aboutsummaryrefslogtreecommitdiff
path: root/llvm/test/CodeGen/AMDGPU/sdiv.ll
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/sdiv.ll')
-rw-r--r--llvm/test/CodeGen/AMDGPU/sdiv.ll784
1 files changed, 398 insertions, 386 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/sdiv.ll b/llvm/test/CodeGen/AMDGPU/sdiv.ll
index 676359fce..5c0f813 100644
--- a/llvm/test/CodeGen/AMDGPU/sdiv.ll
+++ b/llvm/test/CodeGen/AMDGPU/sdiv.ll
@@ -391,144 +391,156 @@ define amdgpu_kernel void @slow_sdiv_i32_3435(ptr addrspace(1) %out, ptr addrspa
define amdgpu_kernel void @sdiv_v2i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {
; GCN-LABEL: sdiv_v2i32:
; GCN: ; %bb.0:
-; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
-; GCN-NEXT: s_mov_b32 s7, 0xf000
-; GCN-NEXT: s_mov_b32 s6, -1
-; GCN-NEXT: s_mov_b32 s10, s6
-; GCN-NEXT: s_mov_b32 s11, s7
+; GCN-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9
+; GCN-NEXT: s_mov_b32 s3, 0xf000
+; GCN-NEXT: s_mov_b32 s2, -1
+; GCN-NEXT: s_mov_b32 s10, s2
+; GCN-NEXT: s_mov_b32 s11, s3
; GCN-NEXT: s_waitcnt lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s8, s2
-; GCN-NEXT: s_mov_b32 s9, s3
+; GCN-NEXT: s_mov_b32 s8, s6
+; GCN-NEXT: s_mov_b32 s9, s7
; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
-; GCN-NEXT: s_mov_b32 s4, s0
-; GCN-NEXT: s_mov_b32 s5, s1
; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_sub_i32_e32 v6, vcc, 0, v2
-; GCN-NEXT: v_sub_i32_e32 v9, vcc, 0, v3
-; GCN-NEXT: v_xor_b32_e32 v4, v0, v2
-; GCN-NEXT: v_xor_b32_e32 v7, v1, v3
-; GCN-NEXT: v_max_i32_e32 v2, v2, v6
-; GCN-NEXT: v_max_i32_e32 v3, v3, v9
-; GCN-NEXT: v_cvt_f32_u32_e32 v6, v2
-; GCN-NEXT: v_cvt_f32_u32_e32 v9, v3
-; GCN-NEXT: v_sub_i32_e32 v5, vcc, 0, v0
-; GCN-NEXT: v_rcp_iflag_f32_e32 v6, v6
-; GCN-NEXT: v_max_i32_e32 v0, v0, v5
-; GCN-NEXT: v_rcp_iflag_f32_e32 v5, v9
-; GCN-NEXT: v_sub_i32_e32 v9, vcc, 0, v2
-; GCN-NEXT: v_mul_f32_e32 v6, 0x4f7ffffe, v6
-; GCN-NEXT: v_mul_f32_e32 v5, 0x4f7ffffe, v5
-; GCN-NEXT: v_cvt_u32_f32_e32 v6, v6
-; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5
-; GCN-NEXT: v_sub_i32_e32 v10, vcc, 0, v3
-; GCN-NEXT: v_mul_lo_u32 v9, v9, v6
-; GCN-NEXT: v_mul_lo_u32 v10, v10, v5
-; GCN-NEXT: v_sub_i32_e32 v8, vcc, 0, v1
-; GCN-NEXT: v_mul_hi_u32 v9, v6, v9
-; GCN-NEXT: v_max_i32_e32 v1, v1, v8
-; GCN-NEXT: v_mul_hi_u32 v8, v5, v10
-; GCN-NEXT: v_ashrrev_i32_e32 v4, 31, v4
-; GCN-NEXT: v_add_i32_e32 v6, vcc, v6, v9
-; GCN-NEXT: v_add_i32_e32 v5, vcc, v5, v8
-; GCN-NEXT: v_mul_hi_u32 v6, v0, v6
-; GCN-NEXT: v_mul_hi_u32 v5, v1, v5
-; GCN-NEXT: v_ashrrev_i32_e32 v7, 31, v7
-; GCN-NEXT: v_mul_lo_u32 v8, v6, v2
-; GCN-NEXT: v_mul_lo_u32 v10, v5, v3
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 1, v6
-; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v8
-; GCN-NEXT: v_sub_i32_e32 v1, vcc, v1, v10
-; GCN-NEXT: v_add_i32_e32 v11, vcc, 1, v5
-; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v2
-; GCN-NEXT: v_cmp_ge_u32_e64 s[2:3], v1, v3
-; GCN-NEXT: v_sub_i32_e32 v8, vcc, v0, v2
-; GCN-NEXT: v_cndmask_b32_e64 v6, v6, v9, s[0:1]
-; GCN-NEXT: v_sub_i32_e32 v9, vcc, v1, v3
-; GCN-NEXT: v_cndmask_b32_e64 v5, v5, v11, s[2:3]
-; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v8, s[0:1]
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 1, v6
-; GCN-NEXT: v_cndmask_b32_e64 v1, v1, v9, s[2:3]
-; GCN-NEXT: v_add_i32_e32 v9, vcc, 1, v5
-; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v0, v2
-; GCN-NEXT: v_cndmask_b32_e32 v0, v6, v8, vcc
-; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v1, v3
-; GCN-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
-; GCN-NEXT: v_xor_b32_e32 v0, v0, v4
-; GCN-NEXT: v_xor_b32_e32 v1, v1, v7
-; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v4
-; GCN-NEXT: v_sub_i32_e32 v1, vcc, v1, v7
-; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; GCN-NEXT: v_readfirstlane_b32 s0, v2
+; GCN-NEXT: s_abs_i32 s1, s0
+; GCN-NEXT: v_cvt_f32_u32_e32 v2, s1
+; GCN-NEXT: s_sub_i32 s6, 0, s1
+; GCN-NEXT: v_readfirstlane_b32 s8, v3
+; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v2
+; GCN-NEXT: v_mul_f32_e32 v2, 0x4f7ffffe, v2
+; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2
+; GCN-NEXT: v_mul_lo_u32 v4, s6, v2
+; GCN-NEXT: v_readfirstlane_b32 s6, v0
+; GCN-NEXT: s_abs_i32 s7, s6
+; GCN-NEXT: s_xor_b32 s0, s6, s0
+; GCN-NEXT: v_mul_hi_u32 v4, v2, v4
+; GCN-NEXT: s_ashr_i32 s6, s0, 31
+; GCN-NEXT: v_add_i32_e32 v0, vcc, v2, v4
+; GCN-NEXT: v_mul_hi_u32 v0, s7, v0
+; GCN-NEXT: v_readfirstlane_b32 s0, v0
+; GCN-NEXT: s_mul_i32 s0, s0, s1
+; GCN-NEXT: s_sub_i32 s0, s7, s0
+; GCN-NEXT: s_sub_i32 s7, s0, s1
