AMDGPU/SI: Add a MachineMemOperand to MIMG instructions

[• tstellarAMD]

Summary

Without a MachineMemOperand, the scheduler was assuming MIMG instructions
were ordered memory references, so no loads or stores could be reordered
across them.

[arsenm]

I thought we needed D13575 for this

lib/Target/AMDGPU/SIMachineFunctionInfo.h

31

Why are these virtual? Are they supposed to be override?

[arsenm]

Can you add a test where this changes the scheduling

[• tstellarAMD]

Add test and address other comment.

[• tstellarAMD]

In D27536#616313, @arsenm wrote:

I thought we needed D13575 for this

Yes, I was planning to commit that one first.

[arsenm]

LGTM

lib/Target/AMDGPU/MIMGInstructions.td

 class MIMG_Helper <dag outs, dag ins, string asm,
                    string dns=""> : MIMG<outs, ins, asm,[]> {
   let mayLoad = 1;
   let mayStore = 0;
   let hasPostISelHook = 1;
   let DecoderNamespace = dns;
   let isAsmParserOnly = !if(!eq(dns,""), 1, 0);
   let AsmMatchConverter = "cvtMIMG";
+  let usesCustomInserter = 1;
 }
 
 class MIMG_NoSampler_Helper <bits<7> op, string asm,
                              RegisterClass dst_rc,
                              RegisterClass addr_rc,
                              string dns=""> : MIMG_Helper <
   (outs dst_rc:$vdata),
   (ins addr_rc:$vaddr, SReg_256:$srsrc,

lib/Target/AMDGPU/SIISelLowering.cpp

static MachineBasicBlock *emitIndirectDst(MachineInstr &MI,
 
   MI.eraseFromParent();
 
   return LoopBB;
 }
 
 MachineBasicBlock *SITargetLowering::EmitInstrWithCustomInserter(
   MachineInstr &MI, MachineBasicBlock *BB) const {
+
+  const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
+  MachineFunction *MF = BB->getParent();
+  SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
+
+  if (TII->isMIMG(MI)) {
+      if (!MI.memoperands_empty())
+        return BB;
+    // Add a memoperand for mimg instructions so that they aren't assumed to
+    // be ordered memory instuctions.
+
+    MachinePointerInfo PtrInfo(MFI->getImagePSV());
+    MachineMemOperand::Flags Flags = MachineMemOperand::MODereferenceable;
+    if (MI.mayStore())
+      Flags |= MachineMemOperand::MOStore;
+
+    if (MI.mayLoad())
+      Flags |= MachineMemOperand::MOLoad;
+
+    auto MMO = MF->getMachineMemOperand(PtrInfo, Flags, 0, 0);
+    MI.addMemOperand(*MF, MMO); 
+    return BB;
+  }
+
   switch (MI.getOpcode()) {
   case AMDGPU::SI_INIT_M0: {
-    const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
     BuildMI(*BB, MI.getIterator(), MI.getDebugLoc(),
             TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
       .addOperand(MI.getOperand(0));
     MI.eraseFromParent();
     return BB;
   }
   case AMDGPU::GET_GROUPSTATICSIZE: {
-    const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
-
-    MachineFunction *MF = BB->getParent();
-    SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
     DebugLoc DL = MI.getDebugLoc();
     BuildMI(*BB, MI, DL, TII->get(AMDGPU::S_MOV_B32))
       .addOperand(MI.getOperand(0))
       .addImm(MFI->getLDSSize());
     MI.eraseFromParent();
     return BB;
   }
   case AMDGPU::SI_INDIRECT_SRC_V1:
   case AMDGPU::SI_INDIRECT_SRC_V2:
   case AMDGPU::SI_INDIRECT_SRC_V4:
   case AMDGPU::SI_INDIRECT_SRC_V8:
   case AMDGPU::SI_INDIRECT_SRC_V16:
     return emitIndirectSrc(MI, *BB, *getSubtarget());
   case AMDGPU::SI_INDIRECT_DST_V1:
   case AMDGPU::SI_INDIRECT_DST_V2:
   case AMDGPU::SI_INDIRECT_DST_V4:
   case AMDGPU::SI_INDIRECT_DST_V8:
   case AMDGPU::SI_INDIRECT_DST_V16:
     return emitIndirectDst(MI, *BB, *getSubtarget());
   case AMDGPU::SI_KILL:
     return splitKillBlock(MI, BB);
   case AMDGPU::V_CNDMASK_B64_PSEUDO: {
     MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
-    const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
 
