1 //===- GVNHoist.cpp - Hoist scalar and load expressions -------------------===//
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 pass hoists expressions from branches to a common dominator. It uses
11 // GVN (global value numbering) to discover expressions computing the same
12 // values. The primary goals of code-hoisting are:
13 // 1. To reduce the code size.
14 // 2. In some cases reduce critical path (by exposing more ILP).
16 // Hoisting may affect the performance in some cases. To mitigate that, hoisting
17 // is disabled in the following cases.
18 // 1. Scalars across calls.
19 // 2. geps when corresponding load/store cannot be hoisted.
20 //===----------------------------------------------------------------------===//
22 #include "llvm/Transforms/Scalar/GVN.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/SmallPtrSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/Transforms/Utils/Local.h"
29 #include "llvm/Transforms/Utils/MemorySSA.h"
33 #define DEBUG_TYPE "gvn-hoist"
35 STATISTIC(NumHoisted, "Number of instructions hoisted");
36 STATISTIC(NumRemoved, "Number of instructions removed");
37 STATISTIC(NumLoadsHoisted, "Number of loads hoisted");
38 STATISTIC(NumLoadsRemoved, "Number of loads removed");
39 STATISTIC(NumStoresHoisted, "Number of stores hoisted");
40 STATISTIC(NumStoresRemoved, "Number of stores removed");
41 STATISTIC(NumCallsHoisted, "Number of calls hoisted");
42 STATISTIC(NumCallsRemoved, "Number of calls removed");
45 MaxHoistedThreshold("gvn-max-hoisted", cl::Hidden, cl::init(-1),
46 cl::desc("Max number of instructions to hoist "
47 "(default unlimited = -1)"));
48 static cl::opt<int> MaxNumberOfBBSInPath(
49 "gvn-hoist-max-bbs", cl::Hidden, cl::init(4),
50 cl::desc("Max number of basic blocks on the path between "
51 "hoisting locations (default = 4, unlimited = -1)"));
53 static cl::opt<int> MaxDepthInBB(
54 "gvn-hoist-max-depth", cl::Hidden, cl::init(100),
55 cl::desc("Hoist instructions from the beginning of the BB up to the "
56 "maximum specified depth (default = 100, unlimited = -1)"));
59 MaxChainLength("gvn-hoist-max-chain-length", cl::Hidden, cl::init(10),
60 cl::desc("Maximum length of dependent chains to hoist "
61 "(default = 10, unlimited = -1)"));
65 // Provides a sorting function based on the execution order of two instructions.
68 DenseMap<const Value *, unsigned> &DFSNumber;
71 SortByDFSIn(DenseMap<const Value *, unsigned> &D) : DFSNumber(D) {}
73 // Returns true when A executes before B.
74 bool operator()(const Instruction *A, const Instruction *B) const {
75 // FIXME: libc++ has a std::sort() algorithm that will call the compare
76 // function on the same element. Once PR20837 is fixed and some more years
77 // pass by and all the buildbots have moved to a corrected std::sort(),
78 // enable the following assert:
82 const BasicBlock *BA = A->getParent();
83 const BasicBlock *BB = B->getParent();
86 ADFS = DFSNumber.lookup(A);
87 BDFS = DFSNumber.lookup(B);
89 ADFS = DFSNumber.lookup(BA);
90 BDFS = DFSNumber.lookup(BB);
97 // A map from a pair of VNs to all the instructions with those VNs.
98 typedef DenseMap<std::pair<unsigned, unsigned>, SmallVector<Instruction *, 4>>
100 // An invalid value number Used when inserting a single value number into
102 enum : unsigned { InvalidVN = ~2U };
104 // Records all scalar instructions candidate for code hoisting.
106 VNtoInsns VNtoScalars;
109 // Inserts I and its value number in VNtoScalars.
110 void insert(Instruction *I, GVN::ValueTable &VN) {
111 // Scalar instruction.
112 unsigned V = VN.lookupOrAdd(I);
113 VNtoScalars[{V, InvalidVN}].push_back(I);
116 const VNtoInsns &getVNTable() const { return VNtoScalars; }
119 // Records all load instructions candidate for code hoisting.
124 // Insert Load and the value number of its memory address in VNtoLoads.
125 void insert(LoadInst *Load, GVN::ValueTable &VN) {
126 if (Load->isSimple()) {
127 unsigned V = VN.lookupOrAdd(Load->getPointerOperand());
128 VNtoLoads[{V, InvalidVN}].push_back(Load);
132 const VNtoInsns &getVNTable() const { return VNtoLoads; }
135 // Records all store instructions candidate for code hoisting.