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 1, v0
+; GCN-NEXT: s_cmp_ge_u32 s0, s1
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT: s_cselect_b32 s0, s7, s0
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 1, v0
+; GCN-NEXT: s_cmp_ge_u32 s0, s1
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: s_abs_i32 s7, s8
+; GCN-NEXT: v_cvt_f32_u32_e32 v3, s7
+; GCN-NEXT: s_mov_b32 s0, s4
+; GCN-NEXT: s_sub_i32 s4, 0, s7
+; GCN-NEXT: s_mov_b32 s1, s5
+; GCN-NEXT: v_rcp_iflag_f32_e32 v3, v3
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc
+; GCN-NEXT: v_xor_b32_e32 v0, s6, v0
+; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s6, v0
+; GCN-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
+; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT: v_mul_lo_u32 v4, s4, v3
+; GCN-NEXT: v_readfirstlane_b32 s4, v1
+; GCN-NEXT: s_xor_b32 s5, s4, s8
+; GCN-NEXT: s_abs_i32 s4, s4
+; GCN-NEXT: v_mul_hi_u32 v1, v3, v4
+; GCN-NEXT: s_ashr_i32 s5, s5, 31
+; GCN-NEXT: v_add_i32_e32 v1, vcc, v3, v1
+; GCN-NEXT: v_mul_hi_u32 v1, s4, v1
+; GCN-NEXT: v_readfirstlane_b32 s6, v1
+; GCN-NEXT: s_mul_i32 s6, s6, s7
+; GCN-NEXT: s_sub_i32 s4, s4, s6
+; GCN-NEXT: s_sub_i32 s6, s4, s7
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 1, v1
+; GCN-NEXT: s_cmp_ge_u32 s4, s7
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: s_cselect_b32 s4, s6, s4
+; GCN-NEXT: v_add_i32_e32 v2, vcc, 1, v1
+; GCN-NEXT: s_cmp_ge_u32 s4, s7
+; GCN-NEXT: s_cselect_b64 vcc, -1, 0
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; GCN-NEXT: v_xor_b32_e32 v1, s5, v1
+; GCN-NEXT: v_subrev_i32_e32 v1, vcc, s5, v1
+; GCN-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: sdiv_v2i32:
; TONGA: ; %bb.0:
-; TONGA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
-; TONGA-NEXT: s_mov_b32 s7, 0xf000
-; TONGA-NEXT: s_mov_b32 s6, -1
-; TONGA-NEXT: s_mov_b32 s10, s6
-; TONGA-NEXT: s_mov_b32 s11, s7
+; TONGA-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24
+; TONGA-NEXT: s_mov_b32 s3, 0xf000
+; TONGA-NEXT: s_mov_b32 s2, -1
+; TONGA-NEXT: s_mov_b32 s10, s2
+; TONGA-NEXT: s_mov_b32 s11, s3
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
-; TONGA-NEXT: s_mov_b32 s8, s2
-; TONGA-NEXT: s_mov_b32 s9, s3
+; TONGA-NEXT: s_mov_b32 s8, s6
+; TONGA-NEXT: s_mov_b32 s9, s7
; TONGA-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
-; TONGA-NEXT: s_mov_b32 s4, s0
-; TONGA-NEXT: s_mov_b32 s5, s1
; TONGA-NEXT: s_waitcnt vmcnt(0)
-; TONGA-NEXT: v_sub_u32_e32 v6, vcc, 0, v2
-; TONGA-NEXT: v_sub_u32_e32 v9, vcc, 0, v3
-; TONGA-NEXT: v_xor_b32_e32 v4, v0, v2
-; TONGA-NEXT: v_xor_b32_e32 v7, v1, v3
-; TONGA-NEXT: v_max_i32_e32 v2, v2, v6
-; TONGA-NEXT: v_max_i32_e32 v3, v3, v9
-; TONGA-NEXT: v_cvt_f32_u32_e32 v6, v2
-; TONGA-NEXT: v_cvt_f32_u32_e32 v9, v3
-; TONGA-NEXT: v_sub_u32_e32 v5, vcc, 0, v0
-; TONGA-NEXT: v_rcp_iflag_f32_e32 v6, v6
-; TONGA-NEXT: v_max_i32_e32 v0, v0, v5
-; TONGA-NEXT: v_rcp_iflag_f32_e32 v5, v9
-; TONGA-NEXT: v_sub_u32_e32 v9, vcc, 0, v2
-; TONGA-NEXT: v_mul_f32_e32 v6, 0x4f7ffffe, v6
-; TONGA-NEXT: v_mul_f32_e32 v5, 0x4f7ffffe, v5
-; TONGA-NEXT: v_cvt_u32_f32_e32 v6, v6
-; TONGA-NEXT: v_cvt_u32_f32_e32 v5, v5
-; TONGA-NEXT: v_sub_u32_e32 v10, vcc, 0, v3
-; TONGA-NEXT: v_mul_lo_u32 v9, v9, v6
-; TONGA-NEXT: v_mul_lo_u32 v10, v10, v5
-; TONGA-NEXT: v_sub_u32_e32 v8, vcc, 0, v1
-; TONGA-NEXT: v_mul_hi_u32 v9, v6, v9
-; TONGA-NEXT: v_max_i32_e32 v1, v1, v8
-; TONGA-NEXT: v_mul_hi_u32 v8, v5, v10
-; TONGA-NEXT: v_ashrrev_i32_e32 v4, 31, v4
-; TONGA-NEXT: v_add_u32_e32 v6, vcc, v6, v9
-; TONGA-NEXT: v_add_u32_e32 v5, vcc, v5, v8
-; TONGA-NEXT: v_mul_hi_u32 v6, v0, v6
-; TONGA-NEXT: v_mul_hi_u32 v5, v1, v5
-; TONGA-NEXT: v_ashrrev_i32_e32 v7, 31, v7
-; TONGA-NEXT: v_mul_lo_u32 v8, v6, v2
-; TONGA-NEXT: v_mul_lo_u32 v10, v5, v3
-; TONGA-NEXT: v_add_u32_e32 v9, vcc, 1, v6
-; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v0, v8
-; TONGA-NEXT: v_sub_u32_e32 v1, vcc, v1, v10
-; TONGA-NEXT: v_add_u32_e32 v11, vcc, 1, v5
-; TONGA-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v2
-; TONGA-NEXT: v_cmp_ge_u32_e64 s[2:3], v1, v3
-; TONGA-NEXT: v_sub_u32_e32 v8, vcc, v0, v2
-; TONGA-NEXT: v_cndmask_b32_e64 v6, v6, v9, s[0:1]
-; TONGA-NEXT: v_sub_u32_e32 v9, vcc, v1, v3
-; TONGA-NEXT: v_cndmask_b32_e64 v5, v5, v11, s[2:3]
-; TONGA-NEXT: v_cndmask_b32_e64 v0, v0, v8, s[0:1]
-; TONGA-NEXT: v_add_u32_e32 v8, vcc, 1, v6
-; TONGA-NEXT: v_cndmask_b32_e64 v1, v1, v9, s[2:3]
-; TONGA-NEXT: v_add_u32_e32 v9, vcc, 1, v5
-; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v0, v2
-; TONGA-NEXT: v_cndmask_b32_e32 v0, v6, v8, vcc
-; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v1, v3
