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/CodeGen/SelectOptimize.h" 14 #include "llvm/ADT/SmallVector.h" 15 #include "llvm/ADT/Statistic.h" 16 #include "llvm/Analysis/BlockFrequencyInfo.h" 17 #include "llvm/Analysis/BranchProbabilityInfo.h" 18 #include "llvm/Analysis/LoopInfo.h" 19 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 20 #include "llvm/Analysis/ProfileSummaryInfo.h" 21 #include "llvm/Analysis/TargetTransformInfo.h" 22 #include "llvm/CodeGen/Passes.h" 23 #include "llvm/CodeGen/TargetLowering.h" 24 #include "llvm/CodeGen/TargetPassConfig.h" 25 #include "llvm/CodeGen/TargetSchedule.h" 26 #include "llvm/CodeGen/TargetSubtargetInfo.h" 27 #include "llvm/IR/BasicBlock.h" 28 #include "llvm/IR/Dominators.h" 29 #include "llvm/IR/Function.h" 30 #include "llvm/IR/IRBuilder.h" 31 #include "llvm/IR/Instruction.h" 32 #include "llvm/IR/PatternMatch.h" 33 #include "llvm/IR/ProfDataUtils.h" 34 #include "llvm/InitializePasses.h" 35 #include "llvm/Pass.h" 36 #include "llvm/Support/ScaledNumber.h" 37 #include "llvm/Target/TargetMachine.h" 38 #include "llvm/Transforms/Utils/SizeOpts.h" 39 #include <algorithm> 40 #include <memory> 41 #include <queue> 42 #include <stack> 43 44 using namespace llvm; 45 using namespace llvm::PatternMatch; 46 47 #define DEBUG_TYPE "select-optimize" 48 49 STATISTIC(NumSelectOptAnalyzed, 50 "Number of select groups considered for conversion to branch"); 51 STATISTIC(NumSelectConvertedExpColdOperand, 52 "Number of select groups converted due to expensive cold operand"); 53 STATISTIC(NumSelectConvertedHighPred, 54 "Number of select groups converted due to high-predictability"); 55 STATISTIC(NumSelectUnPred, 56 "Number of select groups not converted due to unpredictability"); 57 STATISTIC(NumSelectColdBB, 58 "Number of select groups not converted due to cold basic block"); 59 STATISTIC(NumSelectConvertedLoop, 60 "Number of select groups converted due to loop-level analysis"); 61 STATISTIC(NumSelectsConverted, "Number of selects converted"); 62 63 static cl::opt<unsigned> ColdOperandThreshold( 64 "cold-operand-threshold", 65 cl::desc("Maximum frequency of path for an operand to be considered cold."), 66 cl::init(20), cl::Hidden); 67 68 static cl::opt<unsigned> ColdOperandMaxCostMultiplier( 69 "cold-operand-max-cost-multiplier", 70 cl::desc("Maximum cost multiplier of TCC_expensive for the dependence " 71 "slice of a cold operand to be considered inexpensive."), 72 cl::init(1), cl::Hidden); 73 74 static cl::opt<unsigned> 75 GainGradientThreshold("select-opti-loop-gradient-gain-threshold", 76 cl::desc("Gradient gain threshold (%)."), 77 cl::init(25), cl::Hidden); 78 79 static cl::opt<unsigned> 80 GainCycleThreshold("select-opti-loop-cycle-gain-threshold", 81 cl::desc("Minimum gain per loop (in cycles) threshold."), 82 cl::init(4), cl::Hidden); 83 84 static cl::opt<unsigned> GainRelativeThreshold( 85 "select-opti-loop-relative-gain-threshold", 86 cl::desc( 87 "Minimum relative gain per loop threshold (1/X). Defaults to 12.5%"), 88 cl::init(8), cl::Hidden); 89 90 static cl::opt<unsigned> MispredictDefaultRate( 91 "mispredict-default-rate", cl::Hidden, cl::init(25), 92 cl::desc("Default mispredict rate (initialized to 25%).")); 93 94 static cl::opt<bool> 95 DisableLoopLevelHeuristics("disable-loop-level-heuristics", cl::Hidden, 96 cl::init(false), 97 cl::desc("Disable loop-level heuristics.")); 98 99 namespace { 100 101 class SelectOptimizeImpl { 102 const TargetMachine *TM = nullptr; 103 const TargetSubtargetInfo *TSI = nullptr; 104 const TargetLowering *TLI = nullptr; 105 const TargetTransformInfo *TTI = nullptr; 106 const LoopInfo *LI = nullptr; 107 BlockFrequencyInfo *BFI; 108 ProfileSummaryInfo *PSI = nullptr; 109 OptimizationRemarkEmitter *ORE = nullptr; 110 TargetSchedModel TSchedModel; 111 112 public: 113 SelectOptimizeImpl() = default; 114 SelectOptimizeImpl(const TargetMachine *TM) : TM(TM){}; 115 PreservedAnalyses run(Function &F, FunctionAnalysisManager &FAM); 116 bool runOnFunction(Function &F, Pass &P); 117 118 using Scaled64 = ScaledNumber<uint64_t>; 119 120 struct CostInfo { 121 /// Predicated cost (with selects as conditional moves). 122 Scaled64 PredCost; 123 /// Non-predicated cost (with selects converted to branches). 124 Scaled64 NonPredCost; 125 }; 126 127 /// SelectLike is an abstraction over SelectInst and other operations that can 128 /// act like selects. For example Or(Zext(icmp), X) can be treated like 129 /// select(icmp, X|1, X). 130 class SelectLike { 131 SelectLike(Instruction *I) : I(I) {} 132 133 Instruction *I; 134 135 public: 136 /// Match a select or select-like instruction, returning a SelectLike. 137 static SelectLike match(Instruction *I) { 138 // Select instruction are what we are usually looking for. 139 if (isa<SelectInst>(I)) 140 return SelectLike(I); 141 142 // An Or(zext(i1 X), Y) can also be treated like a select, with condition 143 // C and values Y|1 and Y. 144 Value *X; 145 if (PatternMatch::match( 146 I, m_c_Or(m_OneUse(m_ZExt(m_Value(X))), m_Value())) && 147 X->getType()->isIntegerTy(1)) 148 return SelectLike(I); 149 150 return SelectLike(nullptr); 151 } 152 153 bool isValid() { return I; } 154 operator bool() { return isValid(); } 155 156 Instruction *getI() { return I; } 157 const Instruction *getI() const { return I; } 158 159 Type *getType() const { return I->getType(); } 160 161 /// Return the condition for the SelectLike instruction. For example the 162 /// condition of a select or c in `or(zext(c), x)` 163 Value *getCondition() const { 164 if (auto *Sel = dyn_cast<SelectInst>(I)) 165 return Sel->getCondition(); 166 // Or(zext) case 167 if (auto *BO = dyn_cast<BinaryOperator>(I)) { 168 Value *X; 169 if (PatternMatch::match(BO->getOperand(0), 170 m_OneUse(m_ZExt(m_Value(X))))) 171 return X; 172 if (PatternMatch::match(BO->getOperand(1), 173 m_OneUse(m_ZExt(m_Value(X))))) 174 return X; 175 } 176 177 llvm_unreachable("Unhandled case in getCondition"); 178 } 179 180 /// Return the true value for the SelectLike instruction. Note this may not 181 /// exist for all SelectLike instructions. For example, for `or(zext(c), x)` 182 /// the true value would be `or(x,1)`. As this value does not exist, nullptr 183 /// is returned. 