     unsigned Dst = MI.getOperand(0).getReg();
     unsigned Src0 = MI.getOperand(1).getReg();
     unsigned Src1 = MI.getOperand(2).getReg();
     const DebugLoc &DL = MI.getDebugLoc();
     unsigned SrcCond = MI.getOperand(3).getReg();
 
     unsigned DstLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);

lib/Target/AMDGPU/SIMachineFunctionInfo.h

 #include "SIRegisterInfo.h"
 #include <array>
 #include <map>
 
 namespace llvm {
 
 class MachineRegisterInfo;
 
+class AMDGPUImagePseudoSourceValue : public PseudoSourceValue {
+public:
+  explicit AMDGPUImagePseudoSourceValue() :
+    PseudoSourceValue(PseudoSourceValue::TargetCustom) { }
+
+  bool isConstant(const MachineFrameInfo *) const override {
+    // This should probably be true for most images, but we will start by being
+    // conservative.
+    return false;
+  }
+
+  bool isAliased(const MachineFrameInfo *) const override {
+    // FIXME: If we ever change image intrinsics to accept fat pointers, then
+    // this could be true for some cases.
+    return false;
+  }
+
+  bool mayAlias(const MachineFrameInfo*) const override {
+    // FIXME: If we ever change image intrinsics to accept fat pointers, then
+    // this could be true for some cases.
+    return false;
+  }
+};
+
+
 /// This class keeps track of the SPI_SP_INPUT_ADDR config register, which
 /// tells the hardware which interpolation parameters to load.
 class SIMachineFunctionInfo final : public AMDGPUMachineFunction {
   // FIXME: This should be removed and getPreloadedValue moved here.
   friend class SIRegisterInfo;
 
   unsigned TIDReg;
 
class SIMachineFunctionInfo final : public AMDGPUMachineFunction {
   // unit. Minimum - first, maximum - second.
   std::pair<unsigned, unsigned> WavesPerEU;
 
   // Stack object indices for work group IDs.
   std::array<int, 3> DebuggerWorkGroupIDStackObjectIndices;
   // Stack object indices for work item IDs.
   std::array<int, 3> DebuggerWorkItemIDStackObjectIndices;
 
+  std::unique_ptr<AMDGPUImagePseudoSourceValue> ImagePSV;
+
 public:
   // FIXME: Make private
   unsigned LDSWaveSpillSize;
   unsigned PSInputEna;
   std::map<unsigned, unsigned> LaneVGPRs;
   unsigned ScratchOffsetReg;
   unsigned NumUserSGPRs;
   unsigned NumSystemSGPRs;
case 1:
       assert(hasWorkItemIDY());
       return AMDGPU::VGPR1;
     case 2:
       assert(hasWorkItemIDZ());
       return AMDGPU::VGPR2;
     }
     llvm_unreachable("unexpected dimension");
   }
+
+  AMDGPUImagePseudoSourceValue *getImagePSV() {
+    return ImagePSV.get();
+  }
 };
 
 } // End namespace llvm
 
 #endif

lib/Target/AMDGPU/SIMachineFunctionInfo.cpp

: AMDGPUMachineFunction(MF),
     WorkGroupInfoSystemSGPR(AMDGPU::NoRegister),
     PrivateSegmentWaveByteOffsetSystemSGPR(AMDGPU::NoRegister),
     PSInputAddr(0),
     ReturnsVoid(true),
     FlatWorkGroupSizes(0, 0),
     WavesPerEU(0, 0),
     DebuggerWorkGroupIDStackObjectIndices({{0, 0, 0}}),
     DebuggerWorkItemIDStackObjectIndices({{0, 0, 0}}),
+    ImagePSV(llvm::make_unique<AMDGPUImagePseudoSourceValue>()),
     LDSWaveSpillSize(0),
     PSInputEna(0),
     NumUserSGPRs(0),
     NumSystemSGPRs(0),
     HasSpilledSGPRs(false),
     HasSpilledVGPRs(false),
     HasNonSpillStackObjects(false),
     NumSpilledSGPRs(0),

