lshr_b32 s57, s42, 16
; GFX9-NEXT: s_lshr_b32 s58, s41, 16
; GFX9-NEXT: s_lshr_b32 s59, s40, 16
-; GFX9-NEXT: s_lshr_b32 s60, s15, 16
-; GFX9-NEXT: s_lshr_b32 s61, s14, 16
-; GFX9-NEXT: s_lshr_b32 s62, s13, 16
-; GFX9-NEXT: s_lshr_b32 s63, s12, 16
-; GFX9-NEXT: s_lshr_b32 s72, s11, 16
-; GFX9-NEXT: s_lshr_b32 s73, s10, 16
-; GFX9-NEXT: s_lshr_b32 s74, s9, 16
-; GFX9-NEXT: s_lshr_b32 s75, s8, 16
-; GFX9-NEXT: s_lshr_b32 s76, s7, 16
-; GFX9-NEXT: s_lshr_b32 s77, s6, 16
-; GFX9-NEXT: s_lshr_b32 s78, s29, 16
-; GFX9-NEXT: s_lshr_b32 s79, s28, 16
-; GFX9-NEXT: s_lshr_b32 s88, s27, 16
-; GFX9-NEXT: s_lshr_b32 s89, s26, 16
-; GFX9-NEXT: s_lshr_b32 s90, s25, 16
-; GFX9-NEXT: s_lshr_b32 s91, s24, 16
-; GFX9-NEXT: s_lshr_b32 s92, s23, 16
-; GFX9-NEXT: s_lshr_b32 s93, s22, 16
-; GFX9-NEXT: s_lshr_b32 s94, s21, 16
-; GFX9-NEXT: s_lshr_b32 s95, s20, 16
-; GFX9-NEXT: s_lshr_b32 s30, s19, 16
-; GFX9-NEXT: s_lshr_b32 s31, s18, 16
-; GFX9-NEXT: s_lshr_b32 s34, s17, 16
-; GFX9-NEXT: s_lshr_b32 s35, s16, 16
+; GFX9-NEXT: s_lshr_b32 s60, s29, 16
+; GFX9-NEXT: s_lshr_b32 s61, s28, 16
+; GFX9-NEXT: s_lshr_b32 s62, s27, 16
+; GFX9-NEXT: s_lshr_b32 s63, s26, 16
+; GFX9-NEXT: s_lshr_b32 s72, s25, 16
+; GFX9-NEXT: s_lshr_b32 s73, s24, 16
+; GFX9-NEXT: s_lshr_b32 s74, s23, 16
+; GFX9-NEXT: s_lshr_b32 s75, s22, 16
+; GFX9-NEXT: s_lshr_b32 s76, s21, 16
+; GFX9-NEXT: s_lshr_b32 s77, s20, 16
+; GFX9-NEXT: s_lshr_b32 s78, s19, 16
+; GFX9-NEXT: s_lshr_b32 s79, s18, 16
+; GFX9-NEXT: s_lshr_b32 s88, s17, 16
+; GFX9-NEXT: s_lshr_b32 s89, s16, 16
+; GFX9-NEXT: s_lshr_b32 s90, s15, 16
+; GFX9-NEXT: s_lshr_b32 s91, s14, 16
+; GFX9-NEXT: s_lshr_b32 s92, s13, 16
+; GFX9-NEXT: s_lshr_b32 s93, s12, 16
+; GFX9-NEXT: s_lshr_b32 s94, s11, 16
+; GFX9-NEXT: s_lshr_b32 s95, s10, 16
+; GFX9-NEXT: s_lshr_b32 s30, s9, 16
+; GFX9-NEXT: s_lshr_b32 s31, s8, 16
+; GFX9-NEXT: s_lshr_b32 s34, s7, 16
+; GFX9-NEXT: s_lshr_b32 s35, s6, 16
; GFX9-NEXT: s_cbranch_execnz .LBB45_3
; GFX9-NEXT: .LBB45_2: ; %cmp.true
; GFX9-NEXT: s_add_u32 s44, s44, 3
@@ -34183,16 +33776,6 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX9-NEXT: s_addc_u32 s43, s43, 0
; GFX9-NEXT: s_add_u32 s40, s40, 3
; GFX9-NEXT: s_addc_u32 s41, s41, 0
-; GFX9-NEXT: s_add_u32 s14, s14, 3
-; GFX9-NEXT: s_addc_u32 s15, s15, 0
-; GFX9-NEXT: s_add_u32 s12, s12, 3
-; GFX9-NEXT: s_addc_u32 s13, s13, 0
-; GFX9-NEXT: s_add_u32 s10, s10, 3
-; GFX9-NEXT: s_addc_u32 s11, s11, 0
-; GFX9-NEXT: s_add_u32 s8, s8, 3
-; GFX9-NEXT: s_addc_u32 s9, s9, 0
-; GFX9-NEXT: s_add_u32 s6, s6, 3
-; GFX9-NEXT: s_addc_u32 s7, s7, 0
; GFX9-NEXT: s_add_u32 s28, s28, 3
; GFX9-NEXT: s_addc_u32 s29, s29, 0
; GFX9-NEXT: s_add_u32 s26, s26, 3
@@ -34207,61 +33790,71 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX9-NEXT: s_addc_u32 s19, s19, 0
; GFX9-NEXT: s_add_u32 s16, s16, 3
; GFX9-NEXT: s_addc_u32 s17, s17, 0
+; GFX9-NEXT: s_add_u32 s14, s14, 3
+; GFX9-NEXT: s_addc_u32 s15, s15, 0
+; GFX9-NEXT: s_add_u32 s12, s12, 3
+; GFX9-NEXT: s_addc_u32 s13, s13, 0
+; GFX9-NEXT: s_add_u32 s10, s10, 3
+; GFX9-NEXT: s_addc_u32 s11, s11, 0
+; GFX9-NEXT: s_add_u32 s8, s8, 3
+; GFX9-NEXT: s_addc_u32 s9, s9, 0
+; GFX9-NEXT: s_add_u32 s6, s6, 3
+; GFX9-NEXT: s_addc_u32 s7, s7, 0
; GFX9-NEXT: s_lshr_b32 s46, s45, 16
; GFX9-NEXT: s_lshr_b32 s47, s44, 16
; GFX9-NEXT: s_lshr_b32 s56, s43, 16
; GFX9-NEXT: s_lshr_b32 s57, s42, 16
; GFX9-NEXT: s_lshr_b32 s58, s41, 16
; GFX9-NEXT: s_lshr_b32 s59, s40, 16
-; GFX9-NEXT: s_lshr_b32 s60, s15, 16
-; GFX9-NEXT: s_lshr_b32 s61, s14, 16
-; GFX9-NEXT: s_lshr_b32 s62, s13, 16
-; GFX9-NEXT: s_lshr_b32 s63, s12, 16
-; GFX9-NEXT: s_lshr_b32 s72, s11, 16
-; GFX9-NEXT: s_lshr_b32 s73, s10, 16
-; GFX9-NEXT: s_lshr_b32 s74, s9, 16
-; GFX9-NEXT: s_lshr_b32 s75, s8, 16
-; GFX9-NEXT: s_lshr_b32 s76, s7, 16
-; GFX9-NEXT: s_lshr_b32 s77, s6, 16
-; GFX9-NEXT: s_lshr_b32 s78, s29, 16
-; GFX9-NEXT: s_lshr_b32 s79, s28, 16
-; GFX9-NEXT: s_lshr_b32 s88, s27, 16
-; GFX9-NEXT: s_lshr_b32 s89, s26, 16
-; GFX9-NEXT: s_lshr_b32 s90, s25, 16
-; GFX9-NEXT: s_lshr_b32 s91, s24, 16
-; GFX9-NEXT: s_lshr_b32 s92, s23, 16
-; GFX9-NEXT: s_lshr_b32 s93, s22, 16
-; GFX9-NEXT: s_lshr_b32 s94, s21, 16
-; GFX9-NEXT: s_lshr_b32 s95, s20, 16
-; GFX9-NEXT: s_lshr_b32 s30, s19, 16
-; GFX9-NEXT: s_lshr_b32 s31, s18, 16
-; GFX9-NEXT: s_lshr_b32 s34, s17, 16
-; GFX9-NEXT: s_lshr_b32 s35, s16, 16
+; GFX9-NEXT: s_lshr_b32 s60, s29, 16
+; GFX9-NEXT: s_lshr_b32 s61, s28, 16
+; GFX9-NEXT: s_lshr_b32 s62, s27, 16
+; GFX9-NEXT: s_lshr_b32 s63, s26, 16
+; GFX9-NEXT: s_lshr_b32 s72, s25, 16
+; GFX9-NEXT: s_lshr_b32 s73, s24, 16
+; GFX9-NEXT: s_lshr_b32 s74, s23, 16
+; GFX9-NEXT: s_lshr_b32 s75, s22, 16
+; GFX9-NEXT: s_lshr_b32 s76, s21, 16
+; GFX9-NEXT: s_lshr_b32 s77, s20, 16
+; GFX9-NEXT: s_lshr_b32 s78, s19, 16
+; GFX9-NEXT: s_lshr_b32 s79, s18, 16
+; GFX9-NEXT: s_lshr_b32 s88, s17, 16
+; GFX9-NEXT: s_lshr_b32 s89, s16, 16
+; GFX9-NEXT: s_lshr_b32 s90, s15, 16
+; GFX9-NEXT: s_lshr_b32 s91, s14, 16
+; GFX9-NEXT: s_lshr_b32 s92, s13, 16
+; GFX9-NEXT: s_lshr_b32 s93, s12, 16
+; GFX9-NEXT: s_lshr_b32 s94, s11, 16
+; GFX9-NEXT: s_lshr_b32 s95, s10, 16
+; GFX9-NEXT: s_lshr_b32 s30, s9, 16
+; GFX9-NEXT: s_lshr_b32 s31, s8, 16
+; GFX9-NEXT: s_lshr_b32 s34, s7, 16
+; GFX9-NEXT: s_lshr_b32 s35, s6, 16
; GFX9-NEXT: .LBB45_3: ; %end
-; GFX9-NEXT: s_pack_ll_b32_b16 s4, s16, s35
-; GFX9-NEXT: s_pack_ll_b32_b16 s5, s17, s34
-; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s31
-; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s30
-; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s95
-; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s94
-; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s93
-; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s92
-; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s91
-; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s90
-; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s89
-; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s88
-; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s79
-; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s78
-; GFX9-NEXT: s_pack_ll_b32_b16 s6, s6, s77
-; GFX9-NEXT: s_pack_ll_b32_b16 s7, s7, s76
-; GFX9-NEXT: s_pack_ll_b32_b16 s8, s8, s75
-; GFX9-NEXT: s_pack_ll_b32_b16 s9, s9, s74
-; GFX9-NEXT: s_pack_ll_b32_b16 s10, s10, s73
-; GFX9-NEXT: s_pack_ll_b32_b16 s11, s11, s72
-; GFX9-NEXT: s_pack_ll_b32_b16 s12, s12, s63
-; GFX9-NEXT: s_pack_ll_b32_b16 s13, s13, s62
-; GFX9-NEXT: s_pack_ll_b32_b16 s14, s14, s61
-; GFX9-NEXT: s_pack_ll_b32_b16 s15, s15, s60
+; GFX9-NEXT: s_pack_ll_b32_b16 s4, s6, s35
+; GFX9-NEXT: s_pack_ll_b32_b16 s5, s7, s34
+; GFX9-NEXT: s_pack_ll_b32_b16 s6, s8, s31
+; GFX9-NEXT: s_pack_ll_b32_b16 s7, s9, s30
+; GFX9-NEXT: s_pack_ll_b32_b16 s8, s10, s95
+; GFX9-NEXT: s_pack_ll_b32_b16 s9, s11, s94
+; GFX9-NEXT: s_pack_ll_b32_b16 s10, s12, s93
+; GFX9-NEXT: s_pack_ll_b32_b16 s11, s13, s92
+; GFX9-NEXT: s_pack_ll_b32_b16 s12, s14, s91
+; GFX9-NEXT: s_pack_ll_b32_b16 s13, s15, s90
+; GFX9-NEXT: s_pack_ll_b32_b16 s14, s16, s89
+; GFX9-NEXT: s_pack_ll_b32_b16 s15, s17, s88
+; GFX9-NEXT: s_pack_ll_b32_b16 s16, s18, s79
+; GFX9-NEXT: s_pack_ll_b32_b16 s17, s19, s78
+; GFX9-NEXT: s_pack_ll_b32_b16 s18, s20, s77
+; GFX9-NEXT: s_pack_ll_b32_b16 s19, s21, s76
+; GFX9-NEXT: s_pack_ll_b32_b16 s20, s22, s75
+; GFX9-NEXT: s_pack_ll_b32_b16 s21, s23, s74
+; GFX9-NEXT: s_pack_ll_b32_b16 s22, s24, s73
+; GFX9-NEXT: s_pack_ll_b32_b16 s23, s25, s72
+; GFX9-NEXT: s_pack_ll_b32_b16 s24, s26, s63
+; GFX9-NEXT: s_pack_ll_b32_b16 s25, s27, s62
+; GFX9-NEXT: s_pack_ll_b32_b16 s26, s28, s61
+; GFX9-NEXT: s_pack_ll_b32_b16 s27, s29, s60
; GFX9-NEXT: s_pack_ll_b32_b16 s28, s40, s59
; GFX9-NEXT: s_pack_ll_b32_b16 s29, s41, s58
; GFX9-NEXT: s_pack_ll_b32_b16 s40, s42, s57
@@ -34270,28 +33863,28 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX9-NEXT: s_pack_ll_b32_b16 s43, s45, s46
; GFX9-NEXT: v_mov_b32_e32 v0, s4
; GFX9-NEXT: v_mov_b32_e32 v1, s5
-; GFX9-NEXT: v_mov_b32_e32 v2, s16
-; GFX9-NEXT: v_mov_b32_e32 v3, s17
-; GFX9-NEXT: v_mov_b32_e32 v4, s18
-; GFX9-NEXT: v_mov_b32_e32 v5, s19
-; GFX9-NEXT: v_mov_b32_e32 v6, s20
-; GFX9-NEXT: v_mov_b32_e32 v7, s21
-; GFX9-NEXT: v_mov_b32_e32 v8, s22
-; GFX9-NEXT: v_mov_b32_e32 v9, s23
-; GFX9-NEXT: v_mov_b32_e32 v10, s24
-; GFX9-NEXT: v_mov_b32_e32 v11, s25
-; GFX9-NEXT: v_mov_b32_e32 v12, s26
-; GFX9-NEXT: v_mov_b32_e32 v13, s27
-; GFX9-NEXT: v_mov_b32_e32 v14, s6
-; GFX9-NEXT: v_mov_b32_e32 v15, s7
-; GFX9-NEXT: v_mov_b32_e32 v16, s8
-; GFX9-NEXT: v_mov_b32_e32 v17, s9
-; GFX9-NEXT: v_mov_b32_e32 v18, s10
-; GFX9-NEXT: v_mov_b32_e32 v19, s11
-; GFX9-NEXT: v_mov_b32_e32 v20, s12
-; GFX9-NEXT: v_mov_b32_e32 v21, s13
-; GFX9-NEXT: v_mov_b32_e32 v22, s14
-; GFX9-NEXT: v_mov_b32_e32 v23, s15
+; GFX9-NEXT: v_mov_b32_e32 v2, s6
+; GFX9-NEXT: v_mov_b32_e32 v3, s7
+; GFX9-NEXT: v_mov_b32_e32 v4, s8
+; GFX9-NEXT: v_mov_b32_e32 v5, s9
+; GFX9-NEXT: v_mov_b32_e32 v6, s10
+; GFX9-NEXT: v_mov_b32_e32 v7, s11
+; GFX9-NEXT: v_mov_b32_e32 v8, s12
+; GFX9-NEXT: v_mov_b32_e32 v9, s13
+; GFX9-NEXT: v_mov_b32_e32 v10, s14
+; GFX9-NEXT: v_mov_b32_e32 v11, s15
+; GFX9-NEXT: v_mov_b32_e32 v12, s16
+; GFX9-NEXT: v_mov_b32_e32 v13, s17
+; GFX9-NEXT: v_mov_b32_e32 v14, s18
+; GFX9-NEXT: v_mov_b32_e32 v15, s19
+; GFX9-NEXT: v_mov_b32_e32 v16, s20
+; GFX9-NEXT: v_mov_b32_e32 v17, s21
+; GFX9-NEXT: v_mov_b32_e32 v18, s22
+; GFX9-NEXT: v_mov_b32_e32 v19, s23
+; GFX9-NEXT: v_mov_b32_e32 v20, s24
+; GFX9-NEXT: v_mov_b32_e32 v21, s25
+; GFX9-NEXT: v_mov_b32_e32 v22, s26
+; GFX9-NEXT: v_mov_b32_e32 v23, s27
; GFX9-NEXT: v_mov_b32_e32 v24, s28
; GFX9-NEXT: v_mov_b32_e32 v25, s29
; GFX9-NEXT: v_mov_b32_e32 v26, s40
@@ -34343,49 +33936,76 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX11-LABEL: bitcast_v15i64_to_v60f16_scalar:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT: v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
+; GFX11-NEXT: v_dual_mov_b32 v17, s16 :: v_dual_mov_b32 v18, s17
+; GFX11-NEXT: v_dual_mov_b32 v19, s18 :: v_dual_mov_b32 v20, s19
+; GFX11-NEXT: v_dual_mov_b32 v21, s20 :: v_dual_mov_b32 v22, s21
+; GFX11-NEXT: v_dual_mov_b32 v23, s22 :: v_dual_mov_b32 v24, s23
+; GFX11-NEXT: v_dual_mov_b32 v25, s24 :: v_dual_mov_b32 v26, s25
+; GFX11-NEXT: v_dual_mov_b32 v27, s26 :: v_dual_mov_b32 v28, s27
+; GFX11-NEXT: v_dual_mov_b32 v29, s28 :: v_dual_mov_b32 v30, s29
; GFX11-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-NEXT: v_readfirstlane_b32 s4, v0
-; GFX11-NEXT: v_readfirstlane_b32 s5, v1
-; GFX11-NEXT: v_readfirstlane_b32 s6, v2
-; GFX11-NEXT: v_readfirstlane_b32 s7, v3
-; GFX11-NEXT: v_readfirstlane_b32 s8, v4
-; GFX11-NEXT: v_readfirstlane_b32 s9, v5
-; GFX11-NEXT: v_readfirstlane_b32 s10, v6
-; GFX11-NEXT: v_readfirstlane_b32 s11, v7
-; GFX11-NEXT: v_readfirstlane_b32 s12, v8
-; GFX11-NEXT: v_readfirstlane_b32 s13, v9
-; GFX11-NEXT: v_readfirstlane_b32 s15, v10
-; GFX11-NEXT: v_readfirstlane_b32 s14, v11
+; GFX11-NEXT: v_readfirstlane_b32 s0, v13
+; GFX11-NEXT: v_readfirstlane_b32 s1, v14
+; GFX11-NEXT: v_readfirstlane_b32 s2, v15
+; GFX11-NEXT: v_readfirstlane_b32 s3, v16
+; GFX11-NEXT: v_readfirstlane_b32 s4, v17
+; GFX11-NEXT: v_readfirstlane_b32 s5, v18
+; GFX11-NEXT: v_readfirstlane_b32 s6, v19
+; GFX11-NEXT: v_readfirstlane_b32 s7, v20
+; GFX11-NEXT: v_readfirstlane_b32 s8, v21
+; GFX11-NEXT: v_readfirstlane_b32 s9, v22
+; GFX11-NEXT: v_readfirstlane_b32 s10, v23
+; GFX11-NEXT: v_readfirstlane_b32 s11, v24
+; GFX11-NEXT: v_readfirstlane_b32 s12, v25
+; GFX11-NEXT: v_readfirstlane_b32 s13, v26
+; GFX11-NEXT: v_readfirstlane_b32 s14, v27
+; GFX11-NEXT: v_readfirstlane_b32 s15, v28
+; GFX11-NEXT: v_readfirstlane_b32 s16, v29
+; GFX11-NEXT: v_readfirstlane_b32 s17, v30
+; GFX11-NEXT: v_readfirstlane_b32 s18, v0
+; GFX11-NEXT: v_readfirstlane_b32 s19, v1
+; GFX11-NEXT: v_readfirstlane_b32 s20, v2
+; GFX11-NEXT: v_readfirstlane_b32 s21, v3
+; GFX11-NEXT: v_readfirstlane_b32 s22, v4
+; GFX11-NEXT: v_readfirstlane_b32 s23, v5
+; GFX11-NEXT: v_readfirstlane_b32 s24, v6
+; GFX11-NEXT: v_readfirstlane_b32 s25, v7
+; GFX11-NEXT: v_readfirstlane_b32 s26, v8
+; GFX11-NEXT: v_readfirstlane_b32 s27, v9
+; GFX11-NEXT: v_readfirstlane_b32 s29, v10
+; GFX11-NEXT: v_readfirstlane_b32 s28, v11
; GFX11-NEXT: s_mov_b32 s94, 0
; GFX11-NEXT: s_and_b32 s40, vcc_lo, exec_lo
; GFX11-NEXT: s_cbranch_scc0 .LBB45_4
; GFX11-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-NEXT: s_lshr_b32 s40, s14, 16
-; GFX11-NEXT: s_lshr_b32 s41, s15, 16
-; GFX11-NEXT: s_lshr_b32 s42, s13, 16
-; GFX11-NEXT: s_lshr_b32 s43, s12, 16
-; GFX11-NEXT: s_lshr_b32 s44, s11, 16
-; GFX11-NEXT: s_lshr_b32 s45, s10, 16
-; GFX11-NEXT: s_lshr_b32 s46, s9, 16
-; GFX11-NEXT: s_lshr_b32 s47, s8, 16
-; GFX11-NEXT: s_lshr_b32 s56, s7, 16
-; GFX11-NEXT: s_lshr_b32 s57, s6, 16
-; GFX11-NEXT: s_lshr_b32 s58, s5, 16
-; GFX11-NEXT: s_lshr_b32 s59, s4, 16
-; GFX11-NEXT: s_lshr_b32 s60, s29, 16
-; GFX11-NEXT: s_lshr_b32 s61, s28, 16
-; GFX11-NEXT: s_lshr_b32 s62, s27, 16
-; GFX11-NEXT: s_lshr_b32 s63, s26, 16
-; GFX11-NEXT: s_lshr_b32 s72, s25, 16
-; GFX11-NEXT: s_lshr_b32 s73, s24, 16
-; GFX11-NEXT: s_lshr_b32 s74, s23, 16
