1 files changed, 356 insertions, 338 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
index f869b57..053efdc 100644
--- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
@@ -9266,7 +9266,7 @@ define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat %
 ; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v3, v0
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
@@ -9280,20 +9280,22 @@ define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat %
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, v3, v6
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v2
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v7
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.h
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, v3, v7
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -9413,7 +9415,7 @@ define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat %
 ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_ret_bf16:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
@@ -9428,19 +9430,21 @@ define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat %
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, v3, v6
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, v3, v7
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -9767,34 +9771,34 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1)
 ; GFX12-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7fe, v0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
 ; GFX12-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
-; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:    global_load_b32 v5, v[0:1], off
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v4, v3, 0xffff
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_not_b32_e32 v4, v4
 ; GFX12-TRUE16-NEXT:  .LBB33_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, v3, v6
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v2
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v7
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.h
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, v3, v7
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -9920,12 +9924,11 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1)
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7fe, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    global_load_b32 v5, v[0:1], off
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 3, v3
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
@@ -9936,19 +9939,21 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1)
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, v3, v6
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, v3, v7
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -10285,34 +10290,34 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr addrspace(1)
 ; GFX12-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
 ; GFX12-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
-; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:    global_load_b32 v5, v[0:1], off
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v4, v3, 0xffff
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_not_b32_e32 v4, v4
 ; GFX12-TRUE16-NEXT:  .LBB34_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, v3, v6
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v2
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v7
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.h
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, v3, v7
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -10439,12 +10444,11 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr addrspace(1)
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    global_load_b32 v5, v[0:1], off
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 3, v3
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
@@ -10455,19 +10459,21 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr addrspace(1)
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, v3, v6
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, v3, v7
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -10801,7 +10807,7 @@ define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat %
 ; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v3, v0
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
@@ -10815,18 +10821,20 @@ define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat %
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v2
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v7
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v3, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v8, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v3.h
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, v5, v7
@@ -10943,7 +10951,7 @@ define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat %
 ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_noret_bf16:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
@@ -10958,17 +10966,19 @@ define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat %
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
+; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v3, 0x7fff
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v3.h
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, v5, v7
@@ -11283,45 +11293,45 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1)
 ; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7fe, v0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v4
-; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 3, v4
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX12-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 3, v4
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v5, v4, 0xffff
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_not_b32_e32 v5, v5
+; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
+; GFX12-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
+; GFX12-TRUE16-NEXT:    global_load_b32 v4, v[0:1], off
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-TRUE16-NEXT:    v_not_b32_e32 v6, v3
 ; GFX12-TRUE16-NEXT:  .LBB36_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
-; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, v4, v3
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v2, v2, v6
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v7
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
+; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v3, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v8, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v2.h
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, v4, v7
-; GFX12-TRUE16-NEXT:    v_and_or_b32 v2, v3, v5, v2
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v3.h
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, v5, v7
+; GFX12-TRUE16-NEXT:    v_and_or_b32 v3, v4, v6, v3
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    global_inv scope:SCOPE_DEV
-; GFX12-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v3, v2
+; GFX12-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
 ; GFX12-TRUE16-NEXT:    s_or_b32 s0, vcc_lo, s0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
@@ -11433,45 +11443,46 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1)
 ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_add_co_u32 v4, vcc_lo, 0x7fe, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 3, v4
