| Changpeng Fang | 24f035a | 2016-03-01 17:51:23 +0000 | [diff] [blame] | 1 | ; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck %s |
| 2 | |
| 3 | declare 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]+}} |
| Matt Arsenault | 3dbeefa | 2017-03-21 21:39:51 +0000 | [diff] [blame] | 7 | define amdgpu_kernel void @ds_bpermute(i32 addrspace(1)* %out, i32 %index, i32 %src) nounwind { |
| Changpeng Fang | 24f035a | 2016-03-01 17:51:23 +0000 | [diff] [blame] | 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 | |
| Tom Stellard | 92b24f3 | 2016-04-29 14:34:26 +0000 | [diff] [blame] | 13 | ; CHECK-LABEL: {{^}}ds_bpermute_imm_offset: |
| 14 | ; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:4 |
| Matt Arsenault | 3dbeefa | 2017-03-21 21:39:51 +0000 | [diff] [blame] | 15 | define amdgpu_kernel void @ds_bpermute_imm_offset(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind { |
| Tom Stellard | 92b24f3 | 2016-04-29 14:34:26 +0000 | [diff] [blame] | 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 | |
| Tom Stellard | 26a2ab7 | 2016-06-10 00:01:04 +0000 | [diff] [blame] | 22 | ; CHECK-LABEL: {{^}}ds_bpermute_imm_index: |
| 23 | ; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:64 |
| Matt Arsenault | 3dbeefa | 2017-03-21 21:39:51 +0000 | [diff] [blame] | 24 | define amdgpu_kernel void @ds_bpermute_imm_index(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind { |
| Tom Stellard | 26a2ab7 | 2016-06-10 00:01:04 +0000 | [diff] [blame] | 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 | |
| Simon Pilgrim | 8bd2d87 | 2017-09-14 10:38:30 +0000 | [diff] [blame] | 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 |
| 33 | define 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 |
| 44 | define 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 | |
| Changpeng Fang | 24f035a | 2016-03-01 17:51:23 +0000 | [diff] [blame] | 53 | attributes #0 = { nounwind readnone convergent } |