Files
clang-p2996/llvm/lib/Target/SPIRV/SPIRVTargetMachine.cpp
Nathan Gauër 1ed65febd9 [SPIR-V] Add SPIR-V structurizer (#107408)
This commit adds an initial SPIR-V structurizer.
It leverages the previously merged passes, and the convergence region
analysis to determine the correct merge and continue blocks for SPIR-V.

The first part does a branch cleanup (simplifying switches, and
legalizing them), then merge instructions are added to cycles,
convergent and later divergent blocks.
Then comes the important part: splitting critical edges, and making sure
the divergent construct boundaries don't cross.

- we split blocks with multiple headers into 2 blocks.
- we split blocks that are a merge blocks for 2 or more constructs:
SPIR-V spec disallow a merge block to be shared by 2
loop/switch/condition construct.
- we split merge & continue blocks: SPIR-V spec disallow a basic block
to be both a continue block, and a merge block.
- we remove superfluous headers: when a header doesn't bring more info
than the parent on the divergence state, it must be removed.

This PR leverages the merged SPIR-V simulator for testing, as long as
spirv-val. For now, most DXC structurization tests are passing. The
unsupported ones are either caused by unsupported features like switches
on boolean types, or switches in region exits, because the MergeExit
pass doesn't support those yet (there is a FIXME).

This PR is quite large, and the addition not trivial, so I tried to keep
it simple. E.G: as soon as the CFG changes, I recompute the dominator
trees and other structures instead of updating them.

