brintos

brintos / llvm-project-archived public Read only

0
0
Text · 12.3 KiB · 23e1243 Raw
364 lines · cpp
1//===- SpeculativeExecution.cpp ---------------------------------*- C++ -*-===//2//3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.4// See https://llvm.org/LICENSE.txt for license information.5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception6//7//===----------------------------------------------------------------------===//8//9// This pass hoists instructions to enable speculative execution on10// targets where branches are expensive. This is aimed at GPUs. It11// currently works on simple if-then and if-then-else12// patterns.13//14// Removing branches is not the only motivation for this15// pass. E.g. consider this code and assume that there is no16// addressing mode for multiplying by sizeof(*a):17//18//   if (b > 0)19//     c = a[i + 1]20//   if (d > 0)21//     e = a[i + 2]22//23// turns into24//25//   p = &a[i + 1];26//   if (b > 0)27//     c = *p;28//   q = &a[i + 2];29//   if (d > 0)30//     e = *q;31//32// which could later be optimized to33//34//   r = &a[i];35//   if (b > 0)36//     c = r[1];37//   if (d > 0)38//     e = r[2];39//40// Later passes sink back much of the speculated code that did not enable41// further optimization.42//43// This pass is more aggressive than the function SpeculativeyExecuteBB in44// SimplifyCFG. SimplifyCFG will not speculate if no selects are introduced and45// it will speculate at most one instruction. It also will not speculate if46// there is a value defined in the if-block that is only used in the then-block.47// These restrictions make sense since the speculation in SimplifyCFG seems48// aimed at introducing cheap selects, while this pass is intended to do more49// aggressive speculation while counting on later passes to either capitalize on50// that or clean it up.51//52// If the pass was created by calling53// createSpeculativeExecutionIfHasBranchDivergencePass or the54// -spec-exec-only-if-divergent-target option is present, this pass only has an55// effect on targets where TargetTransformInfo::hasBranchDivergence() is true;56// on other targets, it is a nop.57//58// This lets you include this pass unconditionally in the IR pass pipeline, but59// only enable it for relevant targets.60//61//===----------------------------------------------------------------------===//62 63#include "llvm/Transforms/Scalar/SpeculativeExecution.h"64#include "llvm/ADT/SmallPtrSet.h"65#include "llvm/Analysis/GlobalsModRef.h"66#include "llvm/Analysis/TargetTransformInfo.h"67#include "llvm/Analysis/ValueTracking.h"68#include "llvm/IR/Instructions.h"69#include "llvm/IR/Operator.h"70#include "llvm/InitializePasses.h"71#include "llvm/Support/CommandLine.h"72#include "llvm/Support/Debug.h"73#include "llvm/Transforms/Scalar.h"74 75using namespace llvm;76 77#define DEBUG_TYPE "speculative-execution"78 79// The risk that speculation will not pay off increases with the80// number of instructions speculated, so we put a limit on that.81static cl::opt<unsigned> SpecExecMaxSpeculationCost(82    "spec-exec-max-speculation-cost", cl::init(7), cl::Hidden,83    cl::desc("Speculative execution is not applied to basic blocks where "84             "the cost of the instructions to speculatively execute "85             "exceeds this limit."));86 87// Speculating just a few instructions from a larger block tends not88// to be profitable and this limit prevents that. A reason for that is89// that small basic blocks are more likely to be candidates for90// further optimization.91static cl::opt<unsigned> SpecExecMaxNotHoisted(92    "spec-exec-max-not-hoisted", cl::init(5), cl::Hidden,93    cl::desc("Speculative execution is not applied to basic blocks where the "94             "number of instructions that would not be speculatively executed "95             "exceeds this limit."));96 97static cl::opt<bool> SpecExecOnlyIfDivergentTarget(98    "spec-exec-only-if-divergent-target", cl::init(false), cl::Hidden,99    cl::desc("Speculative execution is applied only to targets with divergent "100             "branches, even if the pass was configured to apply only to all "101             "targets."));102 103namespace {104 105class SpeculativeExecutionLegacyPass : public FunctionPass {106public:107  static char ID;108  explicit SpeculativeExecutionLegacyPass(bool OnlyIfDivergentTarget = false)109      : FunctionPass(ID), OnlyIfDivergentTarget(OnlyIfDivergentTarget ||110                                                SpecExecOnlyIfDivergentTarget),111        Impl(OnlyIfDivergentTarget) {}112 113  void getAnalysisUsage(AnalysisUsage &AU) const override;114  bool runOnFunction(Function &F) override;115 116  StringRef getPassName() const override {117    if (OnlyIfDivergentTarget)118      return "Speculatively execute instructions if target has divergent "119             "branches";120    return "Speculatively execute instructions";121  }122 123private:124  // Variable preserved purely for correct name printing.125  const bool OnlyIfDivergentTarget;126 127  SpeculativeExecutionPass Impl;128};129} // namespace130 131char SpeculativeExecutionLegacyPass::ID = 0;132INITIALIZE_PASS_BEGIN(SpeculativeExecutionLegacyPass, "speculative-execution",133                      "Speculatively execute instructions", false, false)134INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)135INITIALIZE_PASS_END(SpeculativeExecutionLegacyPass, "speculative-execution",136                    "Speculatively execute instructions", false, false)137 138void SpeculativeExecutionLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {139  AU.addRequired<TargetTransformInfoWrapperPass>();140  AU.addPreserved<GlobalsAAWrapperPass>();141  AU.setPreservesCFG();142}143 144bool SpeculativeExecutionLegacyPass::runOnFunction(Function &F) {145  if (skipFunction(F))146    return false;147 148  auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);149  return Impl.runImpl(F, TTI);150}151 152bool SpeculativeExecutionPass::runImpl(Function &F, TargetTransformInfo *TTI) {153  if (OnlyIfDivergentTarget && !TTI->hasBranchDivergence(&F)) {154    LLVM_DEBUG(dbgs() << "Not running SpeculativeExecution because "155                         "TTI->hasBranchDivergence() is false.\n");156    return false;157  }158 159  this->TTI = TTI;160  bool Changed = false;161  for (auto& B : F) {162    Changed |= runOnBasicBlock(B);163  }164  return Changed;165}166 167bool SpeculativeExecutionPass::runOnBasicBlock(BasicBlock &B) {168  BranchInst *BI = dyn_cast<BranchInst>(B.getTerminator());169  if (BI == nullptr)170    return false;171 172  if (BI->getNumSuccessors() != 2)173    return false;174  BasicBlock &Succ0 = *BI->getSuccessor(0);175  BasicBlock &Succ1 = *BI->getSuccessor(1);176 177  if (&B == &Succ0 || &B == &Succ1 || &Succ0 == &Succ1) {178    return false;179  }180 181  // Hoist from if-then (triangle).182  if (Succ0.getSinglePredecessor() != nullptr &&183      Succ0.getSingleSuccessor() == &Succ1) {184    return considerHoistingFromTo(Succ0, B);185  }186 187  // Hoist from if-else (triangle).188  if (Succ1.getSinglePredecessor() != nullptr &&189      Succ1.getSingleSuccessor() == &Succ0) {190    return considerHoistingFromTo(Succ1, B);191  }192 193  // Hoist from if-then-else (diamond), but only if it is equivalent to194  // an if-else or if-then due to one of the branches doing nothing.195  if (Succ0.getSinglePredecessor() != nullptr &&196      Succ1.getSinglePredecessor() != nullptr &&197      Succ1.getSingleSuccessor() != nullptr &&198      Succ1.getSingleSuccessor() != &B &&199      Succ1.getSingleSuccessor() == Succ0.getSingleSuccessor()) {200    // If a block has only one instruction, then that is a terminator201    // instruction so that the block does nothing. This does happen.202    if (Succ1.size() == 1) // equivalent to if-then203      return considerHoistingFromTo(Succ0, B);204    if (Succ0.size() == 1) // equivalent to if-else205      return considerHoistingFromTo(Succ1, B);206  }207 208  return false;209}210 211static InstructionCost ComputeSpeculationCost(const Instruction *I,212                                              const TargetTransformInfo &TTI) {213  switch (Operator::getOpcode(I)) {214    case Instruction::GetElementPtr:215    case Instruction::Add:216    case Instruction::Mul:217    case Instruction::And:218    case Instruction::Or:219    case Instruction::Select:220    case Instruction::Shl:221    case Instruction::Sub:222    case Instruction::LShr:223    case Instruction::AShr:224    case Instruction::Xor:225    case Instruction::ZExt:226    case Instruction::SExt:227    case Instruction::Call:228    case Instruction::BitCast:229    case Instruction::PtrToInt:230    case Instruction::PtrToAddr:231    case Instruction::IntToPtr:232    case Instruction::AddrSpaceCast:233    case Instruction::FPToUI:234    case Instruction::FPToSI:235    case