1 //===- ScopHelper.cpp - Some Helper Functions for Scop. ------------------===// 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 // Small functions that help with Scop and LLVM-IR. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "polly/Support/ScopHelper.h" 15 #include "polly/Options.h" 16 #include "polly/ScopInfo.h" 17 #include "polly/Support/SCEVValidator.h" 18 #include "llvm/Analysis/LoopInfo.h" 19 #include "llvm/Analysis/RegionInfo.h" 20 #include "llvm/Analysis/ScalarEvolution.h" 21 #include "llvm/Analysis/ScalarEvolutionExpander.h" 22 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 23 #include "llvm/IR/CFG.h" 24 #include "llvm/IR/IntrinsicInst.h" 25 #include "llvm/Support/Debug.h" 26 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 27 28 using namespace llvm; 29 using namespace polly; 30 31 #define DEBUG_TYPE "polly-scop-helper" 32 33 bool polly::hasInvokeEdge(const PHINode *PN) { 34 for (unsigned i = 0, e = PN->getNumIncomingValues(); i < e; ++i) 35 if (InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i))) 36 if (II->getParent() == PN->getIncomingBlock(i)) 37 return true; 38 39 return false; 40 } 41 42 // Ensures that there is just one predecessor to the entry node from outside the 43 // region. 44 // The identity of the region entry node is preserved. 45 static void simplifyRegionEntry(Region *R, DominatorTree *DT, LoopInfo *LI, 46 RegionInfo *RI) { 47 BasicBlock *EnteringBB = R->getEnteringBlock(); 48 BasicBlock *Entry = R->getEntry(); 49 50 // Before (one of): 51 // 52 // \ / // 53 // EnteringBB // 54 // | \------> // 55 // \ / | // 56 // Entry <--\ Entry <--\ // 57 // / \ / / \ / // 58 // .... .... // 59 60 // Create single entry edge if the region has multiple entry edges. 61 if (!EnteringBB) { 62 SmallVector<BasicBlock *, 4> Preds; 63 for (BasicBlock *P : predecessors(Entry)) 64 if (!R->contains(P)) 65 Preds.push_back(P); 66 67 BasicBlock *NewEntering = 68 SplitBlockPredecessors(Entry, Preds, ".region_entering", DT, LI); 69 70 if (RI) { 71 // The exit block of predecessing regions must be changed to NewEntering 72 for (BasicBlock *ExitPred : predecessors(NewEntering)) { 73 Region *RegionOfPred = RI->getRegionFor(ExitPred); 74 if (RegionOfPred->getExit() != Entry) 75 continue; 76 77 while (!RegionOfPred->isTopLevelRegion() && 78 RegionOfPred->getExit() == Entry) { 79 RegionOfPred->replaceExit(NewEntering); 80 RegionOfPred = RegionOfPred->getParent(); 81 } 82 } 83 84 // Make all ancestors use EnteringBB as entry; there might be edges to it 85 Region *AncestorR = R->getParent(); 86 RI->setRegionFor(NewEntering, AncestorR); 87 while (!AncestorR->isTopLevelRegion() && AncestorR->getEntry() == Entry) { 88 AncestorR->replaceEntry(NewEntering); 89 AncestorR = AncestorR->getParent(); 90 } 91 } 92 93 EnteringBB = NewEntering; 94 } 95 assert(R->getEnteringBlock() == EnteringBB); 96 97 // After: 98 // 99 // \ / // 100 // EnteringBB // 101 // | // 102 // | // 103 // Entry <--\ // 104 // / \ / // 105 // .... // 106 } 107 108 // Ensure that the region has a single block that branches to the exit node. 109 static void simplifyRegionExit(Region *R, DominatorTree *DT, LoopInfo *LI, 110 RegionInfo *RI) { 111 BasicBlock *ExitBB = R->getExit(); 112 BasicBlock *ExitingBB = R->getExitingBlock(); 113 114 // Before: 115 // 116 // (Region) ______/ // 117 // \ | / // 118 // ExitBB // 119 // / \ // 120 121 if (!ExitingBB) { 122 SmallVector<BasicBlock *, 4> Preds; 123 for (BasicBlock *P : predecessors(ExitBB)) 124 if (R->contains(P)) 