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
+; GFX11-TRUE16-NEXT:    global_load_b32 v4, v[0:1], off
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 3, v3
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e64 v5, v4, 0xffff
-; GFX11-TRUE16-NEXT:    v_not_b32_e32 v5, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-TRUE16-NEXT:    v_not_b32_e32 v6, v3
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB36_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
+; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v3, 0x7fff
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v2, v2, v6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v2.h
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v3.h
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, v4, v7
-; GFX11-TRUE16-NEXT:    v_and_or_b32 v2, v3, v5, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, v5, v7
+; GFX11-TRUE16-NEXT:    v_and_or_b32 v3, v4, v6, v3
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    buffer_gl1_inv
 ; GFX11-TRUE16-NEXT:    buffer_gl0_inv
-; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v2
+; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, vcc_lo, s0
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
@@ -11784,45 +11795,45 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr addrspace(1)
 ; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v4
-; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 3, v4
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX12-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 3, v4
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v5, v4, 0xffff
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_not_b32_e32 v5, v5
+; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
+; GFX12-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
+; GFX12-TRUE16-NEXT:    global_load_b32 v4, v[0:1], off
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-TRUE16-NEXT:    v_not_b32_e32 v6, v3
 ; GFX12-TRUE16-NEXT:  .LBB37_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
-; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, v4, v3
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v2, v2, v6
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v7
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
+; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v3, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v8, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v2.h
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, v4, v7
-; GFX12-TRUE16-NEXT:    v_and_or_b32 v2, v3, v5, v2
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v3.h
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, v5, v7
+; GFX12-TRUE16-NEXT:    v_and_or_b32 v3, v4, v6, v3
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    global_inv scope:SCOPE_DEV
-; GFX12-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v3, v2
+; GFX12-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
 ; GFX12-TRUE16-NEXT:    s_or_b32 s0, vcc_lo, s0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
@@ -11935,45 +11946,46 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr addrspace(1)
 ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 3, v4
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
+; GFX11-TRUE16-NEXT:    global_load_b32 v4, v[0:1], off
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 3, v3
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e64 v5, v4, 0xffff
-; GFX11-TRUE16-NEXT:    v_not_b32_e32 v5, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-TRUE16-NEXT:    v_not_b32_e32 v6, v3
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB37_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
+; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v3, 0x7fff
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v2, v2, v6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v2.h
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v3.h
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, v4, v7
-; GFX11-TRUE16-NEXT:    v_and_or_b32 v2, v3, v5, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, v5, v7
+; GFX11-TRUE16-NEXT:    v_and_or_b32 v3, v4, v6, v3
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    buffer_gl1_inv
 ; GFX11-TRUE16-NEXT:    buffer_gl0_inv
-; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v2
+; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, vcc_lo, s0
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
@@ -12287,27 +12299,28 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr addr
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off offset:2046
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:  .LBB38_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v2
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
-; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX12-TRUE16-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v3
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v3.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
+; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-TRUE16-NEXT:    v_add3_u32 v6, v6, v5, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.h, 0
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc_lo
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v5.h
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_and_or_b32 v3, 0xffff0000, v4, v5
+; GFX12-TRUE16-NEXT:    v_and_or_b32 v3, 0xffff0000, v4, v3
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
 ; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
@@ -12407,27 +12420,28 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr addr
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off offset:2046
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB38_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v3
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v3.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
+; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v5, 0x7fff
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc_lo
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v5.h
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_and_or_b32 v3, 0xffff0000, v4, v5
+; GFX11-TRUE16-NEXT:    v_and_or_b32 v3, 0xffff0000, v4, v3
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
@@ -12696,32 +12710,33 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr addr
 ; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off offset:2046
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-TRUE16-NEXT:    global_load_b32 v4, v[0:1], off offset:2046
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:  .LBB39_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v2.l
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v2, v2, v4
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v2
-; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v3
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v3.l
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
+; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-TRUE16-NEXT:    v_add3_u32 v6, v6, v5, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v5, v6, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.h, 0
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v2.h