test/CodeGen/AMDGPU/llvm.amdgcn.image.ll

 ;RUN: llc < %s -march=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck %s
-;RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck %s
+;RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck --check-prefixes=CHECK,VI %s
 
 ;CHECK-LABEL: {{^}}image_load_v4i32:
 ;CHECK: image_load v[0:3], v[0:3], s[0:7] dmask:0xf unorm
 ;CHECK: s_waitcnt vmcnt(0)
 define amdgpu_ps <4 x float> @image_load_v4i32(<8 x i32> inreg %rsrc, <4 x i32> %c) {
 main_body:
   %tex = call <4 x float> @llvm.amdgcn.image.load.v4f32.v4i32.v8i32(<4 x i32> %c, <8 x i32> %rsrc, i32 15, i1 0, i1 0, i1 0, i1 0)
   ret <4 x float> %tex

 define amdgpu_ps void @image_store_wait(<8 x i32> inreg, <8 x i32> inreg, <8 x i32> inreg, <4 x float>, i32) {
 main_body:
   call void @llvm.amdgcn.image.store.v4f32.i32.v8i32(<4 x float> %3, i32 %4, <8 x i32> %0, i32 15, i1 0, i1 0, i1 0, i1 0)
   %data = call <4 x float> @llvm.amdgcn.image.load.v4f32.i32.v8i32(i32 %4, <8 x i32> %1, i32 15, i1 0, i1 0, i1 0, i1 0)
   call void @llvm.amdgcn.image.store.v4f32.i32.v8i32(<4 x float> %data, i32 %4, <8 x i32> %2, i32 15, i1 0, i1 0, i1 0, i1 0)
   ret void
 }
 
+; SI won't merge ds memory operations, because of the signed offset bug, so
+; we only have check lines for VI.
+; VI-LABEL: image_load_mmo
+; VI: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0
+; VI: ds_write2_b32 v{{[0-9]+}}, [[ZERO]], [[ZERO]] offset1:4
+define amdgpu_ps void @image_load_mmo(float addrspace(3)* %lds, <2 x i32> %c, <8 x i32> inreg %rsrc) {
+  store float 0.0, float addrspace(3)* %lds
+  %tex = call float @llvm.amdgcn.image.load.f32.v2i32.v8i32(<2 x i32> %c, <8 x i32> %rsrc, i32 15, i1 0, i1 0, i1 0, i1 0)
+  %tmp2 = getelementptr float, float addrspace(3)* %lds, i32 4
+  store float 0.0, float addrspace(3)* %tmp2
+  call void @llvm.SI.export(i32 15, i32 1, i32 1, i32 0, i32 1, float %tex, float %tex, float %tex, float %tex)
+  ret void
+}
 
 declare float @llvm.amdgcn.image.load.f32.v2i32.v8i32(<2 x i32>, <8 x i32>, i32, i1, i1, i1, i1) #1
 declare <2 x float> @llvm.amdgcn.image.load.v2f32.v4i32.v8i32(<4 x i32>, <8 x i32>, i32, i1, i1, i1, i1) #1
 declare void @llvm.amdgcn.image.store.f32.i32.v8i32(float, i32, <8 x i32>, i32, i1, i1, i1, i1) #0
 declare void @llvm.amdgcn.image.store.v2f32.v4i32.v8i32(<2 x float>, <4 x i32>, <8 x i32>, i32, i1, i1, i1, i1) #0
 
 
 declare void @llvm.amdgcn.image.store.v4f32.i32.v8i32(<4 x float>, i32, <8 x i32>, i32, i1, i1, i1, i1) #0