184 Value *getTrueValue() const { 185 if (auto *Sel = dyn_cast<SelectInst>(I)) 186 return Sel->getTrueValue(); 187 // Or(zext) case - The true value is Or(X), so return nullptr as the value 188 // does not yet exist. 189 if (isa<BinaryOperator>(I)) 190 return nullptr; 191 192 llvm_unreachable("Unhandled case in getTrueValue"); 193 } 194 195 /// Return the false value for the SelectLike instruction. For example the 196 /// getFalseValue of a select or `x` in `or(zext(c), x)` (which is 197 /// `select(c, x|1, x)`) 198 Value *getFalseValue() const { 199 if (auto *Sel = dyn_cast<SelectInst>(I)) 200 return Sel->getFalseValue(); 201 // Or(zext) case - return the operand which is not the zext. 202 if (auto *BO = dyn_cast<BinaryOperator>(I)) { 203 Value *X; 204 if (PatternMatch::match(BO->getOperand(0), 205 m_OneUse(m_ZExt(m_Value(X))))) 206 return BO->getOperand(1); 207 if (PatternMatch::match(BO->getOperand(1), 208 m_OneUse(m_ZExt(m_Value(X))))) 209 return BO->getOperand(0); 210 } 211 212 llvm_unreachable("Unhandled case in getFalseValue"); 213 } 214 215 /// Return the NonPredCost cost of the true op, given the costs in 216 /// InstCostMap. This may need to be generated for select-like instructions. 217 Scaled64 getTrueOpCost(DenseMap<const Instruction *, CostInfo> &InstCostMap, 218 const TargetTransformInfo *TTI) { 219 if (auto *Sel = dyn_cast<SelectInst>(I)) 220 if (auto *I = dyn_cast<Instruction>(Sel->getTrueValue())) 221 return InstCostMap.contains(I) ? InstCostMap[I].NonPredCost 222 : Scaled64::getZero(); 223 224 // Or case - add the cost of an extra Or to the cost of the False case. 225 if (isa<BinaryOperator>(I)) 226 if (auto I = dyn_cast<Instruction>(getFalseValue())) 227 if (InstCostMap.contains(I)) { 228 InstructionCost OrCost = TTI->getArithmeticInstrCost( 229 Instruction::Or, I->getType(), TargetTransformInfo::TCK_Latency, 230 {TargetTransformInfo::OK_AnyValue, 231 TargetTransformInfo::OP_None}, 232 {TTI::OK_UniformConstantValue, TTI::OP_PowerOf2}); 233 return InstCostMap[I].NonPredCost + 234 Scaled64::get(*OrCost.getValue()); 235 } 236 237 return Scaled64::getZero(); 238 } 239 240 /// Return the NonPredCost cost of the false op, given the costs in 241 /// InstCostMap. This may need to be generated for select-like instructions. 242 Scaled64 243 getFalseOpCost(DenseMap<const Instruction *, CostInfo> &InstCostMap, 244 const TargetTransformInfo *TTI) { 245 if (auto *Sel = dyn_cast<SelectInst>(I)) 246 if (auto *I = dyn_cast<Instruction>(Sel->getFalseValue())) 247 return InstCostMap.contains(I) ? InstCostMap[I].NonPredCost 248 : Scaled64::getZero(); 249 250 // Or case - return the cost of the false case 251 if (isa<BinaryOperator>(I)) 252 if (auto I = dyn_cast<Instruction>(getFalseValue())) 253 if (InstCostMap.contains(I)) 254 return InstCostMap[I].NonPredCost; 255 256 return Scaled64::getZero(); 257 } 258 }; 259 260 private: 261 // Select groups consist of consecutive select instructions with the same 262 // condition. 263 using SelectGroup = SmallVector<SelectLike, 2>; 264 using SelectGroups = SmallVector<SelectGroup, 2>; 265 266 // Converts select instructions of a function to conditional jumps when deemed 267 // profitable. Returns true if at least one select was converted. 268 bool optimizeSelects(Function &F); 269 270 // Heuristics for determining which select instructions can be profitably 271 // conveted to branches. Separate heuristics for selects in inner-most loops 272 // and the rest of code regions (base heuristics for non-inner-most loop 273 // regions). 274 void optimizeSelectsBase(Function &F, SelectGroups &ProfSIGroups); 275 void optimizeSelectsInnerLoops(Function &F, SelectGroups &ProfSIGroups); 276 277 // Converts to branches the select groups that were deemed 278 // profitable-to-convert. 279 void convertProfitableSIGroups(SelectGroups &ProfSIGroups); 280 281 // Splits selects of a given basic block into select groups. 282 void collectSelectGroups(BasicBlock &BB, SelectGroups &SIGroups); 283 284 // Determines for which select groups it is profitable converting to branches 285 // (base and inner-most-loop heuristics). 286 void findProfitableSIGroupsBase(SelectGroups &SIGroups, 287 SelectGroups &ProfSIGroups); 288 void findProfitableSIGroupsInnerLoops(const Loop *L, SelectGroups &SIGroups, 289 SelectGroups &ProfSIGroups); 290 291 // Determines if a select group should be converted to a branch (base 292 // heuristics). 293 bool isConvertToBranchProfitableBase(const SelectGroup &ASI); 294 295 // Returns true if there are expensive instructions in the cold value 296 // operand's (if any) dependence slice of any of the selects of the given 297 // group. 298 bool hasExpensiveColdOperand(const SelectGroup &ASI); 299 300 // For a given source instruction, collect its backwards dependence slice 301 // consisting of instructions exclusively computed for producing the operands 302 // of the source instruction. 303 void getExclBackwardsSlice(Instruction *I, std::stack<Instruction *> &Slice, 304 Instruction *SI, bool ForSinking = false); 305 306 // Returns true if the condition of the select is highly predictable. 307 bool isSelectHighlyPredictable(const SelectLike SI); 308 309 // Loop-level checks to determine if a non-predicated version (with branches) 310 // of the given loop is more profitable than its predicated version. 311 bool checkLoopHeuristics(const Loop *L, const CostInfo LoopDepth[2]); 312 313 // Computes instruction and loop-critical-path costs for both the predicated 314 // and non-predicated version of the given loop. 315 bool computeLoopCosts(const Loop *L, const SelectGroups &SIGroups, 316 DenseMap<const Instruction *, CostInfo> &InstCostMap, 317 CostInfo *LoopCost); 318 319 // Returns a set of all the select instructions in the given select groups. 320 SmallDenseMap<const Instruction *, SelectLike, 2> 321 getSImap(const SelectGroups &SIGroups); 322 323 // Returns the latency cost of a given instruction. 324 std::optional<uint64_t> computeInstCost(const Instruction *I); 325 326 // Returns the misprediction cost of a given select when converted to branch. 327 Scaled64 getMispredictionCost(const SelectLike SI, const Scaled64 CondCost); 328 329 // Returns the cost of a branch when the prediction is correct. 