1 //===- HexagonLoopIdiomRecognition.cpp ------------------------------------===// 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 #include "HexagonLoopIdiomRecognition.h" 10 #include "llvm/ADT/APInt.h" 11 #include "llvm/ADT/DenseMap.h" 12 #include "llvm/ADT/SetVector.h" 13 #include "llvm/ADT/SmallPtrSet.h" 14 #include "llvm/ADT/SmallSet.h" 15 #include "llvm/ADT/SmallVector.h" 16 #include "llvm/ADT/StringRef.h" 17 #include "llvm/Analysis/AliasAnalysis.h" 18 #include "llvm/Analysis/InstructionSimplify.h" 19 #include "llvm/Analysis/LoopAnalysisManager.h" 20 #include "llvm/Analysis/LoopInfo.h" 21 #include "llvm/Analysis/LoopPass.h" 22 #include "llvm/Analysis/MemoryLocation.h" 23 #include "llvm/Analysis/ScalarEvolution.h" 24 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 25 #include "llvm/Analysis/TargetLibraryInfo.h" 26 #include "llvm/Analysis/ValueTracking.h" 27 #include "llvm/IR/Attributes.h" 28 #include "llvm/IR/BasicBlock.h" 29 #include "llvm/IR/Constant.h" 30 #include "llvm/IR/Constants.h" 31 #include "llvm/IR/DataLayout.h" 32 #include "llvm/IR/DebugLoc.h" 33 #include "llvm/IR/DerivedTypes.h" 34 #include "llvm/IR/Dominators.h" 35 #include "llvm/IR/Function.h" 36 #include "llvm/IR/IRBuilder.h" 37 #include "llvm/IR/InstrTypes.h" 38 #include "llvm/IR/Instruction.h" 39 #include "llvm/IR/Instructions.h" 40 #include "llvm/IR/IntrinsicInst.h" 41 #include "llvm/IR/Intrinsics.h" 42 #include "llvm/IR/IntrinsicsHexagon.h" 43 #include "llvm/IR/Module.h" 44 #include "llvm/IR/PassManager.h" 45 #include "llvm/IR/PatternMatch.h" 46 #include "llvm/IR/Type.h" 47 #include "llvm/IR/User.h" 48 #include "llvm/IR/Value.h" 49 #include "llvm/InitializePasses.h" 50 #include "llvm/Pass.h" 51 #include "llvm/Support/Casting.h" 52 #include "llvm/Support/CommandLine.h" 53 #include "llvm/Support/Compiler.h" 54 #include "llvm/Support/Debug.h" 55 #include "llvm/Support/ErrorHandling.h" 56 #include "llvm/Support/KnownBits.h" 57 #include "llvm/Support/raw_ostream.h" 58 #include "llvm/TargetParser/Triple.h" 59 #include "llvm/Transforms/Scalar.h" 60 #include "llvm/Transforms/Utils.h" 61 #include "llvm/Transforms/Utils/Local.h" 62 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 63 #include <algorithm> 64 #include <array> 65 #include <cassert> 66 #include <cstdint> 67 #include <cstdlib> 68 #include <deque> 69 #include <functional> 70 #include <iterator> 71 #include <map> 72 #include <set> 73 #include <utility> 74 #include <vector> 75 76 #define DEBUG_TYPE "hexagon-lir" 77 78 using namespace llvm; 79 80 static cl::opt<bool> DisableMemcpyIdiom("disable-memcpy-idiom", 81 cl::Hidden, cl::init(false), 82 cl::desc("Disable generation of memcpy in loop idiom recognition")); 83 84 static cl::opt<bool> DisableMemmoveIdiom("disable-memmove-idiom", 85 cl::Hidden, cl::init(false), 86 cl::desc("Disable generation of memmove in loop idiom recognition")); 87 88 static cl::opt<unsigned> RuntimeMemSizeThreshold("runtime-mem-idiom-threshold", 89 cl::Hidden, cl::init(0), cl::desc("Threshold (in bytes) for the runtime " 90 "check guarding the memmove.")); 91 92 static cl::opt<unsigned> CompileTimeMemSizeThreshold( 93 "compile-time-mem-idiom-threshold", cl::Hidden, cl::init(64), 94 cl::desc("Threshold (in bytes) to perform the transformation, if the " 95 "runtime loop count (mem transfer size) is known at compile-time.")); 96 97 static cl::opt<bool> OnlyNonNestedMemmove("only-nonnested-memmove-idiom", 98 cl::Hidden, cl::init(true), 99 cl::desc("Only enable generating memmove in non-nested loops")); 100 101 static cl::opt<bool> HexagonVolatileMemcpy( 102 "disable-hexagon-volatile-memcpy", cl::Hidden, cl::init(false), 103 cl::desc("Enable Hexagon-specific memcpy for volatile destination.")); 104 105 static cl::opt<unsigned> SimplifyLimit("hlir-simplify-limit", cl::init(10000), 106 cl::Hidden, cl::desc("Maximum number of simplification steps in HLIR")); 107 108 static const char *HexagonVolatileMemcpyName 109 = "hexagon_memcpy_forward_vp4cp4n2"; 110 111 112 namespace llvm { 113 114 void initializeHexagonLoopIdiomRecognizeLegacyPassPass(PassRegistry &); 115 Pass *createHexagonLoopIdiomPass(); 116 117 } // end namespace llvm 118 119 namespace { 120 121 class HexagonLoopIdiomRecognize { 122 public: 123 explicit HexagonLoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT, 124 LoopInfo *LF, const TargetLibraryInfo *TLI, 125 ScalarEvolution *SE) 126 : AA(AA), DT(DT), LF(LF), TLI(TLI), SE(SE) {} 127 128 bool run(Loop *L); 129 130 private: 131 int getSCEVStride(const SCEVAddRecExpr *StoreEv); 132 bool isLegalStore(Loop *CurLoop, StoreInst *SI); 133 void collectStores(Loop *CurLoop, BasicBlock *BB, 134 SmallVectorImpl<StoreInst *> &Stores); 135 bool processCopyingStore(Loop *CurLoop, StoreInst *SI, const SCEV *BECount); 136 bool coverLoop(Loop *L, SmallVectorImpl<Instruction *> &Insts) const; 137 bool runOnLoopBlock(Loop *CurLoop, BasicBlock *BB, const SCEV *BECount, 138 SmallVectorImpl<BasicBlock *> &ExitBlocks); 139 bool runOnCountableLoop(Loop *L); 140 141 AliasAnalysis *AA; 142 const DataLayout *DL; 143 DominatorTree *DT; 144 LoopInfo *LF; 145 const TargetLibraryInfo *TLI; 146 ScalarEvolution *SE; 147 bool HasMemcpy, HasMemmove; 148 }; 149 150 class HexagonLoopIdiomRecognizeLegacyPass : public LoopPass { 151 public: 152 static char ID; 153 154 explicit HexagonLoopIdiomRecognizeLegacyPass() : LoopPass(ID) { 155 initializeHexagonLoopIdiomRecognizeLegacyPassPass( 156 *PassRegistry::getPassRegistry()); 157 } 158 159 StringRef getPassName() const override { 160 return "Recognize Hexagon-specific loop idioms"; 161 } 162 163 void getAnalysisUsage(AnalysisUsage &AU) const override { 164 AU.addRequired<LoopInfoWrapperPass>(); 165 AU.addRequiredID(LoopSimplifyID); 166 AU.addRequiredID(LCSSAID); 167 AU.addRequired<AAResultsWrapperPass>(); 168 AU.addRequired<ScalarEvolutionWrapperPass>(); 169 AU.addRequired<DominatorTreeWrapperPass>(); 170 AU.addRequired<TargetLibraryInfoWrapperPass>(); 171 AU.addPreserved<TargetLibraryInfoWrapperPass>(); 172 } 173 174 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 175 }; 176 177 struct Simplifier { 178 struct Rule { 179 using FuncType = std::function<Value *(Instruction *, LLVMContext &)>; 180 Rule(StringRef N, FuncType F) : Name(N), Fn(F) {} 181 StringRef Name; // For debugging. 182 FuncType Fn; 183 }; 184 185 void addRule(StringRef N, const Rule::FuncType &F) { 186 Rules.push_back(Rule(N, F)); 187 } 188 189 private: 190 struct WorkListType { 191 WorkListType() = default; 192 193 void push_back(Value *V) { 194 // Do not push back duplicates. 195 if (S.insert(V).second) 196 Q.push_back(V); 197 } 198 199 Value *pop_front_val() { 200 Value *V = Q.front(); 201 Q.pop_front(); 202 S.erase(V); 203 return V; 204 } 205 206 bool empty() const { return Q.empty(); } 207 208 private: 209 std::deque<Value *> Q; 210 std::set<Value *> S; 211 }; 212 213 using ValueSetType = std::set<Value *>; 214 215 std::vector<Rule> Rules; 216 217 public: 218 struct Context { 219 using ValueMapType = DenseMap<Value *, Value *>; 220 221 Value *Root; 222 ValueSetType Used; // The set of all cloned values used by Root. 223 ValueSetType Clones; // The set of all cloned values. 224 LLVMContext &Ctx; 225 226 Context(Instruction *Exp) 227 : Ctx(Exp->getParent()->getParent()->getContext()) { 228 initialize(Exp); 229 } 230 231 ~Context() { cleanup(); } 232 233 void print(raw_ostream &OS, const Value *V) const; 234 Value *materialize(BasicBlock *B, BasicBlock::iterator At); 235 236 private: 237 friend struct Simplifier; 238 239 void initialize(Instruction *Exp); 240 void cleanup(); 241 242 template <typename FuncT> void traverse(Value *V, FuncT F); 243 void record(Value *V); 244 void use(Value *V); 245 void unuse(Value *V); 246 247 bool equal(const Instruction *I, const Instruction *J) const; 248 Value *find(Value *Tree, Value *Sub) const; 249 Value *subst(Value *Tree, Value *OldV, Value *NewV); 250 void replace(Value *OldV, Value *NewV); 251 void link(Instruction *I, BasicBlock *B, BasicBlock::iterator At); 252 }; 253 254 Value *simplify(Context &C); 255 }; 256 257 struct PE { 258 PE(const Simplifier::Context &c, Value *v = nullptr) : C(c), V(v) {} 259 260 const Simplifier::Context &C; 261 const Value *V; 262 }; 263 264 LLVM_ATTRIBUTE_USED 265 raw_ostream &operator<<(raw_ostream &OS, const PE &P) { 266 P.C.print(OS, P.V ? P.V : P.C.Root); 267 return OS; 268 } 269 270 } // end anonymous namespace 271 272 char HexagonLoopIdiomRecognizeLegacyPass::ID = 0; 273 274 INITIALIZE_PASS_BEGIN(HexagonLoopIdiomRecognizeLegacyPass, "hexagon-loop-idiom", 275 "Recognize Hexagon-specific loop idioms", false, false) 276 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 277 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 278 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass) 279 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 280 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 281 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 282 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 283 INITIALIZE_PASS_END(HexagonLoopIdiomRecognizeLegacyPass, "hexagon-loop-idiom", 284 "Recognize Hexagon-specific loop idioms", false, false) 285 286 template <typename FuncT> 287 void Simplifier::Context::traverse(Value *V, FuncT F) { 288 WorkListType Q; 289 Q.push_back(V); 290 291 while (!Q.empty()) { 292 Instruction *U = dyn_cast<Instruction>(Q.pop_front_val()); 293 if (!U || U->getParent()) 294 continue; 295 if (!F(U)) 296 continue; 297 for (Value *Op : U->operands()) 298 Q.push_back(Op); 299 } 300 } 301 302 void Simplifier::Context::print(raw_ostream &OS, const Value *V) const { 303 const auto *U = dyn_cast<const Instruction>(V); 304 if (!U) { 305 OS << V << '(' << *V << ')'; 306 return; 307 } 308 309 if (U->getParent()) { 310 OS << U << '('; 311 U->printAsOperand(OS, true); 312 OS << ')'; 313 return; 314 } 315 316 unsigned N = U->getNumOperands(); 317 if (N != 0) 318 OS << U << '('; 319 OS << U->getOpcodeName(); 320 for (const Value *Op : U->operands()) { 321 OS << ' '; 322 print(OS, Op); 323 } 324 if (N != 0) 325 OS << ')'; 326 } 327 328 void Simplifier::Context::initialize(Instruction *Exp) { 329 // Perform a deep clone of the expression, set Root to the root 330 // of the clone, and build a map from the cloned values to the 331 // original ones. 332 ValueMapType M; 333 BasicBlock *Block = Exp->getParent(); 334 WorkListType Q; 335 Q.push_back(Exp); 336 337 while (!Q.empty()) { 338 Value *V = Q.pop_front_val(); 339 if (M.contains(V)) 340 continue; 341 if (Instruction *U = dyn_cast<Instruction>(V)) { 342 if (isa<PHINode>(U) || U->getParent() != Block) 343 continue; 344 for (Value *Op : U->operands()) 345 Q.push_back(Op); 346 M.insert({U, U->clone()}); 347 } 348 } 349 350 for (std::pair<Value*,Value*> P : M) { 351 Instruction *U = cast<Instruction>(P.second); 352 for (unsigned i = 0, n = U->getNumOperands(); i != n; ++i) { 353 auto F = M.find(U->getOperand(i)); 354 if (F != M.end()) 355 U->setOperand(i, F->second); 356 } 357 } 358 359 auto R = M.find(Exp); 360 assert(R != M.end()); 361 Root = R->second; 362 363 record(Root); 364 use(Root); 365 } 366 367 void Simplifier::Context::record(Value *V) { 368 auto Record = [this](Instruction *U) -> bool { 369 Clones.insert(U); 370 return true; 371 }; 372 traverse(V, Record); 373 } 374 375 void Simplifier::Context::use(Value *V) { 376 auto Use = [this](Instruction *U) -> bool { 377 Used.insert(U); 378 return true; 379 }; 380 traverse(V, Use); 381 } 382 383 void Simplifier::Context::unuse(Value *V) { 384 if (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != