AMDGPU: Add skeleton GlobalIsel implementation

[• tstellarAMD]

Summary

This adds the necessary target code to be able to run the ir translator.
Lowering function arguments and returns is a nop and there is no support
for RegBankSelect.

[arsenm]

lib/Target/AMDGPU/AMDGPUCallLowering.h

2

Missing -*- C++ -*-

16

Missing _H

29–30

Should probably update these to use the current start with lowercase naming convention while it's new

lib/Target/AMDGPU/AMDGPUSubtarget.h

112

Typo onwership, although it seems like this comment would only go on the base class

lib/Target/AMDGPU/CMakeLists.txt

19

Weird indent

[• tstellarAMD]

Address some code review comments.

[• tstellarAMD]

lib/Target/AMDGPU/AMDGPUCallLowering.h

29–30

Ok, I can try to fix up the parent in a new class.

lib/Target/AMDGPU/AMDGPUSubtarget.h

112

I dropped this comment.

[arsenm]

LGTM

[qcolombet]

Hi Tom,

Mostly LGTM, but I do not see AMDGPUGISelAccessor.h in the diff.

Cheers,
-Quentin

lib/Target/AMDGPU/AMDGPUCallLowering.h

30–31

By all means, fix that in the base class :).

[qcolombet]

PS: Side question, since I see you used the GISelAccessor trick, would it make sense that we provide the basic GISelAccessor class in GlobalISel header?

[• tstellarAMD]

Update patch after making some changes to core GlobalISel.

[qcolombet]

LGTM.

lib/Target/AMDGPU/AMDGPUCallLowering.h

+//===- lib/Target/AMDGPU/AMDGPUCallLowering.h - Call lowering -*- C++ -*---===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// This file describes how to lower LLVM calls to machine code calls.
+///
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUCALLLOWERING_H
+#define LLVM_LIB_TARGET_AMDGPU_AMDGPUCALLLOWERING_H
+
+#include "llvm/CodeGen/GlobalISel/CallLowering.h"
+
+namespace llvm {
+
+class AMDGPUTargetLowering;
+
+class AMDGPUCallLowering: public CallLowering {
+ public:
+  AMDGPUCallLowering(const AMDGPUTargetLowering &TLI);
+
+  bool lowerReturn(MachineIRBuilder &MIRBuiler, const Value *Val,
+                   unsigned VReg) const override;
+  bool
+  lowerFormalArguments(MachineIRBuilder &MIRBuilder,
+                       const Function::ArgumentListType &Args,
+                       const SmallVectorImpl<unsigned> &VRegs) const override;
+};
+} // End of namespace llvm;
+#endif

lib/Target/AMDGPU/AMDGPUCallLowering.cpp

+//===-- llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp - Call lowering ---===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// This file implements the lowering of LLVM calls to machine code calls for
+/// GlobalISel.
+///
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUCallLowering.h"
+#include "AMDGPUISelLowering.h"
+
+#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+
+using namespace llvm;
+
+#ifndef LLVM_BUILD_GLOBAL_ISEL
+#error "This shouldn't be built without GISel"
+#endif
+
+AMDGPUCallLowering::AMDGPUCallLowering(const AMDGPUTargetLowering &TLI)
+  : CallLowering(&TLI) {
+}
+
+bool AMDGPUCallLowering::lowerReturn(MachineIRBuilder &MIRBuilder,
+                                        const Value *Val, unsigned VReg) const {
+  return true;
+}
+
+bool AMDGPUCallLowering::lowerFormalArguments(
+    MachineIRBuilder &MIRBuilder, const Function::ArgumentListType &Args,
+    const SmallVectorImpl<unsigned> &VRegs) const {
+  // TODO: Implement once there are generic loads/stores.
+  return true;
+}

lib/Target/AMDGPU/AMDGPUSubtarget.h

 
 #include "AMDGPU.h"
 #include "AMDGPUFrameLowering.h"
 #include "AMDGPUInstrInfo.h"
 #include "AMDGPUISelLowering.h"
 #include "AMDGPUSubtarget.h"
 #include "Utils/AMDGPUBaseInfo.h"
 #include "llvm/ADT/StringRef.h"
+#include "llvm/CodeGen/GlobalISel/GISelAccessor.h"
 #include "llvm/Target/TargetSubtargetInfo.h"
 