137 VNtoInsns VNtoStores;
140 // Insert the Store and a hash number of the store address and the stored
141 // value in VNtoStores.
142 void insert(StoreInst *Store, GVN::ValueTable &VN) {
143 if (!Store->isSimple())
145 // Hash the store address and the stored value.
146 Value *Ptr = Store->getPointerOperand();
147 Value *Val = Store->getValueOperand();
148 VNtoStores[{VN.lookupOrAdd(Ptr), VN.lookupOrAdd(Val)}].push_back(Store);
151 const VNtoInsns &getVNTable() const { return VNtoStores; }
154 // Records all call instructions candidate for code hoisting.
156 VNtoInsns VNtoCallsScalars;
157 VNtoInsns VNtoCallsLoads;
158 VNtoInsns VNtoCallsStores;
161 // Insert Call and its value numbering in one of the VNtoCalls* containers.
162 void insert(CallInst *Call, GVN::ValueTable &VN) {
163 // A call that doesNotAccessMemory is handled as a Scalar,
164 // onlyReadsMemory will be handled as a Load instruction,
165 // all other calls will be handled as stores.
166 unsigned V = VN.lookupOrAdd(Call);
167 auto Entry = std::make_pair(V, InvalidVN);
169 if (Call->doesNotAccessMemory())
170 VNtoCallsScalars[Entry].push_back(Call);
171 else if (Call->onlyReadsMemory())
172 VNtoCallsLoads[Entry].push_back(Call);
174 VNtoCallsStores[Entry].push_back(Call);
177 const VNtoInsns &getScalarVNTable() const { return VNtoCallsScalars; }
179 const VNtoInsns &getLoadVNTable() const { return VNtoCallsLoads; }
181 const VNtoInsns &getStoreVNTable() const { return VNtoCallsStores; }
184 typedef DenseMap<const BasicBlock *, bool> BBSideEffectsSet;
185 typedef SmallVector<Instruction *, 4> SmallVecInsn;
186 typedef SmallVectorImpl<Instruction *> SmallVecImplInsn;
188 static void combineKnownMetadata(Instruction *ReplInst, Instruction *I) {
189 static const unsigned KnownIDs[] = {
190 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
191 LLVMContext::MD_noalias, LLVMContext::MD_range,
192 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
193 LLVMContext::MD_invariant_group};
194 combineMetadata(ReplInst, I, KnownIDs);
197 // This pass hoists common computations across branches sharing common
198 // dominator. The primary goal is to reduce the code size, and in some
199 // cases reduce critical path (by exposing more ILP).
202 GVNHoist(DominatorTree *DT, AliasAnalysis *AA, MemoryDependenceResults *MD,
204 : DT(DT), AA(AA), MD(MD), MSSA(MSSA),
209 bool run(Function &F) {
211 VN.setAliasAnalysis(AA);
214 // Perform DFS Numbering of instructions.
216 for (const BasicBlock *BB : depth_first(&F.getEntryBlock())) {
217 DFSNumber[BB] = ++BBI;
219 for (auto &Inst : *BB)
220 DFSNumber[&Inst] = ++I;
225 // FIXME: use lazy evaluation of VN to avoid the fix-point computation.
227 if (MaxChainLength != -1 && ++ChainLength >= MaxChainLength)
230 auto HoistStat = hoistExpressions(F);
231 if (HoistStat.first + HoistStat.second == 0)
234 if (HoistStat.second > 0)
235 // To address a limitation of the current GVN, we need to rerun the
236 // hoisting after we hoisted loads or stores in order to be able to
237 // hoist all scalars dependent on the hoisted ld/st.
250 MemoryDependenceResults *MD;
252 const bool HoistingGeps;
253 DenseMap<const Value *, unsigned> DFSNumber;
254 BBSideEffectsSet BBSideEffects;
257 enum InsKind { Unknown, Scalar, Load, Store };
259 // Return true when there are exception handling in BB.
260 bool hasEH(const BasicBlock *BB) {
261 auto It = BBSideEffects.find(BB);
262 if (It != BBSideEffects.end())
265 if (BB->isEHPad() || BB->hasAddressTaken()) {
266 BBSideEffects[BB] = true;
270 if (BB->getTerminator()->mayThrow()) {
271 BBSideEffects[BB] = true;
275 BBSideEffects[BB] = false;
279 // Return true when a successor of BB dominates A.
