1 //===-- LoopUtils.cpp - Loop Utility functions -------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines common loop utility functions. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/Utils/LoopUtils.h" 15 #include "llvm/ADT/ScopeExit.h" 16 #include "llvm/Analysis/AliasAnalysis.h" 17 #include "llvm/Analysis/BasicAliasAnalysis.h" 18 #include "llvm/Analysis/GlobalsModRef.h" 19 #include "llvm/Analysis/InstructionSimplify.h" 20 #include "llvm/Analysis/LoopInfo.h" 21 #include "llvm/Analysis/LoopPass.h" 22 #include "llvm/Analysis/MustExecute.h" 23 #include "llvm/Analysis/ScalarEvolution.h" 24 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 25 #include "llvm/Analysis/ScalarEvolutionExpander.h" 26 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 27 #include "llvm/Analysis/TargetTransformInfo.h" 28 #include "llvm/Analysis/ValueTracking.h" 29 #include "llvm/IR/DomTreeUpdater.h" 30 #include "llvm/IR/Dominators.h" 31 #include "llvm/IR/Instructions.h" 32 #include "llvm/IR/Module.h" 33 #include "llvm/IR/PatternMatch.h" 34 #include "llvm/IR/ValueHandle.h" 35 #include "llvm/Pass.h" 36 #include "llvm/Support/Debug.h" 37 #include "llvm/Support/KnownBits.h" 38 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 39 40 using namespace llvm; 41 using namespace llvm::PatternMatch; 42 43 #define DEBUG_TYPE "loop-utils" 44 45 bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI, 46 bool PreserveLCSSA) { 47 bool Changed = false; 48 49 // We re-use a vector for the in-loop predecesosrs. 50 SmallVector<BasicBlock *, 4> InLoopPredecessors; 51 52 auto RewriteExit = [&](BasicBlock *BB) { 53 assert(InLoopPredecessors.empty() && 54 "Must start with an empty predecessors list!"); 55 auto Cleanup = make_scope_exit([&] { InLoopPredecessors.clear(); }); 56 57 // See if there are any non-loop predecessors of this exit block and 58 // keep track of the in-loop predecessors. 59 bool IsDedicatedExit = true; 60 for (auto *PredBB : predecessors(BB)) 61 if (L->contains(PredBB)) { 62 if (isa<IndirectBrInst>(PredBB->getTerminator())) 63 // We cannot rewrite exiting edges from an indirectbr. 64 return false; 65 66 InLoopPredecessors.push_back(PredBB); 67 } else { 68 IsDedicatedExit = false; 69 } 70 71 assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!"); 72 73 // Nothing to do if this is already a dedicated exit. 74 if (IsDedicatedExit) 75 return false; 76 77 auto *NewExitBB = SplitBlockPredecessors( 78 BB, InLoopPredecessors, ".loopexit", DT, LI, nullptr, PreserveLCSSA); 79 80 if (!NewExitBB) 81 LLVM_DEBUG( 82 dbgs() << "WARNING: Can't create a dedicated exit block for loop: " 83 << *L << "\n"); 84 else 85 LLVM_DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block " 86 << NewExitBB->getName() << "\n"); 87 return true; 88 }; 89 90 // Walk the exit blocks directly rather than building up a data structure for 91 // them, but only visit each one once. 92 SmallPtrSet<BasicBlock *, 4> Visited; 93 for (auto *BB : L->blocks()) 94 for (auto *SuccBB : successors(BB)) { 95 // We're looking for exit blocks so skip in-loop successors. 96 if (L->contains(SuccBB)) 97 continue; 98 99 // Visit each exit block exactly once. 100 if (!Visited.insert(SuccBB).second) 101 continue; 102 103 Changed |= RewriteExit(SuccBB); 104 } 105 106 return Changed; 107 } 108 109 /// Returns the instructions that use values defined in the loop. 110 SmallVector<Instruction *, 8> llvm::findDefsUsedOutsideOfLoop(Loop *L) { 111 SmallVector<Instruction *, 8> UsedOutside; 112 113 for (auto *Block : L->getBlocks()) 114 // FIXME: I believe that this could use copy_if if the Inst reference could 115 // be adapted into a pointer. 