1 //===- MergedLoadStoreMotion.cpp - merge and hoist/sink load/stores -------===//
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 //===----------------------------------------------------------------------===//
11 //! \brief This pass performs merges of loads and stores on both sides of a
12 // diamond (hammock). It hoists the loads and sinks the stores.
14 // The algorithm iteratively hoists two loads to the same address out of a
15 // diamond (hammock) and merges them into a single load in the header. Similar
16 // it sinks and merges two stores to the tail block (footer). The algorithm
17 // iterates over the instructions of one side of the diamond and attempts to
18 // find a matching load/store on the other side. It hoists / sinks when it
19 // thinks it safe to do so. This optimization helps with eg. hiding load
20 // latencies, triggering if-conversion, and reducing static code size.
22 //===----------------------------------------------------------------------===//
26 // Diamond shaped code before merge:
29 // br %cond, label %if.then, label %if.else
34 // %lt = load %addr_l %le = load %addr_l
35 // <use %lt> <use %le>
37 // store %st, %addr_s store %se, %addr_s
38 // br label %if.end br label %if.end
45 // Diamond shaped code after merge:
49 // br %cond, label %if.then, label %if.else
56 // br label %if.end br label %if.end
61 // %s.sink = phi [%st, if.then], [%se, if.else]
63 // store %s.sink, %addr_s
67 //===----------------------- TODO -----------------------------------------===//
69 // 1) Generalize to regions other than diamonds
70 // 2) Be more aggressive merging memory operations
71 // Note that both changes require register pressure control
73 //===----------------------------------------------------------------------===//
75 #include "llvm/Transforms/Scalar.h"
76 #include "llvm/ADT/SetVector.h"
77 #include "llvm/ADT/SmallPtrSet.h"
78 #include "llvm/ADT/Statistic.h"
79 #include "llvm/Analysis/AliasAnalysis.h"
80 #include "llvm/Analysis/CFG.h"
81 #include "llvm/Analysis/Loads.h"
82 #include "llvm/Analysis/MemoryBuiltins.h"
83 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
84 #include "llvm/IR/Metadata.h"
85 #include "llvm/IR/PatternMatch.h"
86 #include "llvm/Support/Allocator.h"
87 #include "llvm/Support/CommandLine.h"
88 #include "llvm/Support/Debug.h"
89 #include "llvm/Target/TargetLibraryInfo.h"
90 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
91 #include "llvm/Transforms/Utils/SSAUpdater.h"
95 #define DEBUG_TYPE "mldst-motion"
97 //===----------------------------------------------------------------------===//
98 // MergedLoadStoreMotion Pass
99 //===----------------------------------------------------------------------===//
101 EnableMLSM("mlsm", cl::desc("Enable motion of merged load and store"),
105 class MergedLoadStoreMotion : public FunctionPass {
107 MemoryDependenceAnalysis *MD;
110 static char ID; // Pass identification, replacement for typeid
111 explicit MergedLoadStoreMotion(void) : FunctionPass(ID), MD(nullptr) {
112 initializeMergedLoadStoreMotionPass(*PassRegistry::getPassRegistry());
115 bool runOnFunction(Function &F) override;
118 // This transformation requires dominator postdominator info
119 void getAnalysisUsage(AnalysisUsage &AU) const override {
120 AU.addRequired<TargetLibraryInfo>();
121 AU.addRequired<MemoryDependenceAnalysis>();
122 AU.addRequired<AliasAnalysis>();
123 AU.addPreserved<AliasAnalysis>();
129 /// \brief Remove instruction from parent and update memory dependence
132 void removeInstruction(Instruction *Inst);
133 BasicBlock *getDiamondTail(BasicBlock *BB);
134 bool isDiamondHead(BasicBlock *BB);
135 // Routines for hoisting loads
136 bool isLoadHoistBarrier(Instruction *Inst);
137 LoadInst *canHoistFromBlock(BasicBlock *BB, LoadInst *LI);
138 void hoistInstruction(BasicBlock *BB, Instruction *HoistCand,
139 Instruction *ElseInst);
140 bool isSafeToHoist(Instruction *I) const;
141 bool hoistLoad(BasicBlock *BB, LoadInst *HoistCand, LoadInst *ElseInst);
142 bool mergeLoads(BasicBlock *BB);
143 // Routines for sinking stores
144 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI);
145 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1);
146 bool isStoreSinkBarrier(Instruction *Inst);
147 bool sinkStore(BasicBlock *BB, StoreInst *SinkCand, StoreInst *ElseInst);
148 bool mergeStores(BasicBlock *BB);
149 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity,
150 // where Size0 and Size1 are the #instructions on the two sides of
151 // the diamond. The constant chosen here is arbitrary. Compiler Time
152 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl.
153 const int MagicCompileTimeControl = 250;
156 char MergedLoadStoreMotion::ID = 0;
160 /// \brief createMergedLoadStoreMotionPass - The public interface to this file.