-; GFX11-NEXT: s_lshr_b32 s75, s22, 16
-; GFX11-NEXT: s_lshr_b32 s76, s21, 16
-; GFX11-NEXT: s_lshr_b32 s77, s20, 16
-; GFX11-NEXT: s_lshr_b32 s78, s19, 16
-; GFX11-NEXT: s_lshr_b32 s79, s18, 16
-; GFX11-NEXT: s_lshr_b32 s88, s17, 16
-; GFX11-NEXT: s_lshr_b32 s89, s16, 16
+; GFX11-NEXT: s_lshr_b32 s40, s28, 16
+; GFX11-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-NEXT: s_lshr_b32 s43, s26, 16
+; GFX11-NEXT: s_lshr_b32 s44, s25, 16
+; GFX11-NEXT: s_lshr_b32 s45, s24, 16
+; GFX11-NEXT: s_lshr_b32 s46, s23, 16
+; GFX11-NEXT: s_lshr_b32 s47, s22, 16
+; GFX11-NEXT: s_lshr_b32 s56, s21, 16
+; GFX11-NEXT: s_lshr_b32 s57, s20, 16
+; GFX11-NEXT: s_lshr_b32 s58, s19, 16
+; GFX11-NEXT: s_lshr_b32 s59, s18, 16
+; GFX11-NEXT: s_lshr_b32 s60, s17, 16
+; GFX11-NEXT: s_lshr_b32 s61, s16, 16
+; GFX11-NEXT: s_lshr_b32 s62, s15, 16
+; GFX11-NEXT: s_lshr_b32 s63, s14, 16
+; GFX11-NEXT: s_lshr_b32 s72, s13, 16
+; GFX11-NEXT: s_lshr_b32 s73, s12, 16
+; GFX11-NEXT: s_lshr_b32 s74, s11, 16
+; GFX11-NEXT: s_lshr_b32 s75, s10, 16
+; GFX11-NEXT: s_lshr_b32 s76, s9, 16
+; GFX11-NEXT: s_lshr_b32 s77, s8, 16
+; GFX11-NEXT: s_lshr_b32 s78, s7, 16
+; GFX11-NEXT: s_lshr_b32 s79, s6, 16
+; GFX11-NEXT: s_lshr_b32 s88, s5, 16
+; GFX11-NEXT: s_lshr_b32 s89, s4, 16
; GFX11-NEXT: s_lshr_b32 s90, s3, 16
; GFX11-NEXT: s_lshr_b32 s91, s2, 16
; GFX11-NEXT: s_lshr_b32 s92, s1, 16
@@ -34393,20 +34013,8 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s94
; GFX11-NEXT: s_cbranch_vccnz .LBB45_3
; GFX11-NEXT: .LBB45_2: ; %cmp.true
-; GFX11-NEXT: s_add_u32 s15, s15, 3
-; GFX11-NEXT: s_addc_u32 s14, s14, 0
-; GFX11-NEXT: s_add_u32 s12, s12, 3
-; GFX11-NEXT: s_addc_u32 s13, s13, 0
-; GFX11-NEXT: s_add_u32 s10, s10, 3
-; GFX11-NEXT: s_addc_u32 s11, s11, 0
-; GFX11-NEXT: s_add_u32 s8, s8, 3
-; GFX11-NEXT: s_addc_u32 s9, s9, 0
-; GFX11-NEXT: s_add_u32 s6, s6, 3
-; GFX11-NEXT: s_addc_u32 s7, s7, 0
-; GFX11-NEXT: s_add_u32 s4, s4, 3
-; GFX11-NEXT: s_addc_u32 s5, s5, 0
-; GFX11-NEXT: s_add_u32 s28, s28, 3
-; GFX11-NEXT: s_addc_u32 s29, s29, 0
+; GFX11-NEXT: s_add_u32 s29, s29, 3
+; GFX11-NEXT: s_addc_u32 s28, s28, 0
; GFX11-NEXT: s_add_u32 s26, s26, 3
; GFX11-NEXT: s_addc_u32 s27, s27, 0
; GFX11-NEXT: s_add_u32 s24, s24, 3
@@ -34419,36 +34027,48 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX11-NEXT: s_addc_u32 s19, s19, 0
; GFX11-NEXT: s_add_u32 s16, s16, 3
; GFX11-NEXT: s_addc_u32 s17, s17, 0
+; GFX11-NEXT: s_add_u32 s14, s14, 3
+; GFX11-NEXT: s_addc_u32 s15, s15, 0
+; GFX11-NEXT: s_add_u32 s12, s12, 3
+; GFX11-NEXT: s_addc_u32 s13, s13, 0
+; GFX11-NEXT: s_add_u32 s10, s10, 3
+; GFX11-NEXT: s_addc_u32 s11, s11, 0
+; GFX11-NEXT: s_add_u32 s8, s8, 3
+; GFX11-NEXT: s_addc_u32 s9, s9, 0
+; GFX11-NEXT: s_add_u32 s6, s6, 3
+; GFX11-NEXT: s_addc_u32 s7, s7, 0
+; GFX11-NEXT: s_add_u32 s4, s4, 3
+; GFX11-NEXT: s_addc_u32 s5, s5, 0
; GFX11-NEXT: s_add_u32 s2, s2, 3
; GFX11-NEXT: s_addc_u32 s3, s3, 0
; GFX11-NEXT: s_add_u32 s0, s0, 3
; GFX11-NEXT: s_addc_u32 s1, s1, 0
-; GFX11-NEXT: s_lshr_b32 s40, s14, 16
-; GFX11-NEXT: s_lshr_b32 s41, s15, 16
-; GFX11-NEXT: s_lshr_b32 s42, s13, 16
-; GFX11-NEXT: s_lshr_b32 s43, s12, 16
-; GFX11-NEXT: s_lshr_b32 s44, s11, 16
-; GFX11-NEXT: s_lshr_b32 s45, s10, 16
-; GFX11-NEXT: s_lshr_b32 s46, s9, 16
-; GFX11-NEXT: s_lshr_b32 s47, s8, 16
-; GFX11-NEXT: s_lshr_b32 s56, s7, 16
-; GFX11-NEXT: s_lshr_b32 s57, s6, 16
-; GFX11-NEXT: s_lshr_b32 s58, s5, 16
-; GFX11-NEXT: s_lshr_b32 s59, s4, 16
-; GFX11-NEXT: s_lshr_b32 s60, s29, 16
-; GFX11-NEXT: s_lshr_b32 s61, s28, 16
-; GFX11-NEXT: s_lshr_b32 s62, s27, 16
-; GFX11-NEXT: s_lshr_b32 s63, s26, 16
-; GFX11-NEXT: s_lshr_b32 s72, s25, 16
-; GFX11-NEXT: s_lshr_b32 s73, s24, 16
-; GFX11-NEXT: s_lshr_b32 s74, s23, 16
-; GFX11-NEXT: s_lshr_b32 s75, s22, 16
-; GFX11-NEXT: s_lshr_b32 s76, s21, 16
-; GFX11-NEXT: s_lshr_b32 s77, s20, 16
-; GFX11-NEXT: s_lshr_b32 s78, s19, 16
-; GFX11-NEXT: s_lshr_b32 s79, s18, 16
-; GFX11-NEXT: s_lshr_b32 s88, s17, 16
-; GFX11-NEXT: s_lshr_b32 s89, s16, 16
+; GFX11-NEXT: s_lshr_b32 s40, s28, 16
+; GFX11-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-NEXT: s_lshr_b32 s42, s27, 16
+; GFX11-NEXT: s_lshr_b32 s43, s26, 16
+; GFX11-NEXT: s_lshr_b32 s44, s25, 16
+; GFX11-NEXT: s_lshr_b32 s45, s24, 16
+; GFX11-NEXT: s_lshr_b32 s46, s23, 16
+; GFX11-NEXT: s_lshr_b32 s47, s22, 16
+; GFX11-NEXT: s_lshr_b32 s56, s21, 16
+; GFX11-NEXT: s_lshr_b32 s57, s20, 16
+; GFX11-NEXT: s_lshr_b32 s58, s19, 16
+; GFX11-NEXT: s_lshr_b32 s59, s18, 16
+; GFX11-NEXT: s_lshr_b32 s60, s17, 16
+; GFX11-NEXT: s_lshr_b32 s61, s16, 16
+; GFX11-NEXT: s_lshr_b32 s62, s15, 16
+; GFX11-NEXT: s_lshr_b32 s63, s14, 16
+; GFX11-NEXT: s_lshr_b32 s72, s13, 16
+; GFX11-NEXT: s_lshr_b32 s73, s12, 16
+; GFX11-NEXT: s_lshr_b32 s74, s11, 16
+; GFX11-NEXT: s_lshr_b32 s75, s10, 16
+; GFX11-NEXT: s_lshr_b32 s76, s9, 16
+; GFX11-NEXT: s_lshr_b32 s77, s8, 16
+; GFX11-NEXT: s_lshr_b32 s78, s7, 16
+; GFX11-NEXT: s_lshr_b32 s79, s6, 16
+; GFX11-NEXT: s_lshr_b32 s88, s5, 16
+; GFX11-NEXT: s_lshr_b32 s89, s4, 16
; GFX11-NEXT: s_lshr_b32 s90, s3, 16
; GFX11-NEXT: s_lshr_b32 s91, s2, 16
; GFX11-NEXT: s_lshr_b32 s92, s1, 16
@@ -34459,47 +34079,47 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
; GFX11-NEXT: s_pack_ll_b32_b16 s1, s1, s92
; GFX11-NEXT: s_pack_ll_b32_b16 s2, s2, s91
; GFX11-NEXT: s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s59
-; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s58
-; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s57
-; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s56
-; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s47
-; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s46
-; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s45
-; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s44
-; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s43
-; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s42
-; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s41
-; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s40
+; GFX11-NEXT: s_pack_ll_b32_b16 s4, s4, s89
+; GFX11-NEXT: s_pack_ll_b32_b16 s5, s5, s88
+; GFX11-NEXT: s_pack_ll_b32_b16 s6, s6, s79
+; GFX11-NEXT: s_pack_ll_b32_b16 s7, s7, s78
+; GFX11-NEXT: s_pack_ll_b32_b16 s8, s8, s77
+; GFX11-NEXT: s_pack_ll_b32_b16 s9, s9, s76
+; GFX11-NEXT: s_pack_ll_b32_b16 s10, s10, s75
+; GFX11-NEXT: s_pack_ll_b32_b16 s11, s11, s74
+; GFX11-NEXT: s_pack_ll_b32_b16 s12, s12, s73
+; GFX11-NEXT: s_pack_ll_b32_b16 s13, s13, s72
+; GFX11-NEXT: s_pack_ll_b32_b16 s14, s14, s63
+; GFX11-NEXT: s_pack_ll_b32_b16 s15, s15, s62
+; GFX11-NEXT: s_pack_ll_b32_b16 s16, s16, s61
+; GFX11-NEXT: s_pack_ll_b32_b16 s17, s17, s60
+; GFX11-NEXT: s_pack_ll_b32_b16 s18, s18, s59
+; GFX11-NEXT: s_pack_ll_b32_b16 s19, s19, s58
+; GFX11-NEXT: s_pack_ll_b32_b16 s20, s20, s57
+; GFX11-NEXT: s_pack_ll_b32_b16 s21, s21, s56
+; GFX11-NEXT: s_pack_ll_b32_b16 s22, s22, s47
+; GFX11-NEXT: s_pack_ll_b32_b16 s23, s23, s46
+; GFX11-NEXT: s_pack_ll_b32_b16 s24, s24, s45
+; GFX11-NEXT: s_pack_ll_b32_b16 s25, s25, s44
+; GFX11-NEXT: s_pack_ll_b32_b16 s26, s26, s43
+; GFX11-NEXT: s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-NEXT: s_pack_ll_b32_b16 s29, s29, s41
+; GFX11-NEXT: s_pack_ll_b32_b16 s28, s28, s40
; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-NEXT: v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-NEXT: v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-NEXT: v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-NEXT: v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-NEXT: v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-NEXT: v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-NEXT: v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-NEXT: v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-NEXT: v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-NEXT: v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
-; GFX11-NEXT: v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT: v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT: v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT: v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT: v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT: v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT: v_dual_mov_b32 v28, s29 :: v_dual_mov_b32 v29, s28
; GFX11-NEXT: s_setpc_b64 s[30:31]
; GFX11-NEXT: .LBB45_4:
; GFX11-NEXT: ; implicit-def: $sgpr93
@@ -34570,6 +34190,9 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(5)
+; SI-NEXT: v_cvt_f16_f32_e32 v58, v0
+; SI-NEXT: v_cvt_f16_f32_e32 v0, v5
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116
; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32
@@ -34598,23 +34221,12 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:88
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:84
-; SI-NEXT: v_cvt_f16_f32_e32 v58, v0
-; SI-NEXT: v_cvt_f16_f32_e32 v0, v5
; SI-NEXT: v_cvt_f16_f32_e32 v59, v1
; SI-NEXT: v_cvt_f16_f32_e32 v57, v3
-; SI-NEXT: v_cvt_f16_f32_e32 v56, v2
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:104
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:100
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v4
-; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
+; SI-NEXT: v_cvt_f16_f32_e32 v56, v2
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v7
@@ -34624,8 +34236,6 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v9
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v8
@@ -34647,9 +34257,18 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v14
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v17
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:104
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:100
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
+; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v16
@@ -34659,6 +34278,7 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v18
+; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v21
@@ -34696,7 +34316,6 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v61
; SI-NEXT: v_cvt_f16_f32_e32 v61, v49
-; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f16_f32_e32 v49, v55
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -34711,6 +34330,7 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v32
+; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f16_f32_e32 v32, v47
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -35582,7 +35202,6 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: v_mov_b32_e32 v61, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v28
@@ -35661,6 +35280,7 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v60
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v61
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
@@ -35825,6 +35445,9 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT: s_cbranch_execz .LBB46_4
; GFX9-NEXT: ; %bb.3: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
@@ -35839,9 +35462,6 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v38, v61, s6
; GFX9-NEXT: s_movk_i32 s7, 0x200
; GFX9-NEXT: v_perm_b32 v1, v37, v60, s6
@@ -35861,6 +35481,10 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
+; GFX9-NEXT: s_waitcnt vmcnt(14)
+; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: s_waitcnt vmcnt(15)
; GFX9-NEXT: v_perm_b32 v9, v9, v44, s6
; GFX9-NEXT: s_waitcnt vmcnt(14)
@@ -35889,10 +35513,6 @@ define <15 x i64> @bitcast_v60f16_to_v15i64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_perm_b32 v21, v21, v48, s6
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_perm_b32 v22, v22, v39, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
@@ -36213,12 +35833,35 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_cvt_f16_f32_e32 v8, s26
; SI-NEXT: v_cvt_f16_f32_e32 v6, s29
; SI-NEXT: v_cvt_f16_f32_e32 v7, s28
+; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT: v_cvt_f16_f32_e32 v50, v54
; SI-NEXT: v_cvt_f16_f32_e32 v48, v48
; SI-NEXT: v_cvt_f16_f32_e32 v31, v40
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v33
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
@@ -36228,7 +35871,7 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v38
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v44
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -36237,7 +35880,7 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v46
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v47
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -36246,7 +35889,7 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v57
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v58
; SI-NEXT: v_cvt_f16_f32_e32 v58, s16
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
@@ -36257,38 +35900,12 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v60
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_cbranch_scc0 .LBB47_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT: v_or_b32_e32 v3, v10, v3
-; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v34
; SI-NEXT: v_mov_b32_e32 v33, v32
; SI-NEXT: v_or_b32_e32 v10, v32, v10
@@ -36313,12 +35930,12 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v59