330 Scaled64 getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost, 331 const SelectLike SI); 332 333 // Returns true if the target architecture supports lowering a given select. 334 bool isSelectKindSupported(const SelectLike SI); 335 }; 336 337 class SelectOptimize : public FunctionPass { 338 SelectOptimizeImpl Impl; 339 340 public: 341 static char ID; 342 343 SelectOptimize() : FunctionPass(ID) { 344 initializeSelectOptimizePass(*PassRegistry::getPassRegistry()); 345 } 346 347 bool runOnFunction(Function &F) override { 348 return Impl.runOnFunction(F, *this); 349 } 350 351 void getAnalysisUsage(AnalysisUsage &AU) const override { 352 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 353 AU.addRequired<TargetPassConfig>(); 354 AU.addRequired<TargetTransformInfoWrapperPass>(); 355 AU.addRequired<LoopInfoWrapperPass>(); 356 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 357 AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); 358 } 359 }; 360 361 } // namespace 362 363 PreservedAnalyses SelectOptimizePass::run(Function &F, 364 FunctionAnalysisManager &FAM) { 365 SelectOptimizeImpl Impl(TM); 366 return Impl.run(F, FAM); 367 } 368 369 char SelectOptimize::ID = 0; 370 371 INITIALIZE_PASS_BEGIN(SelectOptimize, DEBUG_TYPE, "Optimize selects", false, 372 false) 373 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 374 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 375 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 376 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 377 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 378 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) 379 INITIALIZE_PASS_END(SelectOptimize, DEBUG_TYPE, "Optimize selects", false, 380 false) 381 382 FunctionPass *llvm::createSelectOptimizePass() { return new SelectOptimize(); } 383 384 PreservedAnalyses SelectOptimizeImpl::run(Function &F, 385 FunctionAnalysisManager &FAM) { 386 TSI = TM->getSubtargetImpl(F); 387 TLI = TSI->getTargetLowering(); 388 389 // If none of the select types are supported then skip this pass. 390 // This is an optimization pass. Legality issues will be handled by 391 // instruction selection. 392 if (!TLI->isSelectSupported(TargetLowering::ScalarValSelect) && 393 !TLI->isSelectSupported(TargetLowering::ScalarCondVectorVal) && 394 !TLI->isSelectSupported(TargetLowering::VectorMaskSelect)) 395 return PreservedAnalyses::all(); 396 397 TTI = &FAM.getResult<TargetIRAnalysis>(F); 398 if (!TTI->enableSelectOptimize()) 399 return PreservedAnalyses::all(); 400 401 PSI = FAM.getResult<ModuleAnalysisManagerFunctionProxy>(F) 402 .getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); 403 assert(PSI && "This pass requires module analysis pass `profile-summary`!"); 404 BFI = &FAM.getResult<BlockFrequencyAnalysis>(F); 405 406 // When optimizing for size, selects are preferable over branches. 407 if (F.hasOptSize() || llvm::shouldOptimizeForSize(&F, PSI, BFI)) 408 return PreservedAnalyses::all(); 409 410 LI = &FAM.getResult<LoopAnalysis>(F); 411 ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F); 412 TSchedModel.init(TSI); 413 414 bool Changed = optimizeSelects(F); 415 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); 416 } 417 418 bool SelectOptimizeImpl::runOnFunction(Function &F, Pass &P) { 419 TM = &P.getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); 420 TSI = TM->getSubtargetImpl(F); 421 TLI = TSI->getTargetLowering(); 422 423 // If none of the select types are supported then skip this pass. 424 // This is an optimization pass. Legality issues will be handled by 425 // instruction selection. 426 if (!TLI->isSelectSupported(TargetLowering::ScalarValSelect) && 427 !TLI->isSelectSupported(TargetLowering::ScalarCondVectorVal) && 428 !TLI->isSelectSupported(TargetLowering::VectorMaskSelect)) 429 return false; 430 431 TTI = &P.getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 432 433 if (!TTI->enableSelectOptimize()) 434 return false; 435 436 LI = &P.getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 437 BFI = &P.getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); 438 PSI = &P.getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 439 ORE = &P.getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); 440 TSchedModel.init(TSI); 441 442 // When optimizing for size, selects are preferable over branches. 443 if (F.hasOptSize() || llvm::shouldOptimizeForSize(&F, PSI, BFI)) 444 return false; 445 446 return optimizeSelects(F); 447 } 448 449 bool SelectOptimizeImpl::optimizeSelects(Function &F) { 450 // Determine for which select groups it is profitable converting to branches. 451 SelectGroups ProfSIGroups; 452 // Base heuristics apply only to non-loops and outer loops. 453 optimizeSelectsBase(F, ProfSIGroups); 454 // Separate heuristics for inner-most loops. 455 optimizeSelectsInnerLoops(F, ProfSIGroups); 456 457 // Convert to branches the select groups that were deemed 458 // profitable-to-convert. 459 convertProfitableSIGroups(ProfSIGroups); 460 461 // Code modified if at least one select group was converted. 462 return !ProfSIGroups.empty(); 463 } 464 465 void SelectOptimizeImpl::optimizeSelectsBase(Function &F, 466 SelectGroups &ProfSIGroups) { 467 // Collect all the select groups. 468 SelectGroups SIGroups; 469 for (BasicBlock &BB : F) { 470 // Base heuristics apply only to non-loops and outer loops. 471 Loop *L = LI->getLoopFor(&BB); 472 if (L && L->isInnermost()) 473 continue; 474 collectSelectGroups(BB, SIGroups); 475 } 476 477 // Determine for which select groups it is profitable converting to branches. 478 findProfitableSIGroupsBase(SIGroups, ProfSIGroups); 479 } 480 481 void SelectOptimizeImpl::optimizeSelectsInnerLoops(Function &F, 482 SelectGroups &ProfSIGroups) { 483 SmallVector<Loop *, 4> Loops(LI->begin(), LI->end()); 484 // Need to check size on each iteration as we accumulate child loops. 