nullptr) 385 return; 386 387 auto Unuse = [this](Instruction *U) -> bool { 388 if (!U->use_empty()) 389 return false; 390 Used.erase(U); 391 return true; 392 }; 393 traverse(V, Unuse); 394 } 395 396 Value *Simplifier::Context::subst(Value *Tree, Value *OldV, Value *NewV) { 397 if (Tree == OldV) 398 return NewV; 399 if (OldV == NewV) 400 return Tree; 401 402 WorkListType Q; 403 Q.push_back(Tree); 404 while (!Q.empty()) { 405 Instruction *U = dyn_cast<Instruction>(Q.pop_front_val()); 406 // If U is not an instruction, or it's not a clone, skip it. 407 if (!U || U->getParent()) 408 continue; 409 for (unsigned i = 0, n = U->getNumOperands(); i != n; ++i) { 410 Value *Op = U->getOperand(i); 411 if (Op == OldV) { 412 U->setOperand(i, NewV); 413 unuse(OldV); 414 } else { 415 Q.push_back(Op); 416 } 417 } 418 } 419 return Tree; 420 } 421 422 void Simplifier::Context::replace(Value *OldV, Value *NewV) { 423 if (Root == OldV) { 424 Root = NewV; 425 use(Root); 426 return; 427 } 428 429 // NewV may be a complex tree that has just been created by one of the 430 // transformation rules. We need to make sure that it is commoned with 431 // the existing Root to the maximum extent possible. 432 // Identify all subtrees of NewV (including NewV itself) that have 433 // equivalent counterparts in Root, and replace those subtrees with 434 // these counterparts. 435 WorkListType Q; 436 Q.push_back(NewV); 437 while (!Q.empty()) { 438 Value *V = Q.pop_front_val(); 439 Instruction *U = dyn_cast<Instruction>(V); 440 if (!U || U->getParent()) 441 continue; 442 if (Value *DupV = find(Root, V)) { 443 if (DupV != V) 444 NewV = subst(NewV, V, DupV); 445 } else { 446 for (Value *Op : U->operands()) 447 Q.push_back(Op); 448 } 449 } 450 451 // Now, simply replace OldV with NewV in Root. 452 Root = subst(Root, OldV, NewV); 453 use(Root); 454 } 455 456 void Simplifier::Context::cleanup() { 457 for (Value *V : Clones) { 458 Instruction *U = cast<Instruction>(V); 459 if (!U->getParent()) 460 U->dropAllReferences(); 461 } 462 463 for (Value *V : Clones) { 464 Instruction *U = cast<Instruction>(V); 465 if (!U->getParent()) 466 U->deleteValue(); 467 } 468 } 469 470 bool Simplifier::Context::equal(const Instruction *I, 471 const Instruction *J) const { 472 if (I == J) 473 return true; 474 if (!I->isSameOperationAs(J)) 475 return false; 476 if (isa<PHINode>(I)) 477 return I->isIdenticalTo(J); 478 479 for (unsigned i = 0, n = I->getNumOperands(); i != n; ++i) { 480 Value *OpI = I->getOperand(i), *OpJ = J->getOperand(i); 481 if (OpI == OpJ) 482 continue; 483 auto *InI = dyn_cast<const Instruction>(OpI); 484 auto *InJ = dyn_cast<const Instruction>(OpJ); 485 if (InI && InJ) { 486 if (!equal(InI, InJ)) 487 return false; 488 } else if (InI != InJ || !InI) 489 return false; 490 } 491 return true; 492 } 493 494 Value *Simplifier::Context::find(Value *Tree, Value *Sub) const { 495 Instruction *SubI = dyn_cast<Instruction>(Sub); 496 WorkListType Q; 497 Q.push_back(Tree); 498 499 while (!Q.empty()) { 500 Value *V = Q.pop_front_val(); 501 if (V == Sub) 502 return V; 503 Instruction *U = dyn_cast<Instruction>(V); 504 if (!U || U->getParent()) 505 continue; 506 if (SubI && equal(SubI, U)) 507 return U; 508 assert(!isa<PHINode>(U)); 509 for (Value *Op : U->operands()) 510 Q.push_back(Op); 511 } 512 return nullptr; 513 } 514 515 void Simplifier::Context::link(Instruction *I, BasicBlock *B, 516 BasicBlock::iterator At) { 517 if (I->getParent()) 518 return; 519 520 for (Value *Op : I->operands()) { 521 if (Instruction *OpI = dyn_cast<Instruction>(Op)) 522 link(OpI, B, At); 523 } 524 525 I->insertInto(B, At); 526 } 527 528 Value *Simplifier::Context::materialize(BasicBlock *B, 529 BasicBlock::iterator At) { 530 if (Instruction *RootI = dyn_cast<Instruction>(Root)) 531 link(RootI, B, At); 532 return Root; 533 } 534 535 Value *Simplifier::simplify(Context &C) { 536 WorkListType Q; 537 Q.push_back(C.Root); 538 unsigned Count = 0; 539 const unsigned Limit = SimplifyLimit; 540 541 while (!Q.empty()) { 542 if (Count++ >= Limit) 543 break; 544 Instruction *U = dyn_cast<Instruction>(Q.pop_front_val()); 545 if (!U || U->getParent() || !C.Used.count(U)) 546 continue; 547 bool Changed = false; 548 for (Rule &R : Rules) { 549 Value *W = R.Fn(U, C.Ctx); 550 if (!W) 551 continue; 552 Changed = true; 553 C.record(W); 554 C.replace(U, W); 555 Q.push_back(C.Root); 556 break; 557 } 558 if (!Changed) { 559 for (Value *Op : U->operands()) 560 Q.push_back(Op); 561 } 562 } 563 return Count < Limit ? C.Root : nullptr; 564 } 565 566 //===----------------------------------------------------------------------===// 567 // 568 // Implementation of PolynomialMultiplyRecognize 569 // 570 //===----------------------------------------------------------------------===// 571 572 namespace { 573 574 class PolynomialMultiplyRecognize { 575 public: 576 explicit PolynomialMultiplyRecognize(Loop *loop, const DataLayout &dl, 577 const DominatorTree &dt, const TargetLibraryInfo &tli, 578 ScalarEvolution &se) 579 : CurLoop(loop), DL(dl), DT(dt), TLI(tli), SE(se) {} 580 581 bool recognize(); 582 583 private: 584 using ValueSeq = SetVector<Value *>; 585 586 IntegerType *getPmpyType() const { 587 LLVMContext &Ctx = CurLoop->getHeader()->getParent()->getContext(); 588 return IntegerType::get(Ctx, 32); 589 } 590 591 bool isPromotableTo(Value *V, IntegerType *Ty); 592 void promoteTo(Instruction *In, IntegerType *DestTy, BasicBlock *LoopB); 593 bool promoteTypes(BasicBlock *LoopB, BasicBlock *ExitB); 594 595 Value *getCountIV(BasicBlock *BB); 596 bool findCycle(Value *Out, Value *In, ValueSeq &Cycle); 597 void classifyCycle(Instruction *DivI, ValueSeq &Cycle, ValueSeq &Early, 598 ValueSeq &Late); 599 bool classifyInst(Instruction *UseI, ValueSeq &Early, ValueSeq &Late); 600 bool commutesWithShift(Instruction *I); 601 bool highBitsAreZero(Value *V, unsigned IterCount); 602 bool keepsHighBitsZero(Value *V, unsigned IterCount); 603 bool isOperandShifted(Instruction *I, Value *Op); 604 bool convertShiftsToLeft(BasicBlock *LoopB, BasicBlock *ExitB, 605 unsigned IterCount); 606 void cleanupLoopBody(BasicBlock *LoopB); 607 608 struct ParsedValues { 609 ParsedValues() = default; 610 611 Value *M = nullptr; 612 Value *P = nullptr; 613 Value *Q = nullptr; 614 Value *R = nullptr; 615 Value *X = nullptr; 616 Instruction *Res = nullptr; 617 unsigned IterCount = 0; 618 bool Left = false; 619 bool Inv = false; 620 }; 621 622 bool matchLeftShift(SelectInst *SelI, Value *CIV, ParsedValues &PV); 623 bool matchRightShift(SelectInst *SelI, ParsedValues &PV); 624 bool scanSelect(SelectInst *SI, BasicBlock *LoopB, BasicBlock *PrehB, 625 Value *CIV, ParsedValues &PV, bool PreScan); 626 unsigned getInverseMxN(unsigned QP); 627 Value *generate(BasicBlock::iterator At, ParsedValues &PV); 628 629 void setupPreSimplifier(Simplifier &S); 630 void setupPostSimplifier(Simplifier &S); 631 632 Loop *CurLoop; 633 const DataLayout &DL; 634 const DominatorTree &DT; 635 const TargetLibraryInfo &TLI; 636 ScalarEvolution &SE; 637 }; 638 639 } // end anonymous namespace 640 641 Value *PolynomialMultiplyRecognize::getCountIV(BasicBlock *BB) { 642 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 643 if (std::distance(PI, PE) != 2) 644 return nullptr; 645 BasicBlock *PB = (*PI == BB) ? *std::next(PI) : *PI; 646 647 for (auto I = BB->begin(), E = BB->end(); I != E && isa<PHINode>(I); ++I) { 648 auto *PN = cast<PHINode>(I); 649 Value *InitV = PN->getIncomingValueForBlock(PB); 650 if (!isa<ConstantInt>(InitV) || !cast<ConstantInt>(InitV)->isZero()) 651 continue; 652 Value *IterV = PN->getIncomingValueForBlock(BB); 653 auto *BO = dyn_cast<BinaryOperator>(IterV); 654 if (!BO) 655 continue; 656 if (BO->getOpcode() != Instruction::Add) 657 continue; 658 Value *IncV = nullptr; 659 if (BO->getOperand(0) == PN) 660 IncV = BO->getOperand(1); 661 else if (BO->getOperand(1) == PN) 662 IncV = BO->getOperand(0); 663 if (IncV == nullptr) 664 continue; 665 666 if (auto *T = dyn_cast<ConstantInt>(IncV)) 667 if (T->isOne()) 668 return PN; 669 } 670 return nullptr; 671 } 672 673 static void replaceAllUsesOfWithIn(Value *I, Value *J, BasicBlock *BB) { 674 for (auto UI = I->user_begin(), UE = I->user_end(); UI != UE;) { 675 Use &TheUse = UI.getUse(); 676 ++UI; 677 if (auto *II = dyn_cast<Instruction>(TheUse.getUser())) 678 if (BB == II->getParent()) 679 II->replaceUsesOfWith(I, J); 680 } 681 } 682 683 bool PolynomialMultiplyRecognize::matchLeftShift(SelectInst *SelI, 684 Value *CIV, ParsedValues &PV) { 685 // Match the following: 686 // select (X & (1 << i)) != 0 ? R ^ (Q << i) : R 687 // select (X & (1 << i)) == 0 ? R : R ^ (Q << i) 688 // The condition may also check for equality with the masked value, i.e 689 // select (X & (1 << i)) == (1 << i) ? R ^ (Q << i) : R 690 // select (X & (1 << i)) != (1 << i) ? R : R ^ (Q << i); 691 692 Value *CondV = SelI->getCondition(); 693 Value *TrueV = SelI->getTrueValue(); 694 Value *FalseV = SelI->getFalseValue(); 695 696 using namespace PatternMatch; 697 698 CmpInst::Predicate P; 699 Value *A = nullptr, *B = nullptr, *C = nullptr; 700 701 if (!match(CondV, m_ICmp(P, m_And(m_Value(A), m_Value(B)), m_Value(C))) && 702 !match(CondV, m_ICmp(P, m_Value(C), m_And(m_Value(A), m_Value(B))))) 703 return false; 704 if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE) 705 return false; 706 // Matched: select (A & B) == C ? ... : ... 