; SI-NEXT: v_or_b32_e32 v1, v12, v1
; SI-NEXT: v_or_b32_e32 v2, v11, v2
@@ -36427,12 +36044,10 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_mov_b32_e32 v40, v44
; SI-NEXT: s_cbranch_vccnz .LBB47_5
; SI-NEXT: ; %bb.4: ; %cmp.true
-; SI-NEXT: s_waitcnt expcnt(5)
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT: v_cvt_f32_f16_e32 v0, v59
; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v0, v59
; SI-NEXT: v_cvt_f32_f16_e32 v1, v58
; SI-NEXT: s_waitcnt vmcnt(3)
; SI-NEXT: v_cvt_f32_f16_e32 v8, v33
@@ -37218,304 +36833,115 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v156, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v157, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v158, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v159, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v168, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v169, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v170, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v171, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v172, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v173, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v174, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v175, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v184, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v185, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v186, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v187, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v188, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v189, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v190, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v191, s32
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v25, v7
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v191, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v190, v6 :: v_dual_mov_b32 v185, v4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v186, v3 :: v_dual_mov_b32 v187, v2
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v188, v1 :: v_dual_mov_b32 v189, v0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s28, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s19, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s17, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s1, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s0, 16
-; GFX11-TRUE16-NEXT: s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s40, s0, s40
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s41, s1, s41
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s2, s46
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s3, s45
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s16, s44
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s17, s43
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s18, s4
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s19, s5
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s20, s6
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s21, s7
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s22, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s23, s9
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s25, s11
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s26, s12
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s27, s13
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s28, s14
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s29, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB47_4
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s40 :: v_dual_mov_b32 v5, s0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s41 :: v_dual_mov_b32 v9, s1
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s2 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s3 :: v_dual_mov_b32 v35, s5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v44, s6 :: v_dual_mov_b32 v65, s8
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s7 :: v_dual_mov_b32 v77, s9
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v90, s10 :: v_dual_mov_b32 v119, s12
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v104, s11 :: v_dual_mov_b32 v135, s13
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v152, s14
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v170, s15
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB47_3
; GFX11-TRUE16-NEXT: .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s40 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s41 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v189, 0x200, v189 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v188, 0x200, v188 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v187, 0x200, v187 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v186, 0x200, v186 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v185, 0x200, v185 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v191, 0x200, v191 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v190, 0x200, v190 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v35, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v44, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v54, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v65, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v77, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v90, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v104, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v119, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v135, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v152, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v170, 0x200, s15 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: .LBB47_3: ; %end
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v20 :: v_dual_mov_b32 v6, v27
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v35 :: v_dual_mov_b32 v8, v44
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v77 :: v_dual_mov_b32 v12, v90
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v104
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, v135 :: v_dual_mov_b32 v16, v152
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v170 :: v_dual_mov_b32 v18, v189
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v188 :: v_dual_mov_b32 v20, v187
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v186 :: v_dual_mov_b32 v22, v185
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v191 :: v_dual_mov_b32 v24, v190
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v191, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v190, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v189, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v188, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v187, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v186, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v185, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v184, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v175, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v174, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v173, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v172, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v171, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v170, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v169, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v168, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v159, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v158, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v157, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v156, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:124
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:252
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:316
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v14
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v54 :: v_dual_mov_b32 v10, v65
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v119 :: v_dual_mov_b32 v27, v30
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB47_4:
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, v29 :: v_dual_mov_b32 v65, v28
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v66, v30
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v26 :: v_dual_mov_b32 v54, v25
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v53
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v65 :: v_dual_mov_b32 v29, v64
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v30, v66
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168_vgpr169_vgpr170_vgpr171_vgpr172_vgpr173_vgpr174_vgpr175_vgpr176_vgpr177_vgpr178_vgpr179_vgpr180_vgpr181_vgpr182_vgpr183_vgpr184
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_branch .LBB47_2
;
; GFX11-FAKE16-LABEL: bitcast_v60f16_to_v15i64_scalar:
@@ -39888,40 +39314,53 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:4
; SI-NEXT: v_lshlrev_b32_e32 v44, 16, v2
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v4
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v6, off, s[0:3], s32
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:112
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:108
+; SI-NEXT: v_lshlrev_b32_e32 v38, 16, v4
+; SI-NEXT: s_waitcnt vmcnt(4)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v6
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v8
-; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:104
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:100
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v10
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v45, 16, v6
+; SI-NEXT: v_lshlrev_b32_e32 v62, 16, v8
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:96
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:92
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_lshlrev_b32_e32 v63, 16, v10
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:88
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:84
-; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_lshlrev_b32_e32 v32, 16, v12
+; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt vmcnt(2)
; SI-NEXT: v_lshlrev_b32_e32 v40, 16, v14
-; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:48
@@ -39952,27 +39391,10 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:56
; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:64
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52
-; SI-NEXT: s_waitcnt vmcnt(3)
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: v_lshlrev_b32_e32 v42, 16, v28
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:36
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
+; SI-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; SI-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT: s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
@@ -40057,7 +39479,6 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; SI-NEXT: ; implicit-def: $vgpr30
; SI-NEXT: ; kill: killed $vgpr30
; SI-NEXT: ; implicit-def: $vgpr30
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v55
; SI-NEXT: v_and_b32_e32 v18, 0xffff, v61
; SI-NEXT: ; kill: killed $vgpr30
@@ -40202,7 +39623,6 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_add_i32_e32 v0, vcc, 3, v55
; SI-NEXT: v_add_i32_e32 v18, vcc, 3, v61
; SI-NEXT: v_and_b32_e32 v0, 0xffff, v0
@@ -40329,7 +39749,7 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; SI-NEXT: v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT: v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT: v_add_i32_e32 v26, vcc, s6, v26
-; SI-NEXT: v_add_i32_e32 v27, vcc, s6, v27
+; SI-NEXT: v_add_i32_e32 v27, vcc, 0x30000, v27
; SI-NEXT: v_add_i32_e32 v28, vcc, 0x30000, v28
; SI-NEXT: v_add_i32_e32 v29, vcc, 0x30000, v29
; SI-NEXT: .LBB50_4: ; %end
@@ -40350,7 +39770,7 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v60i16_to_v15f64:
@@ -40632,7 +40052,6 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: v_mov_b32_e32 v61, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v28
@@ -40711,6 +40130,7 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v60
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v61
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
@@ -40875,6 +40295,9 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT: s_cbranch_execz .LBB50_4
; GFX9-NEXT: ; %bb.3: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
@@ -40889,9 +40312,6 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v38, v61, s6
; GFX9-NEXT: v_perm_b32 v1, v37, v60, s6
; GFX9-NEXT: v_perm_b32 v2, v62, v59, s6
@@ -40910,6 +40330,10 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
+; GFX9-NEXT: s_waitcnt vmcnt(14)
+; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: s_waitcnt vmcnt(15)
; GFX9-NEXT: v_perm_b32 v9, v9, v44, s6
; GFX9-NEXT: s_waitcnt vmcnt(14)
@@ -40938,10 +40362,6 @@ define <15 x double> @bitcast_v60i16_to_v15f64(<60 x i16> %a, i32 %b) {
; GFX9-NEXT: v_perm_b32 v21, v21, v48, s6
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_perm_b32 v22, v22, v39, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
