1 //===--- SelectOptimize.cpp - Convert select to branches if profitable ---===// 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 converts selects to conditional jumps when profitable. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ADT/SmallVector.h" 14 #include "llvm/ADT/Statistic.h" 15 #include "llvm/Analysis/BlockFrequencyInfo.h" 16 #include "llvm/Analysis/BranchProbabilityInfo.h" 17 #include "llvm/Analysis/LoopInfo.h" 18 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 19 #include "llvm/Analysis/ProfileSummaryInfo.h" 20 #include "llvm/Analysis/TargetTransformInfo.h" 21 #include "llvm/CodeGen/Passes.h" 22 #include "llvm/CodeGen/TargetLowering.h" 23 #include "llvm/CodeGen/TargetPassConfig.h" 24 #include "llvm/CodeGen/TargetSchedule.h" 25 #include "llvm/CodeGen/TargetSubtargetInfo.h" 26 #include "llvm/IR/BasicBlock.h" 27 #include "llvm/IR/Dominators.h" 28 #include "llvm/IR/Function.h" 29 #include "llvm/IR/IRBuilder.h" 30 #include "llvm/IR/Instruction.h" 31 #include "llvm/IR/PatternMatch.h" 32 #include "llvm/IR/ProfDataUtils.h" 33 #include "llvm/InitializePasses.h" 34 #include "llvm/Pass.h" 35 #include "llvm/Support/ScaledNumber.h" 36 #include "llvm/Target/TargetMachine.h" 37 #include "llvm/Transforms/Utils/SizeOpts.h" 38 #include <algorithm> 39 #include <memory> 40 #include <queue> 41 #include <stack> 42 #include <string> 43 44 using namespace llvm; 45 46 #define DEBUG_TYPE "select-optimize" 47 48 STATISTIC(NumSelectOptAnalyzed, 49 "Number of select groups considered for conversion to branch"); 50 STATISTIC(NumSelectConvertedExpColdOperand, 51 "Number of select groups converted due to expensive cold operand"); 52 STATISTIC(NumSelectConvertedHighPred, 53 "Number of select groups converted due to high-predictability"); 54 STATISTIC(NumSelectUnPred, 55 "Number of select groups not converted due to unpredictability"); 56 STATISTIC(NumSelectColdBB, 57 "Number of select groups not converted due to cold basic block"); 58 STATISTIC(NumSelectConvertedLoop, 59 "Number of select groups converted due to loop-level analysis"); 60 STATISTIC(NumSelectsConverted, "Number of selects converted"); 61 62 static cl::opt<unsigned> ColdOperandThreshold( 63 "cold-operand-threshold", 64 cl::desc("Maximum frequency of path for an operand to be considered cold."), 65 cl::init(20), cl::Hidden); 66 67 static cl::opt<unsigned> ColdOperandMaxCostMultiplier( 68 "cold-operand-max-cost-multiplier", 69 cl::desc("Maximum cost multiplier of TCC_expensive for the dependence " 70 "slice of a cold operand to be considered inexpensive."), 71 cl::init(1), cl::Hidden); 72 73 static cl::opt<unsigned> 74 GainGradientThreshold("select-opti-loop-gradient-gain-threshold", 75 cl::desc("Gradient gain threshold (%)."), 76 cl::init(25), cl::Hidden); 77 78 static cl::opt<unsigned> 79 GainCycleThreshold("select-opti-loop-cycle-gain-threshold", 80 cl::desc("Minimum gain per loop (in cycles) threshold."), 81 cl::init(4), cl::Hidden); 82 83 static cl::opt<unsigned> GainRelativeThreshold( 84 "select-opti-loop-relative-gain-threshold", 85 cl::desc( 86 "Minimum relative gain per loop threshold (1/X). Defaults to 12.5%"), 87 cl::init(8), cl::Hidden); 88 89 static cl::opt<unsigned> MispredictDefaultRate( 90 "mispredict-default-rate", cl::Hidden, cl::init(25), 91 cl::desc("Default mispredict rate (initialized to 25%).")); 92 93 static cl::opt<bool> 94 DisableLoopLevelHeuristics("disable-loop-level-heuristics", cl::Hidden, 95 cl::init(false), 96 cl::desc("Disable loop-level heuristics.")); 97 98 namespace { 99 100 class SelectOptimize : public FunctionPass { 101 const TargetMachine *TM = nullptr; 102 const TargetSubtargetInfo *TSI = nullptr; 103 const TargetLowering *TLI = nullptr; 104 const TargetTransformInfo *TTI = nullptr; 105 const LoopInfo *LI = nullptr; 106 DominatorTree *DT = nullptr; 107 std::unique_ptr<BlockFrequencyInfo> BFI; 108 std::unique_ptr<BranchProbabilityInfo> BPI; 109 ProfileSummaryInfo *PSI = nullptr; 110 OptimizationRemarkEmitter *ORE = nullptr; 111 TargetSchedModel TSchedModel; 112 113 public: 114 static char ID; 115 116 SelectOptimize() : FunctionPass(ID) { 117 initializeSelectOptimizePass(*PassRegistry::getPassRegistry()); 118 } 119 120 bool runOnFunction(Function &F) override; 121 122 void getAnalysisUsage(AnalysisUsage &AU) const override { 123 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 124 AU.addRequired<TargetPassConfig>(); 125 AU.addRequired<TargetTransformInfoWrapperPass>(); 126 AU.addRequired<DominatorTreeWrapperPass>(); 127 AU.addRequired<LoopInfoWrapperPass>(); 128 AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); 129 } 130 131 private: 132 // Select groups consist of consecutive select instructions with the same 133 // condition. 134 using SelectGroup = SmallVector<SelectInst *, 2>; 135 using SelectGroups = SmallVector<SelectGroup, 2>; 136 137 using Scaled64 = ScaledNumber<uint64_t>; 138 139 struct CostInfo { 140 /// Predicated cost (with selects as conditional moves). 141 Scaled64 PredCost; 142 /// Non-predicated cost (with selects converted to branches). 143 Scaled64 NonPredCost; 144 }; 145 146 // Converts select instructions of a function to conditional jumps when deemed 147 // profitable. Returns true if at least one select was converted. 