-; GFX12-TRUE16-NEXT:    v_and_or_b32 v2, 0xffff0000, v3, v5
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v5.h
+; GFX12-TRUE16-NEXT:    v_and_or_b32 v3, 0xffff0000, v4, v3
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    global_inv scope:SCOPE_DEV
-; GFX12-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v3, v2
+; GFX12-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
 ; GFX12-TRUE16-NEXT:    s_or_b32 s0, vcc_lo, s0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
@@ -12811,33 +12826,34 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr addr
 ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off offset:2046
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-TRUE16-NEXT:    global_load_b32 v4, v[0:1], off offset:2046
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB39_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v2.l
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v2, v2, v4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v4
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v3
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v3.l
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
+; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v2, 0x7fff
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v5, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_or_b32 v2, 0xffff0000, v3, v5
+; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v5, 0x7fff
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v5.h
+; GFX11-TRUE16-NEXT:    v_and_or_b32 v3, 0xffff0000, v4, v3
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    buffer_gl1_inv
 ; GFX11-TRUE16-NEXT:    buffer_gl0_inv
-; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v2
+; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, vcc_lo, s0
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
@@ -13095,34 +13111,34 @@ define bfloat @global_system_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1
 ; GFX12-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7fe, v0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
 ; GFX12-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
-; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX12-TRUE16-NEXT:    global_load_b32 v5, v[0:1], off
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v4, v3, 0xffff
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_not_b32_e32 v4, v4
 ; GFX12-TRUE16-NEXT:  .LBB40_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, v3, v6
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v2
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v7
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.h
 ; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, v3, v7
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -13250,12 +13266,11 @@ define bfloat @global_system_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7fe, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
 ; GFX11-TRUE16-NEXT:    global_load_b32 v5, v[0:1], off
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 3, v3
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
@@ -13266,19 +13281,21 @@ define bfloat @global_system_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, v3, v6
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v5, v5, v7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, v3, v7
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -13614,46 +13631,46 @@ define void @global_system_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1
 ; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7fe, v0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v4
-; GFX12-TRUE16-NEXT:    v_and_b32_e32 v4, 3, v4
 ; GFX12-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX12-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 3, v4
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v5, v4, 0xffff
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_not_b32_e32 v5, v5
+; GFX12-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
+; GFX12-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
+; GFX12-TRUE16-NEXT:    global_load_b32 v4, v[0:1], off
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-TRUE16-NEXT:    v_not_b32_e32 v6, v3
 ; GFX12-TRUE16-NEXT:  .LBB41_1: ; %atomicrmw.start
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
-; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, v4, v3
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v2, v2, v6
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX12-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v7
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX12-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
+; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v3, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v8, vcc_lo
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v2.h
 ; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, v4, v7
-; GFX12-TRUE16-NEXT:    v_and_or_b32 v2, v3, v5, v2
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v3.h
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, v5, v7
+; GFX12-TRUE16-NEXT:    v_and_or_b32 v3, v4, v6, v3
 ; GFX12-TRUE16-NEXT:    global_wb scope:SCOPE_SYS
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
 ; GFX12-TRUE16-NEXT:    global_inv scope:SCOPE_SYS
-; GFX12-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v3, v2
+; GFX12-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
 ; GFX12-TRUE16-NEXT:    s_or_b32 s0, vcc_lo, s0
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
@@ -13766,45 +13783,46 @@ define void @global_system_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1
 ; GFX11-TRUE16-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_add_co_u32 v4, vcc_lo, 0x7fe, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 3, v4
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    global_load_b32 v3, v[0:1], off
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, -4, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 3, v3
+; GFX11-TRUE16-NEXT:    global_load_b32 v4, v[0:1], off
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 3, v3
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e64 v5, v4, 0xffff
-; GFX11-TRUE16-NEXT:    v_not_b32_e32 v5, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-TRUE16-NEXT:    v_not_b32_e32 v6, v3
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB41_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v3, v3, v7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v7.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
+; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v3, 0x7fff
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_sub_f32_e32 v2, v2, v6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v2.h
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v3.h
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, v4, v7
-; GFX11-TRUE16-NEXT:    v_and_or_b32 v2, v3, v5, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, v5, v7
+; GFX11-TRUE16-NEXT:    v_and_or_b32 v3, v4, v6, v3
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    buffer_gl1_inv
 ; GFX11-TRUE16-NEXT:    buffer_gl0_inv
-; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v2
+; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, vcc_lo, s0
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0