-; TONGA-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
-; TONGA-NEXT: v_xor_b32_e32 v0, v0, v4
-; TONGA-NEXT: v_xor_b32_e32 v1, v1, v7
-; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v0, v4
-; TONGA-NEXT: v_sub_u32_e32 v1, vcc, v1, v7
-; TONGA-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; TONGA-NEXT: v_readfirstlane_b32 s0, v2
+; TONGA-NEXT: s_abs_i32 s1, s0
+; TONGA-NEXT: v_cvt_f32_u32_e32 v2, s1
+; TONGA-NEXT: s_sub_i32 s6, 0, s1
+; TONGA-NEXT: v_readfirstlane_b32 s8, v3
+; TONGA-NEXT: v_rcp_iflag_f32_e32 v2, v2
+; TONGA-NEXT: v_mul_f32_e32 v2, 0x4f7ffffe, v2
+; TONGA-NEXT: v_cvt_u32_f32_e32 v2, v2
+; TONGA-NEXT: v_mul_lo_u32 v4, s6, v2
+; TONGA-NEXT: v_readfirstlane_b32 s6, v0
+; TONGA-NEXT: s_abs_i32 s7, s6
+; TONGA-NEXT: s_xor_b32 s0, s6, s0
+; TONGA-NEXT: v_mul_hi_u32 v4, v2, v4
+; TONGA-NEXT: s_ashr_i32 s6, s0, 31
+; TONGA-NEXT: v_add_u32_e32 v0, vcc, v2, v4
+; TONGA-NEXT: v_mul_hi_u32 v0, s7, v0
+; TONGA-NEXT: v_readfirstlane_b32 s0, v0
+; TONGA-NEXT: s_mul_i32 s0, s0, s1
+; TONGA-NEXT: s_sub_i32 s0, s7, s0
+; TONGA-NEXT: s_sub_i32 s7, s0, s1
+; TONGA-NEXT: v_add_u32_e32 v2, vcc, 1, v0
+; TONGA-NEXT: s_cmp_ge_u32 s0, s1
+; TONGA-NEXT: s_cselect_b64 vcc, -1, 0
+; TONGA-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc
+; TONGA-NEXT: s_cselect_b32 s0, s7, s0
+; TONGA-NEXT: v_add_u32_e32 v2, vcc, 1, v0
+; TONGA-NEXT: s_cmp_ge_u32 s0, s1
+; TONGA-NEXT: s_cselect_b64 vcc, -1, 0
+; TONGA-NEXT: s_abs_i32 s7, s8
+; TONGA-NEXT: v_cvt_f32_u32_e32 v3, s7
+; TONGA-NEXT: s_mov_b32 s0, s4
+; TONGA-NEXT: s_sub_i32 s4, 0, s7
+; TONGA-NEXT: s_mov_b32 s1, s5
+; TONGA-NEXT: v_rcp_iflag_f32_e32 v3, v3
+; TONGA-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc
+; TONGA-NEXT: v_xor_b32_e32 v0, s6, v0
+; TONGA-NEXT: v_subrev_u32_e32 v0, vcc, s6, v0
+; TONGA-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
+; TONGA-NEXT: v_cvt_u32_f32_e32 v3, v3
+; TONGA-NEXT: v_mul_lo_u32 v4, s4, v3
+; TONGA-NEXT: v_readfirstlane_b32 s4, v1
+; TONGA-NEXT: s_xor_b32 s5, s4, s8
+; TONGA-NEXT: s_abs_i32 s4, s4
+; TONGA-NEXT: v_mul_hi_u32 v1, v3, v4
+; TONGA-NEXT: s_ashr_i32 s5, s5, 31
+; TONGA-NEXT: v_add_u32_e32 v1, vcc, v3, v1
+; TONGA-NEXT: v_mul_hi_u32 v1, s4, v1
+; TONGA-NEXT: v_readfirstlane_b32 s6, v1
+; TONGA-NEXT: s_mul_i32 s6, s6, s7
+; TONGA-NEXT: s_sub_i32 s4, s4, s6
+; TONGA-NEXT: s_sub_i32 s6, s4, s7
+; TONGA-NEXT: v_add_u32_e32 v2, vcc, 1, v1
+; TONGA-NEXT: s_cmp_ge_u32 s4, s7
+; TONGA-NEXT: s_cselect_b64 vcc, -1, 0
+; TONGA-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; TONGA-NEXT: s_cselect_b32 s4, s6, s4
+; TONGA-NEXT: v_add_u32_e32 v2, vcc, 1, v1
+; TONGA-NEXT: s_cmp_ge_u32 s4, s7
+; TONGA-NEXT: s_cselect_b64 vcc, -1, 0
+; TONGA-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
+; TONGA-NEXT: v_xor_b32_e32 v1, s5, v1
+; TONGA-NEXT: v_subrev_u32_e32 v1, vcc, s5, v1
+; TONGA-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: sdiv_v2i32:
@@ -546,44 +558,44 @@ define amdgpu_kernel void @sdiv_v2i32(ptr addrspace(1) %out, ptr addrspace(1) %i
; GFX9-NEXT: v_readfirstlane_b32 s0, v2
; GFX9-NEXT: s_abs_i32 s1, s0
; GFX9-NEXT: v_cvt_f32_u32_e32 v2, s1
-; GFX9-NEXT: v_readfirstlane_b32 s5, v0
-; GFX9-NEXT: s_xor_b32 s0, s5, s0
+; GFX9-NEXT: v_readfirstlane_b32 s4, v0
+; GFX9-NEXT: s_xor_b32 s0, s4, s0
; GFX9-NEXT: s_ashr_i32 s6, s0, 31
; GFX9-NEXT: v_rcp_iflag_f32_e32 v2, v2
; GFX9-NEXT: s_sub_i32 s0, 0, s1
-; GFX9-NEXT: s_abs_i32 s5, s5
-; GFX9-NEXT: v_readfirstlane_b32 s4, v3
+; GFX9-NEXT: s_abs_i32 s4, s4
+; GFX9-NEXT: v_readfirstlane_b32 s5, v3
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v2
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX9-NEXT: v_readfirstlane_b32 s7, v0
; GFX9-NEXT: s_mul_i32 s0, s0, s7
; GFX9-NEXT: s_mul_hi_u32 s0, s7, s0
; GFX9-NEXT: s_add_i32 s7, s7, s0
-; GFX9-NEXT: s_mul_hi_u32 s0, s5, s7
+; GFX9-NEXT: s_mul_hi_u32 s0, s4, s7
; GFX9-NEXT: s_mul_i32 s7, s0, s1
-; GFX9-NEXT: s_sub_i32 s5, s5, s7
+; GFX9-NEXT: s_sub_i32 s4, s4, s7
; GFX9-NEXT: s_add_i32 s10, s0, 1
-; GFX9-NEXT: s_sub_i32 s7, s5, s1
-; GFX9-NEXT: s_cmp_ge_u32 s5, s1
+; GFX9-NEXT: s_sub_i32 s7, s4, s1
+; GFX9-NEXT: s_cmp_ge_u32 s4, s1
; GFX9-NEXT: s_cselect_b32 s0, s10, s0
-; GFX9-NEXT: s_cselect_b32 s5, s7, s5
+; GFX9-NEXT: s_cselect_b32 s4, s7, s4
; GFX9-NEXT: s_add_i32 s7, s0, 1
-; GFX9-NEXT: s_cmp_ge_u32 s5, s1
-; GFX9-NEXT: s_cselect_b32 s5, s7, s0
-; GFX9-NEXT: s_abs_i32 s7, s4
+; GFX9-NEXT: s_cmp_ge_u32 s4, s1
+; GFX9-NEXT: s_cselect_b32 s4, s7, s0
+; GFX9-NEXT: s_abs_i32 s7, s5
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, s7
-; GFX9-NEXT: s_xor_b32 s5, s5, s6
+; GFX9-NEXT: s_xor_b32 s4, s4, s6
; GFX9-NEXT: s_mov_b32 s1, s9