116 for (auto &Inst : *Block) { 117 auto Users = Inst.users(); 118 if (any_of(Users, [&](User *U) { 119 auto *Use = cast<Instruction>(U); 120 return !L->contains(Use->getParent()); 121 })) 122 UsedOutside.push_back(&Inst); 123 } 124 125 return UsedOutside; 126 } 127 128 void llvm::getLoopAnalysisUsage(AnalysisUsage &AU) { 129 // By definition, all loop passes need the LoopInfo analysis and the 130 // Dominator tree it depends on. Because they all participate in the loop 131 // pass manager, they must also preserve these. 132 AU.addRequired<DominatorTreeWrapperPass>(); 133 AU.addPreserved<DominatorTreeWrapperPass>(); 134 AU.addRequired<LoopInfoWrapperPass>(); 135 AU.addPreserved<LoopInfoWrapperPass>(); 136 137 // We must also preserve LoopSimplify and LCSSA. We locally access their IDs 138 // here because users shouldn't directly get them from this header. 139 extern char &LoopSimplifyID; 140 extern char &LCSSAID; 141 AU.addRequiredID(LoopSimplifyID); 142 AU.addPreservedID(LoopSimplifyID); 143 AU.addRequiredID(LCSSAID); 144 AU.addPreservedID(LCSSAID); 145 // This is used in the LPPassManager to perform LCSSA verification on passes 146 // which preserve lcssa form 147 AU.addRequired<LCSSAVerificationPass>(); 148 AU.addPreserved<LCSSAVerificationPass>(); 149 150 // Loop passes are designed to run inside of a loop pass manager which means 151 // that any function analyses they require must be required by the first loop 152 // pass in the manager (so that it is computed before the loop pass manager 153 // runs) and preserved by all loop pasess in the manager. To make this 154 // reasonably robust, the set needed for most loop passes is maintained here. 155 // If your loop pass requires an analysis not listed here, you will need to 156 // carefully audit the loop pass manager nesting structure that results. 157 AU.addRequired<AAResultsWrapperPass>(); 158 AU.addPreserved<AAResultsWrapperPass>(); 159 AU.addPreserved<BasicAAWrapperPass>(); 160 AU.addPreserved<GlobalsAAWrapperPass>(); 161 AU.addPreserved<SCEVAAWrapperPass>(); 162 AU.addRequired<ScalarEvolutionWrapperPass>(); 163 AU.addPreserved<ScalarEvolutionWrapperPass>(); 164 } 165 166 /// Manually defined generic "LoopPass" dependency initialization. This is used 167 /// to initialize the exact set of passes from above in \c 168 /// getLoopAnalysisUsage. It can be used within a loop pass's initialization 169 /// with: 170 /// 171 /// INITIALIZE_PASS_DEPENDENCY(LoopPass) 172 /// 173 /// As-if "LoopPass" were a pass. 174 void llvm::initializeLoopPassPass(PassRegistry &Registry) { 175 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 176 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 177 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 178 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass) 179 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 180 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass) 181 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) 182 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass) 183 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 184 } 185 186 /// Find string metadata for loop 187 /// 188 /// If it has a value (e.g. {"llvm.distribute", 1} return the value as an 189 /// operand or null otherwise. If the string metadata is not found return 190 /// Optional's not-a-value. 191 Optional<const MDOperand *> llvm::findStringMetadataForLoop(Loop *TheLoop, 192 StringRef Name) { 193 MDNode *LoopID = TheLoop->getLoopID(); 194 // Return none if LoopID is false. 195 if (!LoopID) 196 return None; 197 198 // First operand should refer to the loop id itself. 199 assert(LoopID->getNumOperands() > 0 && "requires at least one operand"); 200 assert(LoopID->getOperand(0) == LoopID && "invalid loop id"); 201 202 // Iterate over LoopID operands and look for MDString Metadata 203 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) { 204 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 205 if (!MD) 206 continue; 207 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 208 if (!S) 209 continue; 210 // Return true if MDString holds expected MetaData. 