162 FunctionPass *llvm::createMergedLoadStoreMotionPass() {
163 return new MergedLoadStoreMotion();
166 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotion, "mldst-motion",
167 "MergedLoadStoreMotion", false, false)
168 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
169 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
170 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
171 INITIALIZE_PASS_END(MergedLoadStoreMotion, "mldst-motion",
172 "MergedLoadStoreMotion", false, false)
175 /// \brief Remove instruction from parent and update memory dependence analysis.
177 void MergedLoadStoreMotion::removeInstruction(Instruction *Inst) {
178 // Notify the memory dependence analysis.
180 MD->removeInstruction(Inst);
181 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
182 MD->invalidateCachedPointerInfo(LI->getPointerOperand());
183 if (Inst->getType()->getScalarType()->isPointerTy()) {
184 MD->invalidateCachedPointerInfo(Inst);
187 Inst->eraseFromParent();
191 /// \brief Return tail block of a diamond.
193 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) {
194 assert(isDiamondHead(BB) && "Basic block is not head of a diamond");
195 BranchInst *BI = (BranchInst *)(BB->getTerminator());
196 BasicBlock *Succ0 = BI->getSuccessor(0);
197 BasicBlock *Tail = Succ0->getTerminator()->getSuccessor(0);
202 /// \brief True when BB is the head of a diamond (hammock)
204 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) {
207 if (!isa<BranchInst>(BB->getTerminator()))
209 if (BB->getTerminator()->getNumSuccessors() != 2)
212 BranchInst *BI = (BranchInst *)(BB->getTerminator());
213 BasicBlock *Succ0 = BI->getSuccessor(0);
214 BasicBlock *Succ1 = BI->getSuccessor(1);
216 if (!Succ0->getSinglePredecessor() ||
217 Succ0->getTerminator()->getNumSuccessors() != 1)
219 if (!Succ1->getSinglePredecessor() ||
220 Succ1->getTerminator()->getNumSuccessors() != 1)
223 BasicBlock *Tail = Succ0->getTerminator()->getSuccessor(0);
225 if (Succ1->getTerminator()->getSuccessor(0) != Tail)
231 /// \brief True when instruction is a hoist barrier for a load
233 /// Whenever an instruction could possibly modify the value
234 /// being loaded or protect against the load from happening
235 /// it is considered a hoist barrier.
237 bool MergedLoadStoreMotion::isLoadHoistBarrier(Instruction *Inst) {
238 // FIXME: A call with no side effects should not be a barrier.
239 // Aren't all such calls covered by mayHaveSideEffects() below?
240 // Then this check can be removed.
241 if (isa<CallInst>(Inst))
243 if (isa<TerminatorInst>(Inst))
245 // Note: mayHaveSideEffects covers all instructions that could
246 // trigger a change to state. Eg. in-flight stores have to be executed
247 // before ordered loads or fences, calls could invoke functions that store
248 // data to memory etc.
249 if (Inst->mayHaveSideEffects()) {
252 DEBUG(dbgs() << "No Hoist Barrier\n");
257 /// \brief Decide if a load can be hoisted
259 /// When there is a load in \p BB to the same address as \p LI
260 /// and it can be hoisted from \p BB, return that load.
261 /// Otherwise return Null.
263 LoadInst *MergedLoadStoreMotion::canHoistFromBlock(BasicBlock *BB,
265 LoadInst *I = nullptr;
266 assert(isa<LoadInst>(LI));
267 if (LI->isUsedOutsideOfBlock(LI->getParent()))
270 for (BasicBlock::iterator BBI = BB->begin(), BBE = BB->end(); BBI != BBE;
272 Instruction *Inst = BBI;
274 // Only merge and hoist loads when their result in used only in BB
275 if (isLoadHoistBarrier(Inst))
277 if (!isa<LoadInst>(Inst))
279 if (Inst->isUsedOutsideOfBlock(Inst->getParent()))
282 AliasAnalysis::Location LocLI = AA->getLocation(LI);
283 AliasAnalysis::Location LocInst = AA->getLocation((LoadInst *)Inst);
284 if (AA->isMustAlias(LocLI, LocInst) && LI->getType() == Inst->getType()) {
285 I = (LoadInst *)Inst;
293 /// \brief Merge two equivalent instructions \p HoistCand and \p ElseInst into
296 /// BB is the head of a diamond
298 void MergedLoadStoreMotion::hoistInstruction(BasicBlock *BB,
299 Instruction *HoistCand,
300 Instruction *ElseInst) {
301 DEBUG(dbgs() << " Hoist Instruction into BB \n"; BB->dump();
302 dbgs() << "Instruction Left\n"; HoistCand->dump(); dbgs() << "\n";
303 dbgs() << "Instruction Right\n"; ElseInst->dump(); dbgs() << "\n");
304 // Hoist the instruction.
305 assert(HoistCand->getParent() != BB);
307 // Intersect optional metadata.