707 // select (A & B) != C ? ... : ... 708 709 Value *X = nullptr, *Sh1 = nullptr; 710 // Check (A & B) for (X & (1 << i)): 711 if (match(A, m_Shl(m_One(), m_Specific(CIV)))) { 712 Sh1 = A; 713 X = B; 714 } else if (match(B, m_Shl(m_One(), m_Specific(CIV)))) { 715 Sh1 = B; 716 X = A; 717 } else { 718 // TODO: Could also check for an induction variable containing single 719 // bit shifted left by 1 in each iteration. 720 return false; 721 } 722 723 bool TrueIfZero; 724 725 // Check C against the possible values for comparison: 0 and (1 << i): 726 if (match(C, m_Zero())) 727 TrueIfZero = (P == CmpInst::ICMP_EQ); 728 else if (C == Sh1) 729 TrueIfZero = (P == CmpInst::ICMP_NE); 730 else 731 return false; 732 733 // So far, matched: 734 // select (X & (1 << i)) ? ... : ... 735 // including variations of the check against zero/non-zero value. 736 737 Value *ShouldSameV = nullptr, *ShouldXoredV = nullptr; 738 if (TrueIfZero) { 739 ShouldSameV = TrueV; 740 ShouldXoredV = FalseV; 741 } else { 742 ShouldSameV = FalseV; 743 ShouldXoredV = TrueV; 744 } 745 746 Value *Q = nullptr, *R = nullptr, *Y = nullptr, *Z = nullptr; 747 Value *T = nullptr; 748 if (match(ShouldXoredV, m_Xor(m_Value(Y), m_Value(Z)))) { 749 // Matched: select +++ ? ... : Y ^ Z 750 // select +++ ? Y ^ Z : ... 751 // where +++ denotes previously checked matches. 752 if (ShouldSameV == Y) 753 T = Z; 754 else if (ShouldSameV == Z) 755 T = Y; 756 else 757 return false; 758 R = ShouldSameV; 759 // Matched: select +++ ? R : R ^ T 760 // select +++ ? R ^ T : R 761 // depending on TrueIfZero. 762 763 } else if (match(ShouldSameV, m_Zero())) { 764 // Matched: select +++ ? 0 : ... 765 // select +++ ? ... : 0 766 if (!SelI->hasOneUse()) 767 return false; 768 T = ShouldXoredV; 769 // Matched: select +++ ? 0 : T 770 // select +++ ? T : 0 771 772 Value *U = *SelI->user_begin(); 773 if (!match(U, m_Xor(m_Specific(SelI), m_Value(R))) && 774 !match(U, m_Xor(m_Value(R), m_Specific(SelI)))) 775 return false; 776 // Matched: xor (select +++ ? 0 : T), R 777 // xor (select +++ ? T : 0), R 778 } else 779 return false; 780 781 // The xor input value T is isolated into its own match so that it could 782 // be checked against an induction variable containing a shifted bit 783 // (todo). 784 // For now, check against (Q << i). 785 if (!match(T, m_Shl(m_Value(Q), m_Specific(CIV))) && 786 !match(T, m_Shl(m_ZExt(m_Value(Q)), m_ZExt(m_Specific(CIV))))) 787 return false; 788 // Matched: select +++ ? R : R ^ (Q << i) 789 // select +++ ? R ^ (Q << i) : R 790 791 PV.X = X; 792 PV.Q = Q; 793 PV.R = R; 794 PV.Left = true; 795 return true; 796 } 797 798 bool PolynomialMultiplyRecognize::matchRightShift(SelectInst *SelI, 799 ParsedValues &PV) { 800 // Match the following: 801 // select (X & 1) != 0 ? (R >> 1) ^ Q : (R >> 1) 802 // select (X & 1) == 0 ? (R >> 1) : (R >> 1) ^ Q 803 // The condition may also check for equality with the masked value, i.e 804 // select (X & 1) == 1 ? (R >> 1) ^ Q : (R >> 1) 805 // select (X & 1) != 1 ? (R >> 1) : (R >> 1) ^ Q 806 807 Value *CondV = SelI->getCondition(); 808 Value *TrueV = SelI->getTrueValue(); 809 Value *FalseV = SelI->getFalseValue(); 810 811 using namespace PatternMatch; 812 813 Value *C = nullptr; 814 CmpInst::Predicate P; 815 bool TrueIfZero; 816 817 if (match(CondV, m_ICmp(P, m_Value(C), m_Zero())) || 818 match(CondV, m_ICmp(P, m_Zero(), m_Value(C)))) { 819 if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE) 820 return false; 821 // Matched: select C == 0 ? ... : ... 822 // select C != 0 ? ... : ... 823 TrueIfZero = (P == CmpInst::ICMP_EQ); 824 } else if (match(CondV, m_ICmp(P, m_Value(C), m_One())) || 825 match(CondV, m_ICmp(P, m_One(), m_Value(C)))) { 826 if (P != CmpInst::ICMP_EQ && P != CmpInst::ICMP_NE) 827 return false; 828 // Matched: select C == 1 ? ... : ... 829 // select C != 1 ? ... : ... 830 TrueIfZero = (P == CmpInst::ICMP_NE); 831 } else 832 return false; 833 834 Value *X = nullptr; 835 if (!match(C, m_And(m_Value(X), m_One())) && 836 !match(C, m_And(m_One(), m_Value(X)))) 837 return false; 838 // Matched: select (X & 1) == +++ ? ... : ... 839 // select (X & 1) != +++ ? ... : ... 840 841 Value *R = nullptr, *Q = nullptr; 842 if (TrueIfZero) { 843 // The select's condition is true if the tested bit is 0. 844 // TrueV must be the shift, FalseV must be the xor. 845 if (!match(TrueV, m_LShr(m_Value(R), m_One()))) 846 return false; 847 // Matched: select +++ ? (R >> 1) : ... 848 if (!match(FalseV, m_Xor(m_Specific(TrueV), m_Value(Q))) && 849 !match(FalseV, m_Xor(m_Value(Q), m_Specific(TrueV)))) 850 return false; 851 // Matched: select +++ ? (R >> 1) : (R >> 1) ^ Q 852 // with commuting ^. 853 } else { 854 // The select's condition is true if the tested bit is 1. 855 // TrueV must be the xor, FalseV must be the shift. 856 if (!match(FalseV, m_LShr(m_Value(R), m_One()))) 857 return false; 858 // Matched: select +++ ? ... : (R >> 1) 859 if (!match(TrueV, m_Xor(m_Specific(FalseV), m_Value(Q))) && 860 !match(TrueV, m_Xor(m_Value(Q), m_Specific(FalseV)))) 861 return false; 862 // Matched: select +++ ? (R >> 1) ^ Q : (R >> 1) 863 // with commuting ^. 864 } 865 866 PV.X = X; 867 PV.Q = Q; 868 PV.R = R; 869 PV.Left = false; 870 return true; 871 } 872 873 bool PolynomialMultiplyRecognize::scanSelect(SelectInst *SelI, 874 BasicBlock *LoopB, BasicBlock *PrehB, Value *CIV, ParsedValues &PV, 875 bool PreScan) { 876 using namespace PatternMatch; 877 878 // The basic pattern for R = P.Q is: 879 // for i = 0..31 880 // R = phi (0, R') 881 // if (P & (1 << i)) ; test-bit(P, i) 882 // R' = R ^ (Q << i) 883 // 884 // Similarly, the basic pattern for R = (P/Q).Q - P 885 // for i = 0..31 886 // R = phi(P, R') 887 // if (R & (1 << i)) 888 // R' = R ^ (Q << i) 889 890 // There exist idioms, where instead of Q being shifted left, P is shifted 891 // right. This produces a result that is shifted right by 32 bits (the 892 // non-shifted result is 64-bit). 893 // 894 // For R = P.Q, this would be: 895 // for i = 0..31 896 // R = phi (0, R') 897 // if ((P >> i) & 1) 898 // R' = (R >> 1) ^ Q ; R is cycled through the loop, so it must 899 // else ; be shifted by 1, not i. 900 // R' = R >> 1 901 // 902 // And for the inverse: 903 // for i = 0..31 904 // R = phi (P, R') 905 // if (R & 1) 906 // R' = (R >> 1) ^ Q 907 // else 908 // R' = R >> 1 909 910 // The left-shifting idioms share the same pattern: 911 // select (X & (1 << i)) ? R ^ (Q << i) : R 912 // Similarly for right-shifting idioms: 913 // select (X & 1) ? (R >> 1) ^ Q 914 915 if (matchLeftShift(SelI, CIV, PV)) { 916 // If this is a pre-scan, getting this far is sufficient. 917 if (PreScan) 918 return true; 919 920 // Need to make sure that the SelI goes back into R. 921 auto *RPhi = dyn_cast<PHINode>(PV.R); 922 if (!RPhi) 923 return false; 924 if (SelI != RPhi->getIncomingValueForBlock(LoopB)) 925 return false; 926 PV.Res = SelI; 927 928 // If X is loop invariant, it must be the input polynomial, and the 929 // idiom is the basic polynomial multiply. 930 if (CurLoop->isLoopInvariant(PV.X)) { 931 PV.P = PV.X; 932 PV.Inv = false; 933 } else { 934 // X is not loop invariant. If X == R, this is the inverse pmpy. 935 // Otherwise, check for an xor with an invariant value. If the 936 // variable argument to the xor is R, then this is still a valid 937 // inverse pmpy. 938 PV.Inv = true; 939 if (PV.X != PV.R) { 940 Value *Var = nullptr, *Inv = nullptr, *X1 = nullptr, *X2 = nullptr; 941 if (!match(PV.X, m_Xor(m_Value(X1), m_Value(X2)))) 942 return false; 943 auto *I1 = dyn_cast<Instruction>(X1); 944 auto *I2 = dyn_cast<Instruction>(X2); 945 if (!I1 || I1->getParent() != LoopB) { 946 Var = X2; 947 Inv = X1; 948 } else if (!I2 || I2->getParent() != LoopB) { 949 Var = X1; 950 Inv = X2; 951 } else 952 return false; 953 if (Var != PV.R) 954 return false; 955 PV.M = Inv; 956 } 957 // The input polynomial P still needs to be determined. It will be 958 // the entry value of R. 959 Value *EntryP = RPhi->getIncomingValueForBlock(PrehB); 960 PV.P = EntryP; 961 } 962 963 return true; 964 } 965 966 if (matchRightShift(SelI, PV)) { 967 // If this is an inverse pattern, the Q polynomial must be known at 968 // compile time. 969 if (PV.Inv && !isa<ConstantInt>(PV.Q)) 970 return false; 971 if (PreScan) 972 return true; 973 // There is no exact matching of right-shift pmpy. 974 return false; 975 } 976 977 return false; 978 } 979 980 bool PolynomialMultiplyRecognize::isPromotableTo(Value *Val, 981 IntegerType *DestTy) { 982 IntegerType *T = dyn_cast<IntegerType>(Val->getType()); 983 if (!T || T->getBitWidth() > DestTy->getBitWidth()) 984 return false; 985 if (T->getBitWidth() == DestTy->getBitWidth()) 986 return true; 987 // Non-instructions are promotable. The reason why an instruction may not 988 // be promotable is that it may produce a different result if its operands 989 // and the result are promoted, for example, it may produce more non-zero 990 // bits. While it would still be possible to represent the proper result 991 // in a wider type, it may require adding additional instructions (which 992 // we don't want to do). 993 Instruction *In = dyn_cast<Instruction>(Val); 994 if (!In) 995 return true; 996 // The bitwidth of the source type is smaller than the destination. 997 // Check if the individual operation can be promoted. 998 switch (In->getOpcode()) { 999 case Instruction::PHI: 1000 case Instruction::ZExt: 1001 case Instruction::And: 1002 case Instruction::Or: 1003 case Instruction::Xor: 1004 case Instruction::LShr: // Shift right is ok. 1005 case Instruction::Select: 1006 case Instruction::Trunc: 1007 return true; 1008 case Instruction::ICmp: 1009 if (CmpInst *CI = cast<CmpInst>(In)) 1010 return CI->isEquality() || CI->isUnsigned(); 1011 llvm_unreachable("Cast failed unexpectedly"); 1012 case Instruction::Add: 1013 return In->hasNoSignedWrap() && In->hasNoUnsignedWrap(); 1014 } 1015 return false; 1016 } 1017 1018 void PolynomialMultiplyRecognize::promoteTo(Instruction *In, 1019 IntegerType *DestTy, BasicBlock *LoopB) { 1020 Type *OrigTy = In->getType(); 1021 assert(!OrigTy->isVoidTy() && "Invalid instruction to promote"); 1022 1023 // Leave boolean values alone. 