1 //===-- PredicateInfo.cpp - PredicateInfo Builder--------------------===// 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 implements the PredicateInfo class. 11 // 12 //===----------------------------------------------------------------===// 13 14 #include "llvm/Transforms/Utils/PredicateInfo.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/DepthFirstIterator.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallPtrSet.h" 19 #include "llvm/ADT/Statistic.h" 20 #include "llvm/ADT/StringExtras.h" 21 #include "llvm/Analysis/AssumptionCache.h" 22 #include "llvm/Analysis/CFG.h" 23 #include "llvm/IR/AssemblyAnnotationWriter.h" 24 #include "llvm/IR/DataLayout.h" 25 #include "llvm/IR/Dominators.h" 26 #include "llvm/IR/GlobalVariable.h" 27 #include "llvm/IR/IRBuilder.h" 28 #include "llvm/IR/InstIterator.h" 29 #include "llvm/IR/IntrinsicInst.h" 30 #include "llvm/IR/LLVMContext.h" 31 #include "llvm/IR/Metadata.h" 32 #include "llvm/IR/Module.h" 33 #include "llvm/IR/PatternMatch.h" 34 #include "llvm/Support/Debug.h" 35 #include "llvm/Support/DebugCounter.h" 36 #include "llvm/Support/FormattedStream.h" 37 #include "llvm/Transforms/Utils.h" 38 #include "llvm/Transforms/Utils/OrderedInstructions.h" 39 #include <algorithm> 40 #define DEBUG_TYPE "predicateinfo" 41 using namespace llvm; 42 using namespace PatternMatch; 43 using namespace llvm::PredicateInfoClasses; 44 45 INITIALIZE_PASS_BEGIN(PredicateInfoPrinterLegacyPass, "print-predicateinfo", 46 "PredicateInfo Printer", false, false) 47 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 48 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 49 INITIALIZE_PASS_END(PredicateInfoPrinterLegacyPass, "print-predicateinfo", 50 "PredicateInfo Printer", false, false) 51 static cl::opt<bool> VerifyPredicateInfo( 52 "verify-predicateinfo", cl::init(false), cl::Hidden, 53 cl::desc("Verify PredicateInfo in legacy printer pass.")); 54 DEBUG_COUNTER(RenameCounter, "predicateinfo-rename", 55 "Controls which variables are renamed with predicateinfo"); 56 57 namespace { 58 // Given a predicate info that is a type of branching terminator, get the 59 // branching block. 60 const BasicBlock *getBranchBlock(const PredicateBase *PB) { 61 assert(isa<PredicateWithEdge>(PB) && 62 "Only branches and switches should have PHIOnly defs that " 63 "require branch blocks."); 64 return cast<PredicateWithEdge>(PB)->From; 65 } 66 67 // Given a predicate info that is a type of branching terminator, get the 68 // branching terminator. 69 static Instruction *getBranchTerminator(const PredicateBase *PB) { 70 assert(isa<PredicateWithEdge>(PB) && 71 "Not a predicate info type we know how to get a terminator from."); 72 return cast<PredicateWithEdge>(PB)->From->getTerminator(); 73 } 74 75 // Given a predicate info that is a type of branching terminator, get the 76 // edge this predicate info represents 77 const std::pair<BasicBlock *, BasicBlock *> 78 getBlockEdge(const PredicateBase *PB) { 79 assert(isa<PredicateWithEdge>(PB) && 80 "Not a predicate info type we know how to get an edge from."); 81 const auto *PEdge = cast<PredicateWithEdge>(PB); 82 return std::make_pair(PEdge->From, PEdge->To); 83 } 84 } 85 86 namespace llvm { 87 namespace PredicateInfoClasses { 88 enum LocalNum { 89 // Operations that must appear first in the block. 90 LN_First, 91 // Operations that are somewhere in the middle of the block, and are sorted on 92 // demand. 93 LN_Middle, 94 // Operations that must appear last in a block, like successor phi node uses. 95 LN_Last 96 }; 97 98 // Associate global and local DFS info with defs and uses, so we can sort them 99 // into a global domination ordering. 100 struct ValueDFS { 101 int DFSIn = 0; 102 int DFSOut = 0; 103 unsigned int LocalNum = LN_Middle; 104 // Only one of Def or Use will be set. 105 Value *Def = nullptr; 106 Use *U = nullptr; 107 // Neither PInfo nor EdgeOnly participate in the ordering 108 PredicateBase *PInfo = nullptr; 109 bool EdgeOnly = false; 110 }; 111 112 // Perform a strict weak ordering on instructions and arguments. 