125 Preds.push_back(P); 126 127 // Preds[0] Preds[1] otherBB // 128 // \ | ________/ // 129 // \ | / // 130 // BB // 131 ExitingBB = 132 SplitBlockPredecessors(ExitBB, Preds, ".region_exiting", DT, LI); 133 // Preds[0] Preds[1] otherBB // 134 // \ / / // 135 // BB.region_exiting / // 136 // \ / // 137 // BB // 138 139 if (RI) 140 RI->setRegionFor(ExitingBB, R); 141 142 // Change the exit of nested regions, but not the region itself, 143 R->replaceExitRecursive(ExitingBB); 144 R->replaceExit(ExitBB); 145 } 146 assert(ExitingBB == R->getExitingBlock()); 147 148 // After: 149 // 150 // \ / // 151 // ExitingBB _____/ // 152 // \ / // 153 // ExitBB // 154 // / \ // 155 } 156 157 void polly::simplifyRegion(Region *R, DominatorTree *DT, LoopInfo *LI, 158 RegionInfo *RI) { 159 assert(R && !R->isTopLevelRegion()); 160 assert(!RI || RI == R->getRegionInfo()); 161 assert((!RI || DT) && 162 "RegionInfo requires DominatorTree to be updated as well"); 163 164 simplifyRegionEntry(R, DT, LI, RI); 165 simplifyRegionExit(R, DT, LI, RI); 166 assert(R->isSimple()); 167 } 168 169 // Split the block into two successive blocks. 170 // 171 // Like llvm::SplitBlock, but also preserves RegionInfo 172 static BasicBlock *splitBlock(BasicBlock *Old, Instruction *SplitPt, 173 DominatorTree *DT, llvm::LoopInfo *LI, 174 RegionInfo *RI) { 175 assert(Old && SplitPt); 176 177 // Before: 178 // 179 // \ / // 180 // Old // 181 // / \ // 182 183 BasicBlock *NewBlock = llvm::SplitBlock(Old, SplitPt, DT, LI); 184 185 if (RI) { 186 Region *R = RI->getRegionFor(Old); 187 RI->setRegionFor(NewBlock, R); 188 } 189 190 // After: 191 // 192 // \ / // 193 // Old // 194 // | // 195 // NewBlock // 196 // / \ // 197 198 return NewBlock; 199 } 200 201 void polly::splitEntryBlockForAlloca(BasicBlock *EntryBlock, Pass *P) { 202 // Find first non-alloca instruction. Every basic block has a non-alloc 203 // instruction, as every well formed basic block has a terminator. 204 BasicBlock::iterator I = EntryBlock->begin(); 205 while (isa<AllocaInst>(I)) 206 ++I; 207 208 auto *DTWP = P->getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 209 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr; 210 auto *LIWP = P->getAnalysisIfAvailable<LoopInfoWrapperPass>(); 211 auto *LI = LIWP ? &LIWP->getLoopInfo() : nullptr; 212 RegionInfoPass *RIP = P->getAnalysisIfAvailable<RegionInfoPass>(); 213 RegionInfo *RI = RIP ? &RIP->getRegionInfo() : nullptr; 214 215 // splitBlock updates DT, LI and RI. 216 splitBlock(EntryBlock, &*I, DT, LI, RI); 217 } 218 219 /// The SCEVExpander will __not__ generate any code for an existing SDiv/SRem 220 /// instruction but just use it, if it is referenced as a SCEVUnknown. We want 221 /// however to generate new code if the instruction is in the analyzed region 222 /// and we generate code outside/in front of that region. Hence, we generate the 223 /// code for the SDiv/SRem operands in front of the analyzed region and then 224 /// create a new SDiv/SRem operation there too. 225 struct ScopExpander : SCEVVisitor<ScopExpander, const SCEV *> { 226 friend struct SCEVVisitor<ScopExpander, const SCEV *>; 227 228 explicit ScopExpander(const Region &R, ScalarEvolution &SE, 229 const DataLayout &DL, const char *Name, ValueMapT *VMap) 230 : Expander(SCEVExpander(SE, DL, Name)), SE(SE), Name(Name), R(R), 231 VMap(VMap) {} 232 233 Value *expandCodeFor(const SCEV *E, Type *Ty, Instruction *I) { 234 // If we generate code in the region we will immediately fall back to the 235 // SCEVExpander, otherwise we will stop at all unknowns in the SCEV and if 236 // needed replace them by copies computed in the entering block. 