485 for (unsigned long i = 0; i < Loops.size(); ++i) 486 for (Loop *ChildL : Loops[i]->getSubLoops()) 487 Loops.push_back(ChildL); 488 489 for (Loop *L : Loops) { 490 if (!L->isInnermost()) 491 continue; 492 493 SelectGroups SIGroups; 494 for (BasicBlock *BB : L->getBlocks()) 495 collectSelectGroups(*BB, SIGroups); 496 497 findProfitableSIGroupsInnerLoops(L, SIGroups, ProfSIGroups); 498 } 499 } 500 501 /// If \p isTrue is true, return the true value of \p SI, otherwise return 502 /// false value of \p SI. If the true/false value of \p SI is defined by any 503 /// select instructions in \p Selects, look through the defining select 504 /// instruction until the true/false value is not defined in \p Selects. 505 static Value * 506 getTrueOrFalseValue(SelectOptimizeImpl::SelectLike SI, bool isTrue, 507 const SmallPtrSet<const Instruction *, 2> &Selects, 508 IRBuilder<> &IB) { 509 Value *V = nullptr; 510 for (SelectInst *DefSI = dyn_cast<SelectInst>(SI.getI()); 511 DefSI != nullptr && Selects.count(DefSI); 512 DefSI = dyn_cast<SelectInst>(V)) { 513 assert(DefSI->getCondition() == SI.getCondition() && 514 "The condition of DefSI does not match with SI"); 515 V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue()); 516 } 517 518 if (isa<BinaryOperator>(SI.getI())) { 519 assert(SI.getI()->getOpcode() == Instruction::Or && 520 "Only currently handling Or instructions."); 521 V = SI.getFalseValue(); 522 if (isTrue) 523 V = IB.CreateOr(V, ConstantInt::get(V->getType(), 1)); 524 } 525 526 assert(V && "Failed to get select true/false value"); 527 return V; 528 } 529 530 void SelectOptimizeImpl::convertProfitableSIGroups(SelectGroups &ProfSIGroups) { 531 for (SelectGroup &ASI : ProfSIGroups) { 532 // The code transformation here is a modified version of the sinking 533 // transformation in CodeGenPrepare::optimizeSelectInst with a more 534 // aggressive strategy of which instructions to sink. 535 // 536 // TODO: eliminate the redundancy of logic transforming selects to branches 537 // by removing CodeGenPrepare::optimizeSelectInst and optimizing here 538 // selects for all cases (with and without profile information). 539 540 // Transform a sequence like this: 541 // start: 542 // %cmp = cmp uge i32 %a, %b 543 // %sel = select i1 %cmp, i32 %c, i32 %d 544 // 545 // Into: 546 // start: 547 // %cmp = cmp uge i32 %a, %b 548 // %cmp.frozen = freeze %cmp 549 // br i1 %cmp.frozen, label %select.true, label %select.false 550 // select.true: 551 // br label %select.end 552 // select.false: 553 // br label %select.end 554 // select.end: 555 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ] 556 // 557 // %cmp should be frozen, otherwise it may introduce undefined behavior. 558 // In addition, we may sink instructions that produce %c or %d into the 559 // destination(s) of the new branch. 560 // If the true or false blocks do not contain a sunken instruction, that 561 // block and its branch may be optimized away. In that case, one side of the 562 // first branch will point directly to select.end, and the corresponding PHI 563 // predecessor block will be the start block. 564 565 // Find all the instructions that can be soundly sunk to the true/false 566 // blocks. These are instructions that are computed solely for producing the 567 // operands of the select instructions in the group and can be sunk without 568 // breaking the semantics of the LLVM IR (e.g., cannot sink instructions 569 // with side effects). 570 SmallVector<std::stack<Instruction *>, 2> TrueSlices, FalseSlices; 571 typedef std::stack<Instruction *>::size_type StackSizeType; 572 StackSizeType maxTrueSliceLen = 0, maxFalseSliceLen = 0; 573 for (SelectLike SI : ASI) { 574 // For each select, compute the sinkable dependence chains of the true and 575 // false operands. 576 if (auto *TI = dyn_cast_or_null<Instruction>(SI.getTrueValue())) { 577 std::stack<Instruction *> TrueSlice; 578 getExclBackwardsSlice(TI, TrueSlice, SI.getI(), true); 579 maxTrueSliceLen = std::max(maxTrueSliceLen, TrueSlice.size()); 580 TrueSlices.push_back(TrueSlice); 581 } 582 if (auto *FI = dyn_cast_or_null<Instruction>(SI.getFalseValue())) { 583 if (isa<SelectInst>(SI.getI()) || !FI->hasOneUse()) { 584 std::stack<Instruction *> FalseSlice; 585 getExclBackwardsSlice(FI, FalseSlice, SI.getI(), true); 586 maxFalseSliceLen = std::max(maxFalseSliceLen, FalseSlice.size()); 587 FalseSlices.push_back(FalseSlice); 588 } 589 } 590 } 591 // In the case of multiple select instructions in the same group, the order 592 // of non-dependent instructions (instructions of different dependence 593 // slices) in the true/false blocks appears to affect performance. 594 // Interleaving the slices seems to experimentally be the optimal approach. 595 // This interleaving scheduling allows for more ILP (with a natural downside 596 // of increasing a bit register pressure) compared to a simple ordering of 597 // one whole chain after another. One would expect that this ordering would 598 // not matter since the scheduling in the backend of the compiler would 599 // take care of it, but apparently the scheduler fails to deliver optimal 600 // ILP with a naive ordering here. 601 SmallVector<Instruction *, 2> TrueSlicesInterleaved, FalseSlicesInterleaved; 602 for (StackSizeType IS = 0; IS < maxTrueSliceLen; ++IS) { 603 for (auto &S : TrueSlices) { 604 if (!S.empty()) { 605 TrueSlicesInterleaved.push_back(S.top()); 606 S.pop(); 607 } 608 } 609 } 610 for (StackSizeType IS = 0; IS < maxFalseSliceLen; ++IS) { 611 for (auto &S : FalseSlices) { 612 if (!S.empty()) { 613 FalseSlicesInterleaved.push_back(S.top()); 614 S.pop(); 615 } 616 } 617 } 618 619 // We split the block containing the select(s) into two blocks. 620 SelectLike SI = ASI.front(); 621 SelectLike LastSI = ASI.back(); 622 BasicBlock *StartBlock = SI.getI()->getParent(); 623 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI.getI())); 624 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end"); 625 BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock)); 626 // Delete the unconditional branch that was just created by the split. 627 StartBlock->getTerminator()->eraseFromParent(); 628 629 // Move any debug/pseudo instructions that were in-between the select 630 // group to the newly-created end block. 631 SmallVector<Instruction *, 2> DebugPseudoINS; 632 auto DIt = SI.getI()->getIterator(); 633 while (&*DIt != LastSI.getI()) { 634 if (DIt->isDebugOrPseudoInst()) 635 DebugPseudoINS.push_back(&*DIt); 636 DIt++; 637 } 638 for (auto *DI : DebugPseudoINS) { 639 DI->moveBeforePreserving(&*EndBlock->getFirstInsertionPt()); 640 } 641 642 // Duplicate implementation for DPValues, the non-instruction debug-info 643 // record. Helper lambda for moving DPValues to the end block. 