Instruction::UIToFP:236    case Instruction::SIToFP:237    case Instruction::FPExt:238    case Instruction::FPTrunc:239    case Instruction::FAdd:240    case Instruction::FSub:241    case Instruction::FMul:242    case Instruction::FDiv:243    case Instruction::FRem:244    case Instruction::FNeg:245    case Instruction::ICmp:246    case Instruction::FCmp:247    case Instruction::Trunc:248    case Instruction::Freeze:249    case Instruction::ExtractElement:250    case Instruction::InsertElement:251    case Instruction::ShuffleVector:252    case Instruction::ExtractValue:253    case Instruction::InsertValue:254      return TTI.getInstructionCost(I, TargetTransformInfo::TCK_SizeAndLatency);255 256    default:257      return InstructionCost::getInvalid(); // Disallow anything not explicitly258                                            // listed.259  }260}261 262// Do not hoist any debug info intrinsics.263// ...264// if (cond) {265//   x = y * z;266//   foo();267// }268// ...269// -------- Which then becomes:270// ...271// if.then:272//   %x = mul i32 %y, %z273//   call void @llvm.dbg.value(%x, !"x", !DIExpression())274//   call void foo()275//276// SpeculativeExecution might decide to hoist the 'y * z' calculation277// out of the 'if' block, because it is more efficient that way, so the278// '%x = mul i32 %y, %z' moves to the block above. But it might also279// decide to hoist the 'llvm.dbg.value' call.280// This is incorrect, because even if we've moved the calculation of281// 'y * z', we should not see the value of 'x' change unless we282// actually go inside the 'if' block.283 284bool SpeculativeExecutionPass::considerHoistingFromTo(285    BasicBlock &FromBlock, BasicBlock &ToBlock) {286  SmallPtrSet<const Instruction *, 8> NotHoisted;287  auto HasNoUnhoistedInstr = [&NotHoisted](auto Values) {288    for (const Value *V : Values) {289      if (const auto *I = dyn_cast_or_null<Instruction>(V))290        if (NotHoisted.contains(I))291          return false;292    }293    return true;294  };295  auto AllPrecedingUsesFromBlockHoisted =296      [&HasNoUnhoistedInstr](const User *U) {297        return HasNoUnhoistedInstr(U->operand_values());298      };299 300  InstructionCost TotalSpeculationCost = 0;301  unsigned NotHoistedInstCount = 0;302  for (const auto &I : FromBlock) {303    const InstructionCost Cost = ComputeSpeculationCost(&I, *TTI);304    if (Cost.isValid() && isSafeToSpeculativelyExecute(&I) &&305        AllPrecedingUsesFromBlockHoisted(&I)) {306      TotalSpeculationCost += Cost;307      if (TotalSpeculationCost > SpecExecMaxSpeculationCost)308        return false;  // too much to hoist309    } else {310      NotHoistedInstCount++;311      if (NotHoistedInstCount > SpecExecMaxNotHoisted)312        return false; // too much left behind313      NotHoisted.insert(&I);314    }315  }316 317  for (auto I = FromBlock.begin(); I != FromBlock.end();) {318    // We have to increment I before moving Current as moving Current319    // changes the list that I is iterating through.320    auto Current = I;321    ++I;322    if (!NotHoisted.count(&*Current)) {323      Current->moveBefore(ToBlock.getTerminator()->getIterator());324      Current->dropLocation();325    }326  }327  return true;328}329 330FunctionPass *llvm::createSpeculativeExecutionPass() {331  return new SpeculativeExecutionLegacyPass();332}333 334FunctionPass *llvm::createSpeculativeExecutionIfHasBranchDivergencePass() {335  return new SpeculativeExecutionLegacyPass(/* OnlyIfDivergentTarget = */ true);336}337 338SpeculativeExecutionPass::SpeculativeExecutionPass(bool OnlyIfDivergentTarget)339    : OnlyIfDivergentTarget(OnlyIfDivergentTarget ||340                            SpecExecOnlyIfDivergentTarget) {}341 342PreservedAnalyses SpeculativeExecutionPass::run(Function &F,343                                                FunctionAnalysisManager &AM) {344  auto *TTI = &AM.getResult<TargetIRAnalysis>(F);345 346  bool Changed = runImpl(F, TTI);347 348  if (!Changed)349    return PreservedAnalyses::all();350  PreservedAnalyses PA;351  PA.preserveSet<CFGAnalyses>();352  return PA;353}354 355void SpeculativeExecutionPass::printPipeline(356    raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {357  static_cast<PassInfoMixin<SpeculativeExecutionPass> *>(this)->printPipeline(358      OS, MapClassName2PassName);359  OS << '<';360  if (OnlyIfDivergentTarget)361    OS << "only-if-divergent-target";362  OS << '>';363}364