; GFX9-NEXT: s_sub_i32 s9, 0, s7
; GFX9-NEXT: v_rcp_iflag_f32_e32 v0, v0
-; GFX9-NEXT: s_sub_i32 s5, s5, s6
+; GFX9-NEXT: s_sub_i32 s4, s4, s6
; GFX9-NEXT: s_mov_b32 s0, s8
; GFX9-NEXT: v_readfirstlane_b32 s8, v1
; GFX9-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v0, v0
-; GFX9-NEXT: s_xor_b32 s4, s8, s4
+; GFX9-NEXT: s_xor_b32 s5, s8, s5
; GFX9-NEXT: s_abs_i32 s8, s8
-; GFX9-NEXT: s_ashr_i32 s4, s4, 31
+; GFX9-NEXT: s_ashr_i32 s5, s5, 31
; GFX9-NEXT: v_readfirstlane_b32 s6, v0
; GFX9-NEXT: s_mul_i32 s9, s9, s6
; GFX9-NEXT: s_mul_hi_u32 s9, s6, s9
@@ -599,10 +611,10 @@ define amdgpu_kernel void @sdiv_v2i32(ptr addrspace(1) %out, ptr addrspace(1) %i
; GFX9-NEXT: s_add_i32 s9, s6, 1
; GFX9-NEXT: s_cmp_ge_u32 s8, s7
; GFX9-NEXT: s_cselect_b32 s6, s9, s6
-; GFX9-NEXT: s_xor_b32 s6, s6, s4
-; GFX9-NEXT: s_sub_i32 s4, s6, s4
-; GFX9-NEXT: v_mov_b32_e32 v0, s5
-; GFX9-NEXT: v_mov_b32_e32 v1, s4
+; GFX9-NEXT: s_xor_b32 s6, s6, s5
+; GFX9-NEXT: s_sub_i32 s5, s6, s5
+; GFX9-NEXT: v_mov_b32_e32 v0, s4
+; GFX9-NEXT: v_mov_b32_e32 v1, s5
; GFX9-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GFX9-NEXT: s_endpgm
;
@@ -792,255 +804,255 @@ define amdgpu_kernel void @sdiv_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %i
; GCN-LABEL: sdiv_v4i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
-; GCN-NEXT: s_mov_b32 s7, 0xf000
-; GCN-NEXT: s_mov_b32 s6, -1
-; GCN-NEXT: s_mov_b32 s10, s6
-; GCN-NEXT: s_mov_b32 s11, s7
+; GCN-NEXT: s_mov_b32 s11, 0xf000
+; GCN-NEXT: s_mov_b32 s10, -1
+; GCN-NEXT: s_mov_b32 s6, s10
+; GCN-NEXT: s_mov_b32 s7, s11
; GCN-NEXT: s_waitcnt lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s8, s2
-; GCN-NEXT: s_mov_b32 s9, s3
-; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
-; GCN-NEXT: buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:16
-; GCN-NEXT: s_mov_b32 s4, s0
-; GCN-NEXT: s_mov_b32 s5, s1
+; GCN-NEXT: s_mov_b32 s4, s2
+; GCN-NEXT: s_mov_b32 s5, s3
+; GCN-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0 offset:16
+; GCN-NEXT: buffer_load_dwordx4 v[4:7], off, s[4:7], 0
+; GCN-NEXT: s_mov_b32 s8, s0
+; GCN-NEXT: s_mov_b32 s9, s1
; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_sub_i32_e32 v9, vcc, 0, v0
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_sub_i32_e32 v10, vcc, 0, v4
-; GCN-NEXT: v_xor_b32_e32 v8, v0, v4
-; GCN-NEXT: v_max_i32_e32 v4, v4, v10
-; GCN-NEXT: v_cvt_f32_u32_e32 v10, v4
-; GCN-NEXT: v_sub_i32_e32 v13, vcc, 0, v5
-; GCN-NEXT: v_xor_b32_e32 v11, v1, v5
-; GCN-NEXT: v_rcp_iflag_f32_e32 v10, v10
-; GCN-NEXT: v_max_i32_e32 v5, v5, v13
-; GCN-NEXT: v_cvt_f32_u32_e32 v13, v5
-; GCN-NEXT: v_sub_i32_e32 v16, vcc, 0, v4
-; GCN-NEXT: v_mul_f32_e32 v10, 0x4f7ffffe, v10
-; GCN-NEXT: v_cvt_u32_f32_e32 v10, v10
-; GCN-NEXT: v_rcp_iflag_f32_e32 v13, v13
-; GCN-NEXT: v_sub_i32_e32 v12, vcc, 0, v1
-; GCN-NEXT: v_mul_lo_u32 v16, v16, v10
-; GCN-NEXT: v_mul_f32_e32 v13, 0x4f7ffffe, v13
-; GCN-NEXT: v_cvt_u32_f32_e32 v13, v13
-; GCN-NEXT: v_max_i32_e32 v0, v0, v9
-; GCN-NEXT: v_mul_hi_u32 v16, v10, v16
-; GCN-NEXT: v_max_i32_e32 v1, v1, v12
-; GCN-NEXT: v_sub_i32_e32 v15, vcc, 0, v6
-; GCN-NEXT: v_add_i32_e32 v10, vcc, v10, v16
-; GCN-NEXT: v_sub_i32_e32 v16, vcc, 0, v5
-; GCN-NEXT: v_mul_lo_u32 v16, v16, v13
-; GCN-NEXT: v_mul_hi_u32 v10, v0, v10
-; GCN-NEXT: v_xor_b32_e32 v14, v2, v6
-; GCN-NEXT: v_max_i32_e32 v6, v6, v15
-; GCN-NEXT: v_mul_hi_u32 v12, v13, v16
-; GCN-NEXT: v_cvt_f32_u32_e32 v15, v6
-; GCN-NEXT: v_ashrrev_i32_e32 v8, 31, v8
-; GCN-NEXT: v_ashrrev_i32_e32 v11, 31, v11
-; GCN-NEXT: v_add_i32_e32 v12, vcc, v13, v12
-; GCN-NEXT: v_mul_lo_u32 v13, v10, v4
-; GCN-NEXT: v_mul_hi_u32 v12, v1, v12
-; GCN-NEXT: v_rcp_iflag_f32_e32 v9, v15
-; GCN-NEXT: v_ashrrev_i32_e32 v14, 31, v14
-; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v13
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 1, v10
-; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v4
-; GCN-NEXT: v_cndmask_b32_e64 v10, v10, v13, s[0:1]
-; GCN-NEXT: v_sub_i32_e32 v13, vcc, v0, v4
-; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v13, s[0:1]
-; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v4
-; GCN-NEXT: v_mul_lo_u32 v0, v12, v5
-; GCN-NEXT: v_mul_f32_e32 v9, 0x4f7ffffe, v9
-; GCN-NEXT: v_cvt_u32_f32_e32 v9, v9
-; GCN-NEXT: v_sub_i32_e32 v4, vcc, 0, v6
-; GCN-NEXT: v_sub_i32_e32 v0, vcc, v1, v0
-; GCN-NEXT: v_add_i32_e32 v1, vcc, 1, v12
-; GCN-NEXT: v_cmp_ge_u32_e64 s[2:3], v0, v5
-; GCN-NEXT: v_cndmask_b32_e64 v1, v12, v1, s[2:3]
-; GCN-NEXT: v_sub_i32_e32 v12, vcc, v0, v5
-; GCN-NEXT: v_mul_lo_u32 v4, v4, v9
-; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v12, s[2:3]