280 bool successorDominate(const BasicBlock *BB, const BasicBlock *A) {
281 for (const BasicBlock *Succ : BB->getTerminator()->successors())
282 if (DT->dominates(Succ, A))
288 // Return true when all paths from HoistBB to the end of the function pass
289 // through one of the blocks in WL.
290 bool hoistingFromAllPaths(const BasicBlock *HoistBB,
291 SmallPtrSetImpl<const BasicBlock *> &WL) {
293 // Copy WL as the loop will remove elements from it.
294 SmallPtrSet<const BasicBlock *, 2> WorkList(WL.begin(), WL.end());
296 for (auto It = df_begin(HoistBB), E = df_end(HoistBB); It != E;) {
297 // There exists a path from HoistBB to the exit of the function if we are
298 // still iterating in DF traversal and we removed all instructions from
300 if (WorkList.empty())
303 const BasicBlock *BB = *It;
304 if (WorkList.erase(BB)) {
305 // Stop DFS traversal when BB is in the work list.
310 // Check for end of function, calls that do not return, etc.
311 if (!isGuaranteedToTransferExecutionToSuccessor(BB->getTerminator()))
314 // When reaching the back-edge of a loop, there may be a path through the
315 // loop that does not pass through B or C before exiting the loop.
316 if (successorDominate(BB, HoistBB))
319 // Increment DFS traversal when not skipping children.
326 /* Return true when I1 appears before I2 in the instructions of BB. */
327 bool firstInBB(const Instruction *I1, const Instruction *I2) {
328 assert(I1->getParent() == I2->getParent());
329 unsigned I1DFS = DFSNumber.lookup(I1);
330 unsigned I2DFS = DFSNumber.lookup(I2);
331 assert(I1DFS && I2DFS);
332 return I1DFS < I2DFS;
335 // Return true when there are memory uses of Def in BB.
336 bool hasMemoryUse(const Instruction *NewPt, MemoryDef *Def,
337 const BasicBlock *BB) {
338 const MemorySSA::AccessList *Acc = MSSA->getBlockAccesses(BB);
342 Instruction *OldPt = Def->getMemoryInst();
343 const BasicBlock *OldBB = OldPt->getParent();
344 const BasicBlock *NewBB = NewPt->getParent();
345 bool ReachedNewPt = false;
347 for (const MemoryAccess &MA : *Acc)
348 if (const MemoryUse *MU = dyn_cast<MemoryUse>(&MA)) {
349 Instruction *Insn = MU->getMemoryInst();
351 // Do not check whether MU aliases Def when MU occurs after OldPt.
352 if (BB == OldBB && firstInBB(OldPt, Insn))
355 // Do not check whether MU aliases Def when MU occurs before NewPt.
358 if (firstInBB(Insn, NewPt))
363 if (defClobbersUseOrDef(Def, MU, *AA))
370 // Return true when there are exception handling or loads of memory Def
371 // between Def and NewPt. This function is only called for stores: Def is
372 // the MemoryDef of the store to be hoisted.
374 // Decrement by 1 NBBsOnAllPaths for each block between HoistPt and BB, and
375 // return true when the counter NBBsOnAllPaths reaces 0, except when it is
376 // initialized to -1 which is unlimited.
377 bool hasEHOrLoadsOnPath(const Instruction *NewPt, MemoryDef *Def,
378 int &NBBsOnAllPaths) {
379 const BasicBlock *NewBB = NewPt->getParent();
380 const BasicBlock *OldBB = Def->getBlock();
381 assert(DT->dominates(NewBB, OldBB) && "invalid path");
382 assert(DT->dominates(Def->getDefiningAccess()->getBlock(), NewBB) &&
383 "def does not dominate new hoisting point");
385 // Walk all basic blocks reachable in depth-first iteration on the inverse
386 // CFG from OldBB to NewBB. These blocks are all the blocks that may be
387 // executed between the execution of NewBB and OldBB. Hoisting an expression
388 // from OldBB into NewBB has to be safe on all execution paths.
389 for (auto I = idf_begin(OldBB), E = idf_end(OldBB); I != E;) {
391 // Stop traversal when reaching HoistPt.
396 // Stop walk once the limit is reached.
397 if (NBBsOnAllPaths == 0)
400 // Impossible to hoist with exceptions on the path.
404 // Check that we do not move a store past loads.
