1 files changed, 356 insertions, 338 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
index c987effe..3ee0bb2 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
@@ -9419,7 +9419,7 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 {
 ; 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
@@ -9433,20 +9433,22 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 {
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 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)
@@ -9566,7 +9568,7 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 {
 ; GFX11-TRUE16-LABEL: flat_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
@@ -9581,19 +9583,21 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 {
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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)
@@ -9870,34 +9874,34 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat %
 ; 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:    flat_load_b32 v5, v[0:1]
 ; 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_dscnt 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)
@@ -10023,12 +10027,11 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat %
 ; 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:    flat_load_b32 v5, v[0:1]
 ; 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)
@@ -10039,19 +10042,21 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat %
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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)
@@ -10336,34 +10341,34 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr %ptr, bfloat %
 ; 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:    flat_load_b32 v5, v[0:1]
 ; 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_dscnt 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)
@@ -10490,12 +10495,11 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr %ptr, bfloat %
 ; 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:    flat_load_b32 v5, v[0:1]
 ; 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)
@@ -10506,19 +10510,21 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr %ptr, bfloat %
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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)
@@ -10800,7 +10806,7 @@ define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 {
 ; 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
@@ -10814,18 +10820,20 @@ define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 {
 ; GFX12-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 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
@@ -10942,7 +10950,7 @@ define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 {
 ; GFX11-TRUE16-LABEL: flat_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
@@ -10957,17 +10965,19 @@ define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 {
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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
@@ -11233,45 +11243,45 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr %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_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:    flat_load_b32 v3, v[0:1]
-; 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:    flat_load_b32 v4, v[0:1]
+; 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_dscnt 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:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX12-TRUE16-NEXT:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 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
@@ -11383,45 +11393,46 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr %ptr, bfloat %
 ; GFX11-TRUE16-LABEL: flat_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:    flat_load_b32 v3, v[0:1]
-; 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:    flat_load_b32 v4, v[0:1]
+; 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) lgkmcnt(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:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-TRUE16-NEXT:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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
@@ -11684,45 +11695,45 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr %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_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:    flat_load_b32 v3, v[0:1]
-; 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:    flat_load_b32 v4, v[0:1]
+; 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_dscnt 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:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX12-TRUE16-NEXT:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 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
@@ -11835,45 +11846,46 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr %ptr, bfloat %
 ; GFX11-TRUE16-LABEL: flat_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:    flat_load_b32 v3, v[0:1]
-; 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:    flat_load_b32 v4, v[0:1]
+; 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) lgkmcnt(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:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-TRUE16-NEXT:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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
@@ -12137,27 +12149,28 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr %ptr,
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-TRUE16-NEXT:    flat_load_b32 v3, v[0:1] 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_dscnt 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:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
@@ -12257,27 +12270,28 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr %ptr,
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    flat_load_b32 v3, v[0:1] 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) lgkmcnt(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:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
@@ -12507,32 +12521,33 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr %ptr,
 ; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    flat_load_b32 v3, v[0:1] offset:2046
-; GFX12-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-TRUE16-NEXT:    flat_load_b32 v4, v[0:1] 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_dscnt 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:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX12-TRUE16-NEXT:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 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
@@ -12622,33 +12637,34 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr %ptr,
 ; GFX11-TRUE16-LABEL: flat_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:    flat_load_b32 v3, v[0:1] offset:2046
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-TRUE16-NEXT:    flat_load_b32 v4, v[0:1] 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) lgkmcnt(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:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc
+; GFX11-TRUE16-NEXT:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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
@@ -12869,34 +12885,34 @@ define bfloat @flat_system_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat
 ; 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:    flat_load_b32 v5, v[0:1]
 ; 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_dscnt 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)
@@ -13024,12 +13040,11 @@ define bfloat @flat_system_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat
 ; 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:    flat_load_b32 v5, v[0:1]
 ; 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)
@@ -13040,19 +13055,21 @@ define bfloat @flat_system_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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)
@@ -13336,46 +13353,46 @@ define void @flat_system_atomic_fsub_noret_bf16__offset12b_pos(ptr %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_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:    flat_load_b32 v3, v[0:1]
-; 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:    flat_load_b32 v4, v[0:1]
+; 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_dscnt 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:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SYS
+; GFX12-TRUE16-NEXT:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_SYS
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 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
@@ -13488,45 +13505,46 @@ define void @flat_system_atomic_fsub_noret_bf16__offset12b_pos(ptr %ptr, bfloat
 ; GFX11-TRUE16-LABEL: flat_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:    flat_load_b32 v3, v[0:1]
-; 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:    flat_load_b32 v4, v[0:1]
+; 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) lgkmcnt(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:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-TRUE16-NEXT:    flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(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