211 if (Name.equals(S->getString())) 212 switch (MD->getNumOperands()) { 213 case 1: 214 return nullptr; 215 case 2: 216 return &MD->getOperand(1); 217 default: 218 llvm_unreachable("loop metadata has 0 or 1 operand"); 219 } 220 } 221 return None; 222 } 223 224 /// Does a BFS from a given node to all of its children inside a given loop. 225 /// The returned vector of nodes includes the starting point. 226 SmallVector<DomTreeNode *, 16> 227 llvm::collectChildrenInLoop(DomTreeNode *N, const Loop *CurLoop) { 228 SmallVector<DomTreeNode *, 16> Worklist; 229 auto AddRegionToWorklist = [&](DomTreeNode *DTN) { 230 // Only include subregions in the top level loop. 231 BasicBlock *BB = DTN->getBlock(); 232 if (CurLoop->contains(BB)) 233 Worklist.push_back(DTN); 234 }; 235 236 AddRegionToWorklist(N); 237 238 for (size_t I = 0; I < Worklist.size(); I++) 239 for (DomTreeNode *Child : Worklist[I]->getChildren()) 240 AddRegionToWorklist(Child); 241 242 return Worklist; 243 } 244 245 void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT = nullptr, 246 ScalarEvolution *SE = nullptr, 247 LoopInfo *LI = nullptr) { 248 assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!"); 249 auto *Preheader = L->getLoopPreheader(); 250 assert(Preheader && "Preheader should exist!"); 251 252 // Now that we know the removal is safe, remove the loop by changing the 253 // branch from the preheader to go to the single exit block. 254 // 255 // Because we're deleting a large chunk of code at once, the sequence in which 256 // we remove things is very important to avoid invalidation issues. 257 258 // Tell ScalarEvolution that the loop is deleted. Do this before 259 // deleting the loop so that ScalarEvolution can look at the loop 260 // to determine what it needs to clean up. 261 if (SE) 262 SE->forgetLoop(L); 263 264 auto *ExitBlock = L->getUniqueExitBlock(); 265 assert(ExitBlock && "Should have a unique exit block!"); 266 assert(L->hasDedicatedExits() && "Loop should have dedicated exits!"); 267 268 auto *OldBr = dyn_cast<BranchInst>(Preheader->getTerminator()); 269 assert(OldBr && "Preheader must end with a branch"); 270 assert(OldBr->isUnconditional() && "Preheader must have a single successor"); 271 // Connect the preheader to the exit block. Keep the old edge to the header 272 // around to perform the dominator tree update in two separate steps 273 // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge 274 // preheader -> header. 275 // 276 // 277 // 0. Preheader 1. Preheader 2. Preheader 278 // | | | | 279 // V | V | 280 // Header <--\ | Header <--\ | Header <--\ 281 // | | | | | | | | | | | 282 // | V | | | V | | | V | 283 // | Body --/ | | Body --/ | | Body --/ 284 // V V V V V 285 // Exit Exit Exit 286 // 287 // By doing this is two separate steps we can perform the dominator tree 288 // update without using the batch update API. 289 // 290 // Even when the loop is never executed, we cannot remove the edge from the 291 // source block to the exit block. Consider the case where the unexecuted loop 292 // branches back to an outer loop. If we deleted the loop and removed the edge 293 // coming to this inner loop, this will break the outer loop structure (by 294 // deleting the backedge of the outer loop). If the outer loop is indeed a 295 // non-loop, it will be deleted in a future iteration of loop deletion pass. 296 IRBuilder<> Builder(OldBr); 297 Builder.CreateCondBr(Builder.getFalse(), L->getHeader(), ExitBlock); 298 // Remove the old branch. The conditional branch becomes a new terminator. 299 OldBr->eraseFromParent(); 300 301 // Rewrite phis in the exit block to get their inputs from the Preheader 302 // instead of the exiting block. 