148 bool optimizeSelects(Function &F); 149 150 // Heuristics for determining which select instructions can be profitably 151 // conveted to branches. Separate heuristics for selects in inner-most loops 152 // and the rest of code regions (base heuristics for non-inner-most loop 153 // regions). 154 void optimizeSelectsBase(Function &F, SelectGroups &ProfSIGroups); 155 void optimizeSelectsInnerLoops(Function &F, SelectGroups &ProfSIGroups); 156 157 // Converts to branches the select groups that were deemed 158 // profitable-to-convert. 159 void convertProfitableSIGroups(SelectGroups &ProfSIGroups); 160 161 // Splits selects of a given basic block into select groups. 162 void collectSelectGroups(BasicBlock &BB, SelectGroups &SIGroups); 163 164 // Determines for which select groups it is profitable converting to branches 165 // (base and inner-most-loop heuristics). 166 void findProfitableSIGroupsBase(SelectGroups &SIGroups, 167 SelectGroups &ProfSIGroups); 168 void findProfitableSIGroupsInnerLoops(const Loop *L, SelectGroups &SIGroups, 169 SelectGroups &ProfSIGroups); 170 171 // Determines if a select group should be converted to a branch (base 172 // heuristics). 173 bool isConvertToBranchProfitableBase(const SmallVector<SelectInst *, 2> &ASI); 174 175 // Returns true if there are expensive instructions in the cold value 176 // operand's (if any) dependence slice of any of the selects of the given 177 // group. 178 bool hasExpensiveColdOperand(const SmallVector<SelectInst *, 2> &ASI); 179 180 // For a given source instruction, collect its backwards dependence slice 181 // consisting of instructions exclusively computed for producing the operands 182 // of the source instruction. 183 void getExclBackwardsSlice(Instruction *I, std::stack<Instruction *> &Slice, 184 Instruction *SI, bool ForSinking = false); 185 186 // Returns true if the condition of the select is highly predictable. 187 bool isSelectHighlyPredictable(const SelectInst *SI); 188 189 // Loop-level checks to determine if a non-predicated version (with branches) 190 // of the given loop is more profitable than its predicated version. 191 bool checkLoopHeuristics(const Loop *L, const CostInfo LoopDepth[2]); 192 193 // Computes instruction and loop-critical-path costs for both the predicated 194 // and non-predicated version of the given loop. 195 bool computeLoopCosts(const Loop *L, const SelectGroups &SIGroups, 196 DenseMap<const Instruction *, CostInfo> &InstCostMap, 197 CostInfo *LoopCost); 198 199 // Returns a set of all the select instructions in the given select groups. 200 SmallPtrSet<const Instruction *, 2> getSIset(const SelectGroups &SIGroups); 201 202 // Returns the latency cost of a given instruction. 203 std::optional<uint64_t> computeInstCost(const Instruction *I); 204 205 // Returns the misprediction cost of a given select when converted to branch. 206 Scaled64 getMispredictionCost(const SelectInst *SI, const Scaled64 CondCost); 207 208 // Returns the cost of a branch when the prediction is correct. 209 Scaled64 getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost, 210 const SelectInst *SI); 211 212 // Returns true if the target architecture supports lowering a given select. 213 bool isSelectKindSupported(SelectInst *SI); 214 }; 215 } // namespace 216 217 char SelectOptimize::ID = 0; 218 219 INITIALIZE_PASS_BEGIN(SelectOptimize, DEBUG_TYPE, "Optimize selects", false, 220 false) 221 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 222 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 223 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 224 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 225 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 226 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) 227 INITIALIZE_PASS_END(SelectOptimize, DEBUG_TYPE, "Optimize selects", false, 228 false) 229 230 FunctionPass *llvm::createSelectOptimizePass() { return new SelectOptimize(); } 231 232 bool SelectOptimize::runOnFunction(Function &F) { 233 TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); 234 TSI = TM->getSubtargetImpl(F); 235 TLI = TSI->getTargetLowering(); 236 237 // If none of the select types is supported then skip this pass. 238 // This is an optimization pass. Legality issues will be handled by 239 // instruction selection. 240 if (!TLI->isSelectSupported(TargetLowering::ScalarValSelect) && 241 !TLI->isSelectSupported(TargetLowering::ScalarCondVectorVal) && 242 !TLI->isSelectSupported(TargetLowering::VectorMaskSelect)) 243 return false; 244 245 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 246 247 if (!TTI->enableSelectOptimize()) 248 return false; 249 250 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 251 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 252 BPI.reset(new BranchProbabilityInfo(F, *LI)); 253 BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI)); 254 PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 255 ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); 256 TSchedModel.init(TSI); 257 258 // When optimizing for size, selects are preferable over branches. 