@@ -42097,304 +41517,115 @@ define inreg <15 x double> @bitcast_v60i16_to_v15f64_scalar(<60 x i16> inreg %a,
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v156, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v157, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v158, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v159, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v168, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v169, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v170, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v171, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v172, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v173, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v174, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v175, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v184, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v185, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v186, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v187, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v188, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v189, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v190, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v191, s32
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v25, v7
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v191, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v190, v6 :: v_dual_mov_b32 v185, v4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v186, v3 :: v_dual_mov_b32 v187, v2
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v188, v1 :: v_dual_mov_b32 v189, v0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s28, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s19, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s17, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s1, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s0, 16
-; GFX11-TRUE16-NEXT: s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s40, s0, s40
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s41, s1, s41
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s2, s46
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s3, s45
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s16, s44
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s17, s43
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s18, s4
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s19, s5
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s20, s6
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s21, s7
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s22, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s23, s9
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s25, s11
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s26, s12
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s27, s13
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s28, s14
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s29, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB51_4
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s40 :: v_dual_mov_b32 v5, s0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s41 :: v_dual_mov_b32 v9, s1
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s2 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s3 :: v_dual_mov_b32 v35, s5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v44, s6 :: v_dual_mov_b32 v65, s8
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s7 :: v_dual_mov_b32 v77, s9
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v90, s10 :: v_dual_mov_b32 v119, s12
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v104, s11 :: v_dual_mov_b32 v135, s13
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v152, s14
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v170, s15
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB51_3
; GFX11-TRUE16-NEXT: .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s40, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s41, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v189, v189, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v188, v188, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v187, v187, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v186, v186, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v185, v185, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v191, v191, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v190, v190, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, s4, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v35, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v44, s6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v54, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v65, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v77, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v90, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v104, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v119, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v135, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v152, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v170, s15, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: .LBB51_3: ; %end
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v20 :: v_dual_mov_b32 v6, v27
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v35 :: v_dual_mov_b32 v8, v44
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v77 :: v_dual_mov_b32 v12, v90
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v104
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, v135 :: v_dual_mov_b32 v16, v152
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v170 :: v_dual_mov_b32 v18, v189
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v188 :: v_dual_mov_b32 v20, v187
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v186 :: v_dual_mov_b32 v22, v185
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v191 :: v_dual_mov_b32 v24, v190
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v191, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v190, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v189, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v188, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v187, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v186, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v185, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v184, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v175, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v174, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v173, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v172, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v171, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v170, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v169, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v168, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v159, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v158, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v157, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v156, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:124
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:252
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:316
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v14
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v54 :: v_dual_mov_b32 v10, v65
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v119 :: v_dual_mov_b32 v27, v30
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB51_4:
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, v29 :: v_dual_mov_b32 v65, v28
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v66, v30
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v26 :: v_dual_mov_b32 v54, v25
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v53
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v65 :: v_dual_mov_b32 v29, v64
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v30, v66
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168_vgpr169_vgpr170_vgpr171_vgpr172_vgpr173_vgpr174_vgpr175_vgpr176_vgpr177_vgpr178_vgpr179_vgpr180_vgpr181_vgpr182_vgpr183_vgpr184
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_branch .LBB51_2
;
; GFX11-FAKE16-LABEL: bitcast_v60i16_to_v15f64_scalar:
@@ -43823,23 +43054,21 @@ define inreg <60 x half> @bitcast_v15f64_to_v60f16_scalar(<15 x double> inreg %a
; SI: ; %bb.0:
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v17
-; SI-NEXT: v_readfirstlane_b32 s42, v1
-; SI-NEXT: v_readfirstlane_b32 s43, v2
-; SI-NEXT: v_readfirstlane_b32 s40, v3
-; SI-NEXT: v_readfirstlane_b32 s41, v4
-; SI-NEXT: v_readfirstlane_b32 s14, v5
-; SI-NEXT: v_readfirstlane_b32 s15, v6
-; SI-NEXT: v_readfirstlane_b32 s12, v7
-; SI-NEXT: v_readfirstlane_b32 s13, v8
-; SI-NEXT: v_readfirstlane_b32 s10, v9
-; SI-NEXT: v_readfirstlane_b32 s11, v10
-; SI-NEXT: v_readfirstlane_b32 s8, v11
-; SI-NEXT: v_readfirstlane_b32 s9, v12
-; SI-NEXT: v_readfirstlane_b32 s6, v13
-; SI-NEXT: v_readfirstlane_b32 s7, v14
-; SI-NEXT: v_readfirstlane_b32 s4, v15
-; SI-NEXT: s_and_b64 s[44:45], vcc, exec
-; SI-NEXT: v_readfirstlane_b32 s5, v16
+; SI-NEXT: v_mov_b32_e32 v23, s16
+; SI-NEXT: v_mov_b32_e32 v24, s17
+; SI-NEXT: v_mov_b32_e32 v29, s18
+; SI-NEXT: v_mov_b32_e32 v30, s19
+; SI-NEXT: v_mov_b32_e32 v27, s20
+; SI-NEXT: v_mov_b32_e32 v28, s21
+; SI-NEXT: v_mov_b32_e32 v25, s22
+; SI-NEXT: v_mov_b32_e32 v26, s23
+; SI-NEXT: v_mov_b32_e32 v21, s24
+; SI-NEXT: v_mov_b32_e32 v22, s25
+; SI-NEXT: v_mov_b32_e32 v19, s26
+; SI-NEXT: v_mov_b32_e32 v20, s27
+; SI-NEXT: s_and_b64 s[4:5], vcc, exec
+; SI-NEXT: v_mov_b32_e32 v17, s28
+; SI-NEXT: v_mov_b32_e32 v18, s29
; SI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
@@ -43858,438 +43087,565 @@ define inreg <60 x half> @bitcast_v15f64_to_v60f16_scalar(<15 x double> inreg %a
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT: s_cbranch_scc0 .LBB53_4
; SI-NEXT: ; %bb.1: ; %cmp.false
-; SI-NEXT: s_lshr_b32 s44, s5, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v40, s44
-; SI-NEXT: s_lshr_b32 s44, s4, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v52, s44
-; SI-NEXT: s_lshr_b32 s44, s7, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v48, s44
-; SI-NEXT: s_lshr_b32 s44, s6, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v36, s44
-; SI-NEXT: s_lshr_b32 s44, s9, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v33, s44
-; SI-NEXT: s_lshr_b32 s44, s8, 16
-; SI-NEXT: s_waitcnt expcnt(4)
-; SI-NEXT: v_cvt_f32_f16_e32 v59, s44
-; SI-NEXT: s_lshr_b32 s44, s11, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v25, s44
-; SI-NEXT: s_lshr_b32 s44, s10, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v1, s44
-; SI-NEXT: s_lshr_b32 s44, s13, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v18, s44
-; SI-NEXT: s_lshr_b32 s44, s12, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v17, s44
-; SI-NEXT: s_lshr_b32 s44, s15, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v14, s44
-; SI-NEXT: s_lshr_b32 s44, s14, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v22, s44
-; SI-NEXT: s_lshr_b32 s44, s41, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v24, s44
-; SI-NEXT: s_lshr_b32 s44, s40, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v26, s44
-; SI-NEXT: s_lshr_b32 s44, s43, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v28, s44
-; SI-NEXT: s_lshr_b32 s44, s42, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v30, s44
-; SI-NEXT: s_lshr_b32 s44, s29, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v32, s44
-; SI-NEXT: s_lshr_b32 s44, s28, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v34, s44
-; SI-NEXT: s_lshr_b32 s44, s27, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v37, s44
-; SI-NEXT: s_lshr_b32 s44, s26, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v39, s44
-; SI-NEXT: s_lshr_b32 s44, s25, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v49, s44
-; SI-NEXT: s_lshr_b32 s44, s24, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v51, s44
-; SI-NEXT: s_lshr_b32 s44, s23, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v53, s44
-; SI-NEXT: s_lshr_b32 s44, s22, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v55, s44
-; SI-NEXT: s_lshr_b32 s44, s21, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v41, s44
-; SI-NEXT: s_lshr_b32 s44, s20, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v44, s44
-; SI-NEXT: s_lshr_b32 s44, s19, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v46, s44
-; SI-NEXT: s_lshr_b32 s44, s18, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v56, s44
-; SI-NEXT: s_lshr_b32 s44, s17, 16
-; SI-NEXT: v_cvt_f32_f16_e32 v58, s44
-; SI-NEXT: s_lshr_b32 s44, s16, 16
-; SI-NEXT: s_waitcnt expcnt(3)
-; SI-NEXT: v_cvt_f32_f16_e32 v60, s44
-; SI-NEXT: v_cvt_f32_f16_e32 v10, s5
-; SI-NEXT: v_cvt_f32_f16_e32 v13, s4
-; SI-NEXT: v_cvt_f32_f16_e32 v19, s7
-; SI-NEXT: v_cvt_f32_f16_e32 v29, s6
-; SI-NEXT: v_cvt_f32_f16_e32 v43, s9
-; SI-NEXT: v_cvt_f32_f16_e32 v45, s8
-; SI-NEXT: v_cvt_f32_f16_e32 v47, s11
-; SI-NEXT: s_waitcnt expcnt(2)
