63 lines · plain
1; RUN: opt -S -mtriple=amdgcn-- -data-layout=A5 -passes='amdgpu-promote-alloca,sroa,instcombine' < %s | FileCheck -check-prefix=OPT %s2 3; Show that what the alloca promotion pass will do for non-atomic load/store.4 5; OPT-LABEL: @vector_alloca_not_atomic(6;7; OPT: extractelement <3 x i32> <i32 0, i32 1, i32 2>, i64 %index8define amdgpu_kernel void @vector_alloca_not_atomic(ptr addrspace(1) %out, i64 %index) {9entry:10 %alloca = alloca [3 x i32], addrspace(5)11 %a1 = getelementptr [3 x i32], ptr addrspace(5) %alloca, i32 0, i32 112 %a2 = getelementptr [3 x i32], ptr addrspace(5) %alloca, i32 0, i32 213 store i32 0, ptr addrspace(5) %alloca14 store i32 1, ptr addrspace(5) %a115 store i32 2, ptr addrspace(5) %a216 %tmp = getelementptr [3 x i32], ptr addrspace(5) %alloca, i64 0, i64 %index17 %data = load i32, ptr addrspace(5) %tmp18 store i32 %data, ptr addrspace(1) %out19 ret void20}21 22; OPT-LABEL: @vector_alloca_atomic_read(23;24; OPT: alloca [3 x i32]25; OPT: store i32 0, ptr addrspace(5)26; OPT: store i32 1, ptr addrspace(5)27; OPT: store i32 2, ptr addrspace(5)28; OPT: load atomic i32, ptr addrspace(5)29define amdgpu_kernel void @vector_alloca_atomic_read(ptr addrspace(1) %out, i64 %index) {30entry:31 %alloca = alloca [3 x i32], addrspace(5)32 %a1 = getelementptr [3 x i32], ptr addrspace(5) %alloca, i32 0, i32 133 %a2 = getelementptr [3 x i32], ptr addrspace(5) %alloca, i32 0, i32 234 store i32 0, ptr addrspace(5) %alloca35 store i32 1, ptr addrspace(5) %a136 store i32 2, ptr addrspace(5) %a237 %tmp = getelementptr [3 x i32], ptr addrspace(5) %alloca, i64 0, i64 %index38 %data = load atomic i32, ptr addrspace(5) %tmp acquire, align 439 store i32 %data, ptr addrspace(1) %out40 ret void41}42 43; OPT-LABEL: @vector_alloca_atomic_write(44;45; OPT: alloca [3 x i32]46; OPT: store atomic i32 0, ptr addrspace(5)47; OPT: store atomic i32 1, ptr addrspace(5)48; OPT: store atomic i32 2, ptr addrspace(5)49; OPT: load i32, ptr addrspace(5)50define amdgpu_kernel void @vector_alloca_atomic_write(ptr addrspace(1) %out, i64 %index) {51entry:52 %alloca = alloca [3 x i32], addrspace(5)53 %a1 = getelementptr [3 x i32], ptr addrspace(5) %alloca, i32 0, i32 154 %a2 = getelementptr [3 x i32], ptr addrspace(5) %alloca, i32 0, i32 255 store atomic i32 0, ptr addrspace(5) %alloca release, align 456 store atomic i32 1, ptr addrspace(5) %a1 release, align 457 store atomic i32 2, ptr addrspace(5) %a2 release, align 458 %tmp = getelementptr [3 x i32], ptr addrspace(5) %alloca, i64 0, i64 %index59 %data = load i32, ptr addrspace(5) %tmp60 store i32 %data, ptr addrspace(1) %out61 ret void62}63