---------

Signed-off-by: Nathan Gauër <brioche@google.com>
2024-09-20 11:36:43 +02:00

241 lines
8.3 KiB
C++

//===- SPIRVTargetMachine.cpp - Define TargetMachine for SPIR-V -*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// Implements the info about SPIR-V target spec.
//
//===----------------------------------------------------------------------===//
#include "SPIRVTargetMachine.h"
#include "SPIRV.h"
#include "SPIRVCallLowering.h"
#include "SPIRVGlobalRegistry.h"
#include "SPIRVLegalizerInfo.h"
#include "SPIRVTargetObjectFile.h"
#include "SPIRVTargetTransformInfo.h"
#include "TargetInfo/SPIRVTargetInfo.h"
#include "llvm/CodeGen/GlobalISel/IRTranslator.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
#include "llvm/CodeGen/GlobalISel/Legalizer.h"
#include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/InitializePasses.h"
#include "llvm/MC/TargetRegistry.h"
#include "llvm/Pass.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Transforms/Utils.h"
#include <optional>
using namespace llvm;
extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeSPIRVTarget() {
// Register the target.
RegisterTargetMachine<SPIRVTargetMachine> X(getTheSPIRV32Target());
RegisterTargetMachine<SPIRVTargetMachine> Y(getTheSPIRV64Target());
RegisterTargetMachine<SPIRVTargetMachine> Z(getTheSPIRVLogicalTarget());
PassRegistry &PR = *PassRegistry::getPassRegistry();
initializeGlobalISel(PR);
initializeSPIRVModuleAnalysisPass(PR);
initializeSPIRVConvergenceRegionAnalysisWrapperPassPass(PR);
}
static std::string computeDataLayout(const Triple &TT) {
const auto Arch = TT.getArch();
// TODO: this probably needs to be revisited:
// Logical SPIR-V has no pointer size, so any fixed pointer size would be
// wrong. The choice to default to 32 or 64 is just motivated by another
// memory model used for graphics: PhysicalStorageBuffer64. But it shouldn't
// mean anything.
if (Arch == Triple::spirv32)
return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-"
"v96:128-v192:256-v256:256-v512:512-v1024:1024-G1";
if (TT.getVendor() == Triple::VendorType::AMD &&
TT.getOS() == Triple::OSType::AMDHSA)
return "e-i64:64-v16:16-v24:32-v32:32-v48:64-"
"v96:128-v192:256-v256:256-v512:512-v1024:1024-G1-P4-A0";
return "e-i64:64-v16:16-v24:32-v32:32-v48:64-"
"v96:128-v192:256-v256:256-v512:512-v1024:1024-G1";
}
static Reloc::Model getEffectiveRelocModel(std::optional<Reloc::Model> RM) {
if (!RM)
return Reloc::PIC_;
return *RM;
}
// Pin SPIRVTargetObjectFile's vtables to this file.
SPIRVTargetObjectFile::~SPIRVTargetObjectFile() {}
SPIRVTargetMachine::SPIRVTargetMachine(const Target &T, const Triple &TT,
StringRef CPU, StringRef FS,
const TargetOptions &Options,
std::optional<Reloc::Model> RM,
std::optional<CodeModel::Model> CM,
CodeGenOptLevel OL, bool JIT)
: LLVMTargetMachine(T, computeDataLayout(TT), TT, CPU, FS, Options,
getEffectiveRelocModel(RM),
getEffectiveCodeModel(CM, CodeModel::Small), OL),
TLOF(std::make_unique<SPIRVTargetObjectFile>()),
Subtarget(TT, CPU.str(), FS.str(), *this) {
initAsmInfo();
setGlobalISel(true);
setFastISel(false);
setO0WantsFastISel(false);
setRequiresStructuredCFG(false);
}
namespace {
// SPIR-V Code Generator Pass Configuration Options.
class SPIRVPassConfig : public TargetPassConfig {
public:
SPIRVPassConfig(SPIRVTargetMachine &TM, PassManagerBase &PM)
: TargetPassConfig(TM, PM), TM(TM) {}
SPIRVTargetMachine &getSPIRVTargetMachine() const {
return getTM<SPIRVTargetMachine>();
}
void addIRPasses() override;
void addISelPrepare() override;
bool addIRTranslator() override;
void addPreLegalizeMachineIR() override;
bool addLegalizeMachineIR() override;
bool addRegBankSelect() override;
bool addGlobalInstructionSelect() override;
FunctionPass *createTargetRegisterAllocator(bool) override;
void addFastRegAlloc() override {}
void addOptimizedRegAlloc() override {}
void addPostRegAlloc() override;
void addPreEmitPass() override;
private:
const SPIRVTargetMachine &TM;
};
} // namespace
// We do not use physical registers, and maintain virtual registers throughout
// the entire pipeline, so return nullptr to disable register allocation.
FunctionPass *SPIRVPassConfig::createTargetRegisterAllocator(bool) {
return nullptr;
}
// Disable passes that break from assuming no virtual registers exist.
void SPIRVPassConfig::addPostRegAlloc() {
// Do not work with vregs instead of physical regs.
disablePass(&MachineCopyPropagationID);
disablePass(&PostRAMachineSinkingID);
disablePass(&PostRASchedulerID);
disablePass(&FuncletLayoutID);
disablePass(&StackMapLivenessID);
disablePass(&PatchableFunctionID);
disablePass(&ShrinkWrapID);
disablePass(&LiveDebugValuesID);
disablePass(&MachineLateInstrsCleanupID);
disablePass(&RemoveLoadsIntoFakeUsesID);
// Do not work with OpPhi.
disablePass(&BranchFolderPassID);
disablePass(&MachineBlockPlacementID);
TargetPassConfig::addPostRegAlloc();
}
TargetTransformInfo
SPIRVTargetMachine::getTargetTransformInfo(const Function &F) const {
return TargetTransformInfo(SPIRVTTIImpl(this, F));
}
TargetPassConfig *SPIRVTargetMachine::createPassConfig(PassManagerBase &PM) {
return new SPIRVPassConfig(*this, PM);
}
void SPIRVPassConfig::addIRPasses() {
TargetPassConfig::addIRPasses();
if (TM.getSubtargetImpl()->isVulkanEnv()) {
// 1. Simplify loop for subsequent transformations. After this steps, loops
// have the following properties:
// - loops have a single entry edge (pre-header to loop header).
// - all loop exits are dominated by the loop pre-header.
// - loops have a single back-edge.
addPass(createLoopSimplifyPass());
// 2. Merge the convergence region exit nodes into one. After this step,
// regions are single-entry, single-exit. This will help determine the
// correct merge block.
addPass(createSPIRVMergeRegionExitTargetsPass());
// 3. Structurize.
addPass(createSPIRVStructurizerPass());
}
addPass(createSPIRVRegularizerPass());
addPass(createSPIRVPrepareFunctionsPass(TM));
addPass(createSPIRVStripConvergenceIntrinsicsPass());
}
void SPIRVPassConfig::addISelPrepare() {
addPass(createSPIRVEmitIntrinsicsPass(&getTM<SPIRVTargetMachine>()));
TargetPassConfig::addISelPrepare();
}
bool SPIRVPassConfig::addIRTranslator() {
addPass(new IRTranslator(getOptLevel()));
return false;
}
void SPIRVPassConfig::addPreLegalizeMachineIR() {
addPass(createSPIRVPreLegalizerPass());
}
// Use the default legalizer.
bool SPIRVPassConfig::addLegalizeMachineIR() {
addPass(new Legalizer());
addPass(createSPIRVPostLegalizerPass());
return false;
}
// Do not add the RegBankSelect pass, as we only ever need virtual registers.
bool SPIRVPassConfig::addRegBankSelect() {
disablePass(&RegBankSelect::ID);
return false;
}
static cl::opt<bool> SPVEnableNonSemanticDI(
"spv-emit-nonsemantic-debug-info",
cl::desc("Emit SPIR-V NonSemantic.Shader.DebugInfo.100 instructions"),
cl::Optional, cl::init(false));
void SPIRVPassConfig::addPreEmitPass() {
if (SPVEnableNonSemanticDI) {
addPass(createSPIRVEmitNonSemanticDIPass(&getTM<SPIRVTargetMachine>()));
}
}
namespace {
// A custom subclass of InstructionSelect, which is mostly the same except from
// not requiring RegBankSelect to occur previously.
class SPIRVInstructionSelect : public InstructionSelect {
// We don't use register banks, so unset the requirement for them
MachineFunctionProperties getRequiredProperties() const override {
return InstructionSelect::getRequiredProperties().reset(
MachineFunctionProperties::Property::RegBankSelected);
}
};
} // namespace
// Add the custom SPIRVInstructionSelect from above.
bool SPIRVPassConfig::addGlobalInstructionSelect() {
addPass(new SPIRVInstructionSelect());
return false;
}