1024 if (!In->getType()->isIntegerTy(1)) 1025 In->mutateType(DestTy); 1026 unsigned DestBW = DestTy->getBitWidth(); 1027 1028 // Handle PHIs. 1029 if (PHINode *P = dyn_cast<PHINode>(In)) { 1030 unsigned N = P->getNumIncomingValues(); 1031 for (unsigned i = 0; i != N; ++i) { 1032 BasicBlock *InB = P->getIncomingBlock(i); 1033 if (InB == LoopB) 1034 continue; 1035 Value *InV = P->getIncomingValue(i); 1036 IntegerType *Ty = cast<IntegerType>(InV->getType()); 1037 // Do not promote values in PHI nodes of type i1. 1038 if (Ty != P->getType()) { 1039 // If the value type does not match the PHI type, the PHI type 1040 // must have been promoted. 1041 assert(Ty->getBitWidth() < DestBW); 1042 InV = IRBuilder<>(InB->getTerminator()).CreateZExt(InV, DestTy); 1043 P->setIncomingValue(i, InV); 1044 } 1045 } 1046 } else if (ZExtInst *Z = dyn_cast<ZExtInst>(In)) { 1047 Value *Op = Z->getOperand(0); 1048 if (Op->getType() == Z->getType()) 1049 Z->replaceAllUsesWith(Op); 1050 Z->eraseFromParent(); 1051 return; 1052 } 1053 if (TruncInst *T = dyn_cast<TruncInst>(In)) { 1054 IntegerType *TruncTy = cast<IntegerType>(OrigTy); 1055 Value *Mask = ConstantInt::get(DestTy, (1u << TruncTy->getBitWidth()) - 1); 1056 Value *And = IRBuilder<>(In).CreateAnd(T->getOperand(0), Mask); 1057 T->replaceAllUsesWith(And); 1058 T->eraseFromParent(); 1059 return; 1060 } 1061 1062 // Promote immediates. 1063 for (unsigned i = 0, n = In->getNumOperands(); i != n; ++i) { 1064 if (ConstantInt *CI = dyn_cast<ConstantInt>(In->getOperand(i))) 1065 if (CI->getType()->getBitWidth() < DestBW) 1066 In->setOperand(i, ConstantInt::get(DestTy, CI->getZExtValue())); 1067 } 1068 } 1069 1070 bool PolynomialMultiplyRecognize::promoteTypes(BasicBlock *LoopB, 1071 BasicBlock *ExitB) { 1072 assert(LoopB); 1073 // Skip loops where the exit block has more than one predecessor. The values 1074 // coming from the loop block will be promoted to another type, and so the 1075 // values coming into the exit block from other predecessors would also have 1076 // to be promoted. 1077 if (!ExitB || (ExitB->getSinglePredecessor() != LoopB)) 1078 return false; 1079 IntegerType *DestTy = getPmpyType(); 1080 // Check if the exit values have types that are no wider than the type 1081 // that we want to promote to. 1082 unsigned DestBW = DestTy->getBitWidth(); 1083 for (PHINode &P : ExitB->phis()) { 1084 if (P.getNumIncomingValues() != 1) 1085 return false; 1086 assert(P.getIncomingBlock(0) == LoopB); 1087 IntegerType *T = dyn_cast<IntegerType>(P.getType()); 1088 if (!T || T->getBitWidth() > DestBW) 1089 return false; 1090 } 1091 1092 // Check all instructions in the loop. 1093 for (Instruction &In : *LoopB) 1094 if (!In.isTerminator() && !isPromotableTo(&In, DestTy)) 1095 return false; 1096 1097 // Perform the promotion. 1098 std::vector<Instruction*> LoopIns; 1099 std::transform(LoopB->begin(), LoopB->end(), std::back_inserter(LoopIns), 1100 [](Instruction &In) { return &In; }); 1101 for (Instruction *In : LoopIns) 1102 if (!In->isTerminator()) 1103 promoteTo(In, DestTy, LoopB); 1104 1105 // Fix up the PHI nodes in the exit block. 1106 Instruction *EndI = ExitB->getFirstNonPHI(); 1107 BasicBlock::iterator End = EndI ? EndI->getIterator() : ExitB->end(); 1108 for (auto I = ExitB->begin(); I != End; ++I) { 1109 PHINode *P = dyn_cast<PHINode>(I); 1110 if (!P) 1111 break; 1112 Type *Ty0 = P->getIncomingValue(0)->getType(); 1113 Type *PTy = P->getType(); 1114 if (PTy != Ty0) { 1115 assert(Ty0 == DestTy); 1116 // In order to create the trunc, P must have the promoted type. 1117 P->mutateType(Ty0); 1118 Value *T = IRBuilder<>(ExitB, End).CreateTrunc(P, PTy); 1119 // In order for the RAUW to work, the types of P and T must match. 1120 P->mutateType(PTy); 1121 P->replaceAllUsesWith(T); 1122 // Final update of the P's type. 1123 P->mutateType(Ty0); 1124 cast<Instruction>(T)->setOperand(0, P); 1125 } 1126 } 1127 1128 return true; 1129 } 1130 1131 bool PolynomialMultiplyRecognize::findCycle(Value *Out, Value *In, 1132 ValueSeq &Cycle) { 1133 // Out = ..., In, ... 1134 if (Out == In) 1135 return true; 1136 1137 auto *BB = cast<Instruction>(Out)->getParent(); 1138 bool HadPhi = false; 1139 1140 for (auto *U : Out->users()) { 1141 auto *I = dyn_cast<Instruction>(&*U); 1142 if (I == nullptr || I->getParent() != BB) 1143 continue; 1144 // Make sure that there are no multi-iteration cycles, e.g. 1145 // p1 = phi(p2) 1146 // p2 = phi(p1) 1147 // The cycle p1->p2->p1 would span two loop iterations. 1148 // Check that there is only one phi in the cycle. 1149 bool IsPhi = isa<PHINode>(I); 1150 if (IsPhi && HadPhi) 1151 return false; 1152 HadPhi |= IsPhi; 1153 if (!Cycle.insert(I)) 1154 return false; 1155 if (findCycle(I, In, Cycle)) 1156 break; 1157 Cycle.remove(I); 1158 } 1159 return !Cycle.empty(); 1160 } 1161 1162 void PolynomialMultiplyRecognize::classifyCycle(Instruction *DivI, 1163 ValueSeq &Cycle, ValueSeq &Early, ValueSeq &Late) { 1164 // All the values in the cycle that are between the phi node and the 1165 // divider instruction will be classified as "early", all other values 1166 // will be "late". 1167 1168 bool IsE = true; 1169 unsigned I, N = Cycle.size(); 1170 for (I = 0; I < N; ++I) { 1171 Value *V = Cycle[I]; 1172 if (DivI == V) 1173 IsE = false; 1174 else if (!isa<PHINode>(V)) 1175 continue; 1176 // Stop if found either. 1177 break; 1178 } 1179 // "I" is the index of either DivI or the phi node, whichever was first. 1180 // "E" is "false" or "true" respectively. 1181 ValueSeq &First = !IsE ? Early : Late; 1182 for (unsigned J = 0; J < I; ++J) 1183 First.insert(Cycle[J]); 1184 1185 ValueSeq &Second = IsE ? Early : Late; 1186 Second.insert(Cycle[I]); 1187 for (++I; I < N; ++I) { 1188 Value *V = Cycle[I]; 1189 if (DivI == V || isa<PHINode>(V)) 1190 break; 1191 Second.insert(V); 1192 } 1193 1194 for (; I < N; ++I) 1195 First.insert(Cycle[I]); 1196 } 1197 1198 bool PolynomialMultiplyRecognize::classifyInst(Instruction *UseI, 1199 ValueSeq &Early, ValueSeq &Late) { 1200 // Select is an exception, since the condition value does not have to be 1201 // classified in the same way as the true/false values. The true/false 1202 // values do have to be both early or both late. 1203 if (UseI->getOpcode() == Instruction::Select) { 1204 Value *TV = UseI->getOperand(1), *FV = UseI->getOperand(2); 1205 if (Early.count(TV) || Early.count(FV)) { 1206 if (Late.count(TV) || Late.count(FV)) 1207 return false; 1208 Early.insert(UseI); 1209 } else if (Late.count(TV) || Late.count(FV)) { 1210 if (Early.count(TV) || Early.count(FV)) 1211 return false; 1212 Late.insert(UseI); 1213 } 1214 return true; 1215 } 1216 1217 // Not sure what would be the example of this, but the code below relies 1218 // on having at least one operand. 1219 if (UseI->getNumOperands() == 0) 1220 return true; 1221 1222 bool AE = true, AL = true; 1223 for (auto &I : UseI->operands()) { 1224 if (Early.count(&*I)) 1225 AL = false; 1226 else if (Late.count(&*I)) 1227 AE = false; 1228 } 1229 // If the operands appear "all early" and "all late" at the same time, 1230 // then it means that none of them are actually classified as either. 1231 // This is harmless. 1232 if (AE && AL) 1233 return true; 1234 // Conversely, if they are neither "all early" nor "all late", then 1235 // we have a mixture of early and late operands that is not a known 1236 // exception. 1237 if (!AE && !AL) 1238 return false; 1239 1240 // Check that we have covered the two special cases. 1241 assert(AE != AL); 1242 1243 if (AE) 1244 Early.insert(UseI); 1245 else 1246 Late.insert(UseI); 1247 return true; 1248 } 1249 1250 bool PolynomialMultiplyRecognize::commutesWithShift(Instruction *I) { 1251 switch (I->getOpcode()) { 1252 case Instruction::And: 1253 case Instruction::Or: 1254 case Instruction::Xor: 1255 case Instruction::LShr: 1256 case Instruction::Shl: 1257 case Instruction::Select: 1258 case Instruction::ICmp: 1259 case Instruction::PHI: 1260 break; 1261 default: 1262 return false; 1263 } 1264 return true; 1265 } 1266 1267 bool PolynomialMultiplyRecognize::highBitsAreZero(Value *V, 1268 unsigned IterCount) { 1269 auto *T = dyn_cast<IntegerType>(V->getType()); 1270 if (!T) 1271 return false; 1272 1273 KnownBits Known(T->getBitWidth()); 1274 computeKnownBits(V, Known, DL); 1275 return Known.countMinLeadingZeros() >= IterCount; 1276 } 1277 1278 bool PolynomialMultiplyRecognize::keepsHighBitsZero(Value *V, 1279 unsigned IterCount) { 1280 // Assume that all inputs to the value have the high bits zero. 1281 // Check if the value itself preserves the zeros in the high bits. 1282 if (auto *C = dyn_cast<ConstantInt>(V)) 1283 return C->getValue().countl_zero() >= IterCount; 1284 1285 if (auto *I = dyn_cast<Instruction>(V)) { 1286 switch (I->getOpcode()) { 1287 case Instruction::And: 1288 case Instruction::Or: 1289 case Instruction::Xor: 1290 case Instruction::LShr: 1291 case Instruction::Select: 1292 case Instruction::ICmp: 1293 case Instruction::PHI: 1294 case Instruction::ZExt: 1295 return true; 1296 } 1297 } 1298 1299 return false; 1300 } 1301 1302 bool PolynomialMultiplyRecognize::isOperandShifted(Instruction *I, Value *Op) { 1303 unsigned Opc = I->getOpcode(); 1304 if (Opc == Instruction::Shl || Opc == Instruction::LShr) 1305 return Op != I->getOperand(1); 1306 return true; 1307 } 1308 1309 bool PolynomialMultiplyRecognize::convertShiftsToLeft(BasicBlock *LoopB, 1310 BasicBlock *ExitB, unsigned IterCount) { 1311 Value *CIV = getCountIV(LoopB); 1312 if (CIV == nullptr) 1313 return false; 1314 auto *CIVTy = dyn_cast<IntegerType>(CIV->getType()); 1315 if (CIVTy == nullptr) 1316 return false; 1317 1318 ValueSeq RShifts; 1319 ValueSeq Early, Late, Cycled; 1320 1321 // Find all value cycles that contain logical right shifts by 1. 1322 for (Instruction &I : *LoopB) { 1323 using namespace PatternMatch; 1324 1325 Value *V = nullptr; 1326 if (!match(&I, m_LShr(m_Value(V), m_One()))) 1327 continue; 1328 ValueSeq C; 1329 if (!findCycle(&I, V, C)) 1330 continue; 1331 1332 // Found a cycle. 1333 C.insert(&I); 1334 classifyCycle(&I, C, Early, Late); 1335 Cycled.insert(C.begin(), C.end()); 1336 RShifts.insert(&I); 1337 } 1338 1339 // Find the set of all values affected by the shift cycles, i.e. all 1340 // cycled values, and (recursively) all their users. 1341 ValueSeq Users(Cycled.begin(), Cycled.end()); 1342 for (unsigned i = 0; i < Users.size(); ++i) { 1343 Value *V = Users[i]; 1344 if (!isa<IntegerType>(V->getType())) 1345 return false; 1346 auto *R = cast<Instruction>(V); 1347 // If the instruction does not commute with shifts, the loop cannot 1348 // be unshifted. 