303 for (PHINode &P : ExitBlock->phis()) { 304 // Set the zero'th element of Phi to be from the preheader and remove all 305 // other incoming values. Given the loop has dedicated exits, all other 306 // incoming values must be from the exiting blocks. 307 int PredIndex = 0; 308 P.setIncomingBlock(PredIndex, Preheader); 309 // Removes all incoming values from all other exiting blocks (including 310 // duplicate values from an exiting block). 311 // Nuke all entries except the zero'th entry which is the preheader entry. 312 // NOTE! We need to remove Incoming Values in the reverse order as done 313 // below, to keep the indices valid for deletion (removeIncomingValues 314 // updates getNumIncomingValues and shifts all values down into the operand 315 // being deleted). 316 for (unsigned i = 0, e = P.getNumIncomingValues() - 1; i != e; ++i) 317 P.removeIncomingValue(e - i, false); 318 319 assert((P.getNumIncomingValues() == 1 && 320 P.getIncomingBlock(PredIndex) == Preheader) && 321 "Should have exactly one value and that's from the preheader!"); 322 } 323 324 // Disconnect the loop body by branching directly to its exit. 325 Builder.SetInsertPoint(Preheader->getTerminator()); 326 Builder.CreateBr(ExitBlock); 327 // Remove the old branch. 328 Preheader->getTerminator()->eraseFromParent(); 329 330 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 331 if (DT) { 332 // Update the dominator tree by informing it about the new edge from the 333 // preheader to the exit. 334 DTU.insertEdge(Preheader, ExitBlock); 335 // Inform the dominator tree about the removed edge. 336 DTU.deleteEdge(Preheader, L->getHeader()); 337 } 338 339 // Given LCSSA form is satisfied, we should not have users of instructions 340 // within the dead loop outside of the loop. However, LCSSA doesn't take 341 // unreachable uses into account. We handle them here. 342 // We could do it after drop all references (in this case all users in the 343 // loop will be already eliminated and we have less work to do but according 344 // to API doc of User::dropAllReferences only valid operation after dropping 345 // references, is deletion. So let's substitute all usages of 346 // instruction from the loop with undef value of corresponding type first. 347 for (auto *Block : L->blocks()) 348 for (Instruction &I : *Block) { 349 auto *Undef = UndefValue::get(I.getType()); 350 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E;) { 351 Use &U = *UI; 352 ++UI; 353 if (auto *Usr = dyn_cast<Instruction>(U.getUser())) 354 if (L->contains(Usr->getParent())) 355 continue; 356 // If we have a DT then we can check that uses outside a loop only in 357 // unreachable block. 358 if (DT) 359 assert(!DT->isReachableFromEntry(U) && 360 "Unexpected user in reachable block"); 361 U.set(Undef); 362 } 363 } 364 365 // Remove the block from the reference counting scheme, so that we can 366 // delete it freely later. 367 for (auto *Block : L->blocks()) 368 Block->dropAllReferences(); 369 370 if (LI) { 371 // Erase the instructions and the blocks without having to worry 372 // about ordering because we already dropped the references. 373 // NOTE: This iteration is safe because erasing the block does not remove 374 // its entry from the loop's block list. We do that in the next section. 375 for (Loop::block_iterator LpI = L->block_begin(), LpE = L->block_end(); 376 LpI != LpE; ++LpI) 377 (*LpI)->eraseFromParent(); 378 379 // Finally, the blocks from loopinfo. This has to happen late because 380 // otherwise our loop iterators won't work. 381 382 SmallPtrSet<BasicBlock *, 8> blocks; 383 blocks.insert(L->block_begin(), L->block_end()); 384 for (BasicBlock *BB : blocks) 385 LI->removeBlock(BB); 386 387 // The last step is to update LoopInfo now that we've eliminated this loop. 388 LI->erase(L); 389 } 390 } 391 392 Optional<unsigned> llvm::getLoopEstimatedTripCount(Loop *L) { 393 // Only support loops with a unique exiting block, and a latch. 