405 if (hasMemoryUse(NewPt, Def, *I))
408 // -1 is unlimited number of blocks on all paths.
409 if (NBBsOnAllPaths != -1)
418 // Return true when there are exception handling between HoistPt and BB.
419 // Decrement by 1 NBBsOnAllPaths for each block between HoistPt and BB, and
420 // return true when the counter NBBsOnAllPaths reaches 0, except when it is
421 // initialized to -1 which is unlimited.
422 bool hasEHOnPath(const BasicBlock *HoistPt, const BasicBlock *BB,
423 int &NBBsOnAllPaths) {
424 assert(DT->dominates(HoistPt, BB) && "Invalid path");
426 // Walk all basic blocks reachable in depth-first iteration on
427 // the inverse CFG from BBInsn to NewHoistPt. These blocks are all the
428 // blocks that may be executed between the execution of NewHoistPt and
429 // BBInsn. Hoisting an expression from BBInsn into NewHoistPt has to be safe
430 // on all execution paths.
431 for (auto I = idf_begin(BB), E = idf_end(BB); I != E;) {
433 // Stop traversal when reaching NewHoistPt.
438 // Stop walk once the limit is reached.
439 if (NBBsOnAllPaths == 0)
442 // Impossible to hoist with exceptions on the path.
446 // -1 is unlimited number of blocks on all paths.
447 if (NBBsOnAllPaths != -1)
456 // Return true when it is safe to hoist a memory load or store U from OldPt
458 bool safeToHoistLdSt(const Instruction *NewPt, const Instruction *OldPt,
459 MemoryUseOrDef *U, InsKind K, int &NBBsOnAllPaths) {
461 // In place hoisting is safe.
465 const BasicBlock *NewBB = NewPt->getParent();
466 const BasicBlock *OldBB = OldPt->getParent();
467 const BasicBlock *UBB = U->getBlock();
469 // Check for dependences on the Memory SSA.
470 MemoryAccess *D = U->getDefiningAccess();
471 BasicBlock *DBB = D->getBlock();
472 if (DT->properlyDominates(NewBB, DBB))
473 // Cannot move the load or store to NewBB above its definition in DBB.
476 if (NewBB == DBB && !MSSA->isLiveOnEntryDef(D))
477 if (auto *UD = dyn_cast<MemoryUseOrDef>(D))
478 if (firstInBB(NewPt, UD->getMemoryInst()))
479 // Cannot move the load or store to NewPt above its definition in D.
482 // Check for unsafe hoistings due to side effects.
483 if (K == InsKind::Store) {
484 if (hasEHOrLoadsOnPath(NewPt, dyn_cast<MemoryDef>(U), NBBsOnAllPaths))
486 } else if (hasEHOnPath(NewBB, OldBB, NBBsOnAllPaths))
490 if (DT->properlyDominates(DBB, NewBB))
493 assert(MSSA->locallyDominates(D, U));
496 // No side effects: it is safe to hoist.
500 // Return true when it is safe to hoist scalar instructions from all blocks in
502 bool safeToHoistScalar(const BasicBlock *HoistBB,
503 SmallPtrSetImpl<const BasicBlock *> &WL,
504 int &NBBsOnAllPaths) {
505 // Check that the hoisted expression is needed on all paths.
506 if (!hoistingFromAllPaths(HoistBB, WL))
509 for (const BasicBlock *BB : WL)
510 if (hasEHOnPath(HoistBB, BB, NBBsOnAllPaths))
516 // Each element of a hoisting list contains the basic block where to hoist and
517 // a list of instructions to be hoisted.
518 typedef std::pair<BasicBlock *, SmallVecInsn> HoistingPointInfo;
519 typedef SmallVector<HoistingPointInfo, 4> HoistingPointList;
521 // Partition InstructionsToHoist into a set of candidates which can share a
522 // common hoisting point. The partitions are collected in HPL. IsScalar is
523 // true when the instructions in InstructionsToHoist are scalars. IsLoad is
524 // true when the InstructionsToHoist are loads, false when they are stores.
525 void partitionCandidates(SmallVecImplInsn &InstructionsToHoist,
526 HoistingPointList &HPL, InsKind K) {
527 // No need to sort for two instructions.