-; GCN-NEXT: v_cmp_ge_u32_e64 s[2:3], v0, v5
-; GCN-NEXT: v_sub_i32_e32 v0, vcc, 0, v7
-; GCN-NEXT: v_max_i32_e32 v5, v7, v0
-; GCN-NEXT: v_cvt_f32_u32_e32 v0, v5
-; GCN-NEXT: v_mul_hi_u32 v4, v9, v4
-; GCN-NEXT: v_add_i32_e32 v13, vcc, 1, v10
+; GCN-NEXT: v_readfirstlane_b32 s0, v0
+; GCN-NEXT: v_readfirstlane_b32 s1, v1
+; GCN-NEXT: v_readfirstlane_b32 s2, v2
+; GCN-NEXT: s_abs_i32 s13, s0
+; GCN-NEXT: s_abs_i32 s14, s1
+; GCN-NEXT: s_abs_i32 s15, s2
+; GCN-NEXT: v_cvt_f32_u32_e32 v0, s13
+; GCN-NEXT: v_cvt_f32_u32_e32 v1, s14
+; GCN-NEXT: v_cvt_f32_u32_e32 v2, s15
+; GCN-NEXT: v_readfirstlane_b32 s6, v3
; GCN-NEXT: v_rcp_iflag_f32_e32 v0, v0
-; GCN-NEXT: v_add_i32_e32 v4, vcc, v9, v4
-; GCN-NEXT: v_sub_i32_e32 v9, vcc, 0, v2
-; GCN-NEXT: v_max_i32_e32 v2, v2, v9
-; GCN-NEXT: v_mul_hi_u32 v4, v2, v4
+; GCN-NEXT: v_rcp_iflag_f32_e32 v1, v1
+; GCN-NEXT: v_rcp_iflag_f32_e32 v2, v2
+; GCN-NEXT: s_abs_i32 s17, s6
+; GCN-NEXT: v_cvt_f32_u32_e32 v3, s17
; GCN-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
-; GCN-NEXT: v_cvt_u32_f32_e32 v9, v0
-; GCN-NEXT: v_cndmask_b32_e64 v0, v10, v13, s[0:1]
-; GCN-NEXT: v_xor_b32_e32 v0, v0, v8
-; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v8
-; GCN-NEXT: v_mul_lo_u32 v8, v4, v6
-; GCN-NEXT: v_add_i32_e32 v12, vcc, 1, v1
-; GCN-NEXT: v_sub_i32_e32 v10, vcc, 0, v5
-; GCN-NEXT: v_sub_i32_e32 v2, vcc, v2, v8
-; GCN-NEXT: v_cndmask_b32_e64 v1, v1, v12, s[2:3]
-; GCN-NEXT: v_mul_lo_u32 v10, v10, v9
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 1, v4
-; GCN-NEXT: v_cmp_ge_u32_e64 s[0:1], v2, v6
-; GCN-NEXT: v_xor_b32_e32 v1, v1, v11
-; GCN-NEXT: v_cndmask_b32_e64 v4, v4, v8, s[0:1]
-; GCN-NEXT: v_sub_i32_e32 v8, vcc, v2, v6
-; GCN-NEXT: v_sub_i32_e32 v1, vcc, v1, v11
-; GCN-NEXT: v_cndmask_b32_e64 v2, v2, v8, s[0:1]
-; GCN-NEXT: v_add_i32_e32 v8, vcc, 1, v4
-; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v2, v6
-; GCN-NEXT: v_cndmask_b32_e32 v2, v4, v8, vcc
-; GCN-NEXT: v_mul_hi_u32 v4, v9, v10
-; GCN-NEXT: v_sub_i32_e32 v6, vcc, 0, v3
-; GCN-NEXT: v_max_i32_e32 v6, v3, v6
-; GCN-NEXT: v_add_i32_e32 v4, vcc, v9, v4
-; GCN-NEXT: v_mul_hi_u32 v4, v6, v4
-; GCN-NEXT: v_xor_b32_e32 v2, v2, v14
-; GCN-NEXT: v_sub_i32_e32 v2, vcc, v2, v14
-; GCN-NEXT: v_mul_lo_u32 v8, v4, v5
-; GCN-NEXT: v_xor_b32_e32 v3, v3, v7
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 1, v4
-; GCN-NEXT: v_sub_i32_e32 v6, vcc, v6, v8
-; GCN-NEXT: v_sub_i32_e32 v8, vcc, v6, v5
-; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v6, v5
-; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v7, vcc
-; GCN-NEXT: v_cndmask_b32_e32 v6, v6, v8, vcc
-; GCN-NEXT: v_add_i32_e32 v7, vcc, 1, v4
-; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v6, v5
-; GCN-NEXT: v_ashrrev_i32_e32 v3, 31, v3
-; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v7, vcc
-; GCN-NEXT: v_xor_b32_e32 v4, v4, v3
-; GCN-NEXT: v_sub_i32_e32 v3, vcc, v4, v3
-; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; GCN-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; GCN-NEXT: v_mul_f32_e32 v2, 0x4f7ffffe, v2
+; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0
+; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1
+; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: v_readfirstlane_b32 s3, v4
+; GCN-NEXT: v_readfirstlane_b32 s4, v5
+; GCN-NEXT: v_readfirstlane_b32 s5, v6
+; GCN-NEXT: s_xor_b32 s12, s3, s0
+; GCN-NEXT: s_xor_b32 s0, s4, s1
+; GCN-NEXT: s_xor_b32 s1, s5, s2
+; GCN-NEXT: s_sub_i32 s2, 0, s13
+; GCN-NEXT: s_ashr_i32 s18, s0, 31
+; GCN-NEXT: s_sub_i32 s0, 0, s14
+; GCN-NEXT: s_ashr_i32 s19, s1, 31
+; GCN-NEXT: s_sub_i32 s1, 0, s15
+; GCN-NEXT: v_rcp_iflag_f32_e32 v3, v3
+; GCN-NEXT: v_mul_lo_u32 v4, s2, v0
+; GCN-NEXT: v_mul_lo_u32 v5, s0, v1
+; GCN-NEXT: v_mul_lo_u32 v6, s1, v2
+; GCN-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
+; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3
+; GCN-NEXT: v_mul_hi_u32 v4, v0, v4
+; GCN-NEXT: v_mul_hi_u32 v5, v1, v5
+; GCN-NEXT: v_mul_hi_u32 v6, v2, v6
+; GCN-NEXT: s_sub_i32 s20, 0, s17
+; GCN-NEXT: v_readfirstlane_b32 s7, v7
+; GCN-NEXT: s_abs_i32 s3, s3
+; GCN-NEXT: s_abs_i32 s4, s4
+; GCN-NEXT: s_abs_i32 s5, s5
+; GCN-NEXT: v_mul_lo_u32 v7, s20, v3
+; GCN-NEXT: v_add_i32_e32 v0, vcc, v0, v4
+; GCN-NEXT: v_add_i32_e32 v1, vcc, v1, v5
+; GCN-NEXT: v_add_i32_e32 v2, vcc, v2, v6
+; GCN-NEXT: v_mul_hi_u32 v0, s3, v0
+; GCN-NEXT: v_mul_hi_u32 v1, s4, v1
+; GCN-NEXT: v_mul_hi_u32 v2, s5, v2
+; GCN-NEXT: v_mul_hi_u32 v7, v3, v7
+; GCN-NEXT: v_mul_lo_u32 v4, v0, s13
+; GCN-NEXT: v_mul_lo_u32 v6, v1, s14
+; GCN-NEXT: v_mul_lo_u32 v8, v2, s15
+; GCN-NEXT: s_abs_i32 s16, s7
