Index: llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp =================================================================== --- llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp +++ llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp @@ -2591,6 +2591,14 @@ return DAG.getMergeValues(Ops, DL); } + MachineFunction &MF = DAG.getMachineFunction(); + SIMachineFunctionInfo *MFI = MF.getInfo(); + // If there is a possibilty that flat instruction access scratch memory + // then we need to use the same legalization rules we use for private. + if (AS == AMDGPUAS::FLAT_ADDRESS) + AS = MFI->hasFlatScratchInit() ? + AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS; + unsigned NumElements = MemVT.getVectorNumElements(); switch (AS) { case AMDGPUAS::CONSTANT_ADDRESS: @@ -2890,6 +2898,14 @@ return expandUnalignedStore(Store, DAG); } + MachineFunction &MF = DAG.getMachineFunction(); + SIMachineFunctionInfo *MFI = MF.getInfo(); + // If there is a possibilty that flat instruction access scratch memory + // then we need to use the same legalization rules we use for private. + if (AS == AMDGPUAS::FLAT_ADDRESS) + AS = MFI->hasFlatScratchInit() ? + AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS; + unsigned NumElements = VT.getVectorNumElements(); switch (AS) { case AMDGPUAS::GLOBAL_ADDRESS: Index: llvm/trunk/test/CodeGen/AMDGPU/flat-address-space.ll =================================================================== --- llvm/trunk/test/CodeGen/AMDGPU/flat-address-space.ll +++ llvm/trunk/test/CodeGen/AMDGPU/flat-address-space.ll @@ -1,5 +1,6 @@ -; RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire < %s | FileCheck %s -; RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga < %s | FileCheck %s +; RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire < %s | FileCheck %s +; RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga < %s | FileCheck %s +; RUN: llc -O0 -mtriple=amdgcn-amd-amdhsa -mcpu=fiji < %s | FileCheck -check-prefixes=CHECK,HSA %s ; Disable optimizations in case there are optimizations added that ; specialize away generic pointer accesses. @@ -149,6 +150,28 @@ ret void } +; CHECK-LABEL: flat_scratch_multidword_load: +; HSA: flat_load_dword +; HSA: flat_load_dword +; FIXME: These tests are broken for os = mesa3d, becasue it doesn't initialize flat_scr +define void @flat_scratch_multidword_load() { + %scratch = alloca <2 x i32> + %fptr = addrspacecast <2 x i32>* %scratch to <2 x i32> addrspace(4)* + %ld = load volatile <2 x i32>, <2 x i32> addrspace(4)* %fptr + ret void +} + +; CHECK-LABEL: flat_scratch_multidword_store: +; HSA: flat_store_dword +; HSA: flat_store_dword +; FIXME: These tests are broken for os = mesa3d, becasue it doesn't initialize flat_scr +define void @flat_scratch_multidword_store() { + %scratch = alloca <2 x i32> + %fptr = addrspacecast <2 x i32>* %scratch to <2 x i32> addrspace(4)* + store volatile <2 x i32> zeroinitializer, <2 x i32> addrspace(4)* %fptr + ret void +} + attributes #0 = { nounwind } attributes #1 = { nounwind convergent } attributes #3 = { nounwind readnone }