237 if (!R.contains(I)) 238 E = visit(E); 239 return Expander.expandCodeFor(E, Ty, I); 240 } 241 242 private: 243 SCEVExpander Expander; 244 ScalarEvolution &SE; 245 const char *Name; 246 const Region &R; 247 ValueMapT *VMap; 248 249 const SCEV *visitUnknown(const SCEVUnknown *E) { 250 251 // If a value mapping was given try if the underlying value is remapped. 252 if (VMap) 253 if (Value *NewVal = VMap->lookup(E->getValue())) 254 if (NewVal != E->getValue()) 255 return visit(SE.getSCEV(NewVal)); 256 257 Instruction *Inst = dyn_cast<Instruction>(E->getValue()); 258 if (!Inst || (Inst->getOpcode() != Instruction::SRem && 259 Inst->getOpcode() != Instruction::SDiv)) 260 return E; 261 262 if (!R.contains(Inst)) 263 return E; 264 265 Instruction *StartIP = R.getEnteringBlock()->getTerminator(); 266 267 const SCEV *LHSScev = visit(SE.getSCEV(Inst->getOperand(0))); 268 const SCEV *RHSScev = visit(SE.getSCEV(Inst->getOperand(1))); 269 270 Value *LHS = Expander.expandCodeFor(LHSScev, E->getType(), StartIP); 271 Value *RHS = Expander.expandCodeFor(RHSScev, E->getType(), StartIP); 272 273 Inst = BinaryOperator::Create((Instruction::BinaryOps)Inst->getOpcode(), 274 LHS, RHS, Inst->getName() + Name, StartIP); 275 return SE.getSCEV(Inst); 276 } 277 278 /// The following functions will just traverse the SCEV and rebuild it with 279 /// the new operands returned by the traversal. 280 /// 281 ///{ 282 const SCEV *visitConstant(const SCEVConstant *E) { return E; } 283 const SCEV *visitTruncateExpr(const SCEVTruncateExpr *E) { 284 return SE.getTruncateExpr(visit(E->getOperand()), E->getType()); 285 } 286 const SCEV *visitZeroExtendExpr(const SCEVZeroExtendExpr *E) { 287 return SE.getZeroExtendExpr(visit(E->getOperand()), E->getType()); 288 } 289 const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *E) { 290 return SE.getSignExtendExpr(visit(E->getOperand()), E->getType()); 291 } 292 const SCEV *visitUDivExpr(const SCEVUDivExpr *E) { 293 return SE.getUDivExpr(visit(E->getLHS()), visit(E->getRHS())); 294 } 295 const SCEV *visitAddExpr(const SCEVAddExpr *E) { 296 SmallVector<const SCEV *, 4> NewOps; 297 for (const SCEV *Op : E->operands()) 298 NewOps.push_back(visit(Op)); 299 return SE.getAddExpr(NewOps); 300 } 301 const SCEV *visitMulExpr(const SCEVMulExpr *E) { 302 SmallVector<const SCEV *, 4> NewOps; 303 for (const SCEV *Op : E->operands()) 304 NewOps.push_back(visit(Op)); 305 return SE.getMulExpr(NewOps); 306 } 307 const SCEV *visitUMaxExpr(const SCEVUMaxExpr *E) { 308 SmallVector<const SCEV *, 4> NewOps; 309 for (const SCEV *Op : E->operands()) 310 NewOps.push_back(visit(Op)); 311 return SE.getUMaxExpr(NewOps); 312 } 313 const SCEV *visitSMaxExpr(const SCEVSMaxExpr *E) { 314 SmallVector<const SCEV *, 4> NewOps; 315 for (const SCEV *Op : E->operands()) 316 NewOps.push_back(visit(Op)); 317 return SE.getSMaxExpr(NewOps); 318 } 319 const SCEV *visitAddRecExpr(const SCEVAddRecExpr *E) { 320 SmallVector<const SCEV *, 4> NewOps; 321 for (const SCEV *Op : E->operands()) 322 NewOps.push_back(visit(Op)); 323 return SE.getAddRecExpr(NewOps, E->getLoop(), E->getNoWrapFlags()); 324 } 325 ///} 326 }; 327 328 Value *polly::expandCodeFor(Scop &S, ScalarEvolution &SE, const DataLayout &DL, 329 const char *Name, const SCEV *E, Type *Ty, 330 Instruction *IP, ValueMapT *VMap) { 331 ScopExpander Expander(S.getRegion(), SE, DL, Name, VMap); 332 return Expander.expandCodeFor(E, Ty, IP); 333 } 334 335 bool polly::isErrorBlock(BasicBlock &BB, const Region &R, LoopInfo &LI, 336 const DominatorTree &DT) { 337 338 if (isa<UnreachableInst>(BB.getTerminator())) 339 return true; 340 341 if (LI.isLoopHeader(&BB)) 342 return false; 343 344 // Basic blocks that are always executed are not considered error blocks, 345 // as their execution can not be a rare event. 