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