644 auto TransferDPValues = [&](Instruction &I) { 645 for (auto &DPValue : llvm::make_early_inc_range(I.getDbgValueRange())) { 646 DPValue.removeFromParent(); 647 EndBlock->insertDPValueBefore(&DPValue, 648 EndBlock->getFirstInsertionPt()); 649 } 650 }; 651 652 // Iterate over all instructions in between SI and LastSI, not including 653 // SI itself. These are all the variable assignments that happen "in the 654 // middle" of the select group. 655 auto R = make_range(std::next(SI.getI()->getIterator()), 656 std::next(LastSI.getI()->getIterator())); 657 llvm::for_each(R, TransferDPValues); 658 659 // These are the new basic blocks for the conditional branch. 660 // At least one will become an actual new basic block. 661 BasicBlock *TrueBlock = nullptr, *FalseBlock = nullptr; 662 BranchInst *TrueBranch = nullptr, *FalseBranch = nullptr; 663 if (!TrueSlicesInterleaved.empty()) { 664 TrueBlock = BasicBlock::Create(EndBlock->getContext(), "select.true.sink", 665 EndBlock->getParent(), EndBlock); 666 TrueBranch = BranchInst::Create(EndBlock, TrueBlock); 667 TrueBranch->setDebugLoc(LastSI.getI()->getDebugLoc()); 668 for (Instruction *TrueInst : TrueSlicesInterleaved) 669 TrueInst->moveBefore(TrueBranch); 670 } 671 if (!FalseSlicesInterleaved.empty()) { 672 FalseBlock = 673 BasicBlock::Create(EndBlock->getContext(), "select.false.sink", 674 EndBlock->getParent(), EndBlock); 675 FalseBranch = BranchInst::Create(EndBlock, FalseBlock); 676 FalseBranch->setDebugLoc(LastSI.getI()->getDebugLoc()); 677 for (Instruction *FalseInst : FalseSlicesInterleaved) 678 FalseInst->moveBefore(FalseBranch); 679 } 680 // If there was nothing to sink, then arbitrarily choose the 'false' side 681 // for a new input value to the PHI. 682 if (TrueBlock == FalseBlock) { 683 assert(TrueBlock == nullptr && 684 "Unexpected basic block transform while optimizing select"); 685 686 FalseBlock = BasicBlock::Create(StartBlock->getContext(), "select.false", 687 EndBlock->getParent(), EndBlock); 688 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock); 689 FalseBranch->setDebugLoc(SI.getI()->getDebugLoc()); 690 } 691 692 // Insert the real conditional branch based on the original condition. 693 // If we did not create a new block for one of the 'true' or 'false' paths 694 // of the condition, it means that side of the branch goes to the end block 695 // directly and the path originates from the start block from the point of 696 // view of the new PHI. 697 BasicBlock *TT, *FT; 698 if (TrueBlock == nullptr) { 699 TT = EndBlock; 700 FT = FalseBlock; 701 TrueBlock = StartBlock; 702 } else if (FalseBlock == nullptr) { 703 TT = TrueBlock; 704 FT = EndBlock; 705 FalseBlock = StartBlock; 706 } else { 707 TT = TrueBlock; 708 FT = FalseBlock; 709 } 710 IRBuilder<> IB(SI.getI()); 711 auto *CondFr = IB.CreateFreeze(SI.getCondition(), 712 SI.getCondition()->getName() + ".frozen"); 713 714 SmallPtrSet<const Instruction *, 2> INS; 715 for (auto SI : ASI) 716 INS.insert(SI.getI()); 717 718 // Use reverse iterator because later select may use the value of the 719 // earlier select, and we need to propagate value through earlier select 720 // to get the PHI operand. 721 for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) { 722 SelectLike SI = *It; 723 // The select itself is replaced with a PHI Node. 724 PHINode *PN = PHINode::Create(SI.getType(), 2, ""); 725 PN->insertBefore(EndBlock->begin()); 726 PN->takeName(SI.getI()); 727 PN->addIncoming(getTrueOrFalseValue(SI, true, INS, IB), TrueBlock); 728 PN->addIncoming(getTrueOrFalseValue(SI, false, INS, IB), FalseBlock); 729 PN->setDebugLoc(SI.getI()->getDebugLoc()); 730 SI.getI()->replaceAllUsesWith(PN); 731 INS.erase(SI.getI()); 732 ++NumSelectsConverted; 733 } 734 IB.CreateCondBr(CondFr, TT, FT, SI.getI()); 735 736 // Remove the old select instructions, now that they are not longer used. 737 for (auto SI : ASI) 738 SI.getI()->eraseFromParent(); 739 } 740 } 741 742 void SelectOptimizeImpl::collectSelectGroups(BasicBlock &BB, 743 SelectGroups &SIGroups) { 744 BasicBlock::iterator BBIt = BB.begin(); 745 while (BBIt != BB.end()) { 746 Instruction *I = &*BBIt++; 747 if (SelectLike SI = SelectLike::match(I)) { 748 if (!TTI->shouldTreatInstructionLikeSelect(I)) 749 continue; 750 751 SelectGroup SIGroup; 752 SIGroup.push_back(SI); 753 while (BBIt != BB.end()) { 754 Instruction *NI = &*BBIt; 755 // Debug/pseudo instructions should be skipped and not prevent the 756 // formation of a select group. 757 if (NI->isDebugOrPseudoInst()) { 758 ++BBIt; 759 continue; 760 } 761 // We only allow selects in the same group, not other select-like 762 // instructions. 763 if (!isa<SelectInst>(NI)) 764 break; 765 766 SelectLike NSI = SelectLike::match(NI); 767 if (NSI && SI.getCondition() == NSI.getCondition()) { 768 SIGroup.push_back(NSI); 769 } else 770 break; 771 ++BBIt; 772 } 773 774 // If the select type is not supported, no point optimizing it. 775 // Instruction selection will take care of it. 776 if (!isSelectKindSupported(SI)) 777 continue; 778 779 SIGroups.push_back(SIGroup); 780 } 781 } 782 } 783 784 void SelectOptimizeImpl::findProfitableSIGroupsBase( 785 SelectGroups &SIGroups, SelectGroups &ProfSIGroups) { 786 for (SelectGroup &ASI : SIGroups) { 787 ++NumSelectOptAnalyzed; 788 if (isConvertToBranchProfitableBase(ASI)) 789 ProfSIGroups.push_back(ASI); 790 } 791 } 792 793 static void EmitAndPrintRemark(OptimizationRemarkEmitter *ORE, 794 DiagnosticInfoOptimizationBase &Rem) { 795 LLVM_DEBUG(dbgs() << Rem.getMsg() << "\n"); 796 ORE->emit(Rem); 797 } 798 799 void SelectOptimizeImpl::findProfitableSIGroupsInnerLoops( 800 const Loop *L, SelectGroups &SIGroups, SelectGroups &ProfSIGroups) { 801 NumSelectOptAnalyzed += SIGroups.size(); 802 // For each select group in an inner-most loop, 803 // a branch is more preferable than a select/conditional-move if: 804 // i) conversion to branches for all the select groups of the loop satisfies 805 // loop-level heuristics including reducing the loop's critical path by 806 // some threshold (see SelectOptimizeImpl::checkLoopHeuristics); and 807 // ii) the total cost of the select group is cheaper with a branch compared 808 // to its predicated version. The cost is in terms of latency and the cost 809 // of a select group is the cost of its most expensive select instruction 810 // (assuming infinite resources and thus fully leveraging available ILP). 