lib/Transforms/Scalar/SpeculativeExecution.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-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This pass hoists instructions to enable speculative execution on
  10 // targets where branches are expensive. This is aimed at GPUs. It
  11 // currently works on simple if-then and if-then-else
  12 // patterns.
  13 //
  14 // Removing branches is not the only motivation for this
  15 // pass. E.g. consider this code and assume that there is no
  16 // 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 into
  24 //
  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 to
  33 //
  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 enable
  41 // further optimization.
  42 //
  43 // This pass is more aggressive than the function SpeculativeyExecuteBB in
  44 // SimplifyCFG. SimplifyCFG will not speculate if no selects are introduced and
  45 // it will speculate at most one instruction. It also will not speculate if
  46 // 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 seems
  48 // aimed at introducing cheap selects, while this pass is intended to do more
  49 // aggressive speculation while counting on later passes to either capitalize on
  50 // that or clean it up.
  51 //
  52 // If the pass was created by calling
  53 // createSpeculativeExecutionIfHasBranchDivergencePass or the
  54 // -spec-exec-only-if-divergent-target option is present, this pass only has an
  55 // 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, but
  59 // 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/ValueTracking.h"
  67 #include "llvm/IR/Instructions.h"
  68 #include "llvm/IR/Module.h"
  69 #include "llvm/IR/Operator.h"
  70 #include "llvm/Support/CommandLine.h"
  71 #include "llvm/Support/Debug.h"
  72
  73 using namespace llvm;
  74
  75 #define DEBUG_TYPE "speculative-execution"
  76
  77 // The risk that speculation will not pay off increases with the
  78 // number of instructions speculated, so we put a limit on that.
  79 static cl::opt<unsigned> SpecExecMaxSpeculationCost(
  80     "spec-exec-max-speculation-cost", cl::init(7), cl::Hidden,
  81     cl::desc("Speculative execution is not applied to basic blocks where "
  82              "the cost of the instructions to speculatively execute "
  83              "exceeds this limit."));
  84
  85 // Speculating just a few instructions from a larger block tends not
  86 // to be profitable and this limit prevents that. A reason for that is
  87 // that small basic blocks are more likely to be candidates for
  88 // further optimization.
  89 static cl::opt<unsigned> SpecExecMaxNotHoisted(
  90     "spec-exec-max-not-hoisted", cl::init(5), cl::Hidden,
  91     cl::desc("Speculative execution is not applied to basic blocks where the "
  92              "number of instructions that would not be speculatively executed "
  93              "exceeds this limit."));
  94
  95 static cl::opt<bool> SpecExecOnlyIfDivergentTarget(
  96     "spec-exec-only-if-divergent-target", cl::init(false), cl::Hidden,
  97     cl::desc("Speculative execution is applied only to targets with divergent "
  98              "branches, even if the pass was configured to apply only to all "
  99              "targets."));
 100
 101 namespace {
 102
 103 class SpeculativeExecutionLegacyPass : public FunctionPass {
 104 public:
 105   static char ID;
 106   explicit SpeculativeExecutionLegacyPass(bool OnlyIfDivergentTarget = false)
 107       : FunctionPass(ID), OnlyIfDivergentTarget(OnlyIfDivergentTarget ||
 108                                                 SpecExecOnlyIfDivergentTarget),
 109         Impl(OnlyIfDivergentTarget) {}
 110
 111   void getAnalysisUsage(AnalysisUsage &AU) const override;
 112   bool runOnFunction(Function &F) override;
 113
 114   StringRef getPassName() const override {
 115     if (OnlyIfDivergentTarget)
 116       return "Speculatively execute instructions if target has divergent "
 117              "branches";
 118     return "Speculatively execute instructions";
 119   }
 120
 121 private:
 122   // Variable preserved purely for correct name printing.
 123   const bool OnlyIfDivergentTarget;
 124
 125   SpeculativeExecutionPass Impl;
 126 };
 127 } // namespace
 128
 129 char SpeculativeExecutionLegacyPass::ID = 0;
 130 INITIALIZE_PASS_BEGIN(SpeculativeExecutionLegacyPass, "speculative-execution",
 131                       "Speculatively execute instructions", false, false)
 132 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
 133 INITIALIZE_PASS_END(SpeculativeExecutionLegacyPass, "speculative-execution",
 134                     "Speculatively execute instructions", false, false)
 135
 136 void SpeculativeExecutionLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
 137   AU.addRequired<TargetTransformInfoWrapperPass>();
 138   AU.addPreserved<GlobalsAAWrapperPass>();
 139   AU.setPreservesCFG();
 140 }
 141
 142 bool SpeculativeExecutionLegacyPass::runOnFunction(Function &F) {
 143   if (skipFunction(F))
 144     return false;
 145
 146   auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
 147   return Impl.runImpl(F, TTI);
 148 }
 149
 150 namespace llvm {
 151
 152 bool SpeculativeExecutionPass::runImpl(Function &F, TargetTransformInfo *TTI) {
 153   if (OnlyIfDivergentTarget && !TTI->hasBranchDivergence()) {
 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
 167 bool 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 to
 194   // 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 terminator
 201     // instruction so that the block does nothing. This does happen.