113 static bool valueComesBefore(OrderedInstructions &OI, const Value *A, 114 const Value *B) { 115 auto *ArgA = dyn_cast_or_null<Argument>(A); 116 auto *ArgB = dyn_cast_or_null<Argument>(B); 117 if (ArgA && !ArgB) 118 return true; 119 if (ArgB && !ArgA) 120 return false; 121 if (ArgA && ArgB) 122 return ArgA->getArgNo() < ArgB->getArgNo(); 123 return OI.dfsBefore(cast<Instruction>(A), cast<Instruction>(B)); 124 } 125 126 // This compares ValueDFS structures, creating OrderedBasicBlocks where 127 // necessary to compare uses/defs in the same block. Doing so allows us to walk 128 // the minimum number of instructions necessary to compute our def/use ordering. 129 struct ValueDFS_Compare { 130 OrderedInstructions &OI; 131 ValueDFS_Compare(OrderedInstructions &OI) : OI(OI) {} 132 133 bool operator()(const ValueDFS &A, const ValueDFS &B) const { 134 if (&A == &B) 135 return false; 136 // The only case we can't directly compare them is when they in the same 137 // block, and both have localnum == middle. In that case, we have to use 138 // comesbefore to see what the real ordering is, because they are in the 139 // same basic block. 140 141 bool SameBlock = std::tie(A.DFSIn, A.DFSOut) == std::tie(B.DFSIn, B.DFSOut); 142 143 // We want to put the def that will get used for a given set of phi uses, 144 // before those phi uses. 145 // So we sort by edge, then by def. 146 // Note that only phi nodes uses and defs can come last. 147 if (SameBlock && A.LocalNum == LN_Last && B.LocalNum == LN_Last) 148 return comparePHIRelated(A, B); 149 150 if (!SameBlock || A.LocalNum != LN_Middle || B.LocalNum != LN_Middle) 151 return std::tie(A.DFSIn, A.DFSOut, A.LocalNum, A.Def, A.U) < 152 std::tie(B.DFSIn, B.DFSOut, B.LocalNum, B.Def, B.U); 153 return localComesBefore(A, B); 154 } 155 156 // For a phi use, or a non-materialized def, return the edge it represents. 157 const std::pair<BasicBlock *, BasicBlock *> 158 getBlockEdge(const ValueDFS &VD) const { 159 if (!VD.Def && VD.U) { 160 auto *PHI = cast<PHINode>(VD.U->getUser()); 161 return std::make_pair(PHI->getIncomingBlock(*VD.U), PHI->getParent()); 162 } 163 // This is really a non-materialized def. 164 return ::getBlockEdge(VD.PInfo); 165 } 166 167 // For two phi related values, return the ordering. 168 bool comparePHIRelated(const ValueDFS &A, const ValueDFS &B) const { 169 auto &ABlockEdge = getBlockEdge(A); 170 auto &BBlockEdge = getBlockEdge(B); 171 // Now sort by block edge and then defs before uses. 172 return std::tie(ABlockEdge, A.Def, A.U) < std::tie(BBlockEdge, B.Def, B.U); 173 } 174 175 // Get the definition of an instruction that occurs in the middle of a block. 176 Value *getMiddleDef(const ValueDFS &VD) const { 177 if (VD.Def) 178 return VD.Def; 179 // It's possible for the defs and uses to be null. For branches, the local 180 // numbering will say the placed predicaeinfos should go first (IE 181 // LN_beginning), so we won't be in this function. For assumes, we will end 182 // up here, beause we need to order the def we will place relative to the 183 // assume. So for the purpose of ordering, we pretend the def is the assume 184 // because that is where we will insert the info. 185 if (!VD.U) { 186 assert(VD.PInfo && 187 "No def, no use, and no predicateinfo should not occur"); 188 assert(isa<PredicateAssume>(VD.PInfo) && 189 "Middle of block should only occur for assumes"); 190 return cast<PredicateAssume>(VD.PInfo)->AssumeInst; 191 } 192 return nullptr; 193 } 194 195 // Return either the Def, if it's not null, or the user of the Use, if the def 196 // is null. 197 const Instruction *getDefOrUser(const Value *Def, const Use *U) const { 198 if (Def) 199 return cast<Instruction>(Def); 200 return cast<Instruction>(U->getUser()); 201 } 202 203 // This performs the necessary local basic block ordering checks to tell 204 // whether A comes before B, where both are in the same basic block. 205 bool localComesBefore(const ValueDFS &A, const ValueDFS &B) const { 206 auto *ADef = getMiddleDef(A); 207 auto *BDef = getMiddleDef(B); 208 209 // See if we have real values or uses. If we have real values, we are 210 // guaranteed they are instructions or arguments. No matter what, we are 211 // guaranteed they are in the same block if they are instructions. 