1; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck %s
2
3declare i32 @llvm.amdgcn.ds.bpermute(i32, i32) #0
4
5; FUNC-LABEL: {{^}}ds_bpermute:
6; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
7define amdgpu_kernel void @ds_bpermute(i32 addrspace(1)* %out, i32 %index, i32 %src) nounwind {
8  %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %index, i32 %src) #0
9  store i32 %bpermute, i32 addrspace(1)* %out, align 4
10  ret void
11}
12
13; CHECK-LABEL: {{^}}ds_bpermute_imm_offset:
14; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:4
15define amdgpu_kernel void @ds_bpermute_imm_offset(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind {
16  %index = add i32 %base_index, 4
17  %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %index, i32 %src) #0
18  store i32 %bpermute, i32 addrspace(1)* %out, align 4
19  ret void
20}
21
22; CHECK-LABEL: {{^}}ds_bpermute_imm_index:
23; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:64
24define amdgpu_kernel void @ds_bpermute_imm_index(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind {
25  %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 64, i32 %src) #0
26  store i32 %bpermute, i32 addrspace(1)* %out, align 4
27  ret void
28}
29
30; CHECK-LABEL: {{^}}ds_bpermute_add_shl:
31; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:4
32; CHECK: s_waitcnt lgkmcnt
33define void @ds_bpermute_add_shl(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind {
34  %index = add i32 %base_index, 1
35  %byte_index = shl i32 %index, 2
36  %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %byte_index, i32 %src) #0
37  store i32 %bpermute, i32 addrspace(1)* %out, align 4
38  ret void
39}
40
41; CHECK-LABEL: {{^}}ds_bpermute_or_shl:
42; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:4
43; CHECK: s_waitcnt lgkmcnt
44define void @ds_bpermute_or_shl(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind {
45  %masked = and i32 %base_index, 62
46  %index = or i32 %masked, 1
47  %byte_index = shl i32 %index, 2
48  %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %byte_index, i32 %src) #0
49  store i32 %bpermute, i32 addrspace(1)* %out, align 4
50  ret void
51}
52
53attributes #0 = { nounwind readnone convergent }
54