-; SI-NEXT: v_cvt_f32_f16_e32 v61, s10
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_cvt_f32_f16_e32 v62, s13
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v63, s12
-; SI-NEXT: v_cvt_f32_f16_e32 v23, s15
-; SI-NEXT: v_cvt_f32_f16_e32 v4, s14
-; SI-NEXT: v_cvt_f32_f16_e32 v27, s41
-; SI-NEXT: v_cvt_f32_f16_e32 v11, s40
-; SI-NEXT: v_cvt_f32_f16_e32 v31, s43
-; SI-NEXT: v_cvt_f32_f16_e32 v12, s42
-; SI-NEXT: v_cvt_f32_f16_e32 v35, s29
-; SI-NEXT: v_cvt_f32_f16_e32 v3, s28
-; SI-NEXT: v_cvt_f32_f16_e32 v38, s27
-; SI-NEXT: v_cvt_f32_f16_e32 v15, s26
-; SI-NEXT: v_cvt_f32_f16_e32 v50, s25
-; SI-NEXT: v_cvt_f32_f16_e32 v16, s24
-; SI-NEXT: v_cvt_f32_f16_e32 v54, s23
-; SI-NEXT: v_cvt_f32_f16_e32 v7, s22
-; SI-NEXT: v_cvt_f32_f16_e32 v42, s21
-; SI-NEXT: v_cvt_f32_f16_e32 v8, s20
-; SI-NEXT: v_cvt_f32_f16_e32 v20, s19
-; SI-NEXT: v_cvt_f32_f16_e32 v21, s18
-; SI-NEXT: v_cvt_f32_f16_e32 v57, s17
-; SI-NEXT: v_cvt_f32_f16_e32 v2, s16
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v12
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v16
+; SI-NEXT: v_cvt_f32_f16_e32 v40, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v15
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v11
+; SI-NEXT: v_cvt_f32_f16_e32 v54, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v14
+; SI-NEXT: v_cvt_f32_f16_e32 v52, v31
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v10
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v13
+; SI-NEXT: v_cvt_f32_f16_e32 v51, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v12
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v9
+; SI-NEXT: v_cvt_f32_f16_e32 v49, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v11
+; SI-NEXT: v_cvt_f32_f16_e32 v39, v31
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v8
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v10
+; SI-NEXT: v_cvt_f32_f16_e32 v37, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v9
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v7
+; SI-NEXT: v_cvt_f32_f16_e32 v35, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v8
+; SI-NEXT: v_cvt_f32_f16_e32 v59, v31
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v6
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v7
+; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v5
+; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v6
+; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v4
+; SI-NEXT: v_cvt_f32_f16_e32 v33, v15
+; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v5
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v3
+; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT: v_cvt_f32_f16_e32 v60, v14
+; SI-NEXT: v_cvt_f32_f16_e32 v62, v13
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v2
+; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v4
+; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v1
+; SI-NEXT: v_cvt_f32_f16_e32 v44, v27
+; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v3
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v18
+; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT: v_cvt_f32_f16_e32 v46, v30
+; SI-NEXT: v_cvt_f32_f16_e32 v47, v29
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v17
+; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v2
+; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v20
+; SI-NEXT: v_cvt_f32_f16_e32 v56, v24
+; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v1
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v19
+; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
+; SI-NEXT: v_cvt_f32_f16_e32 v57, v23
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v22
+; SI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v18
+; SI-NEXT: v_cvt_f32_f16_e32 v58, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v17
+; SI-NEXT: v_cvt_f32_f16_e32 v61, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v20
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v21
+; SI-NEXT: v_cvt_f32_f16_e32 v63, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v19
+; SI-NEXT: v_cvt_f32_f16_e32 v32, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v22
+; SI-NEXT: v_cvt_f32_f16_e32 v34, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v21
+; SI-NEXT: v_cvt_f32_f16_e32 v36, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v26
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v26
+; SI-NEXT: v_cvt_f32_f16_e32 v38, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v25
+; SI-NEXT: v_cvt_f32_f16_e32 v48, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v28
+; SI-NEXT: v_cvt_f32_f16_e32 v50, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v27
+; SI-NEXT: v_cvt_f32_f16_e32 v53, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v30
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v25
+; SI-NEXT: v_cvt_f32_f16_e32 v55, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v29
+; SI-NEXT: v_cvt_f32_f16_e32 v41, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v24
+; SI-NEXT: v_cvt_f32_f16_e32 v43, v31
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v23
+; SI-NEXT: v_cvt_f32_f16_e32 v45, v31
+; SI-NEXT: v_cvt_f32_f16_e32 v31, v16
+; SI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v42, v28
; SI-NEXT: s_cbranch_execnz .LBB53_3
; SI-NEXT: .LBB53_2: ; %cmp.true
-; SI-NEXT: v_add_f64 v[1:2], s[16:17], 1.0
+; SI-NEXT: v_add_f64 v[1:2], v[1:2], 1.0
+; SI-NEXT: v_add_f64 v[17:18], v[17:18], 1.0
; SI-NEXT: s_waitcnt expcnt(5)
-; SI-NEXT: v_add_f64 v[57:58], s[18:19], 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v1
+; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v1
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
+; SI-NEXT: v_add_f64 v[19:20], v[19:20], 1.0
+; SI-NEXT: v_add_f64 v[3:4], v[3:4], 1.0
+; SI-NEXT: v_add_f64 v[7:8], v[7:8], 1.0
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v18
+; SI-NEXT: v_lshrrev_b32_e32 v63, 16, v3
+; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
+; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v7
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v17
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v3, v31
+; SI-NEXT: v_add_f64 v[21:22], v[21:22], 1.0
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v20
+; SI-NEXT: v_add_f64 v[5:6], v[5:6], 1.0
+; SI-NEXT: v_lshrrev_b32_e32 v61, 16, v2
+; SI-NEXT: v_lshrrev_b32_e32 v62, 16, v6
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v19
+; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v57
-; SI-NEXT: v_add_f64 v[41:42], s[20:21], 1.0
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; SI-NEXT: v_cvt_f32_f16_e32 v3, v62
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v22
+; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v5
+; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v8
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v58
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v21
+; SI-NEXT: v_cvt_f32_f16_e32 v2, v33
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v41
-; SI-NEXT: v_add_f64 v[53:54], s[22:23], 1.0
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; SI-NEXT: v_cvt_f32_f16_e32 v3, v60
+; SI-NEXT: v_add_f64 v[25:26], v[25:26], 1.0
+; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v4
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v42
-; SI-NEXT: v_add_f64 v[15:16], s[10:11], 1.0
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v26
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v53
-; SI-NEXT: v_add_f64 v[20:21], s[12:13], 1.0
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v15
-; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v20
-; SI-NEXT: v_lshrrev_b32_e32 v18, 16, v21
-; SI-NEXT: v_cvt_f32_f16_e32 v62, v21
-; SI-NEXT: v_cvt_f32_f16_e32 v63, v20
-; SI-NEXT: v_cvt_f32_f16_e32 v20, v58
-; SI-NEXT: v_cvt_f32_f16_e32 v21, v57
-; SI-NEXT: v_lshrrev_b32_e32 v58, 16, v2
-; SI-NEXT: v_cvt_f32_f16_e32 v57, v2
-; SI-NEXT: v_cvt_f32_f16_e32 v2, v1
-; SI-NEXT: v_cvt_f32_f16_e32 v1, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; SI-NEXT: v_add_f64 v[37:38], s[26:27], 1.0
-; SI-NEXT: v_cvt_f32_f16_e32 v61, v15
-; SI-NEXT: v_lshrrev_b32_e32 v55, 16, v38
-; SI-NEXT: v_lshrrev_b32_e32 v44, 16, v37
-; SI-NEXT: v_cvt_f32_f16_e32 v15, v37
-; SI-NEXT: v_cvt_f32_f16_e32 v37, v55
-; SI-NEXT: v_add_f64 v[7:8], s[6:7], 1.0
-; SI-NEXT: v_add_f64 v[34:35], s[28:29], 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v8
-; SI-NEXT: v_cvt_f32_f16_e32 v19, v8
-; SI-NEXT: v_cvt_f32_f16_e32 v8, v41
-; SI-NEXT: v_add_f64 v[30:31], s[42:43], 1.0
-; SI-NEXT: v_add_f64 v[11:12], s[8:9], 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v35
-; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v30
-; SI-NEXT: v_lshrrev_b32_e32 v33, 16, v12
-; SI-NEXT: v_cvt_f32_f16_e32 v43, v12
-; SI-NEXT: v_cvt_f32_f16_e32 v12, v30
-; SI-NEXT: v_cvt_f32_f16_e32 v30, v32
-; SI-NEXT: v_cvt_f32_f16_e32 v32, v39
-; SI-NEXT: v_cvt_f32_f16_e32 v39, v44
-; SI-NEXT: v_add_f64 v[49:50], s[24:25], 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v16
-; SI-NEXT: v_lshrrev_b32_e32 v46, 16, v50
-; SI-NEXT: v_lshrrev_b32_e32 v56, 16, v49
-; SI-NEXT: v_cvt_f32_f16_e32 v47, v16
-; SI-NEXT: v_cvt_f32_f16_e32 v16, v49
-; SI-NEXT: v_cvt_f32_f16_e32 v49, v46
-; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_f64 v[3:4], s[4:5], 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v34
-; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v3
-; SI-NEXT: v_cvt_f32_f16_e32 v13, v3
-; SI-NEXT: v_cvt_f32_f16_e32 v3, v34
-; SI-NEXT: v_cvt_f32_f16_e32 v34, v51
-; SI-NEXT: v_cvt_f32_f16_e32 v51, v56
-; SI-NEXT: v_add_f64 v[26:27], s[40:41], 1.0
-; SI-NEXT: v_add_f64 v[22:23], s[14:15], 1.0
-; SI-NEXT: v_lshrrev_b32_e32 v60, 16, v54
-; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v31
-; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v26
-; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v27
-; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v22
-; SI-NEXT: v_lshrrev_b32_e32 v14, 16, v23
-; SI-NEXT: v_lshrrev_b32_e32 v59, 16, v11
-; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v7
-; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v4
-; SI-NEXT: v_cvt_f32_f16_e32 v10, v4
-; SI-NEXT: v_cvt_f32_f16_e32 v29, v7
-; SI-NEXT: v_cvt_f32_f16_e32 v45, v11
-; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
-; SI-NEXT: v_cvt_f32_f16_e32 v4, v22
-; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
-; SI-NEXT: v_cvt_f32_f16_e32 v11, v26
-; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
-; SI-NEXT: v_cvt_f32_f16_e32 v35, v35
-; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
-; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
-; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
-; SI-NEXT: v_cvt_f32_f16_e32 v7, v53
+; SI-NEXT: v_cvt_f32_f16_e32 v3, v59
+; SI-NEXT: v_mov_b32_e32 v59, v2
+; SI-NEXT: v_cvt_f32_f16_e32 v2, v63
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v25
+; SI-NEXT: v_add_f64 v[55:56], v[23:24], 1.0
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v2, v61
+; SI-NEXT: v_add_f64 v[29:30], v[29:30], 1.0
+; SI-NEXT: v_add_f64 v[41:42], v[27:28], 1.0
+; SI-NEXT: v_add_f64 v[9:10], v[9:10], 1.0
+; SI-NEXT: v_add_f64 v[11:12], v[11:12], 1.0
+; SI-NEXT: v_add_f64 v[13:14], v[13:14], 1.0
+; SI-NEXT: v_add_f64 v[15:16], v[15:16], 1.0
+; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v55
+; SI-NEXT: v_lshrrev_b32_e32 v24, 16, v29
+; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v30
+; SI-NEXT: v_lshrrev_b32_e32 v28, 16, v41
+; SI-NEXT: v_lshrrev_b32_e32 v50, 16, v42
+; SI-NEXT: v_lshrrev_b32_e32 v48, 16, v25
+; SI-NEXT: v_lshrrev_b32_e32 v38, 16, v26
+; SI-NEXT: v_lshrrev_b32_e32 v36, 16, v21
+; SI-NEXT: v_lshrrev_b32_e32 v34, 16, v22
+; SI-NEXT: v_lshrrev_b32_e32 v32, 16, v19
+; SI-NEXT: v_lshrrev_b32_e32 v53, 16, v20
+; SI-NEXT: v_lshrrev_b32_e32 v43, 16, v17
+; SI-NEXT: v_lshrrev_b32_e32 v45, 16, v18
+; SI-NEXT: v_lshrrev_b32_e32 v35, 16, v9
+; SI-NEXT: v_lshrrev_b32_e32 v37, 16, v10
+; SI-NEXT: v_lshrrev_b32_e32 v39, 16, v11
+; SI-NEXT: v_lshrrev_b32_e32 v49, 16, v12
+; SI-NEXT: v_lshrrev_b32_e32 v51, 16, v13
+; SI-NEXT: v_lshrrev_b32_e32 v52, 16, v14
+; SI-NEXT: v_lshrrev_b32_e32 v54, 16, v15
+; SI-NEXT: v_lshrrev_b32_e32 v40, 16, v16
+; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
+; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
+; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
+; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v56
+; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
+; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
+; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
+; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
+; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
+; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
+; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
+; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
; SI-NEXT: v_cvt_f32_f16_e32 v42, v42
+; SI-NEXT: v_cvt_f32_f16_e32 v44, v41
+; SI-NEXT: v_cvt_f32_f16_e32 v46, v30
+; SI-NEXT: v_cvt_f32_f16_e32 v47, v29
+; SI-NEXT: v_cvt_f32_f16_e32 v56, v56
+; SI-NEXT: v_cvt_f32_f16_e32 v57, v55