259 if (F.hasOptSize() || llvm::shouldOptimizeForSize(&F, PSI, BFI.get())) 260 return false; 261 262 return optimizeSelects(F); 263 } 264 265 bool SelectOptimize::optimizeSelects(Function &F) { 266 // Determine for which select groups it is profitable converting to branches. 267 SelectGroups ProfSIGroups; 268 // Base heuristics apply only to non-loops and outer loops. 269 optimizeSelectsBase(F, ProfSIGroups); 270 // Separate heuristics for inner-most loops. 271 optimizeSelectsInnerLoops(F, ProfSIGroups); 272 273 // Convert to branches the select groups that were deemed 274 // profitable-to-convert. 275 convertProfitableSIGroups(ProfSIGroups); 276 277 // Code modified if at least one select group was converted. 278 return !ProfSIGroups.empty(); 279 } 280 281 void SelectOptimize::optimizeSelectsBase(Function &F, 282 SelectGroups &ProfSIGroups) { 283 // Collect all the select groups. 284 SelectGroups SIGroups; 285 for (BasicBlock &BB : F) { 286 // Base heuristics apply only to non-loops and outer loops. 287 Loop *L = LI->getLoopFor(&BB); 288 if (L && L->isInnermost()) 289 continue; 290 collectSelectGroups(BB, SIGroups); 291 } 292 293 // Determine for which select groups it is profitable converting to branches. 294 findProfitableSIGroupsBase(SIGroups, ProfSIGroups); 295 } 296 297 void SelectOptimize::optimizeSelectsInnerLoops(Function &F, 298 SelectGroups &ProfSIGroups) { 299 SmallVector<Loop *, 4> Loops(LI->begin(), LI->end()); 300 // Need to check size on each iteration as we accumulate child loops. 301 for (unsigned long i = 0; i < Loops.size(); ++i) 302 for (Loop *ChildL : Loops[i]->getSubLoops()) 303 Loops.push_back(ChildL); 304 305 for (Loop *L : Loops) { 306 if (!L->isInnermost()) 307 continue; 308 309 SelectGroups SIGroups; 310 for (BasicBlock *BB : L->getBlocks()) 311 collectSelectGroups(*BB, SIGroups); 312 313 findProfitableSIGroupsInnerLoops(L, SIGroups, ProfSIGroups); 314 } 315 } 316 317 /// If \p isTrue is true, return the true value of \p SI, otherwise return 318 /// false value of \p SI. If the true/false value of \p SI is defined by any 319 /// select instructions in \p Selects, look through the defining select 320 /// instruction until the true/false value is not defined in \p Selects. 321 static Value * 322 getTrueOrFalseValue(SelectInst *SI, bool isTrue, 323 const SmallPtrSet<const Instruction *, 2> &Selects) { 324 Value *V = nullptr; 325 for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI); 326 DefSI = dyn_cast<SelectInst>(V)) { 327 assert(DefSI->getCondition() == SI->getCondition() && 328 "The condition of DefSI does not match with SI"); 329 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue()); 330 } 331 assert(V && "Failed to get select true/false value"); 332 return V; 333 } 334 335 void SelectOptimize::convertProfitableSIGroups(SelectGroups &ProfSIGroups) { 336 for (SelectGroup &ASI : ProfSIGroups) { 337 // The code transformation here is a modified version of the sinking 338 // transformation in CodeGenPrepare::optimizeSelectInst with a more 339 // aggressive strategy of which instructions to sink. 340 // 341 // TODO: eliminate the redundancy of logic transforming selects to branches 342 // by removing CodeGenPrepare::optimizeSelectInst and optimizing here 343 // selects for all cases (with and without profile information). 344 345 // Transform a sequence like this: 346 // start: 347 // %cmp = cmp uge i32 %a, %b 348 // %sel = select i1 %cmp, i32 %c, i32 %d 349 // 350 // Into: 351 // start: 352 // %cmp = cmp uge i32 %a, %b 353 // %cmp.frozen = freeze %cmp 354 // br i1 %cmp.frozen, label %select.true, label %select.false 355 // select.true: 356 // br label %select.end 357 // select.false: 358 // br label %select.end 359 // select.end: 360 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ] 361 // 362 // %cmp should be frozen, otherwise it may introduce undefined behavior. 363 // In addition, we may sink instructions that produce %c or %d into the 364 // destination(s) of the new branch. 365 // If the true or false blocks do not contain a sunken instruction, that 366 // block and its branch may be optimized away. In that case, one side of the 367 // first branch will point directly to select.end, and the corresponding PHI 368 // predecessor block will be the start block. 369 370 // Find all the instructions that can be soundly sunk to the true/false 371 // blocks. These are instructions that are computed solely for producing the 372 // operands of the select instructions in the group and can be sunk without 373 // breaking the semantics of the LLVM IR (e.g., cannot sink instructions 374 // with side effects). 375 SmallVector<std::stack<Instruction *>, 2> TrueSlices, FalseSlices; 376 typedef std::stack<Instruction *>::size_type StackSizeType; 377 StackSizeType maxTrueSliceLen = 0, maxFalseSliceLen = 0; 378 for (SelectInst *SI : ASI) { 379 // For each select, compute the sinkable dependence chains of the true and 380 // false operands. 