528 if (InstructionsToHoist.size() > 2) {
529 SortByDFSIn Pred(DFSNumber);
530 std::sort(InstructionsToHoist.begin(), InstructionsToHoist.end(), Pred);
533 int NumBBsOnAllPaths = MaxNumberOfBBSInPath;
535 SmallVecImplInsn::iterator II = InstructionsToHoist.begin();
536 SmallVecImplInsn::iterator Start = II;
537 Instruction *HoistPt = *II;
538 BasicBlock *HoistBB = HoistPt->getParent();
540 if (K != InsKind::Scalar)
541 UD = MSSA->getMemoryAccess(HoistPt);
543 for (++II; II != InstructionsToHoist.end(); ++II) {
544 Instruction *Insn = *II;
545 BasicBlock *BB = Insn->getParent();
546 BasicBlock *NewHoistBB;
547 Instruction *NewHoistPt;
549 if (BB == HoistBB) { // Both are in the same Basic Block.
550 NewHoistBB = HoistBB;
551 NewHoistPt = firstInBB(Insn, HoistPt) ? Insn : HoistPt;
553 // If the hoisting point contains one of the instructions,
554 // then hoist there, otherwise hoist before the terminator.
555 NewHoistBB = DT->findNearestCommonDominator(HoistBB, BB);
556 if (NewHoistBB == BB)
558 else if (NewHoistBB == HoistBB)
559 NewHoistPt = HoistPt;
561 NewHoistPt = NewHoistBB->getTerminator();
564 SmallPtrSet<const BasicBlock *, 2> WL;
568 if (K == InsKind::Scalar) {
569 if (safeToHoistScalar(NewHoistBB, WL, NumBBsOnAllPaths)) {
570 // Extend HoistPt to NewHoistPt.
571 HoistPt = NewHoistPt;
572 HoistBB = NewHoistBB;
576 // When NewBB already contains an instruction to be hoisted, the
577 // expression is needed on all paths.
578 // Check that the hoisted expression is needed on all paths: it is
579 // unsafe to hoist loads to a place where there may be a path not
580 // loading from the same address: for instance there may be a branch on
581 // which the address of the load may not be initialized.
582 if ((HoistBB == NewHoistBB || BB == NewHoistBB ||
583 hoistingFromAllPaths(NewHoistBB, WL)) &&
584 // Also check that it is safe to move the load or store from HoistPt
585 // to NewHoistPt, and from Insn to NewHoistPt.
586 safeToHoistLdSt(NewHoistPt, HoistPt, UD, K, NumBBsOnAllPaths) &&
587 safeToHoistLdSt(NewHoistPt, Insn, MSSA->getMemoryAccess(Insn),
588 K, NumBBsOnAllPaths)) {
589 // Extend HoistPt to NewHoistPt.
590 HoistPt = NewHoistPt;
591 HoistBB = NewHoistBB;
596 // At this point it is not safe to extend the current hoisting to
597 // NewHoistPt: save the hoisting list so far.
598 if (std::distance(Start, II) > 1)
599 HPL.push_back({HoistBB, SmallVecInsn(Start, II)});
601 // Start over from BB.
603 if (K != InsKind::Scalar)
604 UD = MSSA->getMemoryAccess(*Start);
607 NumBBsOnAllPaths = MaxNumberOfBBSInPath;
610 // Save the last partition.
611 if (std::distance(Start, II) > 1)
612 HPL.push_back({HoistBB, SmallVecInsn(Start, II)});
615 // Initialize HPL from Map.
616 void computeInsertionPoints(const VNtoInsns &Map, HoistingPointList &HPL,
618 for (const auto &Entry : Map) {
619 if (MaxHoistedThreshold != -1 && ++HoistedCtr > MaxHoistedThreshold)
622 const SmallVecInsn &V = Entry.second;
626 // Compute the insertion point and the list of expressions to be hoisted.
627 SmallVecInsn InstructionsToHoist;
629 if (!hasEH(I->getParent()))
630 InstructionsToHoist.push_back(I);
632 if (!InstructionsToHoist.empty())
633 partitionCandidates(InstructionsToHoist, HPL, K);
637 // Return true when all operands of Instr are available at insertion point
638 // HoistPt. When limiting the number of hoisted expressions, one could hoist
639 // a load without hoisting its access function. So before hoisting any
640 // expression, make sure that all its operands are available at insert point.
641 bool allOperandsAvailable(const Instruction *I,
642 const BasicBlock *HoistPt) const {
643 for (const Use &Op : I->operands())
644 if (const auto *Inst = dyn_cast<Instruction>(&Op))
645 if (!DT->dominates(Inst->getParent(), HoistPt))
651 // Same as allOperandsAvailable with recursive check for GEP operands.
