54 lines · plain
1; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji < %s | FileCheck %s2 3declare i32 @llvm.amdgcn.ds.bpermute(i32, i32) #04 5; CHECK-LABEL: {{^}}ds_bpermute:6; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}7define amdgpu_kernel void @ds_bpermute(ptr addrspace(1) %out, i32 %index, i32 %src) nounwind {8 %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %index, i32 %src) #09 store i32 %bpermute, ptr addrspace(1) %out, align 410 ret void11}12 13; CHECK-LABEL: {{^}}ds_bpermute_imm_offset:14; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:415define amdgpu_kernel void @ds_bpermute_imm_offset(ptr addrspace(1) %out, i32 %base_index, i32 %src) nounwind {16 %index = add i32 %base_index, 417 %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %index, i32 %src) #018 store i32 %bpermute, ptr addrspace(1) %out, align 419 ret void20}21 22; CHECK-LABEL: {{^}}ds_bpermute_imm_index:23; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:6424define amdgpu_kernel void @ds_bpermute_imm_index(ptr addrspace(1) %out, i32 %base_index, i32 %src) nounwind {25 %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 64, i32 %src) #026 store i32 %bpermute, ptr addrspace(1) %out, align 427 ret void28}29 30; CHECK-LABEL: {{^}}ds_bpermute_add_shl:31; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:432; CHECK: s_waitcnt lgkmcnt33define void @ds_bpermute_add_shl(ptr addrspace(1) %out, i32 %base_index, i32 %src) nounwind {34 %index = add i32 %base_index, 135 %byte_index = shl i32 %index, 236 %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %byte_index, i32 %src) #037 store i32 %bpermute, ptr addrspace(1) %out, align 438 ret void39}40 41; CHECK-LABEL: {{^}}ds_bpermute_or_shl:42; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:443; CHECK: s_waitcnt lgkmcnt44define void @ds_bpermute_or_shl(ptr addrspace(1) %out, i32 %base_index, i32 %src) nounwind {45 %masked = and i32 %base_index, 6246 %index = or i32 %masked, 147 %byte_index = shl i32 %index, 248 %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %byte_index, i32 %src) #049 store i32 %bpermute, ptr addrspace(1) %out, align 450 ret void51}52 53attributes #0 = { nounwind readnone convergent }54