[AMDGPU] allow multi-dword flat scratch access since GFX9

[rampitec]

Summary

This is supported starting with GFX9.

[arsenm]

This also needs to be reproduce for globalisel

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

7385

Should move this into a predicate function inside the subtarget

llvm/test/CodeGen/AMDGPU/flat-address-space.ll

160

What is the alignment situation for this? I thought unaligned scratch access was introduced but we never started using it?

[rampitec]

llvm/test/CodeGen/AMDGPU/flat-address-space.ll

160

Unaligned works too. In fact unaligned works even on VI and I do not have CI to test.

[rampitec]

Added subtarget predicate.

[arsenm]

llvm/test/CodeGen/AMDGPU/flat-address-space.ll

173

Should add some underaligned cases

[rampitec]

llvm/test/CodeGen/AMDGPU/flat-address-space.ll

173

We have it in the same test: @flat_scratch_unaligned_load() and @flat_scratch_unaligned_store(). They are split.
This would be a separate change to allow unaligned flat.

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

public:
   bool hasFlatScratchInsts() const {
     return FlatScratchInsts;
   }
 
   bool hasScalarFlatScratchInsts() const {
     return ScalarFlatScratchInsts;
   }
 
+  bool hasMultiDwordFlatScratchAddressing() const {
+    return getGeneration() >= GFX9;
+  }
+
   bool hasFlatSegmentOffsetBug() const {
     return HasFlatSegmentOffsetBug;
   }
 
   bool hasFlatLgkmVMemCountInOrder() const {
     return getGeneration() > GFX9;
   }
 

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

if (Subtarget->hasLDSMisalignedBug() &&
       Alignment < MemVT.getStoreSize() && MemVT.getSizeInBits() > 32) {
     return SplitVectorLoad(Op, DAG);
   }
 
   MachineFunction &MF = DAG.getMachineFunction();
   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
   // 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)
+  if (AS == AMDGPUAS::FLAT_ADDRESS &&
+      !Subtarget->hasMultiDwordFlatScratchAddressing())
     AS = MFI->hasFlatScratchInit() ?
          AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS;
 
   unsigned NumElements = MemVT.getVectorNumElements();
 
   if (AS == AMDGPUAS::CONSTANT_ADDRESS ||
       AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT) {
     if (!Op->isDivergent() && Alignment >= 4 && NumElements < 32) {
       if (MemVT.isPow2VectorType())
if (Subtarget->hasLDSMisalignedBug() &&
       Store->getAlignment() < VT.getStoreSize() && VT.getSizeInBits() > 32) {
     return SplitVectorStore(Op, DAG);
   }
 
   MachineFunction &MF = DAG.getMachineFunction();
   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
   // 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)
+  if (AS == AMDGPUAS::FLAT_ADDRESS &&
+      !Subtarget->hasMultiDwordFlatScratchAddressing())
     AS = MFI->hasFlatScratchInit() ?
          AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS;
 
   unsigned NumElements = VT.getVectorNumElements();
   if (AS == AMDGPUAS::GLOBAL_ADDRESS ||
       AS == AMDGPUAS::FLAT_ADDRESS) {
     if (NumElements > 4)
       return SplitVectorStore(Op, DAG);

llvm/test/CodeGen/AMDGPU/flat-address-space.ll

 ; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire < %s | FileCheck -check-prefixes=CHECK,CIVI %s
 ; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,CIVI %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,CIVI,HSA %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,CIVI,HSA,CIVI-HSA %s
 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,HSA,GFX9 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,HSA,GFX10 %s
 
 ; CHECK-LABEL: {{^}}store_flat_i32:
 ; CHECK-DAG: s_load_dwordx2 s{{\[}}[[LO_SREG:[0-9]+]]:[[HI_SREG:[0-9]+]]],
 ; CHECK-DAG: s_load_dword s[[SDATA:[0-9]+]],
 ; CHECK: s_waitcnt lgkmcnt(0)
 ; CHECK-DAG: v_mov_b32_e32 v[[DATA:[0-9]+]], s[[SDATA]]
 ; CHECK-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], s[[LO_SREG]]
 ; CHECK-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], s[[HI_SREG]]

 define amdgpu_kernel void @flat_scratch_unaligned_store() {
   %scratch = alloca i32, addrspace(5)
   %fptr = addrspacecast i32 addrspace(5)* %scratch to i32*
   store volatile i32 0, i32* %fptr, align 1
   ret void
 }
 
 ; CHECK-LABEL: flat_scratch_multidword_load:
-; HSA: flat_load_dword
-; HSA: flat_load_dword
+; CIVI-HSA: flat_load_dword v
+; CIVI-HSA: flat_load_dword v
+; GFX9:  flat_load_dwordx2
+; GFX10: flat_load_dwordx2
 ; FIXME: These tests are broken for os = mesa3d, becasue it doesn't initialize flat_scr
 define amdgpu_kernel void @flat_scratch_multidword_load() {
   %scratch = alloca <2 x i32>, addrspace(5)
   %fptr = addrspacecast <2 x i32> addrspace(5)* %scratch to <2 x i32>*
   %ld = load volatile <2 x i32>, <2 x i32>* %fptr
   ret void
 }
 
 ; CHECK-LABEL: flat_scratch_multidword_store:
-; HSA: flat_store_dword
-; HSA: flat_store_dword
+; CIVI-HSA: flat_store_dword v
+; CIVI-HSA: flat_store_dword v
+; GFX9:  flat_store_dwordx2
+; GFX10: flat_store_dwordx2
 ; FIXME: These tests are broken for os = mesa3d, becasue it doesn't initialize flat_scr
 define amdgpu_kernel void @flat_scratch_multidword_store() {
   %scratch = alloca <2 x i32>, addrspace(5)
   %fptr = addrspacecast <2 x i32> addrspace(5)* %scratch to <2 x i32>*
   store volatile <2 x i32> zeroinitializer, <2 x i32>* %fptr
   ret void
 }
 
 ; CHECK-LABEL: {{^}}store_flat_i8_max_offset:
 ; CIVI: flat_store_byte v{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}{{$}}
 ; GFX9: flat_store_byte v{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}} offset:4095{{$}}
 define amdgpu_kernel void @store_flat_i8_max_offset(i8* %fptr, i8 %x) #0 {
   %fptr.offset = getelementptr inbounds i8, i8* %fptr, i64 4095
   store volatile i8 %x, i8* %fptr.offset
   ret void
 }