1349 if (!commutesWithShift(R)) 1350 return false; 1351 for (User *U : R->users()) { 1352 auto *T = cast<Instruction>(U); 1353 // Skip users from outside of the loop. They will be handled later. 1354 // Also, skip the right-shifts and phi nodes, since they mix early 1355 // and late values. 1356 if (T->getParent() != LoopB || RShifts.count(T) || isa<PHINode>(T)) 1357 continue; 1358 1359 Users.insert(T); 1360 if (!classifyInst(T, Early, Late)) 1361 return false; 1362 } 1363 } 1364 1365 if (Users.empty()) 1366 return false; 1367 1368 // Verify that high bits remain zero. 1369 ValueSeq Internal(Users.begin(), Users.end()); 1370 ValueSeq Inputs; 1371 for (unsigned i = 0; i < Internal.size(); ++i) { 1372 auto *R = dyn_cast<Instruction>(Internal[i]); 1373 if (!R) 1374 continue; 1375 for (Value *Op : R->operands()) { 1376 auto *T = dyn_cast<Instruction>(Op); 1377 if (T && T->getParent() != LoopB) 1378 Inputs.insert(Op); 1379 else 1380 Internal.insert(Op); 1381 } 1382 } 1383 for (Value *V : Inputs) 1384 if (!highBitsAreZero(V, IterCount)) 1385 return false; 1386 for (Value *V : Internal) 1387 if (!keepsHighBitsZero(V, IterCount)) 1388 return false; 1389 1390 // Finally, the work can be done. Unshift each user. 1391 IRBuilder<> IRB(LoopB); 1392 std::map<Value*,Value*> ShiftMap; 1393 1394 using CastMapType = std::map<std::pair<Value *, Type *>, Value *>; 1395 1396 CastMapType CastMap; 1397 1398 auto upcast = [] (CastMapType &CM, IRBuilder<> &IRB, Value *V, 1399 IntegerType *Ty) -> Value* { 1400 auto H = CM.find(std::make_pair(V, Ty)); 1401 if (H != CM.end()) 1402 return H->second; 1403 Value *CV = IRB.CreateIntCast(V, Ty, false); 1404 CM.insert(std::make_pair(std::make_pair(V, Ty), CV)); 1405 return CV; 1406 }; 1407 1408 for (auto I = LoopB->begin(), E = LoopB->end(); I != E; ++I) { 1409 using namespace PatternMatch; 1410 1411 if (isa<PHINode>(I) || !Users.count(&*I)) 1412 continue; 1413 1414 // Match lshr x, 1. 1415 Value *V = nullptr; 1416 if (match(&*I, m_LShr(m_Value(V), m_One()))) { 1417 replaceAllUsesOfWithIn(&*I, V, LoopB); 1418 continue; 1419 } 1420 // For each non-cycled operand, replace it with the corresponding 1421 // value shifted left. 1422 for (auto &J : I->operands()) { 1423 Value *Op = J.get(); 1424 if (!isOperandShifted(&*I, Op)) 1425 continue; 1426 if (Users.count(Op)) 1427 continue; 1428 // Skip shifting zeros. 1429 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero()) 1430 continue; 1431 // Check if we have already generated a shift for this value. 1432 auto F = ShiftMap.find(Op); 1433 Value *W = (F != ShiftMap.end()) ? F->second : nullptr; 1434 if (W == nullptr) { 1435 IRB.SetInsertPoint(&*I); 1436 // First, the shift amount will be CIV or CIV+1, depending on 1437 // whether the value is early or late. Instead of creating CIV+1, 1438 // do a single shift of the value. 1439 Value *ShAmt = CIV, *ShVal = Op; 1440 auto *VTy = cast<IntegerType>(ShVal->getType()); 1441 auto *ATy = cast<IntegerType>(ShAmt->getType()); 1442 if (Late.count(&*I)) 1443 ShVal = IRB.CreateShl(Op, ConstantInt::get(VTy, 1)); 1444 // Second, the types of the shifted value and the shift amount 1445 // must match. 1446 if (VTy != ATy) { 1447 if (VTy->getBitWidth() < ATy->getBitWidth()) 1448 ShVal = upcast(CastMap, IRB, ShVal, ATy); 1449 else 1450 ShAmt = upcast(CastMap, IRB, ShAmt, VTy); 1451 } 1452 // Ready to generate the shift and memoize it. 1453 W = IRB.CreateShl(ShVal, ShAmt); 1454 ShiftMap.insert(std::make_pair(Op, W)); 1455 } 1456 I->replaceUsesOfWith(Op, W); 1457 } 1458 } 1459 1460 // Update the users outside of the loop to account for having left 1461 // shifts. They would normally be shifted right in the loop, so shift 1462 // them right after the loop exit. 1463 // Take advantage of the loop-closed SSA form, which has all the post- 1464 // loop values in phi nodes. 1465 IRB.SetInsertPoint(ExitB, ExitB->getFirstInsertionPt()); 1466 for (auto P = ExitB->begin(), Q = ExitB->end(); P != Q; ++P) { 1467 if (!isa<PHINode>(P)) 1468 break; 1469 auto *PN = cast<PHINode>(P); 1470 Value *U = PN->getIncomingValueForBlock(LoopB); 1471 if (!Users.count(U)) 1472 continue; 1473 Value *S = IRB.CreateLShr(PN, ConstantInt::get(PN->getType(), IterCount)); 1474 PN->replaceAllUsesWith(S); 1475 // The above RAUW will create 1476 // S = lshr S, IterCount 1477 // so we need to fix it back into 1478 // S = lshr PN, IterCount 1479 cast<User>(S)->replaceUsesOfWith(S, PN); 1480 } 1481 1482 return true; 1483 } 1484 1485 void PolynomialMultiplyRecognize::cleanupLoopBody(BasicBlock *LoopB) { 1486 for (auto &I : *LoopB) 1487 if (Value *SV = simplifyInstruction(&I, {DL, &TLI, &DT})) 1488 I.replaceAllUsesWith(SV); 1489 1490 for (Instruction &I : llvm::make_early_inc_range(*LoopB)) 1491 RecursivelyDeleteTriviallyDeadInstructions(&I, &TLI); 1492 } 1493 1494 unsigned PolynomialMultiplyRecognize::getInverseMxN(unsigned QP) { 1495 // Arrays of coefficients of Q and the inverse, C. 1496 // Q[i] = coefficient at x^i. 1497 std::array<char,32> Q, C; 1498 1499 for (unsigned i = 0; i < 32; ++i) { 1500 Q[i] = QP & 1; 1501 QP >>= 1; 1502 } 1503 assert(Q[0] == 1); 1504 1505 // Find C, such that 1506 // (Q[n]*x^n + ... + Q[1]*x + Q[0]) * (C[n]*x^n + ... + C[1]*x + C[0]) = 1 1507 // 1508 // For it to have a solution, Q[0] must be 1. Since this is Z2[x], the 1509 // operations * and + are & and ^ respectively. 1510 // 1511 // Find C[i] recursively, by comparing i-th coefficient in the product 1512 // with 0 (or 1 for i=0). 1513 // 1514 // C[0] = 1, since C[0] = Q[0], and Q[0] = 1. 1515 C[0] = 1; 1516 for (unsigned i = 1; i < 32; ++i) { 1517 // Solve for C[i] in: 1518 // C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] ^ C[i]Q[0] = 0 1519 // This is equivalent to 1520 // C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] ^ C[i] = 0 1521 // which is 1522 // C[0]Q[i] ^ C[1]Q[i-1] ^ ... ^ C[i-1]Q[1] = C[i] 1523 unsigned T = 0; 1524 for (unsigned j = 0; j < i; ++j) 1525 T = T ^ (C[j] & Q[i-j]); 1526 C[i] = T; 1527 } 1528 1529 unsigned QV = 0; 1530 for (unsigned i = 0; i < 32; ++i) 1531 if (C[i]) 1532 QV |= (1 << i); 1533 1534 return QV; 1535 } 1536 1537 Value *PolynomialMultiplyRecognize::generate(BasicBlock::iterator At, 1538 ParsedValues &PV) { 1539 IRBuilder<> B(&*At); 1540 Module *M = At->getParent()->getParent()->getParent(); 1541 Function *PMF = Intrinsic::getDeclaration(M, Intrinsic::hexagon_M4_pmpyw); 1542 1543 Value *P = PV.P, *Q = PV.Q, *P0 = P; 1544 unsigned IC = PV.IterCount; 1545 1546 if (PV.M != nullptr) 1547 P0 = P = B.CreateXor(P, PV.M); 1548 1549 // Create a bit mask to clear the high bits beyond IterCount. 1550 auto *BMI = ConstantInt::get(P->getType(), APInt::getLowBitsSet(32, IC)); 1551 1552 if (PV.IterCount != 32) 1553 P = B.CreateAnd(P, BMI); 1554 1555 if (PV.Inv) { 1556 auto *QI = dyn_cast<ConstantInt>(PV.Q); 1557 assert(QI && QI->getBitWidth() <= 32); 1558 1559 // Again, clearing bits beyond IterCount. 1560 unsigned M = (1 << PV.IterCount) - 1; 1561 unsigned Tmp = (QI->getZExtValue() | 1) & M; 1562 unsigned QV = getInverseMxN(Tmp) & M; 1563 auto *QVI = ConstantInt::get(QI->getType(), QV); 1564 P = B.CreateCall(PMF, {P, QVI}); 1565 P = B.CreateTrunc(P, QI->getType()); 1566 if (IC != 32) 1567 P = B.CreateAnd(P, BMI); 1568 } 1569 1570 Value *R = B.CreateCall(PMF, {P, Q}); 1571 1572 if (PV.M != nullptr) 1573 R = B.CreateXor(R, B.CreateIntCast(P0, R->getType(), false)); 1574 1575 return R; 1576 } 1577 1578 static bool hasZeroSignBit(const Value *V) { 1579 if (const auto *CI = dyn_cast<const ConstantInt>(V)) 1580 return (CI->getType()->getSignBit() & CI->getSExtValue()) == 0; 1581 const Instruction *I = dyn_cast<const Instruction>(V); 1582 if (!I) 1583 return false; 1584 switch (I->getOpcode()) { 1585 case Instruction::LShr: 1586 if (const auto SI = dyn_cast<const ConstantInt>(I->getOperand(1))) 1587 return SI->getZExtValue() > 0; 1588 return false; 1589 case Instruction::Or: 1590 case Instruction::Xor: 1591 return hasZeroSignBit(I->getOperand(0)) && 1592 hasZeroSignBit(I->getOperand(1)); 1593 case Instruction::And: 1594 return hasZeroSignBit(I->getOperand(0)) || 1595 hasZeroSignBit(I->getOperand(1)); 1596 } 1597 return false; 1598 } 1599 1600 void PolynomialMultiplyRecognize::setupPreSimplifier(Simplifier &S) { 1601 S.addRule("sink-zext", 1602 // Sink zext past bitwise operations. 1603 [](Instruction *I, LLVMContext &Ctx) -> Value* { 1604 if (I->getOpcode() != Instruction::ZExt) 1605 return nullptr; 1606 Instruction *T = dyn_cast<Instruction>(I->getOperand(0)); 1607 if (!T) 1608 return nullptr; 1609 switch (T->getOpcode()) { 1610 case Instruction::And: 1611 case Instruction::Or: 1612 case Instruction::Xor: 1613 break; 1614 default: 1615 return nullptr; 1616 } 1617 IRBuilder<> B(Ctx); 1618 return B.CreateBinOp(cast<BinaryOperator>(T)->getOpcode(), 1619 B.CreateZExt(T->getOperand(0), I->getType()), 1620 B.CreateZExt(T->getOperand(1), I->getType())); 1621 }); 1622 S.addRule("xor/and -> and/xor", 1623 // (xor (and x a) (and y a)) -> (and (xor x y) a) 1624 [](Instruction *I, LLVMContext &Ctx) -> Value* { 1625 if (I->getOpcode() != Instruction::Xor) 1626 return nullptr; 1627 Instruction *And0 = dyn_cast<Instruction>(I->getOperand(0)); 1628 Instruction *And1 = dyn_cast<Instruction>(I->getOperand(1)); 1629 if (!And0 || !And1) 1630 return nullptr; 1631 if (And0->getOpcode() != Instruction::And || 1632 And1->getOpcode() != Instruction::And) 1633 return nullptr; 1634 if (And0->getOperand(1) != And1->getOperand(1)) 1635 return nullptr; 1636 IRBuilder<> B(Ctx); 1637 return B.CreateAnd(B.CreateXor(And0->getOperand(0), And1->getOperand(0)), 1638 And0->getOperand(1)); 1639 }); 1640 S.addRule("sink binop into select", 1641 // (Op (select c x y) z) -> (select c (Op x z) (Op y z)) 1642 // (Op x (select c y z)) -> (select c (Op x y) (Op x z)) 1643 [](Instruction *I, LLVMContext &Ctx) -> Value* { 1644 BinaryOperator *BO = dyn_cast<BinaryOperator>(I); 1645 if (!BO) 1646 return nullptr; 1647 Instruction::BinaryOps Op = BO->getOpcode(); 1648 if (SelectInst *Sel = dyn_cast<SelectInst>(BO->getOperand(0))) { 1649 IRBuilder<> B(Ctx); 1650 Value *X = Sel->getTrueValue(), *Y = Sel->getFalseValue(); 1651 Value *Z = BO->getOperand(1); 1652 return B.CreateSelect(Sel->getCondition(), 1653 B.CreateBinOp(Op, X, Z), 1654 B.CreateBinOp(Op, Y, Z)); 1655 } 1656 if (SelectInst *Sel = dyn_cast<SelectInst>(BO->getOperand(1))) { 1657 IRBuilder<> B(Ctx); 1658 Value *X = BO->getOperand(0); 1659 Value *Y = Sel->getTrueValue(), *Z = Sel->getFalseValue(); 1660 return B.CreateSelect(Sel->getCondition(), 1661 B.CreateBinOp(Op, X, Y), 1662 B.CreateBinOp(Op, X, Z)); 1663 } 1664 return nullptr; 1665 }); 1666 S.addRule("fold select-select", 1667 // (select c (select c x y) z) -> (select c x z) 1668 // (select c x (select c y z)) -> (select c x z) 1669 [](Instruction *I, LLVMContext &Ctx) -> Value* { 1670 SelectInst *Sel = dyn_cast<SelectInst>(I); 1671 if (!Sel) 1672 return nullptr; 1673 IRBuilder<> B(Ctx); 1674 Value *C = Sel->getCondition(); 1675 if (SelectInst *Sel0 = dyn_cast<SelectInst>(Sel->getTrueValue())) { 1676 if (Sel0->getCondition() == C) 1677 return B.CreateSelect(C, Sel0->getTrueValue(), Sel->getFalseValue()); 1678 } 1679 if (SelectInst *Sel1 = dyn_cast<SelectInst>(Sel->getFalseValue())) { 1680 if (Sel1->getCondition() == C) 1681 return B.CreateSelect(C, Sel->getTrueValue(), Sel1->getFalseValue()); 1682 } 1683 return nullptr; 1684 }); 1685 S.addRule("or-signbit -> xor-signbit", 1686 // (or (lshr x 1) 0x800.0) -> (xor (lshr x 1) 0x800.0) 1687 [](Instruction *I, LLVMContext &Ctx) -> Value* { 1688 if (I->getOpcode() != Instruction::Or) 1689 return nullptr; 1690 ConstantInt *Msb = dyn_cast<ConstantInt>(I->getOperand(1)); 1691 if (!Msb || Msb->getZExtValue() != Msb->getType()->getSignBit()) 1692 return nullptr; 1693 if (!hasZeroSignBit(I->getOperand(0))) 1694 return nullptr; 1695 return IRBuilder<>(Ctx).CreateXor(I->getOperand(0), Msb); 1696 }); 1697 S.addRule("sink lshr into binop", 1698 // (lshr (BitOp x y) c) -> (BitOp (lshr x c) (lshr y c)) 1699 [](Instruction *I, LLVMContext &Ctx) -> Value* { 1700 if (I->getOpcode() != Instruction::LShr) 1701 return nullptr; 1702 BinaryOperator *BitOp = dyn_cast<BinaryOperator>(I->getOperand(0)); 1703 if (!BitOp) 1704 return nullptr; 1705 switch (BitOp->getOpcode()) { 1706 case Instruction::And: 1707 case Instruction::Or: 1708 case Instruction::Xor: 1709 break; 1710 default: 1711 return nullptr; 1712 } 1713 IRBuilder<> B(Ctx); 1714 Value *S = I->getOperand(1); 1715 return B.CreateBinOp(BitOp->getOpcode(), 1716 B.CreateLShr(BitOp->getOperand(0), S), 1717 B.CreateLShr(BitOp->getOperand(1), S)); 1718 }); 1719 S.addRule("expose bitop-const", 1720 // (BitOp1 (BitOp2 x a) b) -> (BitOp2 x (BitOp1 a b)) 1721 [](Instruction *I, LLVMContext &Ctx) -> Value* { 1722 auto IsBitOp = [](unsigned Op) -> bool { 1723 switch (Op) { 1724 case Instruction::And: 1725 case Instruction::Or: 1726 case Instruction::Xor: 1727 return true; 1728 } 1729 return false; 1730 }; 1731 BinaryOperator *BitOp1 = dyn_cast<BinaryOperator>(I); 1732 if (!BitOp1 || !IsBitOp(BitOp1->getOpcode())) 1733 return nullptr; 1734 BinaryOperator *BitOp2 = dyn_cast<BinaryOperator>(BitOp1->getOperand(0)); 1735 if (!BitOp2 || !IsBitOp(BitOp2->getOpcode())) 1736 return nullptr; 1737 ConstantInt *CA = dyn_cast<ConstantInt>(BitOp2->getOperand(1)); 1738 ConstantInt *CB = dyn_cast<ConstantInt>(BitOp1->getOperand(1)); 1739 if (!CA || !CB) 1740 return nullptr; 1741 IRBuilder<> B(Ctx); 1742 Value *X = BitOp2->getOperand(0); 1743 return B.CreateBinOp(BitOp2->getOpcode(), X, 1744 B.CreateBinOp(BitOp1->getOpcode(), CA, CB)); 1745 }); 1746 } 1747 1748 void PolynomialMultiplyRecognize::setupPostSimplifier(Simplifier &S) { 1749 S.addRule("(and (xor (and x a) y) b) -> (and (xor x y) b), if b == b&a", 1750 [](Instruction *I, LLVMContext &Ctx) -> Value* { 1751 if (I->getOpcode() != Instruction::And) 1752 return nullptr; 1753 Instruction *Xor = dyn_cast<Instruction>(I->getOperand(0)); 1754 ConstantInt *C0 = dyn_cast<ConstantInt>(I->getOperand(1)); 1755 if (!Xor || !C0) 1756 return nullptr; 1757 if (Xor->getOpcode() != Instruction::Xor) 1758 return nullptr; 1759 Instruction *And0 = dyn_cast<Instruction>(Xor->getOperand(0)); 1760 Instruction *And1 = dyn_cast<Instruction>(Xor->getOperand(1)); 1761 // Pick the first non-null and. 1762 if (!And0 || And0->getOpcode() != Instruction::And) 1763 std::swap(And0, And1); 1764 ConstantInt *C1 = dyn_cast<ConstantInt>(And0->getOperand(1)); 1765 if (!C1) 1766 return nullptr; 1767 uint32_t V0 = C0->getZExtValue(); 1768 uint32_t V1 = C1->getZExtValue(); 1769 if (V0 != (V0 & V1)) 1770 return nullptr; 1771 IRBuilder<> B(Ctx); 1772 return B.CreateAnd(B.CreateXor(And0->getOperand(0), And1), C0); 1773 }); 1774 } 1775 1776 bool PolynomialMultiplyRecognize::recognize() { 1777 LLVM_DEBUG(dbgs() << "Starting PolynomialMultiplyRecognize on loop\n" 1778 << *CurLoop << '\n'); 1779 // Restrictions: 1780 // - The loop must consist of a single block. 1781 // - The iteration count must be known at compile-time. 1782 // - The loop must have an induction variable starting from 0, and 1783 // incremented in each iteration of the loop. 1784 BasicBlock *LoopB = CurLoop->getHeader(); 1785 LLVM_DEBUG(dbgs() << "Loop header:\n" << *LoopB); 1786 1787 if (LoopB != CurLoop->getLoopLatch()) 1788 return false; 1789 BasicBlock *ExitB = CurLoop->getExitBlock(); 1790 if (ExitB == nullptr) 1791 return false; 1792 BasicBlock *EntryB = CurLoop->getLoopPreheader(); 1793 if (EntryB == nullptr) 1794 return false; 1795 1796 unsigned IterCount = 0; 1797 const SCEV *CT = SE.getBackedgeTakenCount(CurLoop); 1798 if (isa<SCEVCouldNotCompute>(CT)) 1799 return false; 1800 if (auto *CV = dyn_cast<SCEVConstant>(CT)) 1801 IterCount = CV->getValue()->getZExtValue() + 1; 1802 1803 Value *CIV = getCountIV(LoopB); 1804 ParsedValues PV; 1805 Simplifier PreSimp; 1806 PV.IterCount = IterCount; 1807 LLVM_DEBUG(dbgs() << "Loop IV: " << *CIV << "\nIterCount: " << IterCount 1808 << '\n'); 1809 1810 setupPreSimplifier(PreSimp); 1811 1812 // Perform a preliminary scan of select instructions to see if any of them 1813 // looks like a generator of the polynomial multiply steps. Assume that a 1814 // loop can only contain a single transformable operation, so stop the 1815 // traversal after the first reasonable candidate was found. 1816 // XXX: Currently this approach can modify the loop before being 100% sure 1817 // that the transformation can be carried out. 1818 bool FoundPreScan = false; 1819 auto FeedsPHI = [LoopB](const Value *V) -> bool { 1820 for (const Value *U : V->users()) { 1821 if (const auto *P = dyn_cast<const PHINode>(U)) 1822 if (P->getParent() == LoopB) 1823 return true; 1824 } 1825 return false; 1826 }; 1827 for (Instruction &In : *LoopB) { 1828 SelectInst *SI = dyn_cast<SelectInst>(&In); 1829 if (!SI || !FeedsPHI(SI)) 1830 continue; 1831 1832 Simplifier::Context C(SI); 1833 Value *T = PreSimp.simplify(C); 1834 SelectInst *SelI = (T && isa<SelectInst>(T)) ? cast<SelectInst>(T) : SI; 1835 LLVM_DEBUG(dbgs() << "scanSelect(pre-scan): " << PE(C, SelI) << '\n'); 1836 if (scanSelect(SelI, LoopB, EntryB, CIV, PV, true)) { 1837 FoundPreScan = true; 1838 if (SelI != SI) { 1839 Value *NewSel = C.materialize(LoopB, SI->getIterator()); 1840 SI->replaceAllUsesWith(NewSel); 1841 RecursivelyDeleteTriviallyDeadInstructions(SI, &TLI); 1842 } 1843 break; 1844 } 1845 } 1846 1847 if (!FoundPreScan) { 1848 LLVM_DEBUG(dbgs() << "Have not found candidates for pmpy\n"); 1849 return false; 1850 } 1851 1852 if (!PV.Left) { 1853 // The right shift version actually only returns the higher bits of 1854 // the result (each iteration discards the LSB). If we want to convert it 1855 // to a left-shifting loop, the working data type must be at least as 1856 // wide as the target's pmpy instruction. 1857 if (!promoteTypes(LoopB, ExitB)) 1858 return false; 1859 // Run post-promotion simplifications. 1860 Simplifier PostSimp; 1861 setupPostSimplifier(PostSimp); 1862 for (Instruction &In : *LoopB) { 1863 SelectInst *SI = dyn_cast<SelectInst>(&In); 1864 if (!SI || !FeedsPHI(SI)) 1865 continue; 1866 Simplifier::Context C(SI); 1867 Value *T = PostSimp.simplify(C); 1868 SelectInst *SelI = dyn_cast_or_null<SelectInst>(T); 1869 if (SelI != SI) { 1870 Value *NewSel = C.materialize(LoopB, SI->getIterator()); 1871 SI->replaceAllUsesWith(NewSel); 1872 RecursivelyDeleteTriviallyDeadInstructions(SI, &TLI); 1873 } 1874 break; 1875 } 1876 1877 if (!convertShiftsToLeft(LoopB, ExitB, IterCount)) 1878 return false; 1879 cleanupLoopBody(LoopB); 1880 } 1881 1882 // Scan the loop again, find the generating select instruction. 1883 bool FoundScan = false; 1884 for (Instruction &In : *LoopB) { 1885 SelectInst *SelI = dyn_cast<SelectInst>(&In); 1886 if (!SelI) 1887 continue; 1888 LLVM_DEBUG(dbgs() << "scanSelect: " << *SelI << '\n'); 1889 FoundScan = scanSelect(SelI, LoopB, EntryB, CIV, PV, false); 1890 if (FoundScan) 1891 break; 1892 } 1893 assert(FoundScan); 1894 1895 LLVM_DEBUG({ 1896 StringRef PP = (PV.M ? "(P+M)" : "P"); 1897 if (!PV.Inv) 1898 dbgs() << "Found pmpy idiom: R = " << PP << ".Q\n"; 1899 else 1900 dbgs() << "Found inverse pmpy idiom: R = (" << PP << "/Q).Q) + " 1901 << PP << "\n"; 1902 dbgs() << " Res:" << *PV.Res << "\n P:" << *PV.P << "\n"; 1903 if (PV.M) 1904 dbgs() << " M:" << *PV.M << "\n"; 1905 dbgs() << " Q:" << *PV.Q << "\n"; 1906 dbgs() << " Iteration count:" << PV.IterCount << "\n"; 1907 }); 1908 1909 BasicBlock::iterator At(EntryB->getTerminator()); 1910 Value *PM = generate(At, PV); 1911 if (PM == nullptr) 1912 return false; 1913 1914 if (PM->getType() != PV.Res->getType()) 1915 PM = IRBuilder<>(&*At).CreateIntCast(PM, PV.Res->getType(), false); 1916 1917 PV.Res->replaceAllUsesWith(PM); 1918 PV.Res->eraseFromParent(); 1919 return true; 1920 } 1921 1922 int HexagonLoopIdiomRecognize::getSCEVStride(const SCEVAddRecExpr *S) { 1923 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(1))) 1924 return SC->getAPInt().getSExtValue(); 1925 return 0; 1926 } 1927 1928 bool HexagonLoopIdiomRecognize::isLegalStore(Loop *CurLoop, StoreInst *SI) { 1929 // Allow volatile stores if HexagonVolatileMemcpy is enabled. 1930 if (!(SI->isVolatile() && HexagonVolatileMemcpy) && !SI->isSimple()) 1931 return false; 1932 1933 Value *StoredVal = SI->getValueOperand(); 1934 Value *StorePtr = SI->getPointerOperand(); 1935 1936 // Reject stores that are so large that they overflow an unsigned. 1937 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType()); 1938 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0) 1939 return false; 1940 1941 // See if the pointer expression is an AddRec like {base,+,1} on the current 1942 // loop, which indicates a strided store. If we have something else, it's a 1943 // random store we can't handle. 