 #define GET_SUBTARGETINFO_HEADER
 #include "AMDGPUGenSubtargetInfo.inc"
 
 namespace llvm {
 
 class SIMachineFunctionInfo;
private:
   bool FeatureDisable;
   int LDSBankCount;
   unsigned IsaVersion;
   bool EnableSIScheduler;
 
   std::unique_ptr<AMDGPUFrameLowering> FrameLowering;
   std::unique_ptr<AMDGPUTargetLowering> TLInfo;
   std::unique_ptr<AMDGPUInstrInfo> InstrInfo;
+  std::unique_ptr<GISelAccessor> GISel;
   InstrItineraryData InstrItins;
   Triple TargetTriple;
 
 public:
   AMDGPUSubtarget(const Triple &TT, StringRef CPU, StringRef FS,
                   TargetMachine &TM);
   AMDGPUSubtarget &initializeSubtargetDependencies(const Triple &TT,
                                                    StringRef GPU, StringRef FS);
 
+  void setGISelAccessor(GISelAccessor &GISel) {
+    this->GISel.reset(&GISel);
+  }
+
   const AMDGPUFrameLowering *getFrameLowering() const override {
     return FrameLowering.get();
   }
   const AMDGPUInstrInfo *getInstrInfo() const override {
     return InstrInfo.get();
   }
   const AMDGPURegisterInfo *getRegisterInfo() const override {
     return &InstrInfo->getRegisterInfo();
   }
   AMDGPUTargetLowering *getTargetLowering() const override {
     return TLInfo.get();
   }
   const InstrItineraryData *getInstrItineraryData() const override {
     return &InstrItins;
   }
 
+  const CallLowering *getCallLowering() const override;
+
   void ParseSubtargetFeatures(StringRef CPU, StringRef FS);
 
   bool hasVertexCache() const {
     return HasVertexCache;
   }
 
   short getTexVTXClauseSize() const {
     return TexVTXClauseSize;

lib/Target/AMDGPU/AMDGPUSubtarget.cpp

 //===-- AMDGPUSubtarget.cpp - AMDGPU Subtarget Information ----------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 /// \file
 /// \brief Implements the AMDGPU specific subclass of TargetSubtarget.
 //
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPUSubtarget.h"
+#include "AMDGPUCallLowering.h"
 #include "R600ISelLowering.h"
 #include "R600InstrInfo.h"
 #include "R600MachineScheduler.h"
 #include "SIFrameLowering.h"
 #include "SIISelLowering.h"
 #include "SIInstrInfo.h"
 #include "SIMachineFunctionInfo.h"
 #include "llvm/ADT/SmallString.h"
 #include "llvm/CodeGen/MachineScheduler.h"
 
 using namespace llvm;
 
 #define DEBUG_TYPE "amdgpu-subtarget"
 
 #define GET_SUBTARGETINFO_ENUM
 #define GET_SUBTARGETINFO_TARGET_DESC
 #define GET_SUBTARGETINFO_CTOR
 #include "AMDGPUGenSubtargetInfo.inc"
 
+#ifdef LLVM_BUILD_GLOBAL_ISEL
+namespace {
+struct AMDGPUGISelActualAccessor : public GISelAccessor {
+  std::unique_ptr<CallLowering> CallLoweringInfo;
+  const CallLowering *getCallLowering() const override {
+    return CallLoweringInfo.get();
+  }
+};
+} // End anonymous namespace.
+#endif
+
 AMDGPUSubtarget &
 AMDGPUSubtarget::initializeSubtargetDependencies(const Triple &TT,
                                                  StringRef GPU, StringRef FS) {
   // Determine default and user-specified characteristics
   // On SI+, we want FP64 denormals to be on by default. FP32 denormals can be
   // enabled, but some instructions do not respect them and they run at the
   // double precision rate, so don't enable by default.
   //
: AMDGPUGenSubtargetInfo(TT, GPU, FS),
       WavefrontSize(0), CFALUBug(false),
       LocalMemorySize(0), MaxPrivateElementSize(0),
       EnableVGPRSpilling(false), SGPRInitBug(false), IsGCN(false),
       GCN1Encoding(false), GCN3Encoding(false), CIInsts(false),
       HasSMemRealTime(false), Has16BitInsts(false),
       LDSBankCount(0),
       IsaVersion(ISAVersion0_0_0),
       EnableSIScheduler(false), FrameLowering(nullptr),
+      GISel(),
       InstrItins(getInstrItineraryForCPU(GPU)), TargetTriple(TT) {
 