381 if (auto *TI = dyn_cast<Instruction>(SI->getTrueValue())) { 382 std::stack<Instruction *> TrueSlice; 383 getExclBackwardsSlice(TI, TrueSlice, SI, true); 384 maxTrueSliceLen = std::max(maxTrueSliceLen, TrueSlice.size()); 385 TrueSlices.push_back(TrueSlice); 386 } 387 if (auto *FI = dyn_cast<Instruction>(SI->getFalseValue())) { 388 std::stack<Instruction *> FalseSlice; 389 getExclBackwardsSlice(FI, FalseSlice, SI, true); 390 maxFalseSliceLen = std::max(maxFalseSliceLen, FalseSlice.size()); 391 FalseSlices.push_back(FalseSlice); 392 } 393 } 394 // In the case of multiple select instructions in the same group, the order 395 // of non-dependent instructions (instructions of different dependence 396 // slices) in the true/false blocks appears to affect performance. 397 // Interleaving the slices seems to experimentally be the optimal approach. 398 // This interleaving scheduling allows for more ILP (with a natural downside 399 // of increasing a bit register pressure) compared to a simple ordering of 400 // one whole chain after another. One would expect that this ordering would 401 // not matter since the scheduling in the backend of the compiler would 402 // take care of it, but apparently the scheduler fails to deliver optimal 403 // ILP with a naive ordering here. 404 SmallVector<Instruction *, 2> TrueSlicesInterleaved, FalseSlicesInterleaved; 405 for (StackSizeType IS = 0; IS < maxTrueSliceLen; ++IS) { 406 for (auto &S : TrueSlices) { 407 if (!S.empty()) { 408 TrueSlicesInterleaved.push_back(S.top()); 409 S.pop(); 410 } 411 } 412 } 413 for (StackSizeType IS = 0; IS < maxFalseSliceLen; ++IS) { 414 for (auto &S : FalseSlices) { 415 if (!S.empty()) { 416 FalseSlicesInterleaved.push_back(S.top()); 417 S.pop(); 418 } 419 } 420 } 421 422 // We split the block containing the select(s) into two blocks. 423 SelectInst *SI = ASI.front(); 424 SelectInst *LastSI = ASI.back(); 425 BasicBlock *StartBlock = SI->getParent(); 426 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI)); 427 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end"); 428 BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock).getFrequency()); 429 // Delete the unconditional branch that was just created by the split. 430 StartBlock->getTerminator()->eraseFromParent(); 431 432 // Move any debug/pseudo instructions that were in-between the select 433 // group to the newly-created end block. 434 SmallVector<Instruction *, 2> DebugPseudoINS; 435 auto DIt = SI->getIterator(); 436 while (&*DIt != LastSI) { 437 if (DIt->isDebugOrPseudoInst()) 438 DebugPseudoINS.push_back(&*DIt); 439 DIt++; 440 } 441 for (auto *DI : DebugPseudoINS) { 442 DI->moveBefore(&*EndBlock->getFirstInsertionPt()); 443 } 444 445 // These are the new basic blocks for the conditional branch. 446 // At least one will become an actual new basic block. 447 BasicBlock *TrueBlock = nullptr, *FalseBlock = nullptr; 448 BranchInst *TrueBranch = nullptr, *FalseBranch = nullptr; 449 if (!TrueSlicesInterleaved.empty()) { 450 TrueBlock = BasicBlock::Create(LastSI->getContext(), "select.true.sink", 451 EndBlock->getParent(), EndBlock); 452 TrueBranch = BranchInst::Create(EndBlock, TrueBlock); 453 TrueBranch->setDebugLoc(LastSI->getDebugLoc()); 454 for (Instruction *TrueInst : TrueSlicesInterleaved) 455 TrueInst->moveBefore(TrueBranch); 456 } 457 if (!FalseSlicesInterleaved.empty()) { 458 FalseBlock = BasicBlock::Create(LastSI->getContext(), "select.false.sink", 459 EndBlock->getParent(), EndBlock); 460 FalseBranch = BranchInst::Create(EndBlock, FalseBlock); 461 FalseBranch->setDebugLoc(LastSI->getDebugLoc()); 462 for (Instruction *FalseInst : FalseSlicesInterleaved) 463 FalseInst->moveBefore(FalseBranch); 464 } 465 // If there was nothing to sink, then arbitrarily choose the 'false' side 466 // for a new input value to the PHI. 467 if (TrueBlock == FalseBlock) { 468 assert(TrueBlock == nullptr && 469 "Unexpected basic block transform while optimizing select"); 470 471 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false", 472 EndBlock->getParent(), EndBlock); 473 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock); 474 FalseBranch->setDebugLoc(SI->getDebugLoc()); 475 } 476 477 // Insert the real conditional branch based on the original condition. 478 // If we did not create a new block for one of the 'true' or 'false' paths 479 // of the condition, it means that side of the branch goes to the end block 480 // directly and the path originates from the start block from the point of 481 // view of the new PHI. 482 BasicBlock *TT, *FT; 483 if (TrueBlock == nullptr) { 484 TT = EndBlock; 485 FT = FalseBlock; 486 TrueBlock = StartBlock; 487 } else if (FalseBlock == nullptr) { 488 TT = TrueBlock; 489 FT = EndBlock; 490 FalseBlock = StartBlock; 491 } else { 492 TT = TrueBlock; 493 FT = FalseBlock; 494 } 495 IRBuilder<> IB(SI); 496 auto *CondFr = 497 IB.CreateFreeze(SI->getCondition(), SI->getName() + ".frozen"); 498 IB.CreateCondBr(CondFr, TT, FT, SI); 499 500 SmallPtrSet<const Instruction *, 2> INS; 501 INS.insert(ASI.begin(), ASI.end()); 502 // Use reverse iterator because later select may use the value of the 503 // earlier select, and we need to propagate value through earlier select 504 // to get the PHI operand. 