1 ; RUN: opt -mtriple=amdgcn-- -S -structurizecfg -si-annotate-control-flow %s | FileCheck -check-prefix=OPT %s
2 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
5 ; OPT-LABEL: @annotate_unreachable_noloop(
6 ; OPT-NOT: call i1 @llvm.amdgcn.loop
8 ; GCN-LABEL: {{^}}annotate_unreachable_noloop:
12 define amdgpu_kernel void @annotate_unreachable_noloop(<4 x float> addrspace(1)* noalias nocapture readonly %arg) #0 {
14 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
18 %tmp2 = sext i32 %tmp to i64
19 %tmp3 = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %arg, i64 %tmp2
20 %tmp4 = load <4 x float>, <4 x float> addrspace(1)* %tmp3, align 16
21 br i1 undef, label %bb5, label %bb3
24 %tmp6 = extractelement <4 x float> %tmp4, i32 2
25 %tmp7 = fcmp olt float %tmp6, 0.000000e+00
26 br i1 %tmp7, label %bb4, label %bb5 ; crash goes away if these are swapped
31 bb5: ; preds = %bb3, %bb1
36 ; OPT-LABEL: @annotate_ret_noloop(
37 ; OPT-NOT: call i1 @llvm.amdgcn.loop
39 ; GCN-LABEL: {{^}}annotate_ret_noloop:
42 ; GCN: s_and_saveexec_b64
43 ; GCN: ; mask branch [[UNIFIED_RET:BB[0-9]+_[0-9]+]]
44 ; GCN-NEXT: [[UNIFIED_RET]]:
47 define amdgpu_kernel void @annotate_ret_noloop(<4 x float> addrspace(1)* noalias nocapture readonly %arg) #0 {
49 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
53 %tmp2 = sext i32 %tmp to i64
54 %tmp3 = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %arg, i64 %tmp2
55 %tmp4 = load <4 x float>, <4 x float> addrspace(1)* %tmp3, align 16
56 %tmp5 = extractelement <4 x float> %tmp4, i32 1
57 store volatile <4 x float> %tmp4, <4 x float> addrspace(1)* undef
58 %cmp = fcmp ogt float %tmp5, 1.0
59 br i1 %cmp, label %bb5, label %bb3
62 %tmp6 = extractelement <4 x float> %tmp4, i32 2
63 %tmp7 = fcmp olt float %tmp6, 0.000000e+00
64 br i1 %tmp7, label %bb4, label %bb5 ; crash goes away if these are swapped
69 bb5: ; preds = %bb3, %bb1
73 ; OPT-LABEL: @uniform_annotate_ret_noloop(
74 ; OPT-NOT: call i1 @llvm.amdgcn.loop
76 ; GCN-LABEL: {{^}}uniform_annotate_ret_noloop:
80 define amdgpu_kernel void @uniform_annotate_ret_noloop(<4 x float> addrspace(1)* noalias nocapture readonly %arg, i32 %tmp) #0 {
85 %tmp2 = sext i32 %tmp to i64
86 %tmp3 = getelementptr inbounds <4 x float>, <4 x float> addrspace(1)* %arg, i64 %tmp2
87 %tmp4 = load <4 x float>, <4 x float> addrspace(1)* %tmp3, align 16
88 br i1 undef, label %bb5, label %bb3
91 %tmp6 = extractelement <4 x float> %tmp4, i32 2
92 %tmp7 = fcmp olt float %tmp6, 0.000000e+00
93 br i1 %tmp7, label %bb4, label %bb5 ; crash goes away if these are swapped
98 bb5: ; preds = %bb3, %bb1
103 declare i32 @llvm.amdgcn.workitem.id.x() #1
105 attributes #0 = { nounwind }
106 attributes #1 = { nounwind readnone }