1944 auto *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); 1945 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) 1946 return false; 1947 1948 // Check to see if the stride matches the size of the store. If so, then we 1949 // know that every byte is touched in the loop. 1950 int Stride = getSCEVStride(StoreEv); 1951 if (Stride == 0) 1952 return false; 1953 unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); 1954 if (StoreSize != unsigned(std::abs(Stride))) 1955 return false; 1956 1957 // The store must be feeding a non-volatile load. 1958 LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand()); 1959 if (!LI || !LI->isSimple()) 1960 return false; 1961 1962 // See if the pointer expression is an AddRec like {base,+,1} on the current 1963 // loop, which indicates a strided load. If we have something else, it's a 1964 // random load we can't handle. 1965 Value *LoadPtr = LI->getPointerOperand(); 1966 auto *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LoadPtr)); 1967 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) 1968 return false; 1969 1970 // The store and load must share the same stride. 1971 if (StoreEv->getOperand(1) != LoadEv->getOperand(1)) 1972 return false; 1973 1974 // Success. This store can be converted into a memcpy. 1975 return true; 1976 } 1977 1978 /// mayLoopAccessLocation - Return true if the specified loop might access the 1979 /// specified pointer location, which is a loop-strided access. The 'Access' 1980 /// argument specifies what the verboten forms of access are (read or write). 1981 static bool 1982 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L, 1983 const SCEV *BECount, unsigned StoreSize, 1984 AliasAnalysis &AA, 1985 SmallPtrSetImpl<Instruction *> &Ignored) { 1986 // Get the location that may be stored across the loop. Since the access 1987 // is strided positively through memory, we say that the modified location 1988 // starts at the pointer and has infinite size. 1989 LocationSize AccessSize = LocationSize::afterPointer(); 1990 1991 // If the loop iterates a fixed number of times, we can refine the access 1992 // size to be exactly the size of the memset, which is (BECount+1)*StoreSize 1993 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) 1994 AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) * 1995 StoreSize); 1996 1997 // TODO: For this to be really effective, we have to dive into the pointer 1998 // operand in the store. Store to &A[i] of 100 will always return may alias 1999 // with store of &A[100], we need to StoreLoc to be "A" with size of 100, 2000 // which will then no-alias a store to &A[100]. 2001 MemoryLocation StoreLoc(Ptr, AccessSize); 2002 2003 for (auto *B : L->blocks()) 2004 for (auto &I : *B) 2005 if (Ignored.count(&I) == 0 && 2006 isModOrRefSet(AA.getModRefInfo(&I, StoreLoc) & Access)) 2007 return true; 2008 2009 return false; 2010 } 2011 2012 void HexagonLoopIdiomRecognize::collectStores(Loop *CurLoop, BasicBlock *BB, 2013 SmallVectorImpl<StoreInst*> &Stores) { 2014 Stores.clear(); 2015 for (Instruction &I : *BB) 2016 if (StoreInst *SI = dyn_cast<StoreInst>(&I)) 2017 if (isLegalStore(CurLoop, SI)) 2018 Stores.push_back(SI); 2019 } 2020 2021 bool HexagonLoopIdiomRecognize::processCopyingStore(Loop *CurLoop, 2022 StoreInst *SI, const SCEV *BECount) { 2023 assert((SI->isSimple() || (SI->isVolatile() && HexagonVolatileMemcpy)) && 2024 "Expected only non-volatile stores, or Hexagon-specific memcpy" 2025 "to volatile destination."); 2026 2027 Value *StorePtr = SI->getPointerOperand(); 2028 auto *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); 2029 unsigned Stride = getSCEVStride(StoreEv); 2030 unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); 2031 if (Stride != StoreSize) 2032 return false; 2033 2034 // See if the pointer expression is an AddRec like {base,+,1} on the current 2035 // loop, which indicates a strided load. If we have something else, it's a 2036 // random load we can't handle. 2037 auto *LI = cast<LoadInst>(SI->getValueOperand()); 2038 auto *LoadEv = cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand())); 2039 2040 // The trip count of the loop and the base pointer of the addrec SCEV is 2041 // guaranteed to be loop invariant, which means that it should dominate the 2042 // header. This allows us to insert code for it in the preheader. 2043 BasicBlock *Preheader = CurLoop->getLoopPreheader(); 2044 Instruction *ExpPt = Preheader->getTerminator(); 2045 IRBuilder<> Builder(ExpPt); 2046 SCEVExpander Expander(*SE, *DL, "hexagon-loop-idiom"); 2047 2048 Type *IntPtrTy = Builder.getIntPtrTy(*DL, SI->getPointerAddressSpace()); 2049 2050 // Okay, we have a strided store "p[i]" of a loaded value. We can turn 2051 // this into a memcpy/memmove in the loop preheader now if we want. However, 2052 // this would be unsafe to do if there is anything else in the loop that may 2053 // read or write the memory region we're storing to. For memcpy, this 2054 // includes the load that feeds the stores. Check for an alias by generating 2055 // the base address and checking everything. 2056 Value *StoreBasePtr = Expander.expandCodeFor(StoreEv->getStart(), 2057 Builder.getInt8PtrTy(SI->getPointerAddressSpace()), ExpPt); 2058 Value *LoadBasePtr = nullptr; 2059 2060 bool Overlap = false; 2061 bool DestVolatile = SI->isVolatile(); 2062 Type *BECountTy = BECount->getType(); 2063 2064 if (DestVolatile) { 2065 // The trip count must fit in i32, since it is the type of the "num_words" 2066 // argument to hexagon_memcpy_forward_vp4cp4n2. 2067 if (StoreSize != 4 || DL->getTypeSizeInBits(BECountTy) > 32) { 2068 CleanupAndExit: 2069 // If we generated new code for the base pointer, clean up. 2070 Expander.clear(); 2071 if (StoreBasePtr && (LoadBasePtr != StoreBasePtr)) { 2072 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI); 2073 StoreBasePtr = nullptr; 2074 } 2075 if (LoadBasePtr) { 2076 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI); 2077 LoadBasePtr = nullptr; 2078 } 2079 return false; 2080 } 2081 } 2082 2083 SmallPtrSet<Instruction*, 2> Ignore1; 2084 Ignore1.insert(SI); 2085 if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount, 2086 StoreSize, *AA, Ignore1)) { 2087 // Check if the load is the offending instruction. 2088 Ignore1.insert(LI); 2089 if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, 2090 BECount, StoreSize, *AA, Ignore1)) { 2091 // Still bad. Nothing we can do. 2092 goto CleanupAndExit; 2093 } 2094 // It worked with the load ignored. 2095 Overlap = true; 2096 } 2097 2098 if (!Overlap) { 2099 if (DisableMemcpyIdiom || !HasMemcpy) 2100 goto CleanupAndExit; 2101 } else { 2102 // Don't generate memmove if this function will be inlined. This is 2103 // because the caller will undergo this transformation after inlining. 2104 Function *Func = CurLoop->getHeader()->getParent(); 2105 if (Func->hasFnAttribute(Attribute::AlwaysInline)) 2106 goto CleanupAndExit; 2107 2108 // In case of a memmove, the call to memmove will be executed instead 2109 // of the loop, so we need to make sure that there is nothing else in 2110 // the loop than the load, store and instructions that these two depend 2111 // on. 2112 SmallVector<Instruction*,2> Insts; 2113 Insts.push_back(SI); 2114 Insts.push_back(LI); 2115 if (!coverLoop(CurLoop, Insts)) 2116 goto CleanupAndExit; 2117 2118 if (DisableMemmoveIdiom || !HasMemmove) 2119 goto CleanupAndExit; 2120 bool IsNested = CurLoop->getParentLoop() != nullptr; 2121 if (IsNested && OnlyNonNestedMemmove) 2122 goto CleanupAndExit; 2123 } 2124 2125 // For a memcpy, we have to make sure that the input array is not being 2126 // mutated by the loop. 2127 LoadBasePtr = Expander.expandCodeFor(LoadEv->getStart(), 2128 Builder.getInt8PtrTy(LI->getPointerAddressSpace()), ExpPt); 2129 2130 SmallPtrSet<Instruction*, 2> Ignore2; 2131 Ignore2.insert(SI); 2132 if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount, 2133 StoreSize, *AA, Ignore2)) 2134 goto CleanupAndExit; 2135 2136 // Check the stride. 2137 bool StridePos = getSCEVStride(LoadEv) >= 0; 2138 2139 // Currently, the volatile memcpy only emulates traversing memory forward. 2140 if (!StridePos && DestVolatile) 2141 goto CleanupAndExit; 2142 2143 bool RuntimeCheck = (Overlap || DestVolatile); 2144 2145 BasicBlock *ExitB; 2146 if (RuntimeCheck) { 2147 // The runtime check needs a single exit block. 2148 SmallVector<BasicBlock*, 8> ExitBlocks; 2149 CurLoop->getUniqueExitBlocks(ExitBlocks); 2150 if (ExitBlocks.size() != 1) 2151 goto CleanupAndExit; 2152 ExitB = ExitBlocks[0]; 2153 } 2154 2155 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to 2156 // pointer size if it isn't already. 2157 LLVMContext &Ctx = SI->getContext(); 2158 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy); 2159 DebugLoc DLoc = SI->getDebugLoc(); 2160 2161 const SCEV *NumBytesS = 2162 SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW); 2163 if (StoreSize != 1) 2164 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize), 2165 SCEV::FlagNUW); 2166 Value *NumBytes = Expander.expandCodeFor(NumBytesS, IntPtrTy, ExpPt); 2167 if (Instruction *In = dyn_cast<Instruction>(NumBytes)) 2168 if (Value *Simp = simplifyInstruction(In, {*DL, TLI, DT})) 2169 NumBytes = Simp; 2170 2171 CallInst *NewCall; 2172 2173 if (RuntimeCheck) { 2174 unsigned Threshold = RuntimeMemSizeThreshold; 2175 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) { 2176 uint64_t C = CI->getZExtValue(); 2177 if (Threshold != 0 && C < Threshold) 2178 goto CleanupAndExit; 2179 if (C < CompileTimeMemSizeThreshold) 2180 goto CleanupAndExit; 2181 } 2182 2183 BasicBlock *Header = CurLoop->getHeader(); 2184 Function *Func = Header->getParent(); 2185 Loop *ParentL = LF->getLoopFor(Preheader); 2186 StringRef HeaderName = Header->getName(); 2187 2188 // Create a new (empty) preheader, and update the PHI nodes in the 2189 // header to use the new preheader. 2190 BasicBlock *NewPreheader = BasicBlock::Create(Ctx, HeaderName+".rtli.ph", 2191 Func, Header); 2192 if (ParentL) 2193 ParentL->addBasicBlockToLoop(NewPreheader, *LF); 2194 IRBuilder<>(NewPreheader).CreateBr(Header); 2195 for (auto &In : *Header) { 2196 PHINode *PN = dyn_cast<PHINode>(&In); 2197 if (!PN) 2198 break; 2199 int bx = PN->getBasicBlockIndex(Preheader); 2200 if (bx >= 0) 2201 PN->setIncomingBlock(bx, NewPreheader); 2202 } 2203 DT->addNewBlock(NewPreheader, Preheader); 2204 DT->changeImmediateDominator(Header, NewPreheader); 2205 2206 // Check for safe conditions to execute memmove. 