505 for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) { 506 SelectInst *SI = *It; 507 // The select itself is replaced with a PHI Node. 508 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front()); 509 PN->takeName(SI); 510 PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock); 511 PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock); 512 PN->setDebugLoc(SI->getDebugLoc()); 513 514 SI->replaceAllUsesWith(PN); 515 SI->eraseFromParent(); 516 INS.erase(SI); 517 ++NumSelectsConverted; 518 } 519 } 520 } 521 522 static bool isSpecialSelect(SelectInst *SI) { 523 using namespace llvm::PatternMatch; 524 525 // If the select is a logical-and/logical-or then it is better treated as a 526 // and/or by the backend. 527 if (match(SI, m_CombineOr(m_LogicalAnd(m_Value(), m_Value()), 528 m_LogicalOr(m_Value(), m_Value())))) 529 return true; 530 531 return false; 532 } 533 534 void SelectOptimize::collectSelectGroups(BasicBlock &BB, 535 SelectGroups &SIGroups) { 536 BasicBlock::iterator BBIt = BB.begin(); 537 while (BBIt != BB.end()) { 538 Instruction *I = &*BBIt++; 539 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 540 if (isSpecialSelect(SI)) 541 continue; 542 543 SelectGroup SIGroup; 544 SIGroup.push_back(SI); 545 while (BBIt != BB.end()) { 546 Instruction *NI = &*BBIt; 547 SelectInst *NSI = dyn_cast<SelectInst>(NI); 548 if (NSI && SI->getCondition() == NSI->getCondition()) { 549 SIGroup.push_back(NSI); 550 } else if (!NI->isDebugOrPseudoInst()) { 551 // Debug/pseudo instructions should be skipped and not prevent the 552 // formation of a select group. 553 break; 554 } 555 ++BBIt; 556 } 557 558 // If the select type is not supported, no point optimizing it. 559 // Instruction selection will take care of it. 560 if (!isSelectKindSupported(SI)) 561 continue; 562 563 SIGroups.push_back(SIGroup); 564 } 565 } 566 } 567 568 void SelectOptimize::findProfitableSIGroupsBase(SelectGroups &SIGroups, 569 SelectGroups &ProfSIGroups) { 570 for (SelectGroup &ASI : SIGroups) { 571 ++NumSelectOptAnalyzed; 572 if (isConvertToBranchProfitableBase(ASI)) 573 ProfSIGroups.push_back(ASI); 574 } 575 } 576 577 static void EmitAndPrintRemark(OptimizationRemarkEmitter *ORE, 578 DiagnosticInfoOptimizationBase &Rem) { 579 LLVM_DEBUG(dbgs() << Rem.getMsg() << "\n"); 580 ORE->emit(Rem); 581 } 582 583 void SelectOptimize::findProfitableSIGroupsInnerLoops( 584 const Loop *L, SelectGroups &SIGroups, SelectGroups &ProfSIGroups) { 585 NumSelectOptAnalyzed += SIGroups.size(); 586 // For each select group in an inner-most loop, 587 // a branch is more preferable than a select/conditional-move if: 588 // i) conversion to branches for all the select groups of the loop satisfies 589 // loop-level heuristics including reducing the loop's critical path by 590 // some threshold (see SelectOptimize::checkLoopHeuristics); and 591 // ii) the total cost of the select group is cheaper with a branch compared 592 // to its predicated version. The cost is in terms of latency and the cost 593 // of a select group is the cost of its most expensive select instruction 594 // (assuming infinite resources and thus fully leveraging available ILP). 595 596 DenseMap<const Instruction *, CostInfo> InstCostMap; 597 CostInfo LoopCost[2] = {{Scaled64::getZero(), Scaled64::getZero()}, 598 {Scaled64::getZero(), Scaled64::getZero()}}; 599 if (!computeLoopCosts(L, SIGroups, InstCostMap, LoopCost) || 600 !checkLoopHeuristics(L, LoopCost)) { 601 return; 602 } 603 604 for (SelectGroup &ASI : SIGroups) { 605 // Assuming infinite resources, the cost of a group of instructions is the 606 // cost of the most expensive instruction of the group. 607 Scaled64 SelectCost = Scaled64::getZero(), BranchCost = Scaled64::getZero(); 608 for (SelectInst *SI : ASI) { 609 SelectCost = std::max(SelectCost, InstCostMap[SI].PredCost); 610 BranchCost = std::max(BranchCost, InstCostMap[SI].NonPredCost); 611 } 612 if (BranchCost < SelectCost) { 613 OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", ASI.front()); 614 OR << "Profitable to convert to branch (loop analysis). BranchCost=" 615 << BranchCost.toString() << ", SelectCost=" << SelectCost.toString() 616 << ". "; 617 EmitAndPrintRemark(ORE, OR); 618 ++NumSelectConvertedLoop; 619 ProfSIGroups.push_back(ASI); 620 } else { 621 OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", ASI.front()); 622 ORmiss << "Select is more profitable (loop analysis). BranchCost=" 623 << BranchCost.toString() 624 << ", SelectCost=" << SelectCost.toString() << ". "; 625 EmitAndPrintRemark(ORE, ORmiss); 626 } 627 } 628 } 629 630 bool SelectOptimize::isConvertToBranchProfitableBase( 631 const SmallVector<SelectInst *, 2> &ASI) { 632 SelectInst *SI = ASI.front(); 633 LLVM_DEBUG(dbgs() << "Analyzing select group containing " << *SI << "\n"); 634 OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", SI); 635 OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", SI); 636 637 // Skip cold basic blocks. Better to optimize for size for cold blocks. 638 if (PSI->isColdBlock(SI->getParent(), BFI.get())) { 639 ++NumSelectColdBB; 640 ORmiss << "Not converted to branch because of cold basic block. "; 641 EmitAndPrintRemark(ORE, ORmiss); 642 return false; 643 } 644 645 // If unpredictable, branch form is less profitable. 646 if (SI->getMetadata(LLVMContext::MD_unpredictable)) { 647 ++NumSelectUnPred; 648 ORmiss << "Not converted to branch because of unpredictable branch. "; 649 EmitAndPrintRemark(ORE, ORmiss); 650 return false; 651 } 652 653 // If highly predictable, branch form is more profitable, unless a 654 // predictable select is inexpensive in the target architecture. 