+; GCN-NEXT: v_add_i32_e32 v3, vcc, v3, v7
+; GCN-NEXT: v_mul_hi_u32 v3, s16, v3
+; GCN-NEXT: v_sub_i32_e32 v4, vcc, s3, v4
+; GCN-NEXT: v_sub_i32_e32 v6, vcc, s4, v6
+; GCN-NEXT: v_sub_i32_e32 v8, vcc, s5, v8
+; GCN-NEXT: v_add_i32_e32 v5, vcc, 1, v0
+; GCN-NEXT: v_add_i32_e32 v7, vcc, 1, v1
+; GCN-NEXT: v_add_i32_e32 v9, vcc, 1, v2
+; GCN-NEXT: v_cmp_le_u32_e64 s[0:1], s13, v4
+; GCN-NEXT: v_cmp_le_u32_e64 s[2:3], s14, v6
+; GCN-NEXT: v_cmp_le_u32_e64 s[4:5], s15, v8
+; GCN-NEXT: v_subrev_i32_e32 v10, vcc, s13, v4
+; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v5, s[0:1]
+; GCN-NEXT: v_subrev_i32_e32 v5, vcc, s14, v6
+; GCN-NEXT: v_cndmask_b32_e64 v1, v1, v7, s[2:3]
+; GCN-NEXT: v_subrev_i32_e32 v7, vcc, s15, v8
+; GCN-NEXT: v_cndmask_b32_e64 v2, v2, v9, s[4:5]
+; GCN-NEXT: v_cndmask_b32_e64 v4, v4, v10, s[0:1]
+; GCN-NEXT: v_add_i32_e32 v9, vcc, 1, v0
+; GCN-NEXT: v_cndmask_b32_e64 v5, v6, v5, s[2:3]
+; GCN-NEXT: v_add_i32_e32 v6, vcc, 1, v1
+; GCN-NEXT: v_cndmask_b32_e64 v7, v8, v7, s[4:5]
+; GCN-NEXT: v_add_i32_e32 v8, vcc, 1, v2
+; GCN-NEXT: v_cmp_le_u32_e32 vcc, s13, v4
+; GCN-NEXT: v_mul_lo_u32 v4, v3, s17
+; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v9, vcc
+; GCN-NEXT: v_cmp_le_u32_e32 vcc, s14, v5
+; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v6, vcc
+; GCN-NEXT: v_cmp_le_u32_e32 vcc, s15, v7
+; GCN-NEXT: s_ashr_i32 s12, s12, 31
+; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v8, vcc
+; GCN-NEXT: v_xor_b32_e32 v0, s12, v0
+; GCN-NEXT: v_xor_b32_e32 v1, s18, v1
+; GCN-NEXT: v_xor_b32_e32 v2, s19, v2
+; GCN-NEXT: v_sub_i32_e32 v4, vcc, s16, v4
+; GCN-NEXT: v_subrev_i32_e32 v0, vcc, s12, v0
+; GCN-NEXT: v_subrev_i32_e32 v1, vcc, s18, v1
+; GCN-NEXT: v_subrev_i32_e32 v2, vcc, s19, v2
+; GCN-NEXT: v_add_i32_e32 v5, vcc, 1, v3
+; GCN-NEXT: v_subrev_i32_e32 v6, vcc, s17, v4
+; GCN-NEXT: v_cmp_le_u32_e32 vcc, s17, v4
+; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc
+; GCN-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc
+; GCN-NEXT: v_add_i32_e32 v5, vcc, 1, v3
+; GCN-NEXT: s_xor_b32 s0, s7, s6
+; GCN-NEXT: v_cmp_le_u32_e32 vcc, s17, v4
+; GCN-NEXT: s_ashr_i32 s0, s0, 31
+; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc
+; GCN-NEXT: v_xor_b32_e32 v3, s0, v3
+; GCN-NEXT: v_subrev_i32_e32 v3, vcc, s0, v3
+; GCN-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; GCN-NEXT: s_endpgm
;
; TONGA-LABEL: sdiv_v4i32:
; TONGA: ; %bb.0:
; TONGA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
-; TONGA-NEXT: s_mov_b32 s7, 0xf000
-; TONGA-NEXT: s_mov_b32 s6, -1
-; TONGA-NEXT: s_mov_b32 s10, s6
-; TONGA-NEXT: s_mov_b32 s11, s7
+; TONGA-NEXT: s_mov_b32 s11, 0xf000
+; TONGA-NEXT: s_mov_b32 s10, -1
+; TONGA-NEXT: s_mov_b32 s6, s10
+; TONGA-NEXT: s_mov_b32 s7, s11
; TONGA-NEXT: s_waitcnt lgkmcnt(0)
-; TONGA-NEXT: s_mov_b32 s8, s2
-; TONGA-NEXT: s_mov_b32 s9, s3
-; TONGA-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0
-; TONGA-NEXT: buffer_load_dwordx4 v[4:7], off, s[8:11], 0 offset:16
-; TONGA-NEXT: s_mov_b32 s4, s0
-; TONGA-NEXT: s_mov_b32 s5, s1
+; TONGA-NEXT: s_mov_b32 s4, s2
+; TONGA-NEXT: s_mov_b32 s5, s3
+; TONGA-NEXT: buffer_load_dwordx4 v[0:3], off, s[4:7], 0 offset:16
+; TONGA-NEXT: buffer_load_dwordx4 v[4:7], off, s[4:7], 0
+; TONGA-NEXT: s_mov_b32 s8, s0
+; TONGA-NEXT: s_mov_b32 s9, s1
; TONGA-NEXT: s_waitcnt vmcnt(1)
-; TONGA-NEXT: v_sub_u32_e32 v9, vcc, 0, v0
-; TONGA-NEXT: s_waitcnt vmcnt(0)
-; TONGA-NEXT: v_sub_u32_e32 v10, vcc, 0, v4
-; TONGA-NEXT: v_xor_b32_e32 v8, v0, v4
-; TONGA-NEXT: v_max_i32_e32 v4, v4, v10
-; TONGA-NEXT: v_cvt_f32_u32_e32 v10, v4
-; TONGA-NEXT: v_sub_u32_e32 v13, vcc, 0, v5
-; TONGA-NEXT: v_xor_b32_e32 v11, v1, v5
-; TONGA-NEXT: v_rcp_iflag_f32_e32 v10, v10
-; TONGA-NEXT: v_max_i32_e32 v5, v5, v13
-; TONGA-NEXT: v_cvt_f32_u32_e32 v13, v5
-; TONGA-NEXT: v_sub_u32_e32 v16, vcc, 0, v4
-; TONGA-NEXT: v_mul_f32_e32 v10, 0x4f7ffffe, v10
-; TONGA-NEXT: v_cvt_u32_f32_e32 v10, v10
-; TONGA-NEXT: v_rcp_iflag_f32_e32 v13, v13
-; TONGA-NEXT: v_sub_u32_e32 v12, vcc, 0, v1
-; TONGA-NEXT: v_mul_lo_u32 v16, v16, v10
-; TONGA-NEXT: v_mul_f32_e32 v13, 0x4f7ffffe, v13
-; TONGA-NEXT: v_cvt_u32_f32_e32 v13, v13
-; TONGA-NEXT: v_max_i32_e32 v0, v0, v9
-; TONGA-NEXT: v_mul_hi_u32 v16, v10, v16
-; TONGA-NEXT: v_max_i32_e32 v1, v1, v12
-; TONGA-NEXT: v_sub_u32_e32 v15, vcc, 0, v6
-; TONGA-NEXT: v_add_u32_e32 v10, vcc, v10, v16
-; TONGA-NEXT: v_sub_u32_e32 v16, vcc, 0, v5
-; TONGA-NEXT: v_mul_lo_u32 v16, v16, v13
-; TONGA-NEXT: v_mul_hi_u32 v10, v0, v10
-; TONGA-NEXT: v_xor_b32_e32 v14, v2, v6
-; TONGA-NEXT: v_max_i32_e32 v6, v6, v15
-; TONGA-NEXT: v_mul_hi_u32 v12, v13, v16
-; TONGA-NEXT: v_cvt_f32_u32_e32 v15, v6