212 auto *ArgA = dyn_cast_or_null<Argument>(ADef); 213 auto *ArgB = dyn_cast_or_null<Argument>(BDef); 214 215 if (ArgA || ArgB) 216 return valueComesBefore(OI, ArgA, ArgB); 217 218 auto *AInst = getDefOrUser(ADef, A.U); 219 auto *BInst = getDefOrUser(BDef, B.U); 220 return valueComesBefore(OI, AInst, BInst); 221 } 222 }; 223 224 } // namespace PredicateInfoClasses 225 226 bool PredicateInfo::stackIsInScope(const ValueDFSStack &Stack, 227 const ValueDFS &VDUse) const { 228 if (Stack.empty()) 229 return false; 230 // If it's a phi only use, make sure it's for this phi node edge, and that the 231 // use is in a phi node. If it's anything else, and the top of the stack is 232 // EdgeOnly, we need to pop the stack. We deliberately sort phi uses next to 233 // the defs they must go with so that we can know it's time to pop the stack 234 // when we hit the end of the phi uses for a given def. 235 if (Stack.back().EdgeOnly) { 236 if (!VDUse.U) 237 return false; 238 auto *PHI = dyn_cast<PHINode>(VDUse.U->getUser()); 239 if (!PHI) 240 return false; 241 // Check edge 242 BasicBlock *EdgePred = PHI->getIncomingBlock(*VDUse.U); 243 if (EdgePred != getBranchBlock(Stack.back().PInfo)) 244 return false; 245 246 // Use dominates, which knows how to handle edge dominance. 247 return DT.dominates(getBlockEdge(Stack.back().PInfo), *VDUse.U); 248 } 249 250 return (VDUse.DFSIn >= Stack.back().DFSIn && 251 VDUse.DFSOut <= Stack.back().DFSOut); 252 } 253 254 void PredicateInfo::popStackUntilDFSScope(ValueDFSStack &Stack, 255 const ValueDFS &VD) { 256 while (!Stack.empty() && !stackIsInScope(Stack, VD)) 257 Stack.pop_back(); 258 } 259 260 // Convert the uses of Op into a vector of uses, associating global and local 261 // DFS info with each one. 262 void PredicateInfo::convertUsesToDFSOrdered( 263 Value *Op, SmallVectorImpl<ValueDFS> &DFSOrderedSet) { 264 for (auto &U : Op->uses()) { 265 if (auto *I = dyn_cast<Instruction>(U.getUser())) { 266 ValueDFS VD; 267 // Put the phi node uses in the incoming block. 268 BasicBlock *IBlock; 269 if (auto *PN = dyn_cast<PHINode>(I)) { 270 IBlock = PN->getIncomingBlock(U); 271 // Make phi node users appear last in the incoming block 272 // they are from. 273 VD.LocalNum = LN_Last; 274 } else { 275 // If it's not a phi node use, it is somewhere in the middle of the 276 // block. 277 IBlock = I->getParent(); 278 VD.LocalNum = LN_Middle; 279 } 280 DomTreeNode *DomNode = DT.getNode(IBlock); 281 // It's possible our use is in an unreachable block. Skip it if so. 282 if (!DomNode) 283 continue; 284 VD.DFSIn = DomNode->getDFSNumIn(); 285 VD.DFSOut = DomNode->getDFSNumOut(); 286 VD.U = &U; 287 DFSOrderedSet.push_back(VD); 288 } 289 } 290 } 291 292 // Collect relevant operations from Comparison that we may want to insert copies 293 // for. 294 void collectCmpOps(CmpInst *Comparison, SmallVectorImpl<Value *> &CmpOperands) { 295 auto *Op0 = Comparison->getOperand(0); 296 auto *Op1 = Comparison->getOperand(1); 297 if (Op0 == Op1) 298 return; 299 CmpOperands.push_back(Comparison); 300 // Only want real values, not constants. Additionally, operands with one use 301 // are only being used in the comparison, which means they will not be useful 302 // for us to consider for predicateinfo. 303 // 304 if ((isa<Instruction>(Op0) || isa<Argument>(Op0)) && !Op0->hasOneUse()) 305 CmpOperands.push_back(Op0); 306 if ((isa<Instruction>(Op1) || isa<Argument>(Op1)) && !Op1->hasOneUse()) 307 CmpOperands.push_back(Op1); 308 } 309 310 // Add Op, PB to the list of value infos for Op, and mark Op to be renamed. 311 void PredicateInfo::addInfoFor(SmallPtrSetImpl<Value *> &OpsToRename, Value *Op, 312 PredicateBase *PB) { 313 OpsToRename.insert(Op); 314 auto &OperandInfo = getOrCreateValueInfo(Op); 315 AllInfos.push_back(PB); 316 OperandInfo.Infos.push_back(PB); 317 } 318 319 // Process an assume instruction and place relevant operations we want to rename 320 // into OpsToRename. 