655 if (isSelectHighlyPredictable(SI) && TLI->isPredictableSelectExpensive()) { 656 ++NumSelectConvertedHighPred; 657 OR << "Converted to branch because of highly predictable branch. "; 658 EmitAndPrintRemark(ORE, OR); 659 return true; 660 } 661 662 // Look for expensive instructions in the cold operand's (if any) dependence 663 // slice of any of the selects in the group. 664 if (hasExpensiveColdOperand(ASI)) { 665 ++NumSelectConvertedExpColdOperand; 666 OR << "Converted to branch because of expensive cold operand."; 667 EmitAndPrintRemark(ORE, OR); 668 return true; 669 } 670 671 ORmiss << "Not profitable to convert to branch (base heuristic)."; 672 EmitAndPrintRemark(ORE, ORmiss); 673 return false; 674 } 675 676 static InstructionCost divideNearest(InstructionCost Numerator, 677 uint64_t Denominator) { 678 return (Numerator + (Denominator / 2)) / Denominator; 679 } 680 681 bool SelectOptimize::hasExpensiveColdOperand( 682 const SmallVector<SelectInst *, 2> &ASI) { 683 bool ColdOperand = false; 684 uint64_t TrueWeight, FalseWeight, TotalWeight; 685 if (extractBranchWeights(*ASI.front(), TrueWeight, FalseWeight)) { 686 uint64_t MinWeight = std::min(TrueWeight, FalseWeight); 687 TotalWeight = TrueWeight + FalseWeight; 688 // Is there a path with frequency <ColdOperandThreshold% (default:20%) ? 689 ColdOperand = TotalWeight * ColdOperandThreshold > 100 * MinWeight; 690 } else if (PSI->hasProfileSummary()) { 691 OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", ASI.front()); 692 ORmiss << "Profile data available but missing branch-weights metadata for " 693 "select instruction. "; 694 EmitAndPrintRemark(ORE, ORmiss); 695 } 696 if (!ColdOperand) 697 return false; 698 // Check if the cold path's dependence slice is expensive for any of the 699 // selects of the group. 700 for (SelectInst *SI : ASI) { 701 Instruction *ColdI = nullptr; 702 uint64_t HotWeight; 703 if (TrueWeight < FalseWeight) { 704 ColdI = dyn_cast<Instruction>(SI->getTrueValue()); 705 HotWeight = FalseWeight; 706 } else { 707 ColdI = dyn_cast<Instruction>(SI->getFalseValue()); 708 HotWeight = TrueWeight; 709 } 710 if (ColdI) { 711 std::stack<Instruction *> ColdSlice; 712 getExclBackwardsSlice(ColdI, ColdSlice, SI); 713 InstructionCost SliceCost = 0; 714 while (!ColdSlice.empty()) { 715 SliceCost += TTI->getInstructionCost(ColdSlice.top(), 716 TargetTransformInfo::TCK_Latency); 717 ColdSlice.pop(); 718 } 719 // The colder the cold value operand of the select is the more expensive 720 // the cmov becomes for computing the cold value operand every time. Thus, 721 // the colder the cold operand is the more its cost counts. 722 // Get nearest integer cost adjusted for coldness. 723 InstructionCost AdjSliceCost = 724 divideNearest(SliceCost * HotWeight, TotalWeight); 725 if (AdjSliceCost >= 726 ColdOperandMaxCostMultiplier * TargetTransformInfo::TCC_Expensive) 727 return true; 728 } 729 } 730 return false; 731 } 732 733 // Check if it is safe to move LoadI next to the SI. 734 // Conservatively assume it is safe only if there is no instruction 735 // modifying memory in-between the load and the select instruction. 736 static bool isSafeToSinkLoad(Instruction *LoadI, Instruction *SI) { 737 // Assume loads from different basic blocks are unsafe to move. 738 if (LoadI->getParent() != SI->getParent()) 739 return false; 740 auto It = LoadI->getIterator(); 741 while (&*It != SI) { 742 if (It->mayWriteToMemory()) 743 return false; 744 It++; 745 } 746 return true; 747 } 748 749 // For a given source instruction, collect its backwards dependence slice 750 // consisting of instructions exclusively computed for the purpose of producing 751 // the operands of the source instruction. As an approximation 752 // (sufficiently-accurate in practice), we populate this set with the 753 // instructions of the backwards dependence slice that only have one-use and 754 // form an one-use chain that leads to the source instruction. 755 void SelectOptimize::getExclBackwardsSlice(Instruction *I, 756 std::stack<Instruction *> &Slice, 757 Instruction *SI, bool ForSinking) { 758 SmallPtrSet<Instruction *, 2> Visited; 759 std::queue<Instruction *> Worklist; 760 Worklist.push(I); 761 while (!Worklist.empty()) { 762 Instruction *II = Worklist.front(); 763 Worklist.pop(); 764 765 // Avoid cycles. 766 if (!Visited.insert(II).second) 767 continue; 768 769 if (!II->hasOneUse()) 770 continue; 771 772 // Cannot soundly sink instructions with side-effects. 773 // Terminator or phi instructions cannot be sunk. 774 // Avoid sinking other select instructions (should be handled separetely). 775 if (ForSinking && (II->isTerminator() || II->mayHaveSideEffects() || 776 isa<SelectInst>(II) || isa<PHINode>(II))) 777 continue; 778 779 // Avoid sinking loads in order not to skip state-modifying instructions, 780 // that may alias with the loaded address. 781 // Only allow sinking of loads within the same basic block that are 782 // conservatively proven to be safe. 783 if (ForSinking && II->mayReadFromMemory() && !isSafeToSinkLoad(II, SI)) 784 continue; 785 786 // Avoid considering instructions with less frequency than the source 787 // instruction (i.e., avoid colder code regions of the dependence slice). 