346 bool DominatesAllPredecessors = true; 347 for (auto Pred : predecessors(R.getExit())) 348 if (R.contains(Pred) && !DT.dominates(&BB, Pred)) 349 DominatesAllPredecessors = false; 350 351 if (DominatesAllPredecessors) 352 return false; 353 354 // FIXME: This is a simple heuristic to determine if the load is executed 355 // in a conditional. However, we actually would need the control 356 // condition, i.e., the post dominance frontier. Alternatively we 357 // could walk up the dominance tree until we find a block that is 358 // not post dominated by the load and check if it is a conditional 359 // or a loop header. 360 auto *DTNode = DT.getNode(&BB); 361 auto *IDomBB = DTNode->getIDom()->getBlock(); 362 if (LI.isLoopHeader(IDomBB)) 363 return false; 364 365 for (Instruction &Inst : BB) 366 if (CallInst *CI = dyn_cast<CallInst>(&Inst)) { 367 if (isIgnoredIntrinsic(CI)) 368 return false; 369 370 if (!CI->doesNotAccessMemory()) 371 return true; 372 if (CI->doesNotReturn()) 373 return true; 374 } 375 376 return false; 377 } 378 379 Value *polly::getConditionFromTerminator(TerminatorInst *TI) { 380 if (BranchInst *BR = dyn_cast<BranchInst>(TI)) { 381 if (BR->isUnconditional()) 382 return ConstantInt::getTrue(Type::getInt1Ty(TI->getContext())); 383 384 return BR->getCondition(); 385 } 386 387 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) 388 return SI->getCondition(); 389 390 return nullptr; 391 } 392 393 bool polly::isHoistableLoad(LoadInst *LInst, Region &R, LoopInfo &LI, 394 ScalarEvolution &SE) { 395 Loop *L = LI.getLoopFor(LInst->getParent()); 396 const SCEV *PtrSCEV = SE.getSCEVAtScope(LInst->getPointerOperand(), L); 397 while (L && R.contains(L)) { 398 if (!SE.isLoopInvariant(PtrSCEV, L)) 399 return false; 400 L = L->getParentLoop(); 401 } 402 403 return true; 404 } 405 406 bool polly::isIgnoredIntrinsic(const Value *V) { 407 if (auto *IT = dyn_cast<IntrinsicInst>(V)) { 408 switch (IT->getIntrinsicID()) { 409 // Lifetime markers are supported/ignored. 410 case llvm::Intrinsic::lifetime_start: 411 case llvm::Intrinsic::lifetime_end: 412 // Invariant markers are supported/ignored. 413 case llvm::Intrinsic::invariant_start: 414 case llvm::Intrinsic::invariant_end: 415 // Some misc annotations are supported/ignored. 416 case llvm::Intrinsic::var_annotation: 417 case llvm::Intrinsic::ptr_annotation: 418 case llvm::Intrinsic::annotation: 419 case llvm::Intrinsic::donothing: 420 case llvm::Intrinsic::assume: 421 case llvm::Intrinsic::expect: 422 // Some debug info intrisics are supported/ignored. 423 case llvm::Intrinsic::dbg_value: 424 case llvm::Intrinsic::dbg_declare: 425 return true; 426 default: 427 break; 428 } 429 } 430 return false; 431 } 432 433 bool polly::canSynthesize(const Value *V, const llvm::LoopInfo *LI, 434 ScalarEvolution *SE, const Region *R) { 435 if (!V || !SE->isSCEVable(V->getType())) 436 return false; 437 438 if (const SCEV *Scev = SE->getSCEV(const_cast<Value *>(V))) 439 if (!isa<SCEVCouldNotCompute>(Scev)) 440 if (!hasScalarDepsInsideRegion(Scev, R)) 441 return true; 442 443 return false; 444 } 445