From 161e8c39af4b7ff6ddce76ffe7aa78fc1f35583f Mon Sep 17 00:00:00 2001 From: Changpeng Fang Date: Fri, 12 May 2017 20:31:12 +0000 Subject: [PATCH] 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 llvm-svn: 302943 --- llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp | 7 +++++-- llvm/test/CodeGen/AMDGPU/promote-alloca-volatile.ll | 12 ++++++------ 2 files changed, 11 insertions(+), 8 deletions(-) diff --git a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp index 36dcc69..e40f395 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp @@ -397,14 +397,17 @@ static Value* GEPToVectorIndex(GetElementPtrInst *GEP) { // instructions. static bool canVectorizeInst(Instruction *Inst, User *User) { switch (Inst->getOpcode()) { - case Instruction::Load: + case Instruction::Load: { + LoadInst *LI = cast(Inst); + return !LI->isVolatile(); + } case Instruction::BitCast: case Instruction::AddrSpaceCast: return true; case Instruction::Store: { // Must be the stored pointer operand, not a stored value. StoreInst *SI = cast(Inst); - return SI->getPointerOperand() == User; + return (SI->getPointerOperand() == User) && !SI->isVolatile(); } default: return false; diff --git a/llvm/test/CodeGen/AMDGPU/promote-alloca-volatile.ll b/llvm/test/CodeGen/AMDGPU/promote-alloca-volatile.ll index 9c43a6d..d765599 100644 --- a/llvm/test/CodeGen/AMDGPU/promote-alloca-volatile.ll +++ b/llvm/test/CodeGen/AMDGPU/promote-alloca-volatile.ll @@ -1,26 +1,26 @@ ; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -amdgpu-promote-alloca < %s | FileCheck %s ; CHECK-LABEL: @volatile_load( -; CHECK: alloca [5 x i32] +; 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 [5 x i32], align 4 + %stack = alloca [4 x i32], align 4 %tmp = load i32, i32 addrspace(1)* %in, align 4 - %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 %tmp + %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 [5 x i32] +; 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 [5 x i32], align 4 + %stack = alloca [4 x i32], align 4 %tmp = load i32, i32 addrspace(1)* %in, align 4 - %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 %tmp + %arrayidx1 = getelementptr inbounds [4 x i32], [4 x i32]* %stack, i32 0, i32 %tmp store volatile i32 %tmp, i32* %arrayidx1 ret void } -- 2.7.4