788 if (BFI->getBlockFreq(II->getParent()) < BFI->getBlockFreq(I->getParent())) 789 continue; 790 791 // Eligible one-use instruction added to the dependence slice. 792 Slice.push(II); 793 794 // Explore all the operands of the current instruction to expand the slice. 795 for (unsigned k = 0; k < II->getNumOperands(); ++k) 796 if (auto *OpI = dyn_cast<Instruction>(II->getOperand(k))) 797 Worklist.push(OpI); 798 } 799 } 800 801 bool SelectOptimize::isSelectHighlyPredictable(const SelectInst *SI) { 802 uint64_t TrueWeight, FalseWeight; 803 if (extractBranchWeights(*SI, TrueWeight, FalseWeight)) { 804 uint64_t Max = std::max(TrueWeight, FalseWeight); 805 uint64_t Sum = TrueWeight + FalseWeight; 806 if (Sum != 0) { 807 auto Probability = BranchProbability::getBranchProbability(Max, Sum); 808 if (Probability > TTI->getPredictableBranchThreshold()) 809 return true; 810 } 811 } 812 return false; 813 } 814 815 bool SelectOptimize::checkLoopHeuristics(const Loop *L, 816 const CostInfo LoopCost[2]) { 817 // Loop-level checks to determine if a non-predicated version (with branches) 818 // of the loop is more profitable than its predicated version. 819 820 if (DisableLoopLevelHeuristics) 821 return true; 822 823 OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti", 824 L->getHeader()->getFirstNonPHI()); 825 826 if (LoopCost[0].NonPredCost > LoopCost[0].PredCost || 827 LoopCost[1].NonPredCost >= LoopCost[1].PredCost) { 828 ORmissL << "No select conversion in the loop due to no reduction of loop's " 829 "critical path. "; 830 EmitAndPrintRemark(ORE, ORmissL); 831 return false; 832 } 833 834 Scaled64 Gain[2] = {LoopCost[0].PredCost - LoopCost[0].NonPredCost, 835 LoopCost[1].PredCost - LoopCost[1].NonPredCost}; 836 837 // Profitably converting to branches need to reduce the loop's critical path 838 // by at least some threshold (absolute gain of GainCycleThreshold cycles and 839 // relative gain of 12.5%). 840 if (Gain[1] < Scaled64::get(GainCycleThreshold) || 841 Gain[1] * Scaled64::get(GainRelativeThreshold) < LoopCost[1].PredCost) { 842 Scaled64 RelativeGain = Scaled64::get(100) * Gain[1] / LoopCost[1].PredCost; 843 ORmissL << "No select conversion in the loop due to small reduction of " 844 "loop's critical path. Gain=" 845 << Gain[1].toString() 846 << ", RelativeGain=" << RelativeGain.toString() << "%. "; 847 EmitAndPrintRemark(ORE, ORmissL); 848 return false; 849 } 850 851 // If the loop's critical path involves loop-carried dependences, the gradient 852 // of the gain needs to be at least GainGradientThreshold% (defaults to 25%). 853 // This check ensures that the latency reduction for the loop's critical path 854 // keeps decreasing with sufficient rate beyond the two analyzed loop 855 // iterations. 856 if (Gain[1] > Gain[0]) { 857 Scaled64 GradientGain = Scaled64::get(100) * (Gain[1] - Gain[0]) / 858 (LoopCost[1].PredCost - LoopCost[0].PredCost); 859 if (GradientGain < Scaled64::get(GainGradientThreshold)) { 860 ORmissL << "No select conversion in the loop due to small gradient gain. " 861 "GradientGain=" 862 << GradientGain.toString() << "%. "; 863 EmitAndPrintRemark(ORE, ORmissL); 864 return false; 865 } 866 } 867 // If the gain decreases it is not profitable to convert. 868 else if (Gain[1] < Gain[0]) { 869 ORmissL 870 << "No select conversion in the loop due to negative gradient gain. "; 871 EmitAndPrintRemark(ORE, ORmissL); 872 return false; 873 } 874 875 // Non-predicated version of the loop is more profitable than its 876 // predicated version. 877 return true; 878 } 879 880 // Computes instruction and loop-critical-path costs for both the predicated 881 // and non-predicated version of the given loop. 882 // Returns false if unable to compute these costs due to invalid cost of loop 883 // instruction(s). 884 bool SelectOptimize::computeLoopCosts( 885 const Loop *L, const SelectGroups &SIGroups, 886 DenseMap<const Instruction *, CostInfo> &InstCostMap, CostInfo *LoopCost) { 887 LLVM_DEBUG(dbgs() << "Calculating Latency / IPredCost / INonPredCost of loop " 888 << L->getHeader()->getName() << "\n"); 889 const auto &SIset = getSIset(SIGroups); 890 // Compute instruction and loop-critical-path costs across two iterations for 891 // both predicated and non-predicated version. 892 const unsigned Iterations = 2; 893 for (unsigned Iter = 0; Iter < Iterations; ++Iter) { 894 // Cost of the loop's critical path. 895 CostInfo &MaxCost = LoopCost[Iter]; 896 for (BasicBlock *BB : L->getBlocks()) { 897 for (const Instruction &I : *BB) { 898 if (I.isDebugOrPseudoInst()) 899 continue; 900 // Compute the predicated and non-predicated cost of the instruction. 901 Scaled64 IPredCost = Scaled64::getZero(), 902 INonPredCost = Scaled64::getZero(); 903 904 // Assume infinite resources that allow to fully exploit the available 905 // instruction-level parallelism. 