2207 // If stride is positive, copying things from higher to lower addresses 2208 // is equivalent to memmove. For negative stride, it's the other way 2209 // around. Copying forward in memory with positive stride may not be 2210 // same as memmove since we may be copying values that we just stored 2211 // in some previous iteration. 2212 Value *LA = Builder.CreatePtrToInt(LoadBasePtr, IntPtrTy); 2213 Value *SA = Builder.CreatePtrToInt(StoreBasePtr, IntPtrTy); 2214 Value *LowA = StridePos ? SA : LA; 2215 Value *HighA = StridePos ? LA : SA; 2216 Value *CmpA = Builder.CreateICmpULT(LowA, HighA); 2217 Value *Cond = CmpA; 2218 2219 // Check for distance between pointers. Since the case LowA < HighA 2220 // is checked for above, assume LowA >= HighA. 2221 Value *Dist = Builder.CreateSub(LowA, HighA); 2222 Value *CmpD = Builder.CreateICmpSLE(NumBytes, Dist); 2223 Value *CmpEither = Builder.CreateOr(Cond, CmpD); 2224 Cond = CmpEither; 2225 2226 if (Threshold != 0) { 2227 Type *Ty = NumBytes->getType(); 2228 Value *Thr = ConstantInt::get(Ty, Threshold); 2229 Value *CmpB = Builder.CreateICmpULT(Thr, NumBytes); 2230 Value *CmpBoth = Builder.CreateAnd(Cond, CmpB); 2231 Cond = CmpBoth; 2232 } 2233 BasicBlock *MemmoveB = BasicBlock::Create(Ctx, Header->getName()+".rtli", 2234 Func, NewPreheader); 2235 if (ParentL) 2236 ParentL->addBasicBlockToLoop(MemmoveB, *LF); 2237 Instruction *OldT = Preheader->getTerminator(); 2238 Builder.CreateCondBr(Cond, MemmoveB, NewPreheader); 2239 OldT->eraseFromParent(); 2240 Preheader->setName(Preheader->getName()+".old"); 2241 DT->addNewBlock(MemmoveB, Preheader); 2242 // Find the new immediate dominator of the exit block. 2243 BasicBlock *ExitD = Preheader; 2244 for (BasicBlock *PB : predecessors(ExitB)) { 2245 ExitD = DT->findNearestCommonDominator(ExitD, PB); 2246 if (!ExitD) 2247 break; 2248 } 2249 // If the prior immediate dominator of ExitB was dominated by the 2250 // old preheader, then the old preheader becomes the new immediate 2251 // dominator. Otherwise don't change anything (because the newly 2252 // added blocks are dominated by the old preheader). 2253 if (ExitD && DT->dominates(Preheader, ExitD)) { 2254 DomTreeNode *BN = DT->getNode(ExitB); 2255 DomTreeNode *DN = DT->getNode(ExitD); 2256 BN->setIDom(DN); 2257 } 2258 2259 // Add a call to memmove to the conditional block. 2260 IRBuilder<> CondBuilder(MemmoveB); 2261 CondBuilder.CreateBr(ExitB); 2262 CondBuilder.SetInsertPoint(MemmoveB->getTerminator()); 2263 2264 if (DestVolatile) { 2265 Type *Int32Ty = Type::getInt32Ty(Ctx); 2266 Type *Int32PtrTy = Type::getInt32PtrTy(Ctx); 2267 Type *VoidTy = Type::getVoidTy(Ctx); 2268 Module *M = Func->getParent(); 2269 FunctionCallee Fn = M->getOrInsertFunction( 2270 HexagonVolatileMemcpyName, VoidTy, Int32PtrTy, Int32PtrTy, Int32Ty); 2271 2272 const SCEV *OneS = SE->getConstant(Int32Ty, 1); 2273 const SCEV *BECount32 = SE->getTruncateOrZeroExtend(BECount, Int32Ty); 2274 const SCEV *NumWordsS = SE->getAddExpr(BECount32, OneS, SCEV::FlagNUW); 2275 Value *NumWords = Expander.expandCodeFor(NumWordsS, Int32Ty, 2276 MemmoveB->getTerminator()); 2277 if (Instruction *In = dyn_cast<Instruction>(NumWords)) 2278 if (Value *Simp = simplifyInstruction(In, {*DL, TLI, DT})) 2279 NumWords = Simp; 2280 2281 Value *Op0 = (StoreBasePtr->getType() == Int32PtrTy) 2282 ? StoreBasePtr 2283 : CondBuilder.CreateBitCast(StoreBasePtr, Int32PtrTy); 2284 Value *Op1 = (LoadBasePtr->getType() == Int32PtrTy) 2285 ? LoadBasePtr 2286 : CondBuilder.CreateBitCast(LoadBasePtr, Int32PtrTy); 2287 NewCall = CondBuilder.CreateCall(Fn, {Op0, Op1, NumWords}); 2288 } else { 2289 NewCall = CondBuilder.CreateMemMove( 2290 StoreBasePtr, SI->getAlign(), LoadBasePtr, LI->getAlign(), NumBytes); 2291 } 2292 } else { 2293 NewCall = Builder.CreateMemCpy(StoreBasePtr, SI->getAlign(), LoadBasePtr, 2294 LI->getAlign(), NumBytes); 2295 // Okay, the memcpy has been formed. Zap the original store and 2296 // anything that feeds into it. 2297 RecursivelyDeleteTriviallyDeadInstructions(SI, TLI); 2298 } 2299 2300 NewCall->setDebugLoc(DLoc); 2301 2302 LLVM_DEBUG(dbgs() << " Formed " << (Overlap ? "memmove: " : "memcpy: ") 2303 << *NewCall << "\n" 2304 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" 2305 << " from store ptr=" << *StoreEv << " at: " << *SI 2306 << "\n"); 2307 2308 return true; 2309 } 2310 2311 // Check if the instructions in Insts, together with their dependencies 2312 // cover the loop in the sense that the loop could be safely eliminated once 2313 // the instructions in Insts are removed. 2314 bool HexagonLoopIdiomRecognize::coverLoop(Loop *L, 2315 SmallVectorImpl<Instruction*> &Insts) const { 2316 SmallSet<BasicBlock*,8> LoopBlocks; 2317 for (auto *B : L->blocks()) 2318 LoopBlocks.insert(B); 2319 2320 SetVector<Instruction*> Worklist(Insts.begin(), Insts.end()); 2321 2322 // Collect all instructions from the loop that the instructions in Insts 2323 // depend on (plus their dependencies, etc.). These instructions will 2324 // constitute the expression trees that feed those in Insts, but the trees 2325 // will be limited only to instructions contained in the loop. 2326 for (unsigned i = 0; i < Worklist.size(); ++i) { 2327 Instruction *In = Worklist[i]; 2328 for (auto I = In->op_begin(), E = In->op_end(); I != E; ++I) { 2329 Instruction *OpI = dyn_cast<Instruction>(I); 2330 if (!OpI) 2331 continue; 2332 BasicBlock *PB = OpI->getParent(); 2333 if (!LoopBlocks.count(PB)) 2334 continue; 2335 Worklist.insert(OpI); 2336 } 2337 } 2338 2339 // Scan all instructions in the loop, if any of them have a user outside 2340 // of the loop, or outside of the expressions collected above, then either 2341 // the loop has a side-effect visible outside of it, or there are 2342 // instructions in it that are not involved in the original set Insts. 2343 for (auto *B : L->blocks()) { 2344 for (auto &In : *B) { 2345 if (isa<BranchInst>(In) || isa<DbgInfoIntrinsic>(In)) 2346 continue; 2347 if (!Worklist.count(&In) && In.mayHaveSideEffects()) 2348 return false; 2349 for (auto *K : In.users()) { 2350 Instruction *UseI = dyn_cast<Instruction>(K); 2351 if (!UseI) 2352 continue; 2353 BasicBlock *UseB = UseI->getParent(); 2354 if (LF->getLoopFor(UseB) != L) 2355 return false; 2356 } 2357 } 2358 } 2359 2360 return true; 2361 } 2362 2363 /// runOnLoopBlock - Process the specified block, which lives in a counted loop 2364 /// with the specified backedge count. This block is known to be in the current 2365 /// loop and not in any subloops. 2366 bool HexagonLoopIdiomRecognize::runOnLoopBlock(Loop *CurLoop, BasicBlock *BB, 2367 const SCEV *BECount, SmallVectorImpl<BasicBlock*> &ExitBlocks) { 2368 // We can only promote stores in this block if they are unconditionally 2369 // executed in the loop. For a block to be unconditionally executed, it has 2370 // to dominate all the exit blocks of the loop. Verify this now. 2371 auto DominatedByBB = [this,BB] (BasicBlock *EB) -> bool { 2372 return DT->dominates(BB, EB); 2373 }; 2374 if (!all_of(ExitBlocks, DominatedByBB)) 2375 return false; 2376 2377 bool MadeChange = false; 2378 // Look for store instructions, which may be optimized to memset/memcpy. 2379 SmallVector<StoreInst*,8> Stores; 2380 collectStores(CurLoop, BB, Stores); 2381 2382 // Optimize the store into a memcpy, if it feeds an similarly strided load. 2383 for (auto &SI : Stores) 2384 MadeChange |= processCopyingStore(CurLoop, SI, BECount); 2385 2386 return MadeChange; 2387 } 2388 2389 bool HexagonLoopIdiomRecognize::runOnCountableLoop(Loop *L) { 2390 PolynomialMultiplyRecognize PMR(L, *DL, *DT, *TLI, *SE); 2391 if (PMR.recognize()) 2392 return true; 2393 2394 if (!HasMemcpy && !HasMemmove) 2395 return false; 2396 2397 const SCEV *BECount = SE->getBackedgeTakenCount(L); 2398 assert(!isa<SCEVCouldNotCompute>(BECount) && 2399 "runOnCountableLoop() called on a loop without a predictable" 2400 "backedge-taken count"); 2401 2402 SmallVector<BasicBlock *, 8> ExitBlocks; 2403 L->getUniqueExitBlocks(ExitBlocks); 2404 2405 bool Changed = false; 2406 2407 // Scan all the blocks in the loop that are not in subloops. 2408 for (auto *BB : L->getBlocks()) { 2409 // Ignore blocks in subloops. 2410 if (LF->getLoopFor(BB) != L) 2411 continue; 2412 Changed |= runOnLoopBlock(L, BB, BECount, ExitBlocks); 2413 } 2414 2415 return Changed; 2416 } 2417 2418 bool HexagonLoopIdiomRecognize::run(Loop *L) { 2419 const Module &M = *L->getHeader()->getParent()->getParent(); 2420 if (Triple(M.getTargetTriple()).getArch() != Triple::hexagon) 2421 return false; 2422 2423 // If the loop could not be converted to canonical form, it must have an 2424 // indirectbr in it, just give up. 2425 if (!L->getLoopPreheader()) 2426 return false; 2427 2428 // Disable loop idiom recognition if the function's name is a common idiom. 2429 StringRef Name = L->getHeader()->getParent()->getName(); 2430 if (Name == "memset" || Name == "memcpy" || Name == "memmove") 2431 return false; 2432 2433 DL = &L->getHeader()->getModule()->getDataLayout(); 2434 2435 HasMemcpy = TLI->has(LibFunc_memcpy); 2436 HasMemmove = TLI->has(LibFunc_memmove); 2437 2438 if (SE->hasLoopInvariantBackedgeTakenCount(L)) 2439 return runOnCountableLoop(L); 2440 return false; 2441 } 2442 2443 bool HexagonLoopIdiomRecognizeLegacyPass::runOnLoop(Loop *L, 2444 LPPassManager &LPM) { 2445 if (skipLoop(L)) 2446 return false; 2447 2448 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 2449 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2450 auto *LF = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 2451 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( 2452 *L->getHeader()->getParent()); 2453 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 2454 return HexagonLoopIdiomRecognize(AA, DT, LF, TLI, SE).run(L); 2455 } 2456 2457 Pass *llvm::createHexagonLoopIdiomPass() { 2458 return new HexagonLoopIdiomRecognizeLegacyPass(); 2459 } 2460 2461 PreservedAnalyses 2462 HexagonLoopIdiomRecognitionPass::run(Loop &L, LoopAnalysisManager &AM, 2463 LoopStandardAnalysisResults &AR, 2464 LPMUpdater &U) { 2465 return HexagonLoopIdiomRecognize(&AR.AA, &AR.DT, &AR.LI, &AR.TLI, &AR.SE) 2466 .run(&L) 2467 ? getLoopPassPreservedAnalyses() 2468 : PreservedAnalyses::all(); 2469 } 2470