321 void PredicateInfo::processAssume(IntrinsicInst *II, BasicBlock *AssumeBB, 322 SmallPtrSetImpl<Value *> &OpsToRename) { 323 // See if we have a comparison we support 324 SmallVector<Value *, 8> CmpOperands; 325 SmallVector<Value *, 2> ConditionsToProcess; 326 CmpInst::Predicate Pred; 327 Value *Operand = II->getOperand(0); 328 if (m_c_And(m_Cmp(Pred, m_Value(), m_Value()), 329 m_Cmp(Pred, m_Value(), m_Value())) 330 .match(II->getOperand(0))) { 331 ConditionsToProcess.push_back(cast<BinaryOperator>(Operand)->getOperand(0)); 332 ConditionsToProcess.push_back(cast<BinaryOperator>(Operand)->getOperand(1)); 333 ConditionsToProcess.push_back(Operand); 334 } else if (isa<CmpInst>(Operand)) { 335 336 ConditionsToProcess.push_back(Operand); 337 } 338 for (auto Cond : ConditionsToProcess) { 339 if (auto *Cmp = dyn_cast<CmpInst>(Cond)) { 340 collectCmpOps(Cmp, CmpOperands); 341 // Now add our copy infos for our operands 342 for (auto *Op : CmpOperands) { 343 auto *PA = new PredicateAssume(Op, II, Cmp); 344 addInfoFor(OpsToRename, Op, PA); 345 } 346 CmpOperands.clear(); 347 } else if (auto *BinOp = dyn_cast<BinaryOperator>(Cond)) { 348 // Otherwise, it should be an AND. 349 assert(BinOp->getOpcode() == Instruction::And && 350 "Should have been an AND"); 351 auto *PA = new PredicateAssume(BinOp, II, BinOp); 352 addInfoFor(OpsToRename, BinOp, PA); 353 } else { 354 llvm_unreachable("Unknown type of condition"); 355 } 356 } 357 } 358 359 // Process a block terminating branch, and place relevant operations to be 360 // renamed into OpsToRename. 361 void PredicateInfo::processBranch(BranchInst *BI, BasicBlock *BranchBB, 362 SmallPtrSetImpl<Value *> &OpsToRename) { 363 BasicBlock *FirstBB = BI->getSuccessor(0); 364 BasicBlock *SecondBB = BI->getSuccessor(1); 365 SmallVector<BasicBlock *, 2> SuccsToProcess; 366 SuccsToProcess.push_back(FirstBB); 367 SuccsToProcess.push_back(SecondBB); 368 SmallVector<Value *, 2> ConditionsToProcess; 369 370 auto InsertHelper = [&](Value *Op, bool isAnd, bool isOr, Value *Cond) { 371 for (auto *Succ : SuccsToProcess) { 372 // Don't try to insert on a self-edge. This is mainly because we will 373 // eliminate during renaming anyway. 374 if (Succ == BranchBB) 375 continue; 376 bool TakenEdge = (Succ == FirstBB); 377 // For and, only insert on the true edge 378 // For or, only insert on the false edge 379 if ((isAnd && !TakenEdge) || (isOr && TakenEdge)) 380 continue; 381 PredicateBase *PB = 382 new PredicateBranch(Op, BranchBB, Succ, Cond, TakenEdge); 383 addInfoFor(OpsToRename, Op, PB); 384 if (!Succ->getSinglePredecessor()) 385 EdgeUsesOnly.insert({BranchBB, Succ}); 386 } 387 }; 388 389 // Match combinations of conditions. 390 CmpInst::Predicate Pred; 391 bool isAnd = false; 392 bool isOr = false; 393 SmallVector<Value *, 8> CmpOperands; 394 if (match(BI->getCondition(), m_And(m_Cmp(Pred, m_Value(), m_Value()), 395 m_Cmp(Pred, m_Value(), m_Value()))) || 396 match(BI->getCondition(), m_Or(m_Cmp(Pred, m_Value(), m_Value()), 397 m_Cmp(Pred, m_Value(), m_Value())))) { 398 auto *BinOp = cast<BinaryOperator>(BI->getCondition()); 399 if (BinOp->getOpcode() == Instruction::And) 400 isAnd = true; 401 else if (BinOp->getOpcode() == Instruction::Or) 402 isOr = true; 403 ConditionsToProcess.push_back(BinOp->getOperand(0)); 404 ConditionsToProcess.push_back(BinOp->getOperand(1)); 405 ConditionsToProcess.push_back(BI->getCondition()); 406 } else if (isa<CmpInst>(BI->getCondition())) { 407 ConditionsToProcess.push_back(BI->getCondition()); 408 } 409 for (auto Cond : ConditionsToProcess) { 410 if (auto *Cmp = dyn_cast<CmpInst>(Cond)) { 411 collectCmpOps(Cmp, CmpOperands); 412 // Now add our copy infos for our operands 413 for (auto *Op : CmpOperands) 414 InsertHelper(Op, isAnd, isOr, Cmp); 415 } else if (auto *BinOp = dyn_cast<BinaryOperator>(Cond)) { 416 // This must be an AND or an OR. 417 assert((BinOp->getOpcode() == Instruction::And || 418 BinOp->getOpcode() == Instruction::Or) && 419 "Should have been an AND or an OR"); 420 // The actual value of the binop is not subject to the same restrictions 421 // as the comparison. It's either true or false on the true/false branch. 422 InsertHelper(BinOp, false, false, BinOp); 423 } else { 424 llvm_unreachable("Unknown type of condition"); 425 } 426 CmpOperands.clear(); 427 } 428 } 429 // Process a block terminating switch, and place relevant operations to be 430 // renamed into OpsToRename. 431 void PredicateInfo::processSwitch(SwitchInst *SI, BasicBlock *BranchBB, 432 SmallPtrSetImpl<Value *> &OpsToRename) { 433 Value *Op = SI->getCondition(); 434 if ((!isa<Instruction>(Op) && !isa<Argument>(Op)) || Op->hasOneUse()) 435 return; 436 437 // Remember how many outgoing edges there are to every successor. 