; SI-NEXT: v_cvt_f32_f16_e32 v40, v40
+; SI-NEXT: v_cvt_f32_f16_e32 v54, v54
; SI-NEXT: v_cvt_f32_f16_e32 v52, v52
-; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
+; SI-NEXT: v_cvt_f32_f16_e32 v51, v51
+; SI-NEXT: v_cvt_f32_f16_e32 v49, v49
+; SI-NEXT: v_cvt_f32_f16_e32 v39, v39
+; SI-NEXT: v_cvt_f32_f16_e32 v37, v37
+; SI-NEXT: v_cvt_f32_f16_e32 v35, v35
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v2, v58
+; SI-NEXT: v_cvt_f32_f16_e32 v58, v45
+; SI-NEXT: v_cvt_f32_f16_e32 v61, v43
+; SI-NEXT: v_cvt_f32_f16_e32 v63, v53
+; SI-NEXT: v_cvt_f32_f16_e32 v32, v32
+; SI-NEXT: v_cvt_f32_f16_e32 v34, v34
; SI-NEXT: v_cvt_f32_f16_e32 v36, v36
-; SI-NEXT: v_cvt_f32_f16_e32 v33, v33
-; SI-NEXT: v_cvt_f32_f16_e32 v59, v59
-; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
-; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
-; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
-; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
-; SI-NEXT: v_cvt_f32_f16_e32 v22, v6
-; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
-; SI-NEXT: v_cvt_f32_f16_e32 v26, v9
-; SI-NEXT: v_cvt_f32_f16_e32 v28, v28
-; SI-NEXT: v_cvt_f32_f16_e32 v53, v60
-; SI-NEXT: v_cvt_f32_f16_e32 v58, v58
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v55, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v41, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v44, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v46, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v56, v5
-; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_cvt_f32_f16_e32 v60, v5
+; SI-NEXT: v_cvt_f32_f16_e32 v38, v38
+; SI-NEXT: v_cvt_f32_f16_e32 v48, v48
+; SI-NEXT: v_cvt_f32_f16_e32 v50, v50
+; SI-NEXT: v_cvt_f32_f16_e32 v53, v28
+; SI-NEXT: v_cvt_f32_f16_e32 v55, v27
+; SI-NEXT: v_cvt_f32_f16_e32 v41, v24
+; SI-NEXT: v_cvt_f32_f16_e32 v43, v1
+; SI-NEXT: v_cvt_f32_f16_e32 v45, v23
+; SI-NEXT: v_mov_b32_e32 v33, v15
+; SI-NEXT: v_mov_b32_e32 v31, v16
+; SI-NEXT: v_mov_b32_e32 v62, v13
+; SI-NEXT: v_mov_b32_e32 v60, v14
+; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT: .LBB53_3: ; %end
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v60
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v2
-; SI-NEXT: v_cvt_f16_f32_e32 v9, v57
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v58
-; SI-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v45
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_add_i32_e32 v5, vcc, 4, v0
-; SI-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; SI-NEXT: v_or_b32_e32 v6, v9, v6
-; SI-NEXT: buffer_store_dword v6, v5, s[0:3], 0 offen
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v56
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v57
+; SI-NEXT: v_cvt_f16_f32_e32 v3, v56
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v43
+; SI-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v21
-; SI-NEXT: v_add_i32_e32 v9, vcc, 8, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v9, s[0:3], 0 offen
+; SI-NEXT: v_add_i32_e32 v1, vcc, 4, v0
+; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; SI-NEXT: v_or_b32_e32 v2, v3, v2
+; SI-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v41
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v46
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v20
-; SI-NEXT: v_add_i32_e32 v9, vcc, 12, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v9, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v47
+; SI-NEXT: v_add_i32_e32 v3, vcc, 8, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v44
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v8
-; SI-NEXT: v_add_i32_e32 v9, vcc, 16, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v9, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v55
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v46
+; SI-NEXT: v_add_i32_e32 v3, vcc, 12, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v41
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v42
-; SI-NEXT: v_add_i32_e32 v9, vcc, 20, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v9, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v53
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v44
+; SI-NEXT: v_add_i32_e32 v3, vcc, 16, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v55
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v7
-; SI-NEXT: v_add_i32_e32 v7, vcc, 24, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v7, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v50
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v42
+; SI-NEXT: v_add_i32_e32 v3, vcc, 20, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v53
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v54
-; SI-NEXT: v_add_i32_e32 v7, vcc, 28, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v7, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
+; SI-NEXT: v_add_i32_e32 v3, vcc, 24, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v51
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v16
-; SI-NEXT: v_add_i32_e32 v7, vcc, 32, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v7, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v38
+; SI-NEXT: v_add_i32_e32 v3, vcc, 28, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v49
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v50
-; SI-NEXT: v_add_i32_e32 v7, vcc, 36, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v7, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v36
+; SI-NEXT: v_add_i32_e32 v3, vcc, 32, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v39
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v15
-; SI-NEXT: v_add_i32_e32 v7, vcc, 40, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v7, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v34
+; SI-NEXT: v_add_i32_e32 v3, vcc, 36, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v37
-; SI-NEXT: v_cvt_f16_f32_e32 v6, v38
-; SI-NEXT: v_add_i32_e32 v7, vcc, 44, v0
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v5, v6, v5
-; SI-NEXT: buffer_store_dword v5, v7, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v32
+; SI-NEXT: v_add_i32_e32 v3, vcc, 40, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v34
-; SI-NEXT: v_add_i32_e32 v6, vcc, 48, v0
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; SI-NEXT: v_or_b32_e32 v3, v3, v5
-; SI-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v63
+; SI-NEXT: v_add_i32_e32 v3, vcc, 44, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v32
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v35
-; SI-NEXT: v_add_i32_e32 v6, vcc, 52, v0
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v5, v3
-; SI-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v61
+; SI-NEXT: v_add_i32_e32 v3, vcc, 48, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v30
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v12
-; SI-NEXT: v_add_i32_e32 v6, vcc, 56, v0
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v5, v3
-; SI-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v58
+; SI-NEXT: v_add_i32_e32 v3, vcc, 52, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v28
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v31
-; SI-NEXT: v_add_i32_e32 v6, vcc, 60, v0
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v5, v3
-; SI-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 56, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v26
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v11
-; SI-NEXT: v_add_i32_e32 v6, vcc, 64, v0
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v5, v3
-; SI-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 60, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v24
-; SI-NEXT: v_cvt_f16_f32_e32 v5, v27
-; SI-NEXT: v_add_i32_e32 v6, vcc, 0x44, v0
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v5, v3
-; SI-NEXT: buffer_store_dword v3, v6, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 64, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v22
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x48, v0
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x44, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v14
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v23
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x4c, v0
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x48, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v17
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v63
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x50, v0
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x4c, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v18
-; SI-NEXT: v_cvt_f16_f32_e32 v4, v62
-; SI-NEXT: v_add_i32_e32 v5, vcc, 0x54, v0
-; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; SI-NEXT: v_or_b32_e32 v3, v4, v3
-; SI-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x50, v0
+; SI-NEXT: s_waitcnt vmcnt(1)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v61
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x54, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x58, v0
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: buffer_store_dword v1, v4, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v25
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v47
-; SI-NEXT: v_add_i32_e32 v4, vcc, 0x5c, v0
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v35
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x58, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; SI-NEXT: v_or_b32_e32 v1, v3, v1
-; SI-NEXT: buffer_store_dword v1, v4, s[0:3], 0 offen
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v59
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v45
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v37
+; SI-NEXT: v_add_i32_e32 v3, vcc, 0x5c, v0
+; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT: v_or_b32_e32 v1, v2, v1
+; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v39
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x60, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_or_b32_e32 v1, v2, v1
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
+; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v33
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v43
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v49
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x64, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_or_b32_e32 v1, v2, v1
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v36
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v29
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v51
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v62
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x68, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_or_b32_e32 v1, v2, v1
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v48
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v19
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v60
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x6c, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_or_b32_e32 v1, v2, v1
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v52
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v13
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v54
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v33
; SI-NEXT: v_add_i32_e32 v3, vcc, 0x70, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_or_b32_e32 v1, v2, v1
; SI-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v1, v40
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v10
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v31
; SI-NEXT: v_add_i32_e32 v0, vcc, 0x74, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_or_b32_e32 v1, v2, v1
@@ -44313,66 +43669,93 @@ define inreg <60 x half> @bitcast_v15f64_to_v60f16_scalar(<15 x double> inreg %a
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT: s_setpc_b64 s[30:31]
; SI-NEXT: .LBB53_4:
-; SI-NEXT: ; implicit-def: $vgpr2
-; SI-NEXT: ; implicit-def: $vgpr60
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
; SI-NEXT: ; implicit-def: $vgpr57
-; SI-NEXT: ; implicit-def: $vgpr58
-; SI-NEXT: ; implicit-def: $vgpr21
+; SI-NEXT: ; implicit-def: $vgpr45
; SI-NEXT: ; implicit-def: $vgpr56
-; SI-NEXT: ; implicit-def: $vgpr20
-; SI-NEXT: ; implicit-def: $vgpr46
-; SI-NEXT: ; implicit-def: $vgpr8
-; SI-NEXT: ; implicit-def: $vgpr44
-; SI-NEXT: ; implicit-def: $vgpr42
+; SI-NEXT: ; implicit-def: $vgpr43
+; SI-NEXT: ; implicit-def: $vgpr47
; SI-NEXT: ; implicit-def: $vgpr41
-; SI-NEXT: ; implicit-def: $vgpr7
+; SI-NEXT: ; implicit-def: $vgpr46
; SI-NEXT: ; implicit-def: $vgpr55
-; SI-NEXT: ; implicit-def: $vgpr54
+; SI-NEXT: ; implicit-def: $vgpr44
; SI-NEXT: ; implicit-def: $vgpr53
-; SI-NEXT: ; implicit-def: $vgpr16
-; SI-NEXT: ; implicit-def: $vgpr51
+; SI-NEXT: ; implicit-def: $vgpr42
; SI-NEXT: ; implicit-def: $vgpr50
-; SI-NEXT: ; implicit-def: $vgpr49
-; SI-NEXT: ; implicit-def: $vgpr15
-; SI-NEXT: ; implicit-def: $vgpr39
+; SI-NEXT: ; implicit-def: $vgpr48
; SI-NEXT: ; implicit-def: $vgpr38
-; SI-NEXT: ; implicit-def: $vgpr37
-; SI-NEXT: ; implicit-def: $vgpr3
+; SI-NEXT: ; implicit-def: $vgpr36
