llvm/test/CodeGen/AMDGPU/promote-alloca-volatile.ll
Changpeng Fang ed4c8077b0 AMDGPU/SI: Don't promote to vector if the load/store is volatile.
Summary:
  We should not change volatile loads/stores in promoting alloca to vector.

Reviewers:
  arsenm

Differential Revision:
  http://reviews.llvm.org/D33107

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@302943 91177308-0d34-0410-b5e6-96231b3b80d8
2017-05-12 20:31:12 +00:00

46 lines
1.5 KiB
LLVM

; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -amdgpu-promote-alloca < %s | FileCheck %s
; CHECK-LABEL: @volatile_load(
; CHECK: alloca [4 x i32]
; CHECK: load volatile i32, i32*
define amdgpu_kernel void @volatile_load(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
entry:
%stack = alloca [4 x i32], align 4
%tmp = load i32, i32 addrspace(1)* %in, align 4
%arrayidx1 = getelementptr inbounds [4 x i32], [4 x i32]* %stack, i32 0, i32 %tmp
%load = load volatile i32, i32* %arrayidx1
store i32 %load, i32 addrspace(1)* %out
ret void
}
; CHECK-LABEL: @volatile_store(
; CHECK: alloca [4 x i32]
; CHECK: store volatile i32 %tmp, i32*
define amdgpu_kernel void @volatile_store(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
entry:
%stack = alloca [4 x i32], align 4
%tmp = load i32, i32 addrspace(1)* %in, align 4
%arrayidx1 = getelementptr inbounds [4 x i32], [4 x i32]* %stack, i32 0, i32 %tmp
store volatile i32 %tmp, i32* %arrayidx1
ret void
}
; Has on OK non-volatile user but also a volatile user
; CHECK-LABEL: @volatile_and_non_volatile_load(
; CHECK: alloca double
; CHECK: load double
; CHECK: load volatile double
define amdgpu_kernel void @volatile_and_non_volatile_load(double addrspace(1)* nocapture %arg, i32 %arg1) #0 {
bb:
%tmp = alloca double, align 8
store double 0.000000e+00, double* %tmp, align 8
%tmp4 = load double, double* %tmp, align 8
%tmp5 = load volatile double, double* %tmp, align 8
store double %tmp4, double addrspace(1)* %arg
ret void
}
attributes #0 = { nounwind }