438 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; 439 for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) { 440 BasicBlock *TargetBlock = SI->getSuccessor(i); 441 ++SwitchEdges[TargetBlock]; 442 } 443 444 // Now propagate info for each case value 445 for (auto C : SI->cases()) { 446 BasicBlock *TargetBlock = C.getCaseSuccessor(); 447 if (SwitchEdges.lookup(TargetBlock) == 1) { 448 PredicateSwitch *PS = new PredicateSwitch( 449 Op, SI->getParent(), TargetBlock, C.getCaseValue(), SI); 450 addInfoFor(OpsToRename, Op, PS); 451 if (!TargetBlock->getSinglePredecessor()) 452 EdgeUsesOnly.insert({BranchBB, TargetBlock}); 453 } 454 } 455 } 456 457 // Build predicate info for our function 458 void PredicateInfo::buildPredicateInfo() { 459 DT.updateDFSNumbers(); 460 // Collect operands to rename from all conditional branch terminators, as well 461 // as assume statements. 462 SmallPtrSet<Value *, 8> OpsToRename; 463 for (auto DTN : depth_first(DT.getRootNode())) { 464 BasicBlock *BranchBB = DTN->getBlock(); 465 if (auto *BI = dyn_cast<BranchInst>(BranchBB->getTerminator())) { 466 if (!BI->isConditional()) 467 continue; 468 // Can't insert conditional information if they all go to the same place. 469 if (BI->getSuccessor(0) == BI->getSuccessor(1)) 470 continue; 471 processBranch(BI, BranchBB, OpsToRename); 472 } else if (auto *SI = dyn_cast<SwitchInst>(BranchBB->getTerminator())) { 473 processSwitch(SI, BranchBB, OpsToRename); 474 } 475 } 476 for (auto &Assume : AC.assumptions()) { 477 if (auto *II = dyn_cast_or_null<IntrinsicInst>(Assume)) 478 processAssume(II, II->getParent(), OpsToRename); 479 } 480 // Now rename all our operations. 481 renameUses(OpsToRename); 482 } 483 484 // Create a ssa_copy declaration with custom mangling, because 485 // Intrinsic::getDeclaration does not handle overloaded unnamed types properly: 486 // all unnamed types get mangled to the same string. We use the pointer 487 // to the type as name here, as it guarantees unique names for different 488 // types and we remove the declarations when destroying PredicateInfo. 489 // It is a workaround for PR38117, because solving it in a fully general way is 490 // tricky (FIXME). 491 static Function *getCopyDeclaration(Module *M, Type *Ty) { 492 std::string Name = "llvm.ssa.copy." + utostr((uintptr_t) Ty); 493 return cast<Function>(M->getOrInsertFunction( 494 Name, getType(M->getContext(), Intrinsic::ssa_copy, Ty))); 495 } 496 497 // Given the renaming stack, make all the operands currently on the stack real 498 // by inserting them into the IR. Return the last operation's value. 499 Value *PredicateInfo::materializeStack(unsigned int &Counter, 500 ValueDFSStack &RenameStack, 501 Value *OrigOp) { 502 // Find the first thing we have to materialize 503 auto RevIter = RenameStack.rbegin(); 504 for (; RevIter != RenameStack.rend(); ++RevIter) 505 if (RevIter->Def) 506 break; 507 508 size_t Start = RevIter - RenameStack.rbegin(); 509 // The maximum number of things we should be trying to materialize at once 510 // right now is 4, depending on if we had an assume, a branch, and both used 511 // and of conditions. 512 for (auto RenameIter = RenameStack.end() - Start; 513 RenameIter != RenameStack.end(); ++RenameIter) { 514 auto *Op = 515 RenameIter == RenameStack.begin() ? OrigOp : (RenameIter - 1)->Def; 516 ValueDFS &Result = *RenameIter; 517 auto *ValInfo = Result.PInfo; 518 // For edge predicates, we can just place the operand in the block before 519 // the terminator. For assume, we have to place it right before the assume 520 // to ensure we dominate all of our uses. Always insert right before the 521 // relevant instruction (terminator, assume), so that we insert in proper 522 // order in the case of multiple predicateinfo in the same block. 