; SI-NEXT: ; implicit-def: $vgpr34
-; SI-NEXT: ; implicit-def: $vgpr35
; SI-NEXT: ; implicit-def: $vgpr32
-; SI-NEXT: ; implicit-def: $vgpr12
-; SI-NEXT: ; implicit-def: $vgpr30
-; SI-NEXT: ; implicit-def: $vgpr31
-; SI-NEXT: ; implicit-def: $vgpr28
-; SI-NEXT: ; implicit-def: $vgpr11
-; SI-NEXT: ; implicit-def: $vgpr26
-; SI-NEXT: ; implicit-def: $vgpr27
-; SI-NEXT: ; implicit-def: $vgpr24
-; SI-NEXT: ; implicit-def: $vgpr4
-; SI-NEXT: ; implicit-def: $vgpr22
-; SI-NEXT: ; implicit-def: $vgpr23
-; SI-NEXT: ; implicit-def: $vgpr14
; SI-NEXT: ; implicit-def: $vgpr63
-; SI-NEXT: ; implicit-def: $vgpr17
-; SI-NEXT: ; implicit-def: $vgpr62
-; SI-NEXT: ; implicit-def: $vgpr18
; SI-NEXT: ; implicit-def: $vgpr61
-; SI-NEXT: ; implicit-def: $vgpr1
-; SI-NEXT: ; implicit-def: $vgpr47
-; SI-NEXT: ; implicit-def: $vgpr25
-; SI-NEXT: ; implicit-def: $vgpr45
+; SI-NEXT: ; implicit-def: $vgpr58
; SI-NEXT: ; implicit-def: $vgpr59
-; SI-NEXT: ; implicit-def: $vgpr43
-; SI-NEXT: ; implicit-def: $vgpr33
-; SI-NEXT: ; implicit-def: $vgpr29
-; SI-NEXT: ; implicit-def: $vgpr36
-; SI-NEXT: ; implicit-def: $vgpr19
-; SI-NEXT: ; implicit-def: $vgpr48
-; SI-NEXT: ; implicit-def: $vgpr13
+; SI-NEXT: ; implicit-def: $vgpr35
+; SI-NEXT: ; implicit-def: $vgpr37
+; SI-NEXT: ; implicit-def: $vgpr39
+; SI-NEXT: ; implicit-def: $vgpr49
+; SI-NEXT: ; implicit-def: $vgpr62
+; SI-NEXT: ; implicit-def: $vgpr51
+; SI-NEXT: ; implicit-def: $vgpr60
; SI-NEXT: ; implicit-def: $vgpr52
-; SI-NEXT: ; implicit-def: $vgpr10
+; SI-NEXT: ; implicit-def: $vgpr33
+; SI-NEXT: ; implicit-def: $vgpr54
; SI-NEXT: ; implicit-def: $vgpr40
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
+; SI-NEXT: ; kill: killed $vgpr31
+; SI-NEXT: ; implicit-def: $vgpr31
; SI-NEXT: s_branch .LBB53_2
;
; VI-LABEL: bitcast_v15f64_to_v60f16_scalar:
@@ -45262,6 +44645,9 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v62, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v63, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(5)
+; SI-NEXT: v_cvt_f16_f32_e32 v58, v0
+; SI-NEXT: v_cvt_f16_f32_e32 v0, v5
; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116
; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: buffer_load_dword v60, off, s[0:3], s32
@@ -45290,23 +44676,12 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:76
; SI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:88
; SI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:84
-; SI-NEXT: v_cvt_f16_f32_e32 v58, v0
-; SI-NEXT: v_cvt_f16_f32_e32 v0, v5
; SI-NEXT: v_cvt_f16_f32_e32 v59, v1
; SI-NEXT: v_cvt_f16_f32_e32 v57, v3
-; SI-NEXT: v_cvt_f16_f32_e32 v56, v2
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
-; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
-; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:104
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:100
-; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
-; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v4
-; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
+; SI-NEXT: v_cvt_f16_f32_e32 v56, v2
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v7
@@ -45316,8 +44691,6 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v9
-; SI-NEXT: s_waitcnt vmcnt(14)
-; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v8
@@ -45339,9 +44712,18 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v14
+; SI-NEXT: s_waitcnt vmcnt(14)
+; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v17
+; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
+; SI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:92
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:104
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:100
+; SI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:112
+; SI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:108
+; SI-NEXT: v_cvt_f16_f32_e32 v60, v60
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v16
@@ -45351,6 +44733,7 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v18
+; SI-NEXT: v_cvt_f16_f32_e32 v52, v52
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v21
@@ -45388,7 +44771,6 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v61
; SI-NEXT: v_cvt_f16_f32_e32 v61, v49
-; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f16_f32_e32 v49, v55
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -45403,6 +44785,7 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v32
+; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cvt_f16_f32_e32 v32, v47
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -46274,7 +45657,6 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: buffer_store_dword v63, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT: v_mov_b32_e32 v61, v0
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v29
-; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v28, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v28
@@ -46353,6 +45735,7 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_lshrrev_b32_e32 v37, 16, v60
; GFX9-NEXT: v_cmp_ne_u32_e32 vcc, 0, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v38, 16, v61
+; GFX9-NEXT: buffer_store_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT: s_and_saveexec_b64 s[4:5], vcc
@@ -46517,6 +45900,9 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT: s_cbranch_execz .LBB54_4
; GFX9-NEXT: ; %bb.3: ; %cmp.true
+; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
@@ -46531,9 +45917,6 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b32 s6, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v38, v61, s6
; GFX9-NEXT: s_movk_i32 s7, 0x200
; GFX9-NEXT: v_perm_b32 v1, v37, v60, s6
@@ -46553,6 +45936,10 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_pk_add_f16 v6, v6, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v7, v7, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v8, v8, s7 op_sel_hi:[1,0]
+; GFX9-NEXT: s_waitcnt vmcnt(14)
+; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: s_waitcnt vmcnt(15)
; GFX9-NEXT: v_perm_b32 v9, v9, v44, s6
; GFX9-NEXT: s_waitcnt vmcnt(14)
@@ -46581,10 +45968,6 @@ define <15 x double> @bitcast_v60f16_to_v15f64(<60 x half> %a, i32 %b) {
; GFX9-NEXT: v_perm_b32 v21, v21, v48, s6
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_perm_b32 v22, v22, v39, s6
-; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_perm_b32 v23, v24, v23, s6
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT: v_pk_add_f16 v9, v9, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v10, v10, s7 op_sel_hi:[1,0]
; GFX9-NEXT: v_pk_add_f16 v11, v11, s7 op_sel_hi:[1,0]
@@ -46905,12 +46288,35 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
; SI-NEXT: v_cvt_f16_f32_e32 v8, s26
; SI-NEXT: v_cvt_f16_f32_e32 v6, s29
; SI-NEXT: v_cvt_f16_f32_e32 v7, s28
+; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt vmcnt(14)
; SI-NEXT: v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT: v_cvt_f16_f32_e32 v50, v54
; SI-NEXT: v_cvt_f16_f32_e32 v48, v48
; SI-NEXT: v_cvt_f16_f32_e32 v31, v40
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v33
; SI-NEXT: s_and_b64 s[4:5], vcc, exec
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
@@ -46920,7 +46326,7 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v38
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v44
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -46929,7 +46335,7 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v46
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v47
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
@@ -46938,7 +46344,7 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v57
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt vmcnt(14) expcnt(0)
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v58
; SI-NEXT: v_cvt_f16_f32_e32 v58, s16
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
@@ -46949,38 +46355,12 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v0, v60
; SI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v37, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v49, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT: s_cbranch_scc0 .LBB55_2
; SI-NEXT: ; %bb.1: ; %cmp.false
; SI-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT: v_or_b32_e32 v3, v10, v3
-; SI-NEXT: s_waitcnt expcnt(3)
; SI-NEXT: v_lshlrev_b32_e32 v10, 16, v34
; SI-NEXT: v_mov_b32_e32 v33, v32
; SI-NEXT: v_or_b32_e32 v10, v32, v10
@@ -47005,12 +46385,12 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT: v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v59
; SI-NEXT: v_or_b32_e32 v1, v12, v1
; SI-NEXT: v_or_b32_e32 v2, v11, v2
@@ -47119,12 +46499,10 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
; SI-NEXT: v_mov_b32_e32 v40, v44
; SI-NEXT: s_cbranch_vccnz .LBB55_5
; SI-NEXT: ; %bb.4: ; %cmp.true
-; SI-NEXT: s_waitcnt expcnt(5)
; SI-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
-; SI-NEXT: s_waitcnt expcnt(4)
; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
-; SI-NEXT: v_cvt_f32_f16_e32 v0, v59
; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f32_f16_e32 v0, v59
; SI-NEXT: v_cvt_f32_f16_e32 v1, v58
; SI-NEXT: s_waitcnt vmcnt(3)
; SI-NEXT: v_cvt_f32_f16_e32 v8, v33
@@ -47910,304 +47288,115 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:316
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:252
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:192
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v156, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v157, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v158, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v159, s32 offset:64
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v168, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v169, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v170, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v171, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v172, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v173, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v174, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v175, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v184, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v185, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v186, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v187, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v188, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v189, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v190, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v191, s32
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v9 :: v_dual_mov_b32 v25, v7
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v8 :: v_dual_mov_b32 v191, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v190, v6 :: v_dual_mov_b32 v185, v4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v186, v3 :: v_dual_mov_b32 v187, v2
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v188, v1 :: v_dual_mov_b32 v189, v0
-; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s29, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s28, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s27, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s26, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s25, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s22, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s21, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s20, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s19, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s17, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s1, 16
-; GFX11-TRUE16-NEXT: s_lshr_b32 s40, s0, 16
-; GFX11-TRUE16-NEXT: s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s40, s0, s40
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s41, s1, s41
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s2, s46
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s3, s45
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s16, s44
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s17, s43
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s18, s4
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s19, s5
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s20, s6
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s21, s7
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s22, s8
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s23, s9
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s25, s11
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s26, s12
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s27, s13
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s28, s14
-; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s29, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-TRUE16-NEXT: s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT: s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT: s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT: s_pack_ll_b32_b16 s17, s29, s41
; GFX11-TRUE16-NEXT: s_and_b32 s47, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT: s_cbranch_scc0 .LBB55_4