-; TONGA-NEXT: v_ashrrev_i32_e32 v8, 31, v8
-; TONGA-NEXT: v_ashrrev_i32_e32 v11, 31, v11
-; TONGA-NEXT: v_add_u32_e32 v12, vcc, v13, v12
-; TONGA-NEXT: v_mul_lo_u32 v13, v10, v4
-; TONGA-NEXT: v_mul_hi_u32 v12, v1, v12
-; TONGA-NEXT: v_rcp_iflag_f32_e32 v9, v15
-; TONGA-NEXT: v_ashrrev_i32_e32 v14, 31, v14
-; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v0, v13
-; TONGA-NEXT: v_add_u32_e32 v13, vcc, 1, v10
-; TONGA-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v4
-; TONGA-NEXT: v_cndmask_b32_e64 v10, v10, v13, s[0:1]
-; TONGA-NEXT: v_sub_u32_e32 v13, vcc, v0, v4
-; TONGA-NEXT: v_cndmask_b32_e64 v0, v0, v13, s[0:1]
-; TONGA-NEXT: v_cmp_ge_u32_e64 s[0:1], v0, v4
-; TONGA-NEXT: v_mul_lo_u32 v0, v12, v5
-; TONGA-NEXT: v_mul_f32_e32 v9, 0x4f7ffffe, v9
-; TONGA-NEXT: v_cvt_u32_f32_e32 v9, v9
-; TONGA-NEXT: v_sub_u32_e32 v4, vcc, 0, v6
-; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v1, v0
-; TONGA-NEXT: v_add_u32_e32 v1, vcc, 1, v12
-; TONGA-NEXT: v_cmp_ge_u32_e64 s[2:3], v0, v5
-; TONGA-NEXT: v_cndmask_b32_e64 v1, v12, v1, s[2:3]
-; TONGA-NEXT: v_sub_u32_e32 v12, vcc, v0, v5
-; TONGA-NEXT: v_mul_lo_u32 v4, v4, v9
-; TONGA-NEXT: v_cndmask_b32_e64 v0, v0, v12, s[2:3]
-; TONGA-NEXT: v_cmp_ge_u32_e64 s[2:3], v0, v5
-; TONGA-NEXT: v_sub_u32_e32 v0, vcc, 0, v7
-; TONGA-NEXT: v_max_i32_e32 v5, v7, v0
-; TONGA-NEXT: v_cvt_f32_u32_e32 v0, v5
-; TONGA-NEXT: v_mul_hi_u32 v4, v9, v4
-; TONGA-NEXT: v_add_u32_e32 v13, vcc, 1, v10
+; TONGA-NEXT: v_readfirstlane_b32 s0, v0
+; TONGA-NEXT: v_readfirstlane_b32 s1, v1
+; TONGA-NEXT: v_readfirstlane_b32 s2, v2
+; TONGA-NEXT: s_abs_i32 s13, s0
+; TONGA-NEXT: s_abs_i32 s14, s1
+; TONGA-NEXT: s_abs_i32 s15, s2
+; TONGA-NEXT: v_cvt_f32_u32_e32 v0, s13
+; TONGA-NEXT: v_cvt_f32_u32_e32 v1, s14
+; TONGA-NEXT: v_cvt_f32_u32_e32 v2, s15
+; TONGA-NEXT: v_readfirstlane_b32 s6, v3
; TONGA-NEXT: v_rcp_iflag_f32_e32 v0, v0
-; TONGA-NEXT: v_add_u32_e32 v4, vcc, v9, v4
-; TONGA-NEXT: v_sub_u32_e32 v9, vcc, 0, v2
-; TONGA-NEXT: v_max_i32_e32 v2, v2, v9
-; TONGA-NEXT: v_mul_hi_u32 v4, v2, v4
+; TONGA-NEXT: v_rcp_iflag_f32_e32 v1, v1
+; TONGA-NEXT: v_rcp_iflag_f32_e32 v2, v2
+; TONGA-NEXT: s_abs_i32 s17, s6
+; TONGA-NEXT: v_cvt_f32_u32_e32 v3, s17
; TONGA-NEXT: v_mul_f32_e32 v0, 0x4f7ffffe, v0
-; TONGA-NEXT: v_cvt_u32_f32_e32 v9, v0
-; TONGA-NEXT: v_cndmask_b32_e64 v0, v10, v13, s[0:1]
-; TONGA-NEXT: v_xor_b32_e32 v0, v0, v8
-; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v0, v8
-; TONGA-NEXT: v_mul_lo_u32 v8, v4, v6
-; TONGA-NEXT: v_add_u32_e32 v12, vcc, 1, v1
-; TONGA-NEXT: v_sub_u32_e32 v10, vcc, 0, v5
-; TONGA-NEXT: v_sub_u32_e32 v2, vcc, v2, v8
-; TONGA-NEXT: v_cndmask_b32_e64 v1, v1, v12, s[2:3]
-; TONGA-NEXT: v_mul_lo_u32 v10, v10, v9
-; TONGA-NEXT: v_add_u32_e32 v8, vcc, 1, v4
-; TONGA-NEXT: v_cmp_ge_u32_e64 s[0:1], v2, v6
-; TONGA-NEXT: v_xor_b32_e32 v1, v1, v11
-; TONGA-NEXT: v_cndmask_b32_e64 v4, v4, v8, s[0:1]
-; TONGA-NEXT: v_sub_u32_e32 v8, vcc, v2, v6
-; TONGA-NEXT: v_sub_u32_e32 v1, vcc, v1, v11
-; TONGA-NEXT: v_cndmask_b32_e64 v2, v2, v8, s[0:1]
-; TONGA-NEXT: v_add_u32_e32 v8, vcc, 1, v4
-; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v2, v6
-; TONGA-NEXT: v_cndmask_b32_e32 v2, v4, v8, vcc
-; TONGA-NEXT: v_mul_hi_u32 v4, v9, v10
-; TONGA-NEXT: v_sub_u32_e32 v6, vcc, 0, v3
-; TONGA-NEXT: v_max_i32_e32 v6, v3, v6
-; TONGA-NEXT: v_add_u32_e32 v4, vcc, v9, v4
-; TONGA-NEXT: v_mul_hi_u32 v4, v6, v4
-; TONGA-NEXT: v_xor_b32_e32 v2, v2, v14
-; TONGA-NEXT: v_sub_u32_e32 v2, vcc, v2, v14
-; TONGA-NEXT: v_mul_lo_u32 v8, v4, v5
-; TONGA-NEXT: v_xor_b32_e32 v3, v3, v7
-; TONGA-NEXT: v_add_u32_e32 v7, vcc, 1, v4
-; TONGA-NEXT: v_sub_u32_e32 v6, vcc, v6, v8
-; TONGA-NEXT: v_sub_u32_e32 v8, vcc, v6, v5
-; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v6, v5
-; TONGA-NEXT: v_cndmask_b32_e32 v4, v4, v7, vcc
-; TONGA-NEXT: v_cndmask_b32_e32 v6, v6, v8, vcc
-; TONGA-NEXT: v_add_u32_e32 v7, vcc, 1, v4
-; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v6, v5
-; TONGA-NEXT: v_ashrrev_i32_e32 v3, 31, v3
-; TONGA-NEXT: v_cndmask_b32_e32 v4, v4, v7, vcc
-; TONGA-NEXT: v_xor_b32_e32 v4, v4, v3
-; TONGA-NEXT: v_sub_u32_e32 v3, vcc, v4, v3
-; TONGA-NEXT: buffer_store_dwordx4 v[0:3], off, s[4:7], 0
+; TONGA-NEXT: v_mul_f32_e32 v1, 0x4f7ffffe, v1
+; TONGA-NEXT: v_mul_f32_e32 v2, 0x4f7ffffe, v2
+; TONGA-NEXT: v_cvt_u32_f32_e32 v0, v0
+; TONGA-NEXT: v_cvt_u32_f32_e32 v1, v1
+; TONGA-NEXT: v_cvt_u32_f32_e32 v2, v2
+; TONGA-NEXT: s_waitcnt vmcnt(0)
+; TONGA-NEXT: v_readfirstlane_b32 s3, v4
+; TONGA-NEXT: v_readfirstlane_b32 s4, v5
+; TONGA-NEXT: v_readfirstlane_b32 s5, v6
+; TONGA-NEXT: s_xor_b32 s12, s3, s0