811 812 DenseMap<const Instruction *, CostInfo> InstCostMap; 813 CostInfo LoopCost[2] = {{Scaled64::getZero(), Scaled64::getZero()}, 814 {Scaled64::getZero(), Scaled64::getZero()}}; 815 if (!computeLoopCosts(L, SIGroups, InstCostMap, LoopCost) || 816 !checkLoopHeuristics(L, LoopCost)) { 817 return; 818 } 819 820 for (SelectGroup &ASI : SIGroups) { 821 // Assuming infinite resources, the cost of a group of instructions is the 822 // cost of the most expensive instruction of the group. 823 Scaled64 SelectCost = Scaled64::getZero(), BranchCost = Scaled64::getZero(); 824 for (SelectLike SI : ASI) { 825 SelectCost = std::max(SelectCost, InstCostMap[SI.getI()].PredCost); 826 BranchCost = std::max(BranchCost, InstCostMap[SI.getI()].NonPredCost); 827 } 828 if (BranchCost < SelectCost) { 829 OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", ASI.front().getI()); 830 OR << "Profitable to convert to branch (loop analysis). BranchCost=" 831 << BranchCost.toString() << ", SelectCost=" << SelectCost.toString() 832 << ". "; 833 EmitAndPrintRemark(ORE, OR); 834 ++NumSelectConvertedLoop; 835 ProfSIGroups.push_back(ASI); 836 } else { 837 OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", 838 ASI.front().getI()); 839 ORmiss << "Select is more profitable (loop analysis). BranchCost=" 840 << BranchCost.toString() 841 << ", SelectCost=" << SelectCost.toString() << ". "; 842 EmitAndPrintRemark(ORE, ORmiss); 843 } 844 } 845 } 846 847 bool SelectOptimizeImpl::isConvertToBranchProfitableBase( 848 const SelectGroup &ASI) { 849 SelectLike SI = ASI.front(); 850 LLVM_DEBUG(dbgs() << "Analyzing select group containing " << SI.getI() 851 << "\n"); 852 OptimizationRemark OR(DEBUG_TYPE, "SelectOpti", SI.getI()); 853 OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", SI.getI()); 854 855 // Skip cold basic blocks. Better to optimize for size for cold blocks. 856 if (PSI->isColdBlock(SI.getI()->getParent(), BFI)) { 857 ++NumSelectColdBB; 858 ORmiss << "Not converted to branch because of cold basic block. "; 859 EmitAndPrintRemark(ORE, ORmiss); 860 return false; 861 } 862 863 // If unpredictable, branch form is less profitable. 864 if (SI.getI()->getMetadata(LLVMContext::MD_unpredictable)) { 865 ++NumSelectUnPred; 866 ORmiss << "Not converted to branch because of unpredictable branch. "; 867 EmitAndPrintRemark(ORE, ORmiss); 868 return false; 869 } 870 871 // If highly predictable, branch form is more profitable, unless a 872 // predictable select is inexpensive in the target architecture. 873 if (isSelectHighlyPredictable(SI) && TLI->isPredictableSelectExpensive()) { 874 ++NumSelectConvertedHighPred; 875 OR << "Converted to branch because of highly predictable branch. "; 876 EmitAndPrintRemark(ORE, OR); 877 return true; 878 } 879 880 // Look for expensive instructions in the cold operand's (if any) dependence 881 // slice of any of the selects in the group. 882 if (hasExpensiveColdOperand(ASI)) { 883 ++NumSelectConvertedExpColdOperand; 884 OR << "Converted to branch because of expensive cold operand."; 885 EmitAndPrintRemark(ORE, OR); 886 return true; 887 } 888 889 ORmiss << "Not profitable to convert to branch (base heuristic)."; 890 EmitAndPrintRemark(ORE, ORmiss); 891 return false; 892 } 893 894 static InstructionCost divideNearest(InstructionCost Numerator, 895 uint64_t Denominator) { 896 return (Numerator + (Denominator / 2)) / Denominator; 897 } 898 899 static bool extractBranchWeights(const SelectOptimizeImpl::SelectLike SI, 900 uint64_t &TrueVal, uint64_t &FalseVal) { 901 if (isa<SelectInst>(SI.getI())) 902 return extractBranchWeights(*SI.getI(), TrueVal, FalseVal); 903 return false; 904 } 905 906 bool SelectOptimizeImpl::hasExpensiveColdOperand(const SelectGroup &ASI) { 907 bool ColdOperand = false; 908 uint64_t TrueWeight, FalseWeight, TotalWeight; 909 if (extractBranchWeights(ASI.front(), TrueWeight, FalseWeight)) { 910 uint64_t MinWeight = std::min(TrueWeight, FalseWeight); 911 TotalWeight = TrueWeight + FalseWeight; 912 // Is there a path with frequency <ColdOperandThreshold% (default:20%) ? 913 ColdOperand = TotalWeight * ColdOperandThreshold > 100 * MinWeight; 914 } else if (PSI->hasProfileSummary()) { 915 OptimizationRemarkMissed ORmiss(DEBUG_TYPE, "SelectOpti", 916 ASI.front().getI()); 917 ORmiss << "Profile data available but missing branch-weights metadata for " 918 "select instruction. "; 919 EmitAndPrintRemark(ORE, ORmiss); 920 } 921 if (!ColdOperand) 922 return false; 923 // Check if the cold path's dependence slice is expensive for any of the 924 // selects of the group. 925 for (SelectLike SI : ASI) { 926 Instruction *ColdI = nullptr; 927 uint64_t HotWeight; 928 if (TrueWeight < FalseWeight) { 929 ColdI = dyn_cast_or_null<Instruction>(SI.getTrueValue()); 930 HotWeight = FalseWeight; 931 } else { 932 ColdI = dyn_cast_or_null<Instruction>(SI.getFalseValue()); 933 HotWeight = TrueWeight; 934 } 935 if (ColdI) { 936 std::stack<Instruction *> ColdSlice; 937 getExclBackwardsSlice(ColdI, ColdSlice, SI.getI()); 938 InstructionCost SliceCost = 0; 939 while (!ColdSlice.empty()) { 940 SliceCost += TTI->getInstructionCost(ColdSlice.top(), 941 TargetTransformInfo::TCK_Latency); 942 ColdSlice.pop(); 943 } 944 // The colder the cold value operand of the select is the more expensive 945 // the cmov becomes for computing the cold value operand every time. Thus, 946 // the colder the cold operand is the more its cost counts. 947 // Get nearest integer cost adjusted for coldness. 948 InstructionCost AdjSliceCost = 949 divideNearest(SliceCost * HotWeight, TotalWeight); 950 if (AdjSliceCost >= 951 ColdOperandMaxCostMultiplier * TargetTransformInfo::TCC_Expensive) 952 return true; 953 } 954 } 955 return false; 956 } 957 958 // Check if it is safe to move LoadI next to the SI. 959 // Conservatively assume it is safe only if there is no instruction 960 // modifying memory in-between the load and the select instruction. 