 202     if (Succ1.size() == 1) // equivalent to if-then
 203       return considerHoistingFromTo(Succ0, B);
 204     if (Succ0.size() == 1) // equivalent to if-else
 205       return considerHoistingFromTo(Succ1, B);
 206   }
 207
 208   return false;
 209 }
 210
 211 static unsigned 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::IntToPtr:
 231     case Instruction::AddrSpaceCast:
 232     case Instruction::FPToUI:
 233     case Instruction::FPToSI:
 234     case Instruction::UIToFP:
 235     case Instruction::SIToFP:
 236     case Instruction::FPExt:
 237     case Instruction::FPTrunc:
 238     case Instruction::FAdd:
 239     case Instruction::FSub:
 240     case Instruction::FMul:
 241     case Instruction::FDiv:
 242     case Instruction::FRem:
 243     case Instruction::ICmp:
 244     case Instruction::FCmp:
 245       return TTI.getUserCost(I);
 246
 247     default:
 248       return UINT_MAX; // Disallow anything not whitelisted.
 249   }
 250 }
 251
 252 bool SpeculativeExecutionPass::considerHoistingFromTo(
 253     BasicBlock &FromBlock, BasicBlock &ToBlock) {
 254   SmallPtrSet<const Instruction *, 8> NotHoisted;
 255   const auto AllPrecedingUsesFromBlockHoisted = [&NotHoisted](User *U) {
 256     for (Value* V : U->operand_values()) {
 257       if (Instruction *I = dyn_cast<Instruction>(V)) {
 258         if (NotHoisted.count(I) > 0)
 259           return false;
 260       }
 261     }
 262     return true;
 263   };
 264
 265   unsigned TotalSpeculationCost = 0;
 266   for (auto& I : FromBlock) {
 267     const unsigned Cost = ComputeSpeculationCost(&I, *TTI);
 268     if (Cost != UINT_MAX && isSafeToSpeculativelyExecute(&I) &&
 269         AllPrecedingUsesFromBlockHoisted(&I)) {
 270       TotalSpeculationCost += Cost;
 271       if (TotalSpeculationCost > SpecExecMaxSpeculationCost)
 272         return false;  // too much to hoist
 273     } else {
 274       NotHoisted.insert(&I);
 275       if (NotHoisted.size() > SpecExecMaxNotHoisted)
 276         return false; // too much left behind
 277     }
 278   }
 279
 280   if (TotalSpeculationCost == 0)
 281     return false; // nothing to hoist
 282
 283   for (auto I = FromBlock.begin(); I != FromBlock.end();) {
 284     // We have to increment I before moving Current as moving Current
 285     // changes the list that I is iterating through.
 286     auto Current = I;
 287     ++I;
 288     if (!NotHoisted.count(&*Current)) {
 289       Current->moveBefore(ToBlock.getTerminator());
 290     }
 291   }
 292   return true;
 293 }
 294
 295 FunctionPass *createSpeculativeExecutionPass() {
 296   return new SpeculativeExecutionLegacyPass();
 297 }
 298
 299 FunctionPass *createSpeculativeExecutionIfHasBranchDivergencePass() {
 300   return new SpeculativeExecutionLegacyPass(/* OnlyIfDivergentTarget = */ true);
 301 }
 302
 303 SpeculativeExecutionPass::SpeculativeExecutionPass(bool OnlyIfDivergentTarget)
 304     : OnlyIfDivergentTarget(OnlyIfDivergentTarget ||
 305                             SpecExecOnlyIfDivergentTarget) {}
 306
 307 PreservedAnalyses SpeculativeExecutionPass::run(Function &F,
 308                                                 FunctionAnalysisManager &AM) {
 309   auto *TTI = &AM.getResult<TargetIRAnalysis>(F);
 310
 311   bool Changed = runImpl(F, TTI);
 312
 313   if (!Changed)
 314     return PreservedAnalyses::all();
 315   PreservedAnalyses PA;
 316   PA.preserve<GlobalsAA>();
 317   PA.preserveSet<CFGAnalyses>();
 318   return PA;
 319 }
 320 }  // namespace llvm