1 # RUN: llc -run-pass si-insert-waitcnts -march=amdgcn -mcpu=tahiti -o - %s | FileCheck %s
4 define amdgpu_kernel void @vccz_corrupt_workaround(float %cond, i32 addrspace(1)* %out) #0 {
6 %cmp0 = fcmp oeq float %cond, 0.000000e+00
7 br i1 %cmp0, label %if, label %else, !structurizecfg.uniform !0, !amdgpu.uniform !0
10 store volatile i32 100, i32 addrspace(1)* undef
11 br label %done, !structurizecfg.uniform !0
14 store volatile i32 9, i32 addrspace(1)* undef
15 br label %done, !structurizecfg.uniform !0
17 done: ; preds = %if, %else
18 %value = phi i32 [ 0, %if ], [ 1, %else ]
19 store i32 %value, i32 addrspace(1)* %out
23 define amdgpu_kernel void @vccz_corrupt_undef_vcc(float %cond, i32 addrspace(1)* %out) #0 {
25 br i1 undef, label %if, label %else, !structurizecfg.uniform !0, !amdgpu.uniform !0
27 else: ; preds = %entry
28 store volatile i32 100, i32 addrspace(1)* undef
29 br label %done, !structurizecfg.uniform !0
32 store volatile i32 9, i32 addrspace(1)* undef
33 br label %done, !structurizecfg.uniform !0
35 done: ; preds = %if, %else
36 %value = phi i32 [ 0, %if ], [ 1, %else ]
37 store i32 %value, i32 addrspace(1)* %out
41 attributes #0 = { nounwind }
42 attributes #1 = { readnone }
48 # CHECK-LABEL: name: vccz_corrupt_workaround
49 # CHECK: $vcc = V_CMP_EQ_F32
50 # CHECK-NEXT: S_WAITCNT 127
51 # CHECK-NEXT: $vcc = S_MOV_B64 $vcc
52 # CHECK-NEXT: S_CBRANCH_VCCZ %bb.2, implicit killed $vcc
54 name: vccz_corrupt_workaround
56 exposesReturnsTwice: false
58 regBankSelected: false
60 tracksRegLiveness: true
62 - { reg: '$sgpr0_sgpr1' }
64 isFrameAddressTaken: false
65 isReturnAddressTaken: false
74 hasOpaqueSPAdjustment: false
76 hasMustTailInVarArgFunc: false
79 liveins: $sgpr0_sgpr1, $vcc
81 $sgpr0_sgpr1 = S_LOAD_DWORDX2_IMM killed $sgpr0_sgpr1, 11, 0, 0 :: (non-temporal dereferenceable invariant load 8 from `i64 addrspace(4)* undef`)
82 $sgpr7 = S_MOV_B32 61440
84 $vcc = V_CMP_EQ_F32_e64 0, 0, 0, undef $sgpr2, 0, implicit $exec
85 S_CBRANCH_VCCZ %bb.1, implicit killed $vcc
88 liveins: $sgpr6, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr3:0x00000003
90 $vgpr0 = V_MOV_B32_e32 9, implicit $exec
91 BUFFER_STORE_DWORD_OFFSET killed $vgpr0, killed $sgpr4_sgpr5_sgpr6_sgpr7, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `i32 addrspace(1)* undef`)
92 $vgpr0 = V_MOV_B32_e32 0, implicit $exec
96 liveins: $sgpr6, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr3:0x00000003
98 $vgpr0 = V_MOV_B32_e32 100, implicit $exec
99 BUFFER_STORE_DWORD_OFFSET killed $vgpr0, killed $sgpr4_sgpr5_sgpr6_sgpr7, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `i32 addrspace(1)* undef`)
100 $vgpr0 = V_MOV_B32_e32 1, implicit $exec
103 liveins: $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3:0x00000003
105 $sgpr3 = S_MOV_B32 61440
106 $sgpr2 = S_MOV_B32 -1
107 BUFFER_STORE_DWORD_OFFSET killed $vgpr0, killed $sgpr0_sgpr1_sgpr2_sgpr3, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (store 4 into %ir.out)
112 # CHECK-LABEL: name: vccz_corrupt_undef_vcc
113 # CHECK: S_WAITCNT 3855
114 # CHECK-NEXT: $vgpr0 = V_MOV_B32_e32
116 name: vccz_corrupt_undef_vcc
118 exposesReturnsTwice: false
120 regBankSelected: false
122 tracksRegLiveness: true
124 - { reg: '$sgpr0_sgpr1' }
126 isFrameAddressTaken: false
127 isReturnAddressTaken: false
136 hasOpaqueSPAdjustment: false
138 hasMustTailInVarArgFunc: false
141 liveins: $sgpr0_sgpr1
143 $sgpr0_sgpr1 = S_LOAD_DWORDX2_IMM killed $sgpr0_sgpr1, 11, 0, 0 :: (non-temporal dereferenceable invariant load 8 from `i64 addrspace(4)* undef`)
144 $sgpr7 = S_MOV_B32 61440
145 $sgpr6 = S_MOV_B32 -1
146 S_CBRANCH_VCCZ %bb.1, implicit undef $vcc
149 liveins: $sgpr6, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr3:0x00000003
151 $vgpr0 = V_MOV_B32_e32 9, implicit $exec
152 BUFFER_STORE_DWORD_OFFSET killed $vgpr0, killed $sgpr4_sgpr5_sgpr6_sgpr7, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `i32 addrspace(1)* undef`)
153 $vgpr0 = V_MOV_B32_e32 0, implicit $exec
157 liveins: $sgpr6, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr3:0x00000003
159 $vgpr0 = V_MOV_B32_e32 100, implicit $exec
160 BUFFER_STORE_DWORD_OFFSET killed $vgpr0, killed $sgpr4_sgpr5_sgpr6_sgpr7, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `i32 addrspace(1)* undef`)
161 $vgpr0 = V_MOV_B32_e32 1, implicit $exec
164 liveins: $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3:0x00000003
166 $sgpr3 = S_MOV_B32 61440
167 $sgpr2 = S_MOV_B32 -1
168 BUFFER_STORE_DWORD_OFFSET killed $vgpr0, killed $sgpr0_sgpr1_sgpr2_sgpr3, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (store 4 into %ir.out)