961 static bool isSafeToSinkLoad(Instruction *LoadI, Instruction *SI) { 962 // Assume loads from different basic blocks are unsafe to move. 963 if (LoadI->getParent() != SI->getParent()) 964 return false; 965 auto It = LoadI->getIterator(); 966 while (&*It != SI) { 967 if (It->mayWriteToMemory()) 968 return false; 969 It++; 970 } 971 return true; 972 } 973 974 // For a given source instruction, collect its backwards dependence slice 975 // consisting of instructions exclusively computed for the purpose of producing 976 // the operands of the source instruction. As an approximation 977 // (sufficiently-accurate in practice), we populate this set with the 978 // instructions of the backwards dependence slice that only have one-use and 979 // form an one-use chain that leads to the source instruction. 980 void SelectOptimizeImpl::getExclBackwardsSlice(Instruction *I, 981 std::stack<Instruction *> &Slice, 982 Instruction *SI, 983 bool ForSinking) { 984 SmallPtrSet<Instruction *, 2> Visited; 985 std::queue<Instruction *> Worklist; 986 Worklist.push(I); 987 while (!Worklist.empty()) { 988 Instruction *II = Worklist.front(); 989 Worklist.pop(); 990 991 // Avoid cycles. 992 if (!Visited.insert(II).second) 993 continue; 994 995 if (!II->hasOneUse()) 996 continue; 997 998 // Cannot soundly sink instructions with side-effects. 999 // Terminator or phi instructions cannot be sunk. 1000 // Avoid sinking other select instructions (should be handled separetely). 1001 if (ForSinking && (II->isTerminator() || II->mayHaveSideEffects() || 1002 isa<SelectInst>(II) || isa<PHINode>(II))) 1003 continue; 1004 1005 // Avoid sinking loads in order not to skip state-modifying instructions, 1006 // that may alias with the loaded address. 1007 // Only allow sinking of loads within the same basic block that are 1008 // conservatively proven to be safe. 1009 if (ForSinking && II->mayReadFromMemory() && !isSafeToSinkLoad(II, SI)) 1010 continue; 1011 1012 // Avoid considering instructions with less frequency than the source 1013 // instruction (i.e., avoid colder code regions of the dependence slice). 1014 if (BFI->getBlockFreq(II->getParent()) < BFI->getBlockFreq(I->getParent())) 1015 continue; 1016 1017 // Eligible one-use instruction added to the dependence slice. 1018 Slice.push(II); 1019 1020 // Explore all the operands of the current instruction to expand the slice. 1021 for (unsigned k = 0; k < II->getNumOperands(); ++k) 1022 if (auto *OpI = dyn_cast<Instruction>(II->getOperand(k))) 1023 Worklist.push(OpI); 1024 } 1025 } 1026 1027 bool SelectOptimizeImpl::isSelectHighlyPredictable(const SelectLike SI) { 1028 uint64_t TrueWeight, FalseWeight; 1029 if (extractBranchWeights(SI, TrueWeight, FalseWeight)) { 1030 uint64_t Max = std::max(TrueWeight, FalseWeight); 1031 uint64_t Sum = TrueWeight + FalseWeight; 1032 if (Sum != 0) { 1033 auto Probability = BranchProbability::getBranchProbability(Max, Sum); 1034 if (Probability > TTI->getPredictableBranchThreshold()) 1035 return true; 1036 } 1037 } 1038 return false; 1039 } 1040 1041 bool SelectOptimizeImpl::checkLoopHeuristics(const Loop *L, 1042 const CostInfo LoopCost[2]) { 1043 // Loop-level checks to determine if a non-predicated version (with branches) 1044 // of the loop is more profitable than its predicated version. 1045 1046 if (DisableLoopLevelHeuristics) 1047 return true; 1048 1049 OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti", 1050 L->getHeader()->getFirstNonPHI()); 1051 1052 if (LoopCost[0].NonPredCost > LoopCost[0].PredCost || 1053 LoopCost[1].NonPredCost >= LoopCost[1].PredCost) { 1054 ORmissL << "No select conversion in the loop due to no reduction of loop's " 1055 "critical path. "; 1056 EmitAndPrintRemark(ORE, ORmissL); 1057 return false; 1058 } 1059 1060 Scaled64 Gain[2] = {LoopCost[0].PredCost - LoopCost[0].NonPredCost, 1061 LoopCost[1].PredCost - LoopCost[1].NonPredCost}; 1062 1063 // Profitably converting to branches need to reduce the loop's critical path 1064 // by at least some threshold (absolute gain of GainCycleThreshold cycles and 1065 // relative gain of 12.5%). 1066 if (Gain[1] < Scaled64::get(GainCycleThreshold) || 1067 Gain[1] * Scaled64::get(GainRelativeThreshold) < LoopCost[1].PredCost) { 1068 Scaled64 RelativeGain = Scaled64::get(100) * Gain[1] / LoopCost[1].PredCost; 1069 ORmissL << "No select conversion in the loop due to small reduction of " 1070 "loop's critical path. Gain=" 1071 << Gain[1].toString() 1072 << ", RelativeGain=" << RelativeGain.toString() << "%. "; 1073 EmitAndPrintRemark(ORE, ORmissL); 1074 return false; 1075 } 1076 1077 // If the loop's critical path involves loop-carried dependences, the gradient 1078 // of the gain needs to be at least GainGradientThreshold% (defaults to 25%). 1079 // This check ensures that the latency reduction for the loop's critical path 1080 // keeps decreasing with sufficient rate beyond the two analyzed loop 1081 // iterations. 1082 if (Gain[1] > Gain[0]) { 1083 Scaled64 GradientGain = Scaled64::get(100) * (Gain[1] - Gain[0]) / 1084 (LoopCost[1].PredCost - LoopCost[0].PredCost); 1085 if (GradientGain < Scaled64::get(GainGradientThreshold)) { 1086 ORmissL << "No select conversion in the loop due to small gradient gain. " 1087 "GradientGain=" 1088 << GradientGain.toString() << "%. "; 1089 EmitAndPrintRemark(ORE, ORmissL); 1090 return false; 1091 } 1092 } 1093 // If the gain decreases it is not profitable to convert. 1094 else if (Gain[1] < Gain[0]) { 1095 ORmissL 1096 << "No select conversion in the loop due to negative gradient gain. "; 1097 EmitAndPrintRemark(ORE, ORmissL); 1098 return false; 1099 } 1100 1101 // Non-predicated version of the loop is more profitable than its 1102 // predicated version. 1103 return true; 1104 } 1105 1106 // Computes instruction and loop-critical-path costs for both the predicated 1107 // and non-predicated version of the given loop. 1108 // Returns false if unable to compute these costs due to invalid cost of loop 1109 // instruction(s). 