652 bool allGepOperandsAvailable(const Instruction *I,
653 const BasicBlock *HoistPt) const {
654 for (const Use &Op : I->operands())
655 if (const auto *Inst = dyn_cast<Instruction>(&Op))
656 if (!DT->dominates(Inst->getParent(), HoistPt)) {
657 if (const GetElementPtrInst *GepOp =
658 dyn_cast<GetElementPtrInst>(Inst)) {
659 if (!allGepOperandsAvailable(GepOp, HoistPt))
661 // Gep is available if all operands of GepOp are available.
663 // Gep is not available if it has operands other than GEPs that are
664 // defined in blocks not dominating HoistPt.
671 // Make all operands of the GEP available.
672 void makeGepsAvailable(Instruction *Repl, BasicBlock *HoistPt,
673 const SmallVecInsn &InstructionsToHoist,
674 Instruction *Gep) const {
675 assert(allGepOperandsAvailable(Gep, HoistPt) &&
676 "GEP operands not available");
678 Instruction *ClonedGep = Gep->clone();
679 for (unsigned i = 0, e = Gep->getNumOperands(); i != e; ++i)
680 if (Instruction *Op = dyn_cast<Instruction>(Gep->getOperand(i))) {
682 // Check whether the operand is already available.
683 if (DT->dominates(Op->getParent(), HoistPt))
686 // As a GEP can refer to other GEPs, recursively make all the operands
687 // of this GEP available at HoistPt.
688 if (GetElementPtrInst *GepOp = dyn_cast<GetElementPtrInst>(Op))
689 makeGepsAvailable(ClonedGep, HoistPt, InstructionsToHoist, GepOp);
692 // Copy Gep and replace its uses in Repl with ClonedGep.
693 ClonedGep->insertBefore(HoistPt->getTerminator());
695 // Conservatively discard any optimization hints, they may differ on the
697 ClonedGep->dropUnknownNonDebugMetadata();
699 // If we have optimization hints which agree with each other along different
700 // paths, preserve them.
701 for (const Instruction *OtherInst : InstructionsToHoist) {
702 const GetElementPtrInst *OtherGep;
703 if (auto *OtherLd = dyn_cast<LoadInst>(OtherInst))
704 OtherGep = cast<GetElementPtrInst>(OtherLd->getPointerOperand());
706 OtherGep = cast<GetElementPtrInst>(
707 cast<StoreInst>(OtherInst)->getPointerOperand());
708 ClonedGep->andIRFlags(OtherGep);
711 // Replace uses of Gep with ClonedGep in Repl.
712 Repl->replaceUsesOfWith(Gep, ClonedGep);
715 // In the case Repl is a load or a store, we make all their GEPs
716 // available: GEPs are not hoisted by default to avoid the address
717 // computations to be hoisted without the associated load or store.
718 bool makeGepOperandsAvailable(Instruction *Repl, BasicBlock *HoistPt,
719 const SmallVecInsn &InstructionsToHoist) const {
720 // Check whether the GEP of a ld/st can be synthesized at HoistPt.
721 GetElementPtrInst *Gep = nullptr;
722 Instruction *Val = nullptr;
723 if (auto *Ld = dyn_cast<LoadInst>(Repl)) {
724 Gep = dyn_cast<GetElementPtrInst>(Ld->getPointerOperand());
725 } else if (auto *St = dyn_cast<StoreInst>(Repl)) {
726 Gep = dyn_cast<GetElementPtrInst>(St->getPointerOperand());
727 Val = dyn_cast<Instruction>(St->getValueOperand());
728 // Check that the stored value is available.
730 if (isa<GetElementPtrInst>(Val)) {
731 // Check whether we can compute the GEP at HoistPt.
732 if (!allGepOperandsAvailable(Val, HoistPt))
734 } else if (!DT->dominates(Val->getParent(), HoistPt))
739 // Check whether we can compute the Gep at HoistPt.
740 if (!Gep || !allGepOperandsAvailable(Gep, HoistPt))
743 makeGepsAvailable(Repl, HoistPt, InstructionsToHoist, Gep);
745 if (Val && isa<GetElementPtrInst>(Val))
746 makeGepsAvailable(Repl, HoistPt, InstructionsToHoist, Val);
751 std::pair<unsigned, unsigned> hoist(HoistingPointList &HPL) {
752 unsigned NI = 0, NL = 0, NS = 0, NC = 0, NR = 0;
753 for (const HoistingPointInfo &HP : HPL) {
754 // Find out whether we already have one of the instructions in HoistPt,
755 // in which case we do not have to move it.
