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