523 if (isa<PredicateWithEdge>(ValInfo)) { 524 IRBuilder<> B(getBranchTerminator(ValInfo)); 525 Function *IF = getCopyDeclaration(F.getParent(), Op->getType()); 526 if (IF->user_begin() == IF->user_end()) 527 CreatedDeclarations.insert(IF); 528 CallInst *PIC = 529 B.CreateCall(IF, Op, Op->getName() + "." + Twine(Counter++)); 530 PredicateMap.insert({PIC, ValInfo}); 531 Result.Def = PIC; 532 } else { 533 auto *PAssume = dyn_cast<PredicateAssume>(ValInfo); 534 assert(PAssume && 535 "Should not have gotten here without it being an assume"); 536 IRBuilder<> B(PAssume->AssumeInst); 537 Function *IF = getCopyDeclaration(F.getParent(), Op->getType()); 538 if (IF->user_begin() == IF->user_end()) 539 CreatedDeclarations.insert(IF); 540 CallInst *PIC = B.CreateCall(IF, Op); 541 PredicateMap.insert({PIC, ValInfo}); 542 Result.Def = PIC; 543 } 544 } 545 return RenameStack.back().Def; 546 } 547 548 // Instead of the standard SSA renaming algorithm, which is O(Number of 549 // instructions), and walks the entire dominator tree, we walk only the defs + 550 // uses. The standard SSA renaming algorithm does not really rely on the 551 // dominator tree except to order the stack push/pops of the renaming stacks, so 552 // that defs end up getting pushed before hitting the correct uses. This does 553 // not require the dominator tree, only the *order* of the dominator tree. The 554 // complete and correct ordering of the defs and uses, in dominator tree is 555 // contained in the DFS numbering of the dominator tree. So we sort the defs and 556 // uses into the DFS ordering, and then just use the renaming stack as per 557 // normal, pushing when we hit a def (which is a predicateinfo instruction), 558 // popping when we are out of the dfs scope for that def, and replacing any uses 559 // with top of stack if it exists. In order to handle liveness without 560 // propagating liveness info, we don't actually insert the predicateinfo 561 // instruction def until we see a use that it would dominate. Once we see such 562 // a use, we materialize the predicateinfo instruction in the right place and 563 // use it. 564 // 565 // TODO: Use this algorithm to perform fast single-variable renaming in 566 // promotememtoreg and memoryssa. 567 void PredicateInfo::renameUses(SmallPtrSetImpl<Value *> &OpSet) { 568 // Sort OpsToRename since we are going to iterate it. 569 SmallVector<Value *, 8> OpsToRename(OpSet.begin(), OpSet.end()); 570 auto Comparator = [&](const Value *A, const Value *B) { 571 return valueComesBefore(OI, A, B); 572 }; 573 llvm::sort(OpsToRename.begin(), OpsToRename.end(), Comparator); 574 ValueDFS_Compare Compare(OI); 575 // Compute liveness, and rename in O(uses) per Op. 576 for (auto *Op : OpsToRename) { 577 LLVM_DEBUG(dbgs() << "Visiting " << *Op << "\n"); 578 unsigned Counter = 0; 579 SmallVector<ValueDFS, 16> OrderedUses; 580 const auto &ValueInfo = getValueInfo(Op); 581 // Insert the possible copies into the def/use list. 582 // They will become real copies if we find a real use for them, and never 583 // created otherwise. 584 for (auto &PossibleCopy : ValueInfo.Infos) { 585 ValueDFS VD; 586 // Determine where we are going to place the copy by the copy type. 587 // The predicate info for branches always come first, they will get 588 // materialized in the split block at the top of the block. 589 // The predicate info for assumes will be somewhere in the middle, 590 // it will get materialized in front of the assume. 591 if (const auto *PAssume = dyn_cast<PredicateAssume>(PossibleCopy)) { 592 VD.LocalNum = LN_Middle; 593 DomTreeNode *DomNode = DT.getNode(PAssume->AssumeInst->getParent()); 594 if (!DomNode) 595 continue; 596 VD.DFSIn = DomNode->getDFSNumIn(); 597 VD.DFSOut = DomNode->getDFSNumOut(); 598 VD.PInfo = PossibleCopy; 599 OrderedUses.push_back(VD); 600 } else if (isa<PredicateWithEdge>(PossibleCopy)) { 601 // If we can only do phi uses, we treat it like it's in the branch 602 // block, and handle it specially. We know that it goes last, and only 603 // dominate phi uses. 604 auto BlockEdge = getBlockEdge(PossibleCopy); 605 if (EdgeUsesOnly.count(BlockEdge)) { 606 VD.LocalNum = LN_Last; 607 auto *DomNode = DT.getNode(BlockEdge.first); 608 if (DomNode) { 609 VD.DFSIn = DomNode->getDFSNumIn(); 610 VD.DFSOut = DomNode->getDFSNumOut(); 611 VD.PInfo = PossibleCopy; 612 VD.EdgeOnly = true; 613 OrderedUses.push_back(VD); 614 } 615 } else { 616 // Otherwise, we are in the split block (even though we perform 617 // insertion in the branch block). 618 // Insert a possible copy at the split block and before the branch. 