1 files changed, 52 insertions, 88 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
index b8762d13e132..f0e16150c9e7 100644
--- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
@@ -16350,11 +16350,10 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16(ptr addrspace(1) %ptr,
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v3.h
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-TRUE16-NEXT:    v_bfi_b32 v5, 0xffff, v3, v5
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
 ; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
@@ -16472,7 +16471,6 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16(ptr addrspace(1) %ptr,
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB50_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -16495,10 +16493,9 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16(ptr addrspace(1) %ptr,
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v3, 0x7fff
 ; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v5, 0xffff, v3, v5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
@@ -16510,7 +16507,6 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16(ptr addrspace(1) %ptr,
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB50_1
 ; GFX11-TRUE16-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x2
 ; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v3
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -16854,11 +16850,10 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_pos(ptr addr
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v3.h
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-TRUE16-NEXT:    v_bfi_b32 v5, 0xffff, v3, v5
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
 ; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
@@ -16976,7 +16971,6 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_pos(ptr addr
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB51_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -16999,10 +16993,9 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_pos(ptr addr
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v3, 0x7fff
 ; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v5, 0xffff, v3, v5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
@@ -17014,7 +17007,6 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_pos(ptr addr
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB51_1
 ; GFX11-TRUE16-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x2
 ; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v3
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -17360,11 +17352,10 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_neg(ptr addr
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v3.h
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-TRUE16-NEXT:    v_bfi_b32 v5, 0xffff, v3, v5
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
 ; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
@@ -17482,7 +17473,6 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_neg(ptr addr
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB52_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -17505,10 +17495,9 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_neg(ptr addr
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v3, 0x7fff
 ; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v5, 0xffff, v3, v5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
@@ -17520,7 +17509,6 @@ define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_neg(ptr addr
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB52_1
 ; GFX11-TRUE16-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x2
 ; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v3
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -17865,14 +17853,12 @@ define void @global_agent_atomic_fsub_noret_v2bf16(ptr addrspace(1) %ptr, <2 x b
 ; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
 ; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v2.h
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_bfi_b32 v2, 0xffff, v2, v6
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v8, v10, vcc_lo
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v7.h
 ; 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:    s_wait_loadcnt 0x0
@@ -17988,7 +17974,6 @@ define void @global_agent_atomic_fsub_noret_v2bf16(ptr addrspace(1) %ptr, <2 x b
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB53_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -18006,13 +17991,12 @@ define void @global_agent_atomic_fsub_noret_v2bf16(ptr addrspace(1) %ptr, <2 x b
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v2, 0xffff, v2, v6
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v8, v10, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v7.h
 ; 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:    s_waitcnt vmcnt(0)
@@ -18025,7 +18009,6 @@ define void @global_agent_atomic_fsub_noret_v2bf16(ptr addrspace(1) %ptr, <2 x b
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB53_1
 ; GFX11-TRUE16-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x2
 ; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -18353,14 +18336,12 @@ define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace(
 ; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
 ; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v2.h
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_bfi_b32 v2, 0xffff, v2, v6
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v8, v10, vcc_lo
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v7.h
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
 ; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
@@ -18476,7 +18457,6 @@ define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace(
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB54_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -18494,13 +18474,12 @@ define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace(
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v2, 0xffff, v2, v6
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v8, v10, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v7.h
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
@@ -18513,7 +18492,6 @@ define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace(
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB54_1
 ; GFX11-TRUE16-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x2
 ; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -18844,14 +18822,12 @@ define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_neg(ptr addrspace(
 ; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
 ; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v2.h
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_bfi_b32 v2, 0xffff, v2, v6
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v8, v10, vcc_lo
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v7.h
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
 ; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-TRUE16-NEXT:    s_wait_loadcnt 0x0
@@ -18967,7 +18943,6 @@ define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_neg(ptr addrspace(
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB55_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -18985,13 +18960,12 @@ define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_neg(ptr addrspace(
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v2, 0xffff, v2, v6
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v8, v10, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v7.h
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
@@ -19004,7 +18978,6 @@ define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_neg(ptr addrspace(
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB55_1
 ; GFX11-TRUE16-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x2
 ; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -19348,11 +19321,10 @@ define <2 x bfloat> @global_system_atomic_fsub_ret_v2bf16__offset12b_pos(ptr add
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v3.h
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
 ; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-TRUE16-NEXT:    v_bfi_b32 v5, 0xffff, v3, v5
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
 ; GFX12-TRUE16-NEXT:    global_wb scope:SCOPE_SYS
 ; GFX12-TRUE16-NEXT:    s_wait_storecnt 0x0
 ; GFX12-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN scope:SCOPE_SYS
@@ -19472,7 +19444,6 @@ define <2 x bfloat> @global_system_atomic_fsub_ret_v2bf16__offset12b_pos(ptr add
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB56_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -19495,10 +19466,9 @@ define <2 x bfloat> @global_system_atomic_fsub_ret_v2bf16__offset12b_pos(ptr add
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v3, 0x7fff
 ; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v7, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v5, 0xffff, v3, v5
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v3.h
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
@@ -19510,7 +19480,6 @@ define <2 x bfloat> @global_system_atomic_fsub_ret_v2bf16__offset12b_pos(ptr add
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB56_1
 ; GFX11-TRUE16-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x2
 ; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v3
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -19853,14 +19822,12 @@ define void @global_system_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace
 ; GFX12-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
 ; GFX12-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v9, vcc_lo
 ; GFX12-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v2.h
 ; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffd
-; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_bfi_b32 v2, 0xffff, v2, v6
+; GFX12-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v8, v10, vcc_lo
+; GFX12-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v7.h
 ; 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 offset:2044 th:TH_ATOMIC_RETURN scope:SCOPE_SYS
@@ -19978,7 +19945,6 @@ define void @global_system_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x1
 ; GFX11-TRUE16-NEXT:    .p2align 6
 ; GFX11-TRUE16-NEXT:  .LBB57_1: ; %atomicrmw.start
 ; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
@@ -19996,13 +19962,12 @@ define void @global_system_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v2, 0x7fff
 ; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v9, vcc_lo
 ; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v2, 0xffff, v2, v6
+; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v8, v10, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v7.h
 ; GFX11-TRUE16-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-TRUE16-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
@@ -20015,7 +19980,6 @@ define void @global_system_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_cbranch_execnz .LBB57_1
 ; GFX11-TRUE16-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-TRUE16-NEXT:    s_set_inst_prefetch_distance 0x2
 ; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;