85 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 52; RUN: opt < %s -passes=amdgpu-lower-intrinsics -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -codegen-opt-level=0 | FileCheck --check-prefixes=GFX11,GFX11-NOOPT %s3; RUN: opt < %s -passes=amdgpu-lower-intrinsics -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -codegen-opt-level=1 -mattr=+wavefrontsize32 | FileCheck --check-prefixes=GFX11,OPT-WAVE32,GFX11-OPT-WAVE32 %s4; RUN: opt < %s -passes=amdgpu-lower-intrinsics -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -codegen-opt-level=1 -mattr=+wavefrontsize64 | FileCheck --check-prefixes=GFX11,OPT-WAVE64,GFX11-OPT-WAVE64 %s5; RUN: opt < %s -passes=amdgpu-lower-intrinsics -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -codegen-opt-level=0 | FileCheck --check-prefixes=GFX12,GFX12-NOOPT %s6; RUN: opt < %s -passes=amdgpu-lower-intrinsics -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -codegen-opt-level=1 -mattr=+wavefrontsize32 | FileCheck --check-prefixes=GFX12,OPT-WAVE32,GFX12-OPT-WAVE32 %s7; RUN: opt < %s -passes=amdgpu-lower-intrinsics -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -codegen-opt-level=1 -mattr=+wavefrontsize64 | FileCheck --check-prefixes=GFX12,OPT-WAVE64,GFX12-OPT-WAVE64 %s8 9define amdgpu_kernel void @barrier() {10; GFX11-LABEL: define amdgpu_kernel void @barrier(11; GFX11-SAME: ) #[[ATTR0:[0-9]+]] {12; GFX11-NEXT: call void @llvm.amdgcn.s.barrier()13; GFX11-NEXT: ret void14;15; GFX12-LABEL: define amdgpu_kernel void @barrier(16; GFX12-SAME: ) #[[ATTR0:[0-9]+]] {17; GFX12-NEXT: call void @llvm.amdgcn.s.barrier.signal(i32 -1)18; GFX12-NEXT: call void @llvm.amdgcn.s.barrier.wait(i16 -1)19; GFX12-NEXT: ret void20;21 call void @llvm.amdgcn.s.barrier()22 ret void23}24 25define amdgpu_kernel void @barrier_32threads() "amdgpu-flat-work-group-size"="32,32" {26; GFX11-NOOPT-LABEL: define amdgpu_kernel void @barrier_32threads(27; GFX11-NOOPT-SAME: ) #[[ATTR1:[0-9]+]] {28; GFX11-NOOPT-NEXT: call void @llvm.amdgcn.s.barrier()29; GFX11-NOOPT-NEXT: ret void30;31; OPT-WAVE32-LABEL: define amdgpu_kernel void @barrier_32threads(32; OPT-WAVE32-SAME: ) #[[ATTR1:[0-9]+]] {33; OPT-WAVE32-NEXT: call void @llvm.amdgcn.wave.barrier()34; OPT-WAVE32-NEXT: ret void35;36; OPT-WAVE64-LABEL: define amdgpu_kernel void @barrier_32threads(37; OPT-WAVE64-SAME: ) #[[ATTR1:[0-9]+]] {38; OPT-WAVE64-NEXT: call void @llvm.amdgcn.wave.barrier()39; OPT-WAVE64-NEXT: ret void40;41; GFX12-NOOPT-LABEL: define amdgpu_kernel void @barrier_32threads(42; GFX12-NOOPT-SAME: ) #[[ATTR1:[0-9]+]] {43; GFX12-NOOPT-NEXT: call void @llvm.amdgcn.s.barrier.signal(i32 -1)44; GFX12-NOOPT-NEXT: call void @llvm.amdgcn.s.barrier.wait(i16 -1)45; GFX12-NOOPT-NEXT: ret void46;47 call void @llvm.amdgcn.s.barrier()48 ret void49}50 51define amdgpu_kernel void @barrier_64threads() "amdgpu-flat-work-group-size"="64,64" {52; GFX11-NOOPT-LABEL: define amdgpu_kernel void @barrier_64threads(53; GFX11-NOOPT-SAME: ) #[[ATTR2:[0-9]+]] {54; GFX11-NOOPT-NEXT: call void @llvm.amdgcn.s.barrier()55; GFX11-NOOPT-NEXT: ret void56;57; GFX11-OPT-WAVE32-LABEL: define amdgpu_kernel void @barrier_64threads(58; GFX11-OPT-WAVE32-SAME: ) #[[ATTR2:[0-9]+]] {59; GFX11-OPT-WAVE32-NEXT: call void @llvm.amdgcn.s.barrier()60; GFX11-OPT-WAVE32-NEXT: ret void61;62; OPT-WAVE64-LABEL: define amdgpu_kernel void @barrier_64threads(63; OPT-WAVE64-SAME: ) #[[ATTR2:[0-9]+]] {64; OPT-WAVE64-NEXT: call void @llvm.amdgcn.wave.barrier()65; OPT-WAVE64-NEXT: ret void66;67; GFX12-NOOPT-LABEL: define amdgpu_kernel void @barrier_64threads(68; GFX12-NOOPT-SAME: ) #[[ATTR2:[0-9]+]] {69; GFX12-NOOPT-NEXT: call void @llvm.amdgcn.s.barrier.signal(i32 -1)70; GFX12-NOOPT-NEXT: call void @llvm.amdgcn.s.barrier.wait(i16 -1)71; GFX12-NOOPT-NEXT: ret void72;73; GFX12-OPT-WAVE32-LABEL: define amdgpu_kernel void @barrier_64threads(74; GFX12-OPT-WAVE32-SAME: ) #[[ATTR2:[0-9]+]] {75; GFX12-OPT-WAVE32-NEXT: call void @llvm.amdgcn.s.barrier.signal(i32 -1)76; GFX12-OPT-WAVE32-NEXT: call void @llvm.amdgcn.s.barrier.wait(i16 -1)77; GFX12-OPT-WAVE32-NEXT: ret void78;79 call void @llvm.amdgcn.s.barrier()80 ret void81}82;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:83; GFX11-OPT-WAVE64: {{.*}}84; GFX12-OPT-WAVE64: {{.*}}85