619 VD.LocalNum = LN_First; 620 auto *DomNode = DT.getNode(BlockEdge.second); 621 if (DomNode) { 622 VD.DFSIn = DomNode->getDFSNumIn(); 623 VD.DFSOut = DomNode->getDFSNumOut(); 624 VD.PInfo = PossibleCopy; 625 OrderedUses.push_back(VD); 626 } 627 } 628 } 629 } 630 631 convertUsesToDFSOrdered(Op, OrderedUses); 632 // Here we require a stable sort because we do not bother to try to 633 // assign an order to the operands the uses represent. Thus, two 634 // uses in the same instruction do not have a strict sort order 635 // currently and will be considered equal. We could get rid of the 636 // stable sort by creating one if we wanted. 637 std::stable_sort(OrderedUses.begin(), OrderedUses.end(), Compare); 638 SmallVector<ValueDFS, 8> RenameStack; 639 // For each use, sorted into dfs order, push values and replaces uses with 640 // top of stack, which will represent the reaching def. 641 for (auto &VD : OrderedUses) { 642 // We currently do not materialize copy over copy, but we should decide if 643 // we want to. 644 bool PossibleCopy = VD.PInfo != nullptr; 645 if (RenameStack.empty()) { 646 LLVM_DEBUG(dbgs() << "Rename Stack is empty\n"); 647 } else { 648 LLVM_DEBUG(dbgs() << "Rename Stack Top DFS numbers are (" 649 << RenameStack.back().DFSIn << "," 650 << RenameStack.back().DFSOut << ")\n"); 651 } 652 653 LLVM_DEBUG(dbgs() << "Current DFS numbers are (" << VD.DFSIn << "," 654 << VD.DFSOut << ")\n"); 655 656 bool ShouldPush = (VD.Def || PossibleCopy); 657 bool OutOfScope = !stackIsInScope(RenameStack, VD); 658 if (OutOfScope || ShouldPush) { 659 // Sync to our current scope. 660 popStackUntilDFSScope(RenameStack, VD); 661 if (ShouldPush) { 662 RenameStack.push_back(VD); 663 } 664 } 665 // If we get to this point, and the stack is empty we must have a use 666 // with no renaming needed, just skip it. 667 if (RenameStack.empty()) 668 continue; 669 // Skip values, only want to rename the uses 670 if (VD.Def || PossibleCopy) 671 continue; 672 if (!DebugCounter::shouldExecute(RenameCounter)) { 673 LLVM_DEBUG(dbgs() << "Skipping execution due to debug counter\n"); 674 continue; 675 } 676 ValueDFS &Result = RenameStack.back(); 677 678 // If the possible copy dominates something, materialize our stack up to 679 // this point. This ensures every comparison that affects our operation 680 // ends up with predicateinfo. 681 if (!Result.Def) 682 Result.Def = materializeStack(Counter, RenameStack, Op); 683 684 LLVM_DEBUG(dbgs() << "Found replacement " << *Result.Def << " for " 685 << *VD.U->get() << " in " << *(VD.U->getUser()) 686 << "\n"); 687 assert(DT.dominates(cast<Instruction>(Result.Def), *VD.U) && 688 "Predicateinfo def should have dominated this use"); 689 VD.U->set(Result.Def); 690 } 691 } 692 } 693 694 PredicateInfo::ValueInfo &PredicateInfo::getOrCreateValueInfo(Value *Operand) { 695 auto OIN = ValueInfoNums.find(Operand); 696 if (OIN == ValueInfoNums.end()) { 697 // This will grow it 698 ValueInfos.resize(ValueInfos.size() + 1); 699 // This will use the new size and give us a 0 based number of the info 700 auto InsertResult = ValueInfoNums.insert({Operand, ValueInfos.size() - 1}); 701 assert(InsertResult.second && "Value info number already existed?"); 702 return ValueInfos[InsertResult.first->second]; 703 } 704 return ValueInfos[OIN->second]; 705 } 706 707 const PredicateInfo::ValueInfo & 708 PredicateInfo::getValueInfo(Value *Operand) const { 709 auto OINI = ValueInfoNums.lookup(Operand); 710 assert(OINI != 0 && "Operand was not really in the Value Info Numbers"); 711 assert(OINI < ValueInfos.size() && 712 "Value Info Number greater than size of Value Info Table"); 713 return ValueInfos[OINI]; 714 } 715 716 PredicateInfo::PredicateInfo(Function &F, DominatorTree &DT, 717 AssumptionCache &AC) 718 : F(F), DT(DT), AC(AC), OI(&DT) { 719 // Push an empty operand info so that we can detect 0 as not finding one 720 ValueInfos.resize(1); 721 buildPredicateInfo(); 722 } 723 724 // Remove all declarations we created . The PredicateInfo consumers are 725 // responsible for remove the ssa_copy calls created. 726 PredicateInfo::~PredicateInfo() { 727 // Collect function pointers in set first, as SmallSet uses a SmallVector 728 // internally and we have to remove the asserting value handles first. 