756 BasicBlock *HoistPt = HP.first;
757 const SmallVecInsn &InstructionsToHoist = HP.second;
758 Instruction *Repl = nullptr;
759 for (Instruction *I : InstructionsToHoist)
760 if (I->getParent() == HoistPt)
761 // If there are two instructions in HoistPt to be hoisted in place:
762 // update Repl to be the first one, such that we can rename the uses
763 // of the second based on the first.
764 if (!Repl || firstInBB(I, Repl))
767 // Keep track of whether we moved the instruction so we know whether we
768 // should move the MemoryAccess.
769 bool MoveAccess = true;
771 // Repl is already in HoistPt: it remains in place.
772 assert(allOperandsAvailable(Repl, HoistPt) &&
773 "instruction depends on operands that are not available");
776 // When we do not find Repl in HoistPt, select the first in the list
777 // and move it to HoistPt.
778 Repl = InstructionsToHoist.front();
780 // We can move Repl in HoistPt only when all operands are available.
781 // The order in which hoistings are done may influence the availability
783 if (!allOperandsAvailable(Repl, HoistPt)) {
785 // When HoistingGeps there is nothing more we can do to make the
786 // operands available: just continue.
790 // When not HoistingGeps we need to copy the GEPs.
791 if (!makeGepOperandsAvailable(Repl, HoistPt, InstructionsToHoist))
795 // Move the instruction at the end of HoistPt.
796 Instruction *Last = HoistPt->getTerminator();
797 MD->removeInstruction(Repl);
798 Repl->moveBefore(Last);
800 DFSNumber[Repl] = DFSNumber[Last]++;
803 MemoryAccess *NewMemAcc = MSSA->getMemoryAccess(Repl);
806 if (MemoryUseOrDef *OldMemAcc =
807 dyn_cast_or_null<MemoryUseOrDef>(NewMemAcc)) {
808 // The definition of this ld/st will not change: ld/st hoisting is
809 // legal when the ld/st is not moved past its current definition.
810 MemoryAccess *Def = OldMemAcc->getDefiningAccess();
812 MSSA->createMemoryAccessInBB(Repl, Def, HoistPt, MemorySSA::End);
813 OldMemAcc->replaceAllUsesWith(NewMemAcc);
814 MSSA->removeMemoryAccess(OldMemAcc);
818 if (isa<LoadInst>(Repl))
820 else if (isa<StoreInst>(Repl))
822 else if (isa<CallInst>(Repl))
827 // Remove and rename all other instructions.
828 for (Instruction *I : InstructionsToHoist)
831 if (auto *ReplacementLoad = dyn_cast<LoadInst>(Repl)) {
832 ReplacementLoad->setAlignment(
833 std::min(ReplacementLoad->getAlignment(),
834 cast<LoadInst>(I)->getAlignment()));
836 } else if (auto *ReplacementStore = dyn_cast<StoreInst>(Repl)) {
837 ReplacementStore->setAlignment(
838 std::min(ReplacementStore->getAlignment(),
839 cast<StoreInst>(I)->getAlignment()));
841 } else if (auto *ReplacementAlloca = dyn_cast<AllocaInst>(Repl)) {
842 ReplacementAlloca->setAlignment(
843 std::max(ReplacementAlloca->getAlignment(),
844 cast<AllocaInst>(I)->getAlignment()));
845 } else if (isa<CallInst>(Repl)) {
850 // Update the uses of the old MSSA access with NewMemAcc.
851 MemoryAccess *OldMA = MSSA->getMemoryAccess(I);
852 OldMA->replaceAllUsesWith(NewMemAcc);
853 MSSA->removeMemoryAccess(OldMA);
857 combineKnownMetadata(Repl, I);
858 I->replaceAllUsesWith(Repl);
859 // Also invalidate the Alias Analysis cache.
860 MD->removeInstruction(I);
861 I->eraseFromParent();
864 // Remove MemorySSA phi nodes with the same arguments.