   initializeSubtargetDependencies(TT, GPU, FS);
 
   const unsigned MaxStackAlign = 64 * 16; // Maximum stack alignment (long16)
 
   if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
     InstrInfo.reset(new R600InstrInfo(*this));
     TLInfo.reset(new R600TargetLowering(TM, *this));
 
     // FIXME: Should have R600 specific FrameLowering
     FrameLowering.reset(new AMDGPUFrameLowering(
                           TargetFrameLowering::StackGrowsUp,
                           MaxStackAlign,
                           0));
   } else {
     InstrInfo.reset(new SIInstrInfo(*this));
     TLInfo.reset(new SITargetLowering(TM, *this));
     FrameLowering.reset(new SIFrameLowering(
                           TargetFrameLowering::StackGrowsUp,
                           MaxStackAlign,
                           0));
-  }
+#ifndef LLVM_BUILD_GLOBAL_ISEL
+    GISelAccessor *GISel = new GISelAccessor();
+#else
+    AMDGPUGISelActualAccessor *GISel =
+        new AMDGPUGISelActualAccessor();
+    GISel->CallLoweringInfo.reset(
+        new AMDGPUCallLowering(*getTargetLowering()));
+#endif
+    setGISelAccessor(*GISel);
+  }
+
+const CallLowering *AMDGPUSubtarget::getCallLowering() const {
+  assert(GISel && "Access to GlobalISel APIs not set");
+  return GISel->getCallLowering();
 }
 
 unsigned AMDGPUSubtarget::getStackEntrySize() const {
   assert(getGeneration() <= NORTHERN_ISLANDS);
   switch(getWavefrontSize()) {
   case 16:
     return 8;
   case 32:

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

 #include "AMDGPU.h"
 #include "AMDGPUTargetTransformInfo.h"
 #include "R600ISelLowering.h"
 #include "R600InstrInfo.h"
 #include "R600MachineScheduler.h"
 #include "SIISelLowering.h"
 #include "SIInstrInfo.h"
 #include "llvm/Analysis/Passes.h"
+#include "llvm/CodeGen/GlobalISel/IRTranslator.h"
 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
 #include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/IR/Verifier.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/Support/CommandLine.h"

 };
 
 class GCNPassConfig final : public AMDGPUPassConfig {
 public:
   GCNPassConfig(TargetMachine *TM, PassManagerBase &PM)
     : AMDGPUPassConfig(TM, PM) { }
   bool addPreISel() override;
   bool addInstSelector() override;
+#ifdef LLVM_BUILD_GLOBAL_ISEL
+  bool addIRTranslator() override;
+  bool addRegBankSelect() override;
+#endif
   void addFastRegAlloc(FunctionPass *RegAllocPass) override;
   void addOptimizedRegAlloc(FunctionPass *RegAllocPass) override;
   void addPreRegAlloc() override;
   void addPostRegAlloc() override;
   void addPreSched2() override;
   void addPreEmitPass() override;
 };
 

 bool GCNPassConfig::addInstSelector() {
   AMDGPUPassConfig::addInstSelector();
   addPass(createSILowerI1CopiesPass());
   addPass(&SIFixSGPRCopiesID);
   addPass(createSIFoldOperandsPass());
   return false;
 }
 
+#ifdef LLVM_BUILD_GLOBAL_ISEL
+bool GCNPassConfig::addIRTranslator() {
+  addPass(new IRTranslator());
+  return false;
+}
+
+bool GCNPassConfig::addRegBankSelect() {
+  return false;
+}
+#endif
+
 void GCNPassConfig::addPreRegAlloc() {
   const AMDGPUSubtarget &ST = *getAMDGPUTargetMachine().getSubtargetImpl();
 