394 if (!L->getExitingBlock()) 395 return None; 396 397 // Get the branch weights for the loop's backedge. 398 BranchInst *LatchBR = 399 dyn_cast<BranchInst>(L->getLoopLatch()->getTerminator()); 400 if (!LatchBR || LatchBR->getNumSuccessors() != 2) 401 return None; 402 403 assert((LatchBR->getSuccessor(0) == L->getHeader() || 404 LatchBR->getSuccessor(1) == L->getHeader()) && 405 "At least one edge out of the latch must go to the header"); 406 407 // To estimate the number of times the loop body was executed, we want to 408 // know the number of times the backedge was taken, vs. the number of times 409 // we exited the loop. 410 uint64_t TrueVal, FalseVal; 411 if (!LatchBR->extractProfMetadata(TrueVal, FalseVal)) 412 return None; 413 414 if (!TrueVal || !FalseVal) 415 return 0; 416 417 // Divide the count of the backedge by the count of the edge exiting the loop, 418 // rounding to nearest. 419 if (LatchBR->getSuccessor(0) == L->getHeader()) 420 return (TrueVal + (FalseVal / 2)) / FalseVal; 421 else 422 return (FalseVal + (TrueVal / 2)) / TrueVal; 423 } 424 425 bool llvm::hasIterationCountInvariantInParent(Loop *InnerLoop, 426 ScalarEvolution &SE) { 427 Loop *OuterL = InnerLoop->getParentLoop(); 428 if (!OuterL) 429 return true; 430 431 // Get the backedge taken count for the inner loop 432 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch(); 433 const SCEV *InnerLoopBECountSC = SE.getExitCount(InnerLoop, InnerLoopLatch); 434 if (isa<SCEVCouldNotCompute>(InnerLoopBECountSC) || 435 !InnerLoopBECountSC->getType()->isIntegerTy()) 436 return false; 437 438 // Get whether count is invariant to the outer loop 439 ScalarEvolution::LoopDisposition LD = 440 SE.getLoopDisposition(InnerLoopBECountSC, OuterL); 441 if (LD != ScalarEvolution::LoopInvariant) 442 return false; 443 444 return true; 445 } 446 447 /// Adds a 'fast' flag to floating point operations. 448 static Value *addFastMathFlag(Value *V) { 449 if (isa<FPMathOperator>(V)) { 450 FastMathFlags Flags; 451 Flags.setFast(); 452 cast<Instruction>(V)->setFastMathFlags(Flags); 453 } 454 return V; 455 } 456 457 Value *llvm::createMinMaxOp(IRBuilder<> &Builder, 458 RecurrenceDescriptor::MinMaxRecurrenceKind RK, 459 Value *Left, Value *Right) { 460 CmpInst::Predicate P = CmpInst::ICMP_NE; 461 switch (RK) { 462 default: 463 llvm_unreachable("Unknown min/max recurrence kind"); 464 case RecurrenceDescriptor::MRK_UIntMin: 465 P = CmpInst::ICMP_ULT; 466 break; 467 case RecurrenceDescriptor::MRK_UIntMax: 468 P = CmpInst::ICMP_UGT; 469 break; 470 case RecurrenceDescriptor::MRK_SIntMin: 471 P = CmpInst::ICMP_SLT; 472 break; 473 case RecurrenceDescriptor::MRK_SIntMax: 474 P = CmpInst::ICMP_SGT; 475 break; 476 case RecurrenceDescriptor::MRK_FloatMin: 477 P = CmpInst::FCMP_OLT; 478 break; 479 case RecurrenceDescriptor::MRK_FloatMax: 480 P = CmpInst::FCMP_OGT; 481 break; 482 } 483 484 // We only match FP sequences that are 'fast', so we can unconditionally 485 // set it on any generated instructions. 486 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 487 FastMathFlags FMF; 488 FMF.setFast(); 489 Builder.setFastMathFlags(FMF); 490 491 Value *Cmp; 492 if (RK == RecurrenceDescriptor::MRK_FloatMin || 493 RK == RecurrenceDescriptor::MRK_FloatMax) 494 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp"); 495 else 496 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp"); 497 498 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select"); 499 return Select; 500 } 501 502 // Helper to generate an ordered reduction. 