866 SmallPtrSet<MemoryPhi *, 4> UsePhis;
867 for (User *U : NewMemAcc->users())
868 if (MemoryPhi *Phi = dyn_cast<MemoryPhi>(U))
871 for (auto *Phi : UsePhis) {
872 auto In = Phi->incoming_values();
873 if (all_of(In, [&](Use &U) { return U == NewMemAcc; })) {
874 Phi->replaceAllUsesWith(NewMemAcc);
875 MSSA->removeMemoryAccess(Phi);
881 NumHoisted += NL + NS + NC + NI;
883 NumLoadsHoisted += NL;
884 NumStoresHoisted += NS;
885 NumCallsHoisted += NC;
886 return {NI, NL + NC + NS};
889 // Hoist all expressions. Returns Number of scalars hoisted
890 // and number of non-scalars hoisted.
891 std::pair<unsigned, unsigned> hoistExpressions(Function &F) {
896 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) {
897 int InstructionNb = 0;
898 for (Instruction &I1 : *BB) {
899 // Only hoist the first instructions in BB up to MaxDepthInBB. Hoisting
900 // deeper may increase the register pressure and compilation time.
901 if (MaxDepthInBB != -1 && InstructionNb++ >= MaxDepthInBB)
904 // Do not value number terminator instructions.
905 if (isa<TerminatorInst>(&I1))
908 if (auto *Load = dyn_cast<LoadInst>(&I1))
910 else if (auto *Store = dyn_cast<StoreInst>(&I1))
911 SI.insert(Store, VN);
912 else if (auto *Call = dyn_cast<CallInst>(&I1)) {
913 if (auto *Intr = dyn_cast<IntrinsicInst>(Call)) {
914 if (isa<DbgInfoIntrinsic>(Intr) ||
915 Intr->getIntrinsicID() == Intrinsic::assume)
918 if (Call->mayHaveSideEffects())
921 if (Call->isConvergent())
925 } else if (HoistingGeps || !isa<GetElementPtrInst>(&I1))
926 // Do not hoist scalars past calls that may write to memory because
927 // that could result in spills later. geps are handled separately.
928 // TODO: We can relax this for targets like AArch64 as they have more
929 // registers than X86.
934 HoistingPointList HPL;
935 computeInsertionPoints(II.getVNTable(), HPL, InsKind::Scalar);
936 computeInsertionPoints(LI.getVNTable(), HPL, InsKind::Load);
937 computeInsertionPoints(SI.getVNTable(), HPL, InsKind::Store);
938 computeInsertionPoints(CI.getScalarVNTable(), HPL, InsKind::Scalar);
939 computeInsertionPoints(CI.getLoadVNTable(), HPL, InsKind::Load);
940 computeInsertionPoints(CI.getStoreVNTable(), HPL, InsKind::Store);
945 class GVNHoistLegacyPass : public FunctionPass {
949 GVNHoistLegacyPass() : FunctionPass(ID) {
950 initializeGVNHoistLegacyPassPass(*PassRegistry::getPassRegistry());
953 bool runOnFunction(Function &F) override {
956 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
957 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
958 auto &MD = getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
959 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
961 GVNHoist G(&DT, &AA, &MD, &MSSA);
965 void getAnalysisUsage(AnalysisUsage &AU) const override {
966 AU.addRequired<DominatorTreeWrapperPass>();
967 AU.addRequired<AAResultsWrapperPass>();
968 AU.addRequired<MemoryDependenceWrapperPass>();
969 AU.addRequired<MemorySSAWrapperPass>();
970 AU.addPreserved<DominatorTreeWrapperPass>();
971 AU.addPreserved<MemorySSAWrapperPass>();
976 PreservedAnalyses GVNHoistPass::run(Function &F, FunctionAnalysisManager &AM) {
977 DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F);
978 AliasAnalysis &AA = AM.getResult<AAManager>(F);
979 MemoryDependenceResults &MD = AM.getResult<MemoryDependenceAnalysis>(F);
980 MemorySSA &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA();
981 GVNHoist G(&DT, &AA, &MD, &MSSA);
983 return PreservedAnalyses::all();
985 PreservedAnalyses PA;
986 PA.preserve<DominatorTreeAnalysis>();
987 PA.preserve<MemorySSAAnalysis>();
991 char GVNHoistLegacyPass::ID = 0;
992 INITIALIZE_PASS_BEGIN(GVNHoistLegacyPass, "gvn-hoist",
993 "Early GVN Hoisting of Expressions", false, false)
994 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
995 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
996 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
997 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
998 INITIALIZE_PASS_END(GVNHoistLegacyPass, "gvn-hoist",
999 "Early GVN Hoisting of Expressions", false, false)
1001 FunctionPass *llvm::createGVNHoistPass() { return new GVNHoistLegacyPass(); }