906 // InstCost = InstLatency + max(Op1Cost, Op2Cost, … OpNCost) 907 for (const Use &U : I.operands()) { 908 auto UI = dyn_cast<Instruction>(U.get()); 909 if (!UI) 910 continue; 911 if (InstCostMap.count(UI)) { 912 IPredCost = std::max(IPredCost, InstCostMap[UI].PredCost); 913 INonPredCost = std::max(INonPredCost, InstCostMap[UI].NonPredCost); 914 } 915 } 916 auto ILatency = computeInstCost(&I); 917 if (!ILatency) { 918 OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti", &I); 919 ORmissL << "Invalid instruction cost preventing analysis and " 920 "optimization of the inner-most loop containing this " 921 "instruction. "; 922 EmitAndPrintRemark(ORE, ORmissL); 923 return false; 924 } 925 IPredCost += Scaled64::get(*ILatency); 926 INonPredCost += Scaled64::get(*ILatency); 927 928 // For a select that can be converted to branch, 929 // compute its cost as a branch (non-predicated cost). 930 // 931 // BranchCost = PredictedPathCost + MispredictCost 932 // PredictedPathCost = TrueOpCost * TrueProb + FalseOpCost * FalseProb 933 // MispredictCost = max(MispredictPenalty, CondCost) * MispredictRate 934 if (SIset.contains(&I)) { 935 auto SI = cast<SelectInst>(&I); 936 937 Scaled64 TrueOpCost = Scaled64::getZero(), 938 FalseOpCost = Scaled64::getZero(); 939 if (auto *TI = dyn_cast<Instruction>(SI->getTrueValue())) 940 if (InstCostMap.count(TI)) 941 TrueOpCost = InstCostMap[TI].NonPredCost; 942 if (auto *FI = dyn_cast<Instruction>(SI->getFalseValue())) 943 if (InstCostMap.count(FI)) 944 FalseOpCost = InstCostMap[FI].NonPredCost; 945 Scaled64 PredictedPathCost = 946 getPredictedPathCost(TrueOpCost, FalseOpCost, SI); 947 948 Scaled64 CondCost = Scaled64::getZero(); 949 if (auto *CI = dyn_cast<Instruction>(SI->getCondition())) 950 if (InstCostMap.count(CI)) 951 CondCost = InstCostMap[CI].NonPredCost; 952 Scaled64 MispredictCost = getMispredictionCost(SI, CondCost); 953 954 INonPredCost = PredictedPathCost + MispredictCost; 955 } 956 LLVM_DEBUG(dbgs() << " " << ILatency << "/" << IPredCost << "/" 957 << INonPredCost << " for " << I << "\n"); 958 959 InstCostMap[&I] = {IPredCost, INonPredCost}; 960 MaxCost.PredCost = std::max(MaxCost.PredCost, IPredCost); 961 MaxCost.NonPredCost = std::max(MaxCost.NonPredCost, INonPredCost); 962 } 963 } 964 LLVM_DEBUG(dbgs() << "Iteration " << Iter + 1 965 << " MaxCost = " << MaxCost.PredCost << " " 966 << MaxCost.NonPredCost << "\n"); 967 } 968 return true; 969 } 970 971 SmallPtrSet<const Instruction *, 2> 972 SelectOptimize::getSIset(const SelectGroups &SIGroups) { 973 SmallPtrSet<const Instruction *, 2> SIset; 974 for (const SelectGroup &ASI : SIGroups) 975 for (const SelectInst *SI : ASI) 976 SIset.insert(SI); 977 return SIset; 978 } 979 980 std::optional<uint64_t> SelectOptimize::computeInstCost(const Instruction *I) { 981 InstructionCost ICost = 982 TTI->getInstructionCost(I, TargetTransformInfo::TCK_Latency); 983 if (auto OC = ICost.getValue()) 984 return std::optional<uint64_t>(*OC); 985 return std::nullopt; 986 } 987 988 ScaledNumber<uint64_t> 989 SelectOptimize::getMispredictionCost(const SelectInst *SI, 990 const Scaled64 CondCost) { 991 uint64_t MispredictPenalty = TSchedModel.getMCSchedModel()->MispredictPenalty; 992 993 // Account for the default misprediction rate when using a branch 994 // (conservatively set to 25% by default). 995 uint64_t MispredictRate = MispredictDefaultRate; 996 // If the select condition is obviously predictable, then the misprediction 997 // rate is zero. 998 if (isSelectHighlyPredictable(SI)) 999 MispredictRate = 0; 1000 1001 // CondCost is included to account for cases where the computation of the 1002 // condition is part of a long dependence chain (potentially loop-carried) 1003 // that would delay detection of a misprediction and increase its cost. 1004 Scaled64 MispredictCost = 1005 std::max(Scaled64::get(MispredictPenalty), CondCost) * 1006 Scaled64::get(MispredictRate); 1007 MispredictCost /= Scaled64::get(100); 1008 1009 return MispredictCost; 1010 } 1011 1012 // Returns the cost of a branch when the prediction is correct. 1013 // TrueCost * TrueProbability + FalseCost * FalseProbability. 1014 ScaledNumber<uint64_t> 1015 SelectOptimize::getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost, 1016 const SelectInst *SI) { 1017 Scaled64 PredPathCost; 1018 uint64_t TrueWeight, FalseWeight; 1019 if (extractBranchWeights(*SI, TrueWeight, FalseWeight)) { 1020 uint64_t SumWeight = TrueWeight + FalseWeight; 1021 if (SumWeight != 0) { 1022 PredPathCost = TrueCost * Scaled64::get(TrueWeight) + 1023 FalseCost * Scaled64::get(FalseWeight); 1024 PredPathCost /= Scaled64::get(SumWeight); 1025 return PredPathCost; 1026 } 1027 } 1028 // Without branch weight metadata, we assume 75% for the one path and 25% for 1029 // the other, and pick the result with the biggest cost. 1030 PredPathCost = std::max(TrueCost * Scaled64::get(3) + FalseCost, 1031 FalseCost * Scaled64::get(3) + TrueCost); 1032 PredPathCost /= Scaled64::get(4); 1033 return PredPathCost; 1034 } 1035 1036 bool SelectOptimize::isSelectKindSupported(SelectInst *SI) { 1037 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1); 1038 if (VectorCond) 1039 return false; 1040 TargetLowering::SelectSupportKind SelectKind; 1041 if (SI->getType()->isVectorTy()) 1042 SelectKind = TargetLowering::ScalarCondVectorVal; 1043 else 1044 SelectKind = TargetLowering::ScalarValSelect; 1045 return TLI->isSelectSupported(SelectKind); 1046 } 1047