503 Value * 504 llvm::getOrderedReduction(IRBuilder<> &Builder, Value *Acc, Value *Src, 505 unsigned Op, 506 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind, 507 ArrayRef<Value *> RedOps) { 508 unsigned VF = Src->getType()->getVectorNumElements(); 509 510 // Extract and apply reduction ops in ascending order: 511 // e.g. ((((Acc + Scl[0]) + Scl[1]) + Scl[2]) + ) ... + Scl[VF-1] 512 Value *Result = Acc; 513 for (unsigned ExtractIdx = 0; ExtractIdx != VF; ++ExtractIdx) { 514 Value *Ext = 515 Builder.CreateExtractElement(Src, Builder.getInt32(ExtractIdx)); 516 517 if (Op != Instruction::ICmp && Op != Instruction::FCmp) { 518 Result = Builder.CreateBinOp((Instruction::BinaryOps)Op, Result, Ext, 519 "bin.rdx"); 520 } else { 521 assert(MinMaxKind != RecurrenceDescriptor::MRK_Invalid && 522 "Invalid min/max"); 523 Result = createMinMaxOp(Builder, MinMaxKind, Result, Ext); 524 } 525 526 if (!RedOps.empty()) 527 propagateIRFlags(Result, RedOps); 528 } 529 530 return Result; 531 } 532 533 // Helper to generate a log2 shuffle reduction. 534 Value * 535 llvm::getShuffleReduction(IRBuilder<> &Builder, Value *Src, unsigned Op, 536 RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind, 537 ArrayRef<Value *> RedOps) { 538 unsigned VF = Src->getType()->getVectorNumElements(); 539 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles 540 // and vector ops, reducing the set of values being computed by half each 541 // round. 542 assert(isPowerOf2_32(VF) && 543 "Reduction emission only supported for pow2 vectors!"); 544 Value *TmpVec = Src; 545 SmallVector<Constant *, 32> ShuffleMask(VF, nullptr); 546 for (unsigned i = VF; i != 1; i >>= 1) { 547 // Move the upper half of the vector to the lower half. 548 for (unsigned j = 0; j != i / 2; ++j) 549 ShuffleMask[j] = Builder.getInt32(i / 2 + j); 550 551 // Fill the rest of the mask with undef. 552 std::fill(&ShuffleMask[i / 2], ShuffleMask.end(), 553 UndefValue::get(Builder.getInt32Ty())); 554 555 Value *Shuf = Builder.CreateShuffleVector( 556 TmpVec, UndefValue::get(TmpVec->getType()), 557 ConstantVector::get(ShuffleMask), "rdx.shuf"); 558 559 if (Op != Instruction::ICmp && Op != Instruction::FCmp) { 560 // Floating point operations had to be 'fast' to enable the reduction. 561 TmpVec = addFastMathFlag(Builder.CreateBinOp((Instruction::BinaryOps)Op, 562 TmpVec, Shuf, "bin.rdx")); 563 } else { 564 assert(MinMaxKind != RecurrenceDescriptor::MRK_Invalid && 565 "Invalid min/max"); 566 TmpVec = createMinMaxOp(Builder, MinMaxKind, TmpVec, Shuf); 567 } 568 if (!RedOps.empty()) 569 propagateIRFlags(TmpVec, RedOps); 570 } 571 // The result is in the first element of the vector. 572 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0)); 573 } 574 575 /// Create a simple vector reduction specified by an opcode and some 576 /// flags (if generating min/max reductions). 577 Value *llvm::createSimpleTargetReduction( 578 IRBuilder<> &Builder, const TargetTransformInfo *TTI, unsigned Opcode, 579 Value *Src, TargetTransformInfo::ReductionFlags Flags, 580 ArrayRef<Value *> RedOps) { 581 assert(isa<VectorType>(Src->getType()) && "Type must be a vector"); 582 583 Value *ScalarUdf = UndefValue::get(Src->getType()->getVectorElementType()); 584 std::function<Value *()> BuildFunc; 585 using RD = RecurrenceDescriptor; 586 RD::MinMaxRecurrenceKind MinMaxKind = RD::MRK_Invalid; 587 // TODO: Support creating ordered reductions. 