| ; RUN: opt -S -mtriple=amdgpu-amd-amdhsa --amdgpu-annotate-uniform < %s | FileCheck -check-prefix=OPT %s | |
| ; RUN: opt -S -mtriple=amdgpu-amd-amdhsa -passes=amdgpu-annotate-uniform < %s | FileCheck -check-prefix=OPT %s | |
| ; OPT-LABEL: @amdgpu_noclobber_global( | |
| ; OPT-NEXT: %load = load i32, ptr addrspace(1) %in, align 4, !amdgpu.noclobber !0 | |
| define amdgpu_kernel void @amdgpu_noclobber_global( ptr addrspace(1) %in, ptr addrspace(1) %out) { | |
| %load = load i32, ptr addrspace(1) %in, align 4 | |
| store i32 %load, ptr addrspace(1) %out, align 4 | |
| ret void | |
| } | |
| ; OPT-LABEL: @amdgpu_noclobber_local( | |
| ; OPT-NEXT: %load = load i32, ptr addrspace(3) %in, align 4 | |
| define amdgpu_kernel void @amdgpu_noclobber_local( ptr addrspace(3) %in, ptr addrspace(1) %out) { | |
| %load = load i32, ptr addrspace(3) %in, align 4 | |
| store i32 %load, ptr addrspace(1) %out, align 4 | |
| ret void | |
| } | |
| ; OPT-LABEL: @amdgpu_noclobber_private( | |
| ; OPT-NEXT: %load = load i32, ptr addrspace(5) %in, align 4 | |
| define amdgpu_kernel void @amdgpu_noclobber_private( ptr addrspace(5) %in, ptr addrspace(1) %out) { | |
| %load = load i32, ptr addrspace(5) %in, align 4 | |
| store i32 %load, ptr addrspace(1) %out, align 4 | |
| ret void | |
| } | |
| ; OPT-LABEL: @amdgpu_noclobber_flat( | |
| ; OPT-NEXT: %load = load i32, ptr addrspace(4) %in, align 4 | |
| define amdgpu_kernel void @amdgpu_noclobber_flat( ptr addrspace(4) %in, ptr addrspace(1) %out) { | |
| %load = load i32, ptr addrspace(4) %in, align 4 | |
| store i32 %load, ptr addrspace(1) %out, align 4 | |
| ret void | |
| } | |
| ; A prefetch does not clobber the memory it prefetches. | |
| ; OPT-LABEL: @amdgpu_noclobber_global_after_prefetch( | |
| ; OPT-NEXT: call void @llvm.amdgcn.global.prefetch(ptr addrspace(1) %in, i32 8) | |
| ; OPT-NEXT: %load = load i32, ptr addrspace(1) %in, align 4, !amdgpu.noclobber !0 | |
| define amdgpu_kernel void @amdgpu_noclobber_global_after_prefetch(ptr addrspace(1) %in, ptr addrspace(1) %out) { | |
| call void @llvm.amdgcn.global.prefetch(ptr addrspace(1) %in, i32 8) | |
| %load = load i32, ptr addrspace(1) %in, align 4 | |
| store i32 %load, ptr addrspace(1) %out, align 4 | |
| ret void | |
| } |