1 ; RUN: opt -S --amdgpu-annotate-uniform < %s | FileCheck -check-prefix=OPT %s
2 target datalayout = "A5"
4 ; "load vaddr" depends on the store, so we should not mark vaddr as amdgpu.noclobber.
6 ; OPT-LABEL: @store_clobbers_load(
7 ; OPT: %zero = load <4 x i32>, ptr addrspace(1) %input, align 16{{$}}
8 define amdgpu_kernel void @store_clobbers_load(ptr addrspace(1) %input, ptr addrspace(1) %out, i32 %index) {
10 store i32 0, ptr addrspace(1) %input
11 %zero = load <4 x i32>, ptr addrspace(1) %input, align 16
12 %one = insertelement <4 x i32> %zero, i32 1, i32 1
13 %two = insertelement <4 x i32> %one, i32 2, i32 2
14 %three = insertelement <4 x i32> %two, i32 3, i32 3
15 store <4 x i32> %three, ptr addrspace(1) %input, align 16
16 %rslt = extractelement <4 x i32> %three, i32 %index
17 store i32 %rslt, ptr addrspace(1) %out, align 4
22 declare i32 @llvm.amdgcn.workitem.id.x()
23 @lds0 = addrspace(3) global [512 x i32] undef, align 4
25 ; To check that %arrayidx0 is not marked as amdgpu.noclobber.
27 ; OPT-LABEL: @atomicrmw_clobbers_load(
28 ; OPT: %arrayidx0 = getelementptr inbounds [512 x i32], ptr addrspace(3) @lds0, i32 0, i32 %idx.0, !amdgpu.uniform !0
29 ; OPT-NEXT: %val = atomicrmw xchg ptr addrspace(3) %arrayidx0, i32 3 seq_cst
31 define amdgpu_kernel void @atomicrmw_clobbers_load(ptr addrspace(1) %out0, ptr addrspace(1) %out1) {
32 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
33 %idx.0 = add nsw i32 %tid.x, 2
34 %arrayidx0 = getelementptr inbounds [512 x i32], ptr addrspace(3) @lds0, i32 0, i32 %idx.0
35 %val = atomicrmw xchg ptr addrspace(3) %arrayidx0, i32 3 seq_cst
36 %load = load i32, ptr addrspace(3) %arrayidx0, align 4
37 store i32 %val, ptr addrspace(1) %out0, align 4
38 store i32 %load, ptr addrspace(1) %out1, align 4