summaryrefslogtreecommitdiff
path: root/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll')
-rw-r--r--llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll62
1 files changed, 22 insertions, 40 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
index 6da80262951e..11ed43d73763 100644
--- a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
@@ -6651,13 +6651,12 @@ define <2 x bfloat> @local_atomic_fmin_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bf
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v6, v8, 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 v2.l, v2.h
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX12-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v5, v4
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE
; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
@@ -6771,7 +6770,6 @@ define <2 x bfloat> @local_atomic_fmin_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bf
; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; 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: .LBB24_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -6794,12 +6792,11 @@ define <2 x bfloat> @local_atomic_fmin_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bf
; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v6, v8, 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 v2.l, v2.h
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v5, v4
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: buffer_gl0_inv
; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
@@ -6808,7 +6805,6 @@ define <2 x bfloat> @local_atomic_fmin_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bf
; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB24_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, v2
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -7133,13 +7129,12 @@ define <2 x bfloat> @local_atomic_fmin_ret_v2bf16__offset(ptr addrspace(3) %ptr,
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v6, v8, 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 v2.l, v2.h
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX12-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v5, v4 offset:65532
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE
; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
@@ -7253,7 +7248,6 @@ define <2 x bfloat> @local_atomic_fmin_ret_v2bf16__offset(ptr addrspace(3) %ptr,
; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; 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: .LBB25_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -7276,12 +7270,11 @@ define <2 x bfloat> @local_atomic_fmin_ret_v2bf16__offset(ptr addrspace(3) %ptr,
; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v6, v8, 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 v2.l, v2.h
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v5, v4 offset:65532
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: buffer_gl0_inv
; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
@@ -7290,7 +7283,6 @@ define <2 x bfloat> @local_atomic_fmin_ret_v2bf16__offset(ptr addrspace(3) %ptr,
; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB25_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, v2
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -7615,11 +7607,9 @@ define void @local_atomic_fmin_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat>
; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v5
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v5, v3
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE
; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
@@ -7730,7 +7720,6 @@ define void @local_atomic_fmin_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat>
; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; 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: .LBB26_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -7751,11 +7740,10 @@ define void @local_atomic_fmin_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat>
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v8, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v5, v3
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: buffer_gl0_inv
; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
@@ -7765,7 +7753,6 @@ define void @local_atomic_fmin_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat>
; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB26_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]
;
@@ -8078,11 +8065,9 @@ define void @local_atomic_fmin_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b
; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd
; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v5
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v5, v3 offset:65532
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE
; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
@@ -8193,7 +8178,6 @@ define void @local_atomic_fmin_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b
; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; 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: .LBB27_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -8214,11 +8198,10 @@ define void @local_atomic_fmin_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v8, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v5, v3 offset:65532
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: buffer_gl0_inv
; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
@@ -8228,7 +8211,6 @@ define void @local_atomic_fmin_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b
; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB27_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]
;
generated by cgit v1.2.3 (git 2.25.1) at 2026-05-31 14:55:25 +0000