1; RUN: llc -march=amdgcn -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GCN %s
2
3; GCN-LABEL: name:            s_ctlz_i32
4; GCN: S_FLBIT_I32_B32
5define amdgpu_kernel void @s_ctlz_i32(i32 addrspace(1)* noalias %out, i32 %val) nounwind {
6  %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone
7  store i32 %ctlz, i32 addrspace(1)* %out, align 4
8  ret void
9}
10; GCN-LABEL: name:            v_ctlz_i32
11; GCN: V_FFBH_U32_e64
12define amdgpu_kernel void @v_ctlz_i32(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind {
13  %tid = call i32 @llvm.amdgcn.workitem.id.x()
14  %in.gep = getelementptr i32, i32 addrspace(1)* %valptr, i32 %tid
15  %val = load i32, i32 addrspace(1)* %in.gep, align 4
16  %ctlz = call i32 @llvm.ctlz.i32(i32 %val, i1 false) nounwind readnone
17  store i32 %ctlz, i32 addrspace(1)* %out, align 4
18  ret void
19}
20
21; GCN-LABEL: name:            s_cttz_i32
22; GCN: S_FF1_I32_B32
23define amdgpu_kernel void @s_cttz_i32(i32 addrspace(1)* noalias %out, i32 %val) nounwind {
24  %cttz = call i32 @llvm.cttz.i32(i32 %val, i1 false) nounwind readnone
25  store i32 %cttz, i32 addrspace(1)* %out, align 4
26  ret void
27}
28
29; GCN-LABEL: name:            v_cttz_i32
30; GCN: V_FFBL_B32_e64
31define amdgpu_kernel void @v_cttz_i32(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) nounwind {
32  %tid = call i32 @llvm.amdgcn.workitem.id.x()
33  %in.gep = getelementptr i32, i32 addrspace(1)* %valptr, i32 %tid
34  %val = load i32, i32 addrspace(1)* %in.gep, align 4
35  %cttz = call i32 @llvm.cttz.i32(i32 %val, i1 false) nounwind readnone
36  store i32 %cttz, i32 addrspace(1)* %out, align 4
37  ret void
38}
39
40; GCN-LABEL: name:            s_flbit
41; GCN: S_FLBIT_I32
42define amdgpu_kernel void @s_flbit(i32 addrspace(1)* noalias %out, i32 %val) #0 {
43  %r = call i32 @llvm.amdgcn.sffbh.i32(i32 %val)
44  store i32 %r, i32 addrspace(1)* %out, align 4
45  ret void
46}
47
48; GCN-LABEL: name:            v_flbit
49; GCN: V_FFBH_I32_e64
50define amdgpu_kernel void @v_flbit(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %valptr) #0 {
51  %tid = call i32 @llvm.amdgcn.workitem.id.x()
52  %in.gep = getelementptr i32, i32 addrspace(1)* %valptr, i32 %tid
53  %val = load i32, i32 addrspace(1)* %in.gep, align 4
54  %r = call i32 @llvm.amdgcn.sffbh.i32(i32 %val)
55  store i32 %r, i32 addrspace(1)* %out, align 4
56  ret void
57}
58
59
60declare i32 @llvm.ctlz.i32(i32, i1) nounwind readnone
61declare i32 @llvm.cttz.i32(i32, i1) nounwind readnone
62declare i32 @llvm.amdgcn.sffbh.i32(i32)
63declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
64
65