llvm-project/llvm/lib/Target/SPIRV/SPIRVPostLegalizer.cpp
Vyacheslav Levytskyy 54cc4141e4
[SPIR-V] Rework duplicate tracker and tracking of IR entities and types to improve compile-time performance (#130605)
This PR is to thoroughly rework duplicate tracker implementation and
tracking of IR entities and types. These are legacy parts of the project
resulting in an extremely bloated intermediate representation and
computational delays due to inefficient data flow and structure choices.

Main results of the rework:

1) Improved compile-time performance. The reference binary LLVM IR used
to measure speed gains in
https://github.com/llvm/llvm-project/pull/120415 shows ~x5 speed up also
after this PR. The timing before this PR is ~42s and after this PR it's
~7.5s. In total this PR and the previous overhaul of the module analysis
in https://github.com/llvm/llvm-project/pull/120415 results in ~x25
speed improvement.
```
$ time llc -O0 -mtriple=spirv64v1.6-unknown-unknown _group_barrier_phi.bc -o 1 --filetype=obj

real    0m7.545s
user    0m6.685s
sys     0m0.859s
```

2) Less bloated intermediate representation of internal translation
steps. Elimination of `spv_track_constant` intrinsic usage for scalar
constants, rework of `spv_assign_name`, removal of the gMIR `GET_XXX`
pseudo code and a smaller number of generated `ASSIGN_TYPE` pseudo codes
substantially decrease volume of data generated during translation.

3) Simpler code and easier maintenance. The duplicate tracker
implementation is simplified, as well as other features.

4) Numerous fixes of issues and logical flaws in different passes. The
main achievement is rework of the duplicate tracker itself that had
never guaranteed a correct caching of LLVM IR entities, rarely and
randomly returning stale/incorrect records (like, remove an instruction
from gMIR but still refer to it). Other fixes comprise consistent
generation of OpConstantNull, assigning types to newly created
registers, creation of integer/bool types, and other minor fixes.

5) Numerous fixes of LIT tests: mainly CHECK-DAG to properly reflect
SPIR-V spec guarantees, `{{$}}` at the end of constants to avoid
matching of substrings, and XFAILS for `SPV_INTEL_long_composites` test
cases, because the feature is not completed in full yet and doesn't
generate a requested by the extension sequence of instructions.