+; TONGA-NEXT: s_xor_b32 s0, s4, s1
+; TONGA-NEXT: s_xor_b32 s1, s5, s2
+; TONGA-NEXT: s_sub_i32 s2, 0, s13
+; TONGA-NEXT: s_ashr_i32 s18, s0, 31
+; TONGA-NEXT: s_sub_i32 s0, 0, s14
+; TONGA-NEXT: s_ashr_i32 s19, s1, 31
+; TONGA-NEXT: s_sub_i32 s1, 0, s15
+; TONGA-NEXT: v_rcp_iflag_f32_e32 v3, v3
+; TONGA-NEXT: v_mul_lo_u32 v4, s2, v0
+; TONGA-NEXT: v_mul_lo_u32 v5, s0, v1
+; TONGA-NEXT: v_mul_lo_u32 v6, s1, v2
+; TONGA-NEXT: v_mul_f32_e32 v3, 0x4f7ffffe, v3
+; TONGA-NEXT: v_cvt_u32_f32_e32 v3, v3
+; TONGA-NEXT: v_mul_hi_u32 v4, v0, v4
+; TONGA-NEXT: v_mul_hi_u32 v5, v1, v5
+; TONGA-NEXT: v_mul_hi_u32 v6, v2, v6
+; TONGA-NEXT: s_sub_i32 s20, 0, s17
+; TONGA-NEXT: v_readfirstlane_b32 s7, v7
+; TONGA-NEXT: s_abs_i32 s3, s3
+; TONGA-NEXT: s_abs_i32 s4, s4
+; TONGA-NEXT: s_abs_i32 s5, s5
+; TONGA-NEXT: v_mul_lo_u32 v7, s20, v3
+; TONGA-NEXT: v_add_u32_e32 v0, vcc, v0, v4
+; TONGA-NEXT: v_add_u32_e32 v1, vcc, v1, v5
+; TONGA-NEXT: v_add_u32_e32 v2, vcc, v2, v6
+; TONGA-NEXT: v_mul_hi_u32 v0, s3, v0
+; TONGA-NEXT: v_mul_hi_u32 v1, s4, v1
+; TONGA-NEXT: v_mul_hi_u32 v2, s5, v2
+; TONGA-NEXT: v_mul_hi_u32 v7, v3, v7
+; TONGA-NEXT: v_mul_lo_u32 v4, v0, s13
+; TONGA-NEXT: v_mul_lo_u32 v6, v1, s14
+; TONGA-NEXT: v_mul_lo_u32 v8, v2, s15
+; TONGA-NEXT: s_abs_i32 s16, s7
+; TONGA-NEXT: v_add_u32_e32 v3, vcc, v3, v7
+; TONGA-NEXT: v_mul_hi_u32 v3, s16, v3
+; TONGA-NEXT: v_sub_u32_e32 v4, vcc, s3, v4
+; TONGA-NEXT: v_sub_u32_e32 v6, vcc, s4, v6
+; TONGA-NEXT: v_sub_u32_e32 v8, vcc, s5, v8
+; TONGA-NEXT: v_add_u32_e32 v5, vcc, 1, v0
+; TONGA-NEXT: v_add_u32_e32 v7, vcc, 1, v1
+; TONGA-NEXT: v_add_u32_e32 v9, vcc, 1, v2
+; TONGA-NEXT: v_cmp_le_u32_e64 s[0:1], s13, v4
+; TONGA-NEXT: v_cmp_le_u32_e64 s[2:3], s14, v6
+; TONGA-NEXT: v_cmp_le_u32_e64 s[4:5], s15, v8
+; TONGA-NEXT: v_subrev_u32_e32 v10, vcc, s13, v4
+; TONGA-NEXT: v_cndmask_b32_e64 v0, v0, v5, s[0:1]
+; TONGA-NEXT: v_subrev_u32_e32 v5, vcc, s14, v6
+; TONGA-NEXT: v_cndmask_b32_e64 v1, v1, v7, s[2:3]
+; TONGA-NEXT: v_subrev_u32_e32 v7, vcc, s15, v8
+; TONGA-NEXT: v_cndmask_b32_e64 v2, v2, v9, s[4:5]
+; TONGA-NEXT: v_cndmask_b32_e64 v4, v4, v10, s[0:1]
+; TONGA-NEXT: v_add_u32_e32 v9, vcc, 1, v0
+; TONGA-NEXT: v_cndmask_b32_e64 v5, v6, v5, s[2:3]
+; TONGA-NEXT: v_add_u32_e32 v6, vcc, 1, v1
+; TONGA-NEXT: v_cndmask_b32_e64 v7, v8, v7, s[4:5]
+; TONGA-NEXT: v_add_u32_e32 v8, vcc, 1, v2
+; TONGA-NEXT: v_cmp_le_u32_e32 vcc, s13, v4
+; TONGA-NEXT: v_mul_lo_u32 v4, v3, s17
+; TONGA-NEXT: v_cndmask_b32_e32 v0, v0, v9, vcc
+; TONGA-NEXT: v_cmp_le_u32_e32 vcc, s14, v5
+; TONGA-NEXT: v_cndmask_b32_e32 v1, v1, v6, vcc
+; TONGA-NEXT: v_cmp_le_u32_e32 vcc, s15, v7
+; TONGA-NEXT: s_ashr_i32 s12, s12, 31
+; TONGA-NEXT: v_cndmask_b32_e32 v2, v2, v8, vcc
+; TONGA-NEXT: v_xor_b32_e32 v0, s12, v0
+; TONGA-NEXT: v_xor_b32_e32 v1, s18, v1
+; TONGA-NEXT: v_xor_b32_e32 v2, s19, v2
+; TONGA-NEXT: v_sub_u32_e32 v4, vcc, s16, v4
+; TONGA-NEXT: v_subrev_u32_e32 v0, vcc, s12, v0
+; TONGA-NEXT: v_subrev_u32_e32 v1, vcc, s18, v1
+; TONGA-NEXT: v_subrev_u32_e32 v2, vcc, s19, v2
+; TONGA-NEXT: v_add_u32_e32 v5, vcc, 1, v3
+; TONGA-NEXT: v_subrev_u32_e32 v6, vcc, s17, v4
+; TONGA-NEXT: v_cmp_le_u32_e32 vcc, s17, v4
+; TONGA-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc
+; TONGA-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc
+; TONGA-NEXT: v_add_u32_e32 v5, vcc, 1, v3
+; TONGA-NEXT: s_xor_b32 s0, s7, s6
+; TONGA-NEXT: v_cmp_le_u32_e32 vcc, s17, v4
+; TONGA-NEXT: s_ashr_i32 s0, s0, 31
+; TONGA-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc
+; TONGA-NEXT: v_xor_b32_e32 v3, s0, v3
+; TONGA-NEXT: v_subrev_u32_e32 v3, vcc, s0, v3
+; TONGA-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; TONGA-NEXT: s_endpgm
;
; GFX9-LABEL: sdiv_v4i32:
@@ -2002,7 +2014,7 @@ define amdgpu_kernel void @v_sdiv_i25(ptr addrspace(1) %out, ptr addrspace(1) %i
; GCN-NEXT: v_cmp_ge_u32_e32 vcc, v1, v2
; GCN-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GCN-NEXT: v_xor_b32_e32 v1, v1, v0
-; GCN-NEXT: v_sub_i32_e32 v0, vcc, v1, v0
+; GCN-NEXT: v_subrev_i32_e32 v0, vcc, v0, v1
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 25
; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-NEXT: s_endpgm
@@ -2049,7 +2061,7 @@ define amdgpu_kernel void @v_sdiv_i25(ptr addrspace(1) %out, ptr addrspace(1) %i
; TONGA-NEXT: v_cmp_ge_u32_e32 vcc, v1, v2
; TONGA-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; TONGA-NEXT: v_xor_b32_e32 v1, v1, v0
-; TONGA-NEXT: v_sub_u32_e32 v0, vcc, v1, v0
+; TONGA-NEXT: v_subrev_u32_e32 v0, vcc, v0, v1
; TONGA-NEXT: v_bfe_i32 v0, v0, 0, 25
; TONGA-NEXT: buffer_store_dword v0, off, s[0:3], 0
; TONGA-NEXT: s_endpgm