729 SmallPtrSet<Function *, 20> FunctionPtrs; 730 for (auto &F : CreatedDeclarations) 731 FunctionPtrs.insert(&*F); 732 CreatedDeclarations.clear(); 733 734 for (Function *F : FunctionPtrs) { 735 assert(F->user_begin() == F->user_end() && 736 "PredicateInfo consumer did not remove all SSA copies."); 737 F->eraseFromParent(); 738 } 739 } 740 741 void PredicateInfo::verifyPredicateInfo() const {} 742 743 char PredicateInfoPrinterLegacyPass::ID = 0; 744 745 PredicateInfoPrinterLegacyPass::PredicateInfoPrinterLegacyPass() 746 : FunctionPass(ID) { 747 initializePredicateInfoPrinterLegacyPassPass( 748 *PassRegistry::getPassRegistry()); 749 } 750 751 void PredicateInfoPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const { 752 AU.setPreservesAll(); 753 AU.addRequiredTransitive<DominatorTreeWrapperPass>(); 754 AU.addRequired<AssumptionCacheTracker>(); 755 } 756 757 // Replace ssa_copy calls created by PredicateInfo with their operand. 758 static void replaceCreatedSSACopys(PredicateInfo &PredInfo, Function &F) { 759 for (auto I = inst_begin(F), E = inst_end(F); I != E;) { 760 Instruction *Inst = &*I++; 761 const auto *PI = PredInfo.getPredicateInfoFor(Inst); 762 auto *II = dyn_cast<IntrinsicInst>(Inst); 763 if (!PI || !II || II->getIntrinsicID() != Intrinsic::ssa_copy) 764 continue; 765 766 Inst->replaceAllUsesWith(II->getOperand(0)); 767 Inst->eraseFromParent(); 768 } 769 } 770 771 bool PredicateInfoPrinterLegacyPass::runOnFunction(Function &F) { 772 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 773 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 774 auto PredInfo = make_unique<PredicateInfo>(F, DT, AC); 775 PredInfo->print(dbgs()); 776 if (VerifyPredicateInfo) 777 PredInfo->verifyPredicateInfo(); 778 779 replaceCreatedSSACopys(*PredInfo, F); 780 return false; 781 } 782 783 PreservedAnalyses PredicateInfoPrinterPass::run(Function &F, 784 FunctionAnalysisManager &AM) { 785 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 786 auto &AC = AM.getResult<AssumptionAnalysis>(F); 787 OS << "PredicateInfo for function: " << F.getName() << "\n"; 788 auto PredInfo = make_unique<PredicateInfo>(F, DT, AC); 789 PredInfo->print(OS); 790 791 replaceCreatedSSACopys(*PredInfo, F); 792 return PreservedAnalyses::all(); 793 } 794 795 /// An assembly annotator class to print PredicateInfo information in 796 /// comments. 797 class PredicateInfoAnnotatedWriter : public AssemblyAnnotationWriter { 798 friend class PredicateInfo; 799 const PredicateInfo *PredInfo; 800 801 public: 802 PredicateInfoAnnotatedWriter(const PredicateInfo *M) : PredInfo(M) {} 803 804 virtual void emitBasicBlockStartAnnot(const BasicBlock *BB, 805 formatted_raw_ostream &OS) {} 806 807 virtual void emitInstructionAnnot(const Instruction *I, 808 formatted_raw_ostream &OS) { 809 if (const auto *PI = PredInfo->getPredicateInfoFor(I)) { 810 OS << "; Has predicate info\n"; 811 if (const auto *PB = dyn_cast<PredicateBranch>(PI)) { 812 OS << "; branch predicate info { TrueEdge: " << PB->TrueEdge 813 << " Comparison:" << *PB->Condition << " Edge: ["; 814 PB->From->printAsOperand(OS); 815 OS << ","; 816 PB->To->printAsOperand(OS); 817 OS << "] }\n"; 818 } else if (const auto *PS = dyn_cast<PredicateSwitch>(PI)) { 819 OS << "; switch predicate info { CaseValue: " << *PS->CaseValue 820 << " Switch:" << *PS->Switch << " Edge: ["; 821 PS->From->printAsOperand(OS); 822 OS << ","; 823 PS->To->printAsOperand(OS); 824 OS << "] }\n"; 825 } else if (const auto *PA = dyn_cast<PredicateAssume>(PI)) { 826 OS << "; assume predicate info {" 827 << " Comparison:" << *PA->Condition << " }\n"; 828 } 829 } 830 } 831 }; 832 833 void PredicateInfo::print(raw_ostream &OS) const { 834 PredicateInfoAnnotatedWriter Writer(this); 835 F.print(OS, &Writer); 836 } 837 838 void PredicateInfo::dump() const { 839 PredicateInfoAnnotatedWriter Writer(this); 840 F.print(dbgs(), &Writer); 841 } 842 843 PreservedAnalyses PredicateInfoVerifierPass::run(Function &F, 844 FunctionAnalysisManager &AM) { 845 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 846 auto &AC = AM.getResult<AssumptionAnalysis>(F); 847 make_unique<PredicateInfo>(F, DT, AC)->verifyPredicateInfo(); 848 849 return PreservedAnalyses::all(); 850 } 851 } 852