6) New test cases are added.
2025-03-26 17:58:10 +01:00

181 lines
6.4 KiB
C++

//===-- SPIRVPostLegalizer.cpp - ammend info after legalization -*- C++ -*-===//
//
// which may appear after the legalizer pass
//
// 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
//
//===----------------------------------------------------------------------===//
//
// The pass partially apply pre-legalization logic to new instructions inserted
// as a result of legalization:
// - assigns SPIR-V types to registers for new instructions.
//
//===----------------------------------------------------------------------===//
#include "SPIRV.h"
#include "SPIRVSubtarget.h"
#include "SPIRVUtils.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/CodeGen/MachinePostDominators.h"
#include "llvm/IR/Attributes.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/IntrinsicsSPIRV.h"
#include <stack>
#define DEBUG_TYPE "spirv-postlegalizer"
using namespace llvm;
namespace {
class SPIRVPostLegalizer : public MachineFunctionPass {
public:
static char ID;
SPIRVPostLegalizer() : MachineFunctionPass(ID) {
initializeSPIRVPostLegalizerPass(*PassRegistry::getPassRegistry());
}
bool runOnMachineFunction(MachineFunction &MF) override;
};
} // namespace
namespace llvm {
// Defined in SPIRVPreLegalizer.cpp.
extern void insertAssignInstr(Register Reg, Type *Ty, SPIRVType *SpirvTy,
SPIRVGlobalRegistry *GR, MachineIRBuilder &MIB,
MachineRegisterInfo &MRI);
extern void processInstr(MachineInstr &MI, MachineIRBuilder &MIB,
MachineRegisterInfo &MRI, SPIRVGlobalRegistry *GR,
SPIRVType *KnownResType);
} // namespace llvm
static bool mayBeInserted(unsigned Opcode) {
switch (Opcode) {
case TargetOpcode::G_SMAX:
case TargetOpcode::G_UMAX:
case TargetOpcode::G_SMIN:
case TargetOpcode::G_UMIN:
case TargetOpcode::G_FMINNUM:
case TargetOpcode::G_FMINIMUM:
case TargetOpcode::G_FMAXNUM:
case TargetOpcode::G_FMAXIMUM:
return true;
default:
return isTypeFoldingSupported(Opcode);
}
}
static void processNewInstrs(MachineFunction &MF, SPIRVGlobalRegistry *GR,
MachineIRBuilder MIB) {
MachineRegisterInfo &MRI = MF.getRegInfo();
for (MachineBasicBlock &MBB : MF) {
for (MachineInstr &I : MBB) {
const unsigned Opcode = I.getOpcode();
if (Opcode == TargetOpcode::G_UNMERGE_VALUES) {
unsigned ArgI = I.getNumOperands() - 1;
Register SrcReg = I.getOperand(ArgI).isReg()
? I.getOperand(ArgI).getReg()
: Register(0);
SPIRVType *DefType =
SrcReg.isValid() ? GR->getSPIRVTypeForVReg(SrcReg) : nullptr;
if (!DefType || DefType->getOpcode() != SPIRV::OpTypeVector)
report_fatal_error(
"cannot select G_UNMERGE_VALUES with a non-vector argument");
SPIRVType *ScalarType =
GR->getSPIRVTypeForVReg(DefType->getOperand(1).getReg());
for (unsigned i = 0; i < I.getNumDefs(); ++i) {
Register ResVReg = I.getOperand(i).getReg();
SPIRVType *ResType = GR->getSPIRVTypeForVReg(ResVReg);
if (!ResType) {
// There was no "assign type" actions, let's fix this now
ResType = ScalarType;
setRegClassType(ResVReg, ResType, GR, &MRI, *GR->CurMF, true);
}
}
} else if (mayBeInserted(Opcode) && I.getNumDefs() == 1 &&
I.getNumOperands() > 1 && I.getOperand(1).isReg()) {
// Legalizer may have added a new instructions and introduced new
// registers, we must decorate them as if they were introduced in a
// non-automatic way
Register ResVReg = I.getOperand(0).getReg();
// Check if the register defined by the instruction is newly generated
// or already processed
// Check if we have type defined for operands of the new instruction
bool IsKnownReg = MRI.getRegClassOrNull(ResVReg);
SPIRVType *ResVType = GR->getSPIRVTypeForVReg(
IsKnownReg ? ResVReg : I.getOperand(1).getReg());
if (!ResVType)
continue;
// Set type & class
if (!IsKnownReg)
setRegClassType(ResVReg, ResVType, GR, &MRI, *GR->CurMF, true);
// If this is a simple operation that is to be reduced by TableGen
// definition we must apply some of pre-legalizer rules here
if (isTypeFoldingSupported(Opcode)) {
processInstr(I, MIB, MRI, GR, GR->getSPIRVTypeForVReg(ResVReg));
if (IsKnownReg && MRI.hasOneUse(ResVReg)) {
MachineInstr &UseMI = *MRI.use_instr_begin(ResVReg);
if (UseMI.getOpcode() == SPIRV::ASSIGN_TYPE)
continue;
}
insertAssignInstr(ResVReg, nullptr, ResVType, GR, MIB, MRI);
}
}
}
}
}
// Do a preorder traversal of the CFG starting from the BB |Start|.
// point. Calls |op| on each basic block encountered during the traversal.
void visit(MachineFunction &MF, MachineBasicBlock &Start,
std::function<void(MachineBasicBlock *)> op) {
std::stack<MachineBasicBlock *> ToVisit;
SmallPtrSet<MachineBasicBlock *, 8> Seen;
ToVisit.push(&Start);
Seen.insert(ToVisit.top());
while (ToVisit.size() != 0) {
MachineBasicBlock *MBB = ToVisit.top();
ToVisit.pop();
op(MBB);
for (auto Succ : MBB->successors()) {
if (Seen.contains(Succ))
continue;
ToVisit.push(Succ);
Seen.insert(Succ);
}
}
}
// Do a preorder traversal of the CFG starting from the given function's entry
// point. Calls |op| on each basic block encountered during the traversal.
void visit(MachineFunction &MF, std::function<void(MachineBasicBlock *)> op) {
visit(MF, *MF.begin(), op);
}
bool SPIRVPostLegalizer::runOnMachineFunction(MachineFunction &MF) {
// Initialize the type registry.
const SPIRVSubtarget &ST = MF.getSubtarget<SPIRVSubtarget>();
SPIRVGlobalRegistry *GR = ST.getSPIRVGlobalRegistry();
GR->setCurrentFunc(MF);
MachineIRBuilder MIB(MF);
processNewInstrs(MF, GR, MIB);
return true;
}
INITIALIZE_PASS(SPIRVPostLegalizer, DEBUG_TYPE, "SPIRV post legalizer", false,
false)
char SPIRVPostLegalizer::ID = 0;
FunctionPass *llvm::createSPIRVPostLegalizerPass() {
return new SPIRVPostLegalizer();
}