1110 bool SelectOptimizeImpl::computeLoopCosts( 1111 const Loop *L, const SelectGroups &SIGroups, 1112 DenseMap<const Instruction *, CostInfo> &InstCostMap, CostInfo *LoopCost) { 1113 LLVM_DEBUG(dbgs() << "Calculating Latency / IPredCost / INonPredCost of loop " 1114 << L->getHeader()->getName() << "\n"); 1115 const auto &SImap = getSImap(SIGroups); 1116 // Compute instruction and loop-critical-path costs across two iterations for 1117 // both predicated and non-predicated version. 1118 const unsigned Iterations = 2; 1119 for (unsigned Iter = 0; Iter < Iterations; ++Iter) { 1120 // Cost of the loop's critical path. 1121 CostInfo &MaxCost = LoopCost[Iter]; 1122 for (BasicBlock *BB : L->getBlocks()) { 1123 for (const Instruction &I : *BB) { 1124 if (I.isDebugOrPseudoInst()) 1125 continue; 1126 // Compute the predicated and non-predicated cost of the instruction. 1127 Scaled64 IPredCost = Scaled64::getZero(), 1128 INonPredCost = Scaled64::getZero(); 1129 1130 // Assume infinite resources that allow to fully exploit the available 1131 // instruction-level parallelism. 1132 // InstCost = InstLatency + max(Op1Cost, Op2Cost, … OpNCost) 1133 for (const Use &U : I.operands()) { 1134 auto UI = dyn_cast<Instruction>(U.get()); 1135 if (!UI) 1136 continue; 1137 if (InstCostMap.count(UI)) { 1138 IPredCost = std::max(IPredCost, InstCostMap[UI].PredCost); 1139 INonPredCost = std::max(INonPredCost, InstCostMap[UI].NonPredCost); 1140 } 1141 } 1142 auto ILatency = computeInstCost(&I); 1143 if (!ILatency) { 1144 OptimizationRemarkMissed ORmissL(DEBUG_TYPE, "SelectOpti", &I); 1145 ORmissL << "Invalid instruction cost preventing analysis and " 1146 "optimization of the inner-most loop containing this " 1147 "instruction. "; 1148 EmitAndPrintRemark(ORE, ORmissL); 1149 return false; 1150 } 1151 IPredCost += Scaled64::get(*ILatency); 1152 INonPredCost += Scaled64::get(*ILatency); 1153 1154 // For a select that can be converted to branch, 1155 // compute its cost as a branch (non-predicated cost). 1156 // 1157 // BranchCost = PredictedPathCost + MispredictCost 1158 // PredictedPathCost = TrueOpCost * TrueProb + FalseOpCost * FalseProb 1159 // MispredictCost = max(MispredictPenalty, CondCost) * MispredictRate 1160 if (SImap.contains(&I)) { 1161 auto SI = SImap.at(&I); 1162 Scaled64 TrueOpCost = SI.getTrueOpCost(InstCostMap, TTI); 1163 Scaled64 FalseOpCost = SI.getFalseOpCost(InstCostMap, TTI); 1164 Scaled64 PredictedPathCost = 1165 getPredictedPathCost(TrueOpCost, FalseOpCost, SI); 1166 1167 Scaled64 CondCost = Scaled64::getZero(); 1168 if (auto *CI = dyn_cast<Instruction>(SI.getCondition())) 1169 if (InstCostMap.count(CI)) 1170 CondCost = InstCostMap[CI].NonPredCost; 1171 Scaled64 MispredictCost = getMispredictionCost(SI, CondCost); 1172 1173 INonPredCost = PredictedPathCost + MispredictCost; 1174 } 1175 LLVM_DEBUG(dbgs() << " " << ILatency << "/" << IPredCost << "/" 1176 << INonPredCost << " for " << I << "\n"); 1177 1178 InstCostMap[&I] = {IPredCost, INonPredCost}; 1179 MaxCost.PredCost = std::max(MaxCost.PredCost, IPredCost); 1180 MaxCost.NonPredCost = std::max(MaxCost.NonPredCost, INonPredCost); 1181 } 1182 } 1183 LLVM_DEBUG(dbgs() << "Iteration " << Iter + 1 1184 << " MaxCost = " << MaxCost.PredCost << " " 1185 << MaxCost.NonPredCost << "\n"); 1186 } 1187 return true; 1188 } 1189 1190 SmallDenseMap<const Instruction *, SelectOptimizeImpl::SelectLike, 2> 1191 SelectOptimizeImpl::getSImap(const SelectGroups &SIGroups) { 1192 SmallDenseMap<const Instruction *, SelectLike, 2> SImap; 1193 for (const SelectGroup &ASI : SIGroups) 1194 for (SelectLike SI : ASI) 1195 SImap.try_emplace(SI.getI(), SI); 1196 return SImap; 1197 } 1198 1199 std::optional<uint64_t> 1200 SelectOptimizeImpl::computeInstCost(const Instruction *I) { 1201 InstructionCost ICost = 1202 TTI->getInstructionCost(I, TargetTransformInfo::TCK_Latency); 1203 if (auto OC = ICost.getValue()) 1204 return std::optional<uint64_t>(*OC); 1205 return std::nullopt; 1206 } 1207 1208 ScaledNumber<uint64_t> 1209 SelectOptimizeImpl::getMispredictionCost(const SelectLike SI, 1210 const Scaled64 CondCost) { 1211 uint64_t MispredictPenalty = TSchedModel.getMCSchedModel()->MispredictPenalty; 1212 1213 // Account for the default misprediction rate when using a branch 1214 // (conservatively set to 25% by default). 1215 uint64_t MispredictRate = MispredictDefaultRate; 1216 // If the select condition is obviously predictable, then the misprediction 1217 // rate is zero. 1218 if (isSelectHighlyPredictable(SI)) 1219 MispredictRate = 0; 1220 1221 // CondCost is included to account for cases where the computation of the 1222 // condition is part of a long dependence chain (potentially loop-carried) 1223 // that would delay detection of a misprediction and increase its cost. 1224 Scaled64 MispredictCost = 1225 std::max(Scaled64::get(MispredictPenalty), CondCost) * 1226 Scaled64::get(MispredictRate); 1227 MispredictCost /= Scaled64::get(100); 1228 1229 return MispredictCost; 1230 } 1231 1232 // Returns the cost of a branch when the prediction is correct. 1233 // TrueCost * TrueProbability + FalseCost * FalseProbability. 1234 ScaledNumber<uint64_t> 1235 SelectOptimizeImpl::getPredictedPathCost(Scaled64 TrueCost, Scaled64 FalseCost, 1236 const SelectLike SI) { 1237 Scaled64 PredPathCost; 1238 uint64_t TrueWeight, FalseWeight; 1239 if (extractBranchWeights(SI, TrueWeight, FalseWeight)) { 1240 uint64_t SumWeight = TrueWeight + FalseWeight; 1241 if (SumWeight != 0) { 1242 PredPathCost = TrueCost * Scaled64::get(TrueWeight) + 1243 FalseCost * Scaled64::get(FalseWeight); 1244 PredPathCost /= Scaled64::get(SumWeight); 1245 return PredPathCost; 1246 } 1247 } 1248 // Without branch weight metadata, we assume 75% for the one path and 25% for 1249 // the other, and pick the result with the biggest cost. 1250 PredPathCost = std::max(TrueCost * Scaled64::get(3) + FalseCost, 1251 FalseCost * Scaled64::get(3) + TrueCost); 1252 PredPathCost /= Scaled64::get(4); 1253 return PredPathCost; 1254 } 1255 1256 bool SelectOptimizeImpl::isSelectKindSupported(const SelectLike SI) { 1257 bool VectorCond = !SI.getCondition()->getType()->isIntegerTy(1); 1258 if (VectorCond) 1259 return false; 1260 TargetLowering::SelectSupportKind SelectKind; 1261 if (SI.getType()->isVectorTy()) 1262 SelectKind = TargetLowering::ScalarCondVectorVal; 1263 else 1264 SelectKind = TargetLowering::ScalarValSelect; 1265 return TLI->isSelectSupported(SelectKind); 1266 } 1267