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s40 :: v_dual_mov_b32 v5, s0
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s41 :: v_dual_mov_b32 v9, s1
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s2 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, s3 :: v_dual_mov_b32 v35, s5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v44, s6 :: v_dual_mov_b32 v65, s8
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v54, s7 :: v_dual_mov_b32 v77, s9
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v90, s10 :: v_dual_mov_b32 v119, s12
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v104, s11 :: v_dual_mov_b32 v135, s13
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v152, s14
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v170, s15
-; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s40
; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB55_3
; GFX11-TRUE16-NEXT: .LBB55_2: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s40 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s41 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v189, 0x200, v189 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v188, 0x200, v188 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v187, 0x200, v187 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v186, 0x200, v186 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v185, 0x200, v185 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v191, 0x200, v191 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v190, 0x200, v190 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v35, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v44, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v54, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v65, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v77, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v90, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v104, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v119, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v135, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v152, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v170, 0x200, s15 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: .LBB55_3: ; %end
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v5
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v20 :: v_dual_mov_b32 v6, v27
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v35 :: v_dual_mov_b32 v8, v44
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v77 :: v_dual_mov_b32 v12, v90
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v104
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v15, v135 :: v_dual_mov_b32 v16, v152
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v170 :: v_dual_mov_b32 v18, v189
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v188 :: v_dual_mov_b32 v20, v187
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v186 :: v_dual_mov_b32 v22, v185
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v191 :: v_dual_mov_b32 v24, v190
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v191, off, s32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v190, off, s32 offset:4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v189, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_b32 v188, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v187, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v186, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v185, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v184, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v175, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v174, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v173, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v172, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v171, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v170, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v169, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v168, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v159, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v158, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v157, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v156, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:124
-; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:160
-; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:168
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:176
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:184
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:192
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:200
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:240
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:248
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:252
-; GFX11-TRUE16-NEXT: s_clause 0xf
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:256
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:260
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:264
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:300
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:304
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:308
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:312
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:316
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v14
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v54 :: v_dual_mov_b32 v10, v65
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v119 :: v_dual_mov_b32 v27, v30
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB55_4:
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v64, v29 :: v_dual_mov_b32 v65, v28
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v66, v30
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v53, v26 :: v_dual_mov_b32 v54, v25
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v54 :: v_dual_mov_b32 v26, v53
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_vgpr29_vgpr30_vgpr31_vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v65 :: v_dual_mov_b32 v29, v64
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v30, v66
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168_vgpr169_vgpr170_vgpr171_vgpr172_vgpr173_vgpr174_vgpr175_vgpr176_vgpr177_vgpr178_vgpr179_vgpr180_vgpr181_vgpr182_vgpr183_vgpr184
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_branch .LBB55_2
;
; GFX11-FAKE16-LABEL: bitcast_v60f16_to_v15f64_scalar:
@@ -51893,27 +51082,27 @@ define <60 x i16> @bitcast_v60f16_to_v60i16(<60 x half> %a, i32 %b) {
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_cvt_f16_f32_e32 v55, v3
; SI-NEXT: v_cvt_f16_f32_e32 v3, v22
-; SI-NEXT: v_cvt_f16_f32_e32 v60, v2
+; SI-NEXT: v_cvt_f16_f32_e32 v40, v4
; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v1, v5
-; SI-NEXT: v_cvt_f16_f32_e32 v2, v18
; SI-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; SI-NEXT: v_cvt_f16_f32_e32 v40, v4
-; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; SI-NEXT: s_waitcnt expcnt(1)
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v6
; SI-NEXT: v_cvt_f16_f32_e32 v53, v8
; SI-NEXT: v_cvt_f16_f32_e32 v8, v10
+; SI-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
+; SI-NEXT: s_waitcnt expcnt(0)
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v6
; SI-NEXT: v_cvt_f16_f32_e32 v49, v12
; SI-NEXT: v_cvt_f16_f32_e32 v6, v13
; SI-NEXT: v_cvt_f16_f32_e32 v37, v15
+; SI-NEXT: v_cvt_f16_f32_e32 v60, v2
+; SI-NEXT: v_cvt_f16_f32_e32 v2, v18
; SI-NEXT: v_cvt_f16_f32_e32 v52, v7
; SI-NEXT: v_cvt_f16_f32_e32 v7, v9
; SI-NEXT: v_cvt_f16_f32_e32 v48, v11
; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
; SI-NEXT: v_cvt_f16_f32_e32 v38, v16
+; SI-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_cvt_f16_f32_e32 v2, v19
; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
@@ -53259,6 +52448,8 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v1
; SI-NEXT: v_cvt_f32_f16_e32 v1, v44
+; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
@@ -53285,10 +52476,13 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v1
; SI-NEXT: v_cvt_f32_f16_e32 v1, v50
+; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; SI-NEXT: s_waitcnt vmcnt(0)
+; SI-NEXT: v_mov_b32_e32 v51, v11
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
@@ -53300,8 +52494,26 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v3, v26
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v50
+; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v1
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v8
+; SI-NEXT: v_mov_b32_e32 v8, v48
+; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v1
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v44
+; SI-NEXT: v_lshr_b64 v[44:45], v[29:30], 16
+; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v1
+; SI-NEXT: v_cvt_f32_f16_e32 v1, v38
+; SI-NEXT: v_cvt_f32_f16_e32 v38, v43
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
+; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
+; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
@@ -53329,17 +52541,11 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
; SI-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; SI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
-; SI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT: v_or_b32_e32 v18, v3, v5
; SI-NEXT: v_cvt_f32_f16_e32 v5, v37
; SI-NEXT: v_cvt_f32_f16_e32 v3, v16
-; SI-NEXT: s_waitcnt vmcnt(2)
-; SI-NEXT: v_mov_b32_e32 v51, v11
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT: v_cvt_f16_f32_e32 v55, v5
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
@@ -53382,52 +52588,32 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_cvt_f32_f16_e32 v3, v6
; SI-NEXT: v_lshr_b64 v[58:59], v[34:35], 16
; SI-NEXT: v_add_f32_e32 v5, 0x38000000, v5
-; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
-; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT: v_cvt_f32_f16_e32 v1, v50
-; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v1
-; SI-NEXT: v_cvt_f32_f16_e32 v1, v8
-; SI-NEXT: v_mov_b32_e32 v8, v48
; SI-NEXT: v_cvt_f16_f32_e32 v48, v5
-; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v48
-; SI-NEXT: v_or_b32_e32 v6, v3, v5
-; SI-NEXT: v_cvt_f32_f16_e32 v3, v4
-; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v1
-; SI-NEXT: s_waitcnt vmcnt(1)
-; SI-NEXT: v_cvt_f32_f16_e32 v1, v44
-; SI-NEXT: v_cvt_f32_f16_e32 v4, v31
; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
-; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
-; SI-NEXT: v_cvt_f16_f32_e32 v60, v4
+; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v48
; SI-NEXT: v_mov_b32_e32 v59, v48
-; SI-NEXT: v_lshlrev_b32_e32 v56, 16, v1
-; SI-NEXT: v_cvt_f32_f16_e32 v1, v38
-; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v60
-; SI-NEXT: v_or_b32_e32 v4, v3, v4
-; SI-NEXT: v_cvt_f32_f16_e32 v38, v43
-; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
-; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
-; SI-NEXT: v_lshr_b64 v[47:48], v[17:18], 16
-; SI-NEXT: s_waitcnt vmcnt(0)
-; SI-NEXT: v_lshr_b64 v[44:45], v[29:30], 16
+; SI-NEXT: v_or_b32_e32 v6, v3, v5
; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; SI-NEXT: v_cvt_f32_f16_e32 v1, v24
; SI-NEXT: v_cvt_f32_f16_e32 v24, v8
; SI-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; SI-NEXT: v_cvt_f32_f16_e32 v3, v4
+; SI-NEXT: v_cvt_f32_f16_e32 v4, v31
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
+; SI-NEXT: v_add_f32_e32 v3, 0x38000000, v3
+; SI-NEXT: v_add_f32_e32 v4, 0x38000000, v4
+; SI-NEXT: v_cvt_f16_f32_e32 v60, v4
+; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
; SI-NEXT: v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
+; SI-NEXT: v_lshlrev_b32_e32 v4, 16, v60
+; SI-NEXT: v_or_b32_e32 v4, v3, v4
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v1
; SI-NEXT: v_cvt_f32_f16_e32 v1, v20
; SI-NEXT: v_cvt_f32_f16_e32 v20, v39
+; SI-NEXT: v_lshr_b64 v[47:48], v[17:18], 16
; SI-NEXT: v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT: v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT: v_cvt_f16_f32_e32 v31, v20
@@ -53524,14 +52710,15 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_mov_b32_e32 v32, v41
; SI-NEXT: v_lshr_b64 v[40:41], v[21:22], 16
; SI-NEXT: v_lshr_b64 v[20:21], v[11:12], 16
-; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshr_b64 v[20:21], v[56:57], 16
; SI-NEXT: buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT: buffer_store_dword v21, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
+; SI-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT: v_mov_b32_e32 v11, v24
+; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshr_b64 v[8:9], v[9:10], 16
; SI-NEXT: v_mov_b32_e32 v39, v31
; SI-NEXT: v_mov_b32_e32 v31, v60
@@ -53541,7 +52728,6 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
; SI-NEXT: v_mov_b32_e32 v37, v55
; SI-NEXT: v_lshr_b64 v[55:56], v[5:6], 16
; SI-NEXT: v_lshr_b64 v[24:25], v[3:4], 16
-; SI-NEXT: s_waitcnt expcnt(0)
; SI-NEXT: v_lshr_b64 v[20:21], v[1:2], 16
; SI-NEXT: .LBB59_3: ; %end
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v58