   // This needs to be run directly before register allocation because
   // earlier passes might recompute live intervals.
   // TODO: handle CodeGenOpt::None; fast RA ignores spill weights set by the pass
   if (getOptLevel() > CodeGenOpt::None) {
     insertPass(&MachineSchedulerID, &SIFixControlFlowLiveIntervalsID);

lib/Target/AMDGPU/CMakeLists.txt

 set(LLVM_TARGET_DEFINITIONS AMDGPU.td)
 
 tablegen(LLVM AMDGPUGenRegisterInfo.inc -gen-register-info)
 tablegen(LLVM AMDGPUGenInstrInfo.inc -gen-instr-info)
 tablegen(LLVM AMDGPUGenDAGISel.inc -gen-dag-isel)
 tablegen(LLVM AMDGPUGenCallingConv.inc -gen-callingconv)
 tablegen(LLVM AMDGPUGenSubtargetInfo.inc -gen-subtarget)
 tablegen(LLVM AMDGPUGenIntrinsics.inc -gen-tgt-intrinsic)
 tablegen(LLVM AMDGPUGenMCCodeEmitter.inc -gen-emitter)
 tablegen(LLVM AMDGPUGenDFAPacketizer.inc -gen-dfa-packetizer)
 tablegen(LLVM AMDGPUGenAsmWriter.inc -gen-asm-writer)
 tablegen(LLVM AMDGPUGenAsmMatcher.inc -gen-asm-matcher)
 tablegen(LLVM AMDGPUGenDisassemblerTables.inc -gen-disassembler)
 add_public_tablegen_target(AMDGPUCommonTableGen)
 
+# List of all GlobalISel files.
+set(GLOBAL_ISEL_FILES
+  AMDGPUCallLowering.cpp
+  )
+
+# Add GlobalISel files to the dependencies if the user wants to build it.
+if(LLVM_BUILD_GLOBAL_ISEL)
+  set(GLOBAL_ISEL_BUILD_FILES ${GLOBAL_ISEL_FILES})
+else()
+  set(GLOBAL_ISEL_BUILD_FILES"")
+  set(LLVM_OPTIONAL_SOURCES LLVMGlobalISel ${GLOBAL_ISEL_FILES})
+endif()
+
+
 add_llvm_target(AMDGPUCodeGen
   AMDILCFGStructurizer.cpp
   AMDGPUAlwaysInlinePass.cpp
   AMDGPUAnnotateKernelFeatures.cpp
   AMDGPUAnnotateUniformValues.cpp
   AMDGPUAsmPrinter.cpp
   AMDGPUFrameLowering.cpp
   AMDGPUTargetObjectFile.cpp
add_llvm_target(AMDGPUCodeGen
   SILowerControlFlow.cpp
   SILowerI1Copies.cpp
   SIMachineFunctionInfo.cpp
   SIMachineScheduler.cpp
   SIRegisterInfo.cpp
   SIShrinkInstructions.cpp
   SITypeRewriter.cpp
   SIWholeQuadMode.cpp
+  ${GLOBAL_ISEL_BUILD_FILES}
   )
 
 add_subdirectory(AsmParser)
 add_subdirectory(InstPrinter)
 add_subdirectory(Disassembler)
 add_subdirectory(TargetInfo)
 add_subdirectory(MCTargetDesc)
 add_subdirectory(Utils)

test/CodeGen/AMDGPU/GlobalISel/amdgpu-irtranslator.ll

+; RUN: llc -march=amdgcn -mcpu=fiji -O0 -stop-after=irtranslator -global-isel %s -o - 2>&1 | FileCheck %s
+; REQUIRES: global-isel
+; This file checks that the translation from llvm IR to generic MachineInstr
+; is correct.
+
+; Tests for add.
+; CHECK: name: addi32
+; CHECK: G_ADD i32
+define i32 @addi32(i32 %arg1, i32 %arg2) {
+  %res = add i32 %arg1, %arg2
+  ret i32 %res
+}