588 FastMathFlags FMFFast; 589 FMFFast.setFast(); 590 591 switch (Opcode) { 592 case Instruction::Add: 593 BuildFunc = [&]() { return Builder.CreateAddReduce(Src); }; 594 break; 595 case Instruction::Mul: 596 BuildFunc = [&]() { return Builder.CreateMulReduce(Src); }; 597 break; 598 case Instruction::And: 599 BuildFunc = [&]() { return Builder.CreateAndReduce(Src); }; 600 break; 601 case Instruction::Or: 602 BuildFunc = [&]() { return Builder.CreateOrReduce(Src); }; 603 break; 604 case Instruction::Xor: 605 BuildFunc = [&]() { return Builder.CreateXorReduce(Src); }; 606 break; 607 case Instruction::FAdd: 608 BuildFunc = [&]() { 609 auto Rdx = Builder.CreateFAddReduce(ScalarUdf, Src); 610 cast<CallInst>(Rdx)->setFastMathFlags(FMFFast); 611 return Rdx; 612 }; 613 break; 614 case Instruction::FMul: 615 BuildFunc = [&]() { 616 auto Rdx = Builder.CreateFMulReduce(ScalarUdf, Src); 617 cast<CallInst>(Rdx)->setFastMathFlags(FMFFast); 618 return Rdx; 619 }; 620 break; 621 case Instruction::ICmp: 622 if (Flags.IsMaxOp) { 623 MinMaxKind = Flags.IsSigned ? RD::MRK_SIntMax : RD::MRK_UIntMax; 624 BuildFunc = [&]() { 625 return Builder.CreateIntMaxReduce(Src, Flags.IsSigned); 626 }; 627 } else { 628 MinMaxKind = Flags.IsSigned ? RD::MRK_SIntMin : RD::MRK_UIntMin; 629 BuildFunc = [&]() { 630 return Builder.CreateIntMinReduce(Src, Flags.IsSigned); 631 }; 632 } 633 break; 634 case Instruction::FCmp: 635 if (Flags.IsMaxOp) { 636 MinMaxKind = RD::MRK_FloatMax; 637 BuildFunc = [&]() { return Builder.CreateFPMaxReduce(Src, Flags.NoNaN); }; 638 } else { 639 MinMaxKind = RD::MRK_FloatMin; 640 BuildFunc = [&]() { return Builder.CreateFPMinReduce(Src, Flags.NoNaN); }; 641 } 642 break; 643 default: 644 llvm_unreachable("Unhandled opcode"); 645 break; 646 } 647 if (TTI->useReductionIntrinsic(Opcode, Src->getType(), Flags)) 648 return BuildFunc(); 649 return getShuffleReduction(Builder, Src, Opcode, MinMaxKind, RedOps); 650 } 651 652 /// Create a vector reduction using a given recurrence descriptor. 653 Value *llvm::createTargetReduction(IRBuilder<> &B, 654 const TargetTransformInfo *TTI, 655 RecurrenceDescriptor &Desc, Value *Src, 656 bool NoNaN) { 657 // TODO: Support in-order reductions based on the recurrence descriptor. 658 using RD = RecurrenceDescriptor; 659 RD::RecurrenceKind RecKind = Desc.getRecurrenceKind(); 660 TargetTransformInfo::ReductionFlags Flags; 661 Flags.NoNaN = NoNaN; 662 switch (RecKind) { 663 case RD::RK_FloatAdd: 664 return createSimpleTargetReduction(B, TTI, Instruction::FAdd, Src, Flags); 665 case RD::RK_FloatMult: 666 return createSimpleTargetReduction(B, TTI, Instruction::FMul, Src, Flags); 667 case RD::RK_IntegerAdd: 668 return createSimpleTargetReduction(B, TTI, Instruction::Add, Src, Flags); 669 case RD::RK_IntegerMult: 670 return createSimpleTargetReduction(B, TTI, Instruction::Mul, Src, Flags); 671 case RD::RK_IntegerAnd: 672 return createSimpleTargetReduction(B, TTI, Instruction::And, Src, Flags); 673 case RD::RK_IntegerOr: 674 return createSimpleTargetReduction(B, TTI, Instruction::Or, Src, Flags); 675 case RD::RK_IntegerXor: 676 return createSimpleTargetReduction(B, TTI, Instruction::Xor, Src, Flags); 677 case RD::RK_IntegerMinMax: { 678 RD::MinMaxRecurrenceKind MMKind = Desc.getMinMaxRecurrenceKind(); 679 Flags.IsMaxOp = (MMKind == RD::MRK_SIntMax || MMKind == RD::MRK_UIntMax); 680 Flags.IsSigned = (MMKind == RD::MRK_SIntMax || MMKind == RD::MRK_SIntMin); 681 return createSimpleTargetReduction(B, TTI, Instruction::ICmp, Src, Flags); 682 } 683 case RD::RK_FloatMinMax: { 684 Flags.IsMaxOp = Desc.getMinMaxRecurrenceKind() == RD::MRK_FloatMax; 685 return createSimpleTargetReduction(B, TTI, Instruction::FCmp, Src, Flags); 686 } 687 default: 688 llvm_unreachable("Unhandled RecKind"); 689 } 690 } 691 692 void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue) { 693 auto *VecOp = dyn_cast<Instruction>(I); 694 if (!VecOp) 695 return; 696 auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(VL[0]) 697 : dyn_cast<Instruction>(OpValue); 698 if (!Intersection) 699 return; 700 const unsigned Opcode = Intersection->getOpcode(); 701 VecOp->copyIRFlags(Intersection); 702 for (auto *V : VL) { 703 auto *Instr = dyn_cast<Instruction>(V); 704 if (!Instr) 705 continue; 706 if (OpValue == nullptr || Opcode == Instr->getOpcode()) 707 VecOp->andIRFlags(V); 708 } 709 } 710