308 HoistCand->intersectOptionalDataWith(ElseInst);
309 HoistCand->dropUnknownMetadata();
311 // Prepend point for instruction insert
312 Instruction *HoistPt = BB->getTerminator();
314 // Merged instruction
315 Instruction *HoistedInst = HoistCand->clone();
317 // Notify AA of the new value.
318 if (isa<LoadInst>(HoistCand))
319 AA->copyValue(HoistCand, HoistedInst);
321 // Hoist instruction.
322 HoistedInst->insertBefore(HoistPt);
324 HoistCand->replaceAllUsesWith(HoistedInst);
325 removeInstruction(HoistCand);
326 // Replace the else block instruction.
327 ElseInst->replaceAllUsesWith(HoistedInst);
328 removeInstruction(ElseInst);
332 /// \brief Return true if no operand of \p I is defined in I's parent block
334 bool MergedLoadStoreMotion::isSafeToHoist(Instruction *I) const {
335 BasicBlock *Parent = I->getParent();
336 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
337 Instruction *Instr = dyn_cast<Instruction>(I->getOperand(i));
338 if (Instr && Instr->getParent() == Parent)
345 /// \brief Merge two equivalent loads and GEPs and hoist into diamond head
347 bool MergedLoadStoreMotion::hoistLoad(BasicBlock *BB, LoadInst *L0,
349 // Only one definition?
350 Instruction *A0 = dyn_cast<Instruction>(L0->getPointerOperand());
351 Instruction *A1 = dyn_cast<Instruction>(L1->getPointerOperand());
352 if (A0 && A1 && A0->isIdenticalTo(A1) && isSafeToHoist(A0) &&
353 A0->hasOneUse() && (A0->getParent() == L0->getParent()) &&
354 A1->hasOneUse() && (A1->getParent() == L1->getParent()) &&
355 isa<GetElementPtrInst>(A0)) {
356 DEBUG(dbgs() << "Hoist Instruction into BB \n"; BB->dump();
357 dbgs() << "Instruction Left\n"; L0->dump(); dbgs() << "\n";
358 dbgs() << "Instruction Right\n"; L1->dump(); dbgs() << "\n");
359 hoistInstruction(BB, A0, A1);
360 hoistInstruction(BB, L0, L1);
367 /// \brief Try to hoist two loads to same address into diamond header
369 /// Starting from a diamond head block, iterate over the instructions in one
370 /// successor block and try to match a load in the second successor.
372 bool MergedLoadStoreMotion::mergeLoads(BasicBlock *BB) {
373 bool MergedLoads = false;
374 assert(isDiamondHead(BB));
375 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
376 BasicBlock *Succ0 = BI->getSuccessor(0);
377 BasicBlock *Succ1 = BI->getSuccessor(1);
378 // #Instructions in Succ1 for Compile Time Control
379 int Size1 = Succ1->size();
381 for (BasicBlock::iterator BBI = Succ0->begin(), BBE = Succ0->end();
384 Instruction *I = BBI;
386 if (isLoadHoistBarrier(I))
389 // Only move non-simple (atomic, volatile) loads.
390 if (!isa<LoadInst>(I))
393 LoadInst *L0 = (LoadInst *)I;
398 if (NLoads * Size1 >= MagicCompileTimeControl)
400 if (LoadInst *L1 = canHoistFromBlock(Succ1, L0)) {
401 bool Res = hoistLoad(BB, L0, L1);
403 // Don't attempt to hoist above loads that had not been hoisted.
412 /// \brief True when instruction is sink barrier for a store
414 bool MergedLoadStoreMotion::isStoreSinkBarrier(Instruction *Inst) {
415 if (isa<CallInst>(Inst))
417 if (isa<TerminatorInst>(Inst) && !isa<BranchInst>(Inst))
419 // Note: mayHaveSideEffects covers all instructions that could
420 // trigger a change to state. Eg. in-flight stores have to be executed
421 // before ordered loads or fences, calls could invoke functions that store
422 // data to memory etc.
423 if (!isa<StoreInst>(Inst) && Inst->mayHaveSideEffects()) {
426 DEBUG(dbgs() << "No Sink Barrier\n");
431 /// \brief Check if \p BB contains a store to the same address as \p SI
433 /// \return The store in \p when it is safe to sink. Otherwise return Null.
435 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB,
438 DEBUG(dbgs() << "can Sink? : "; SI->dump(); dbgs() << "\n");
439 for (BasicBlock::reverse_iterator RBI = BB->rbegin(), RBE = BB->rend();
441 Instruction *Inst = &*RBI;
443 // Only move loads if they are used in the block.
444 if (isStoreSinkBarrier(Inst))
446 if (isa<StoreInst>(Inst)) {
447 AliasAnalysis::Location LocSI = AA->getLocation(SI);
448 AliasAnalysis::Location LocInst = AA->getLocation((StoreInst *)Inst);
449 if (AA->isMustAlias(LocSI, LocInst)) {
450 I = (StoreInst *)Inst;
459 /// \brief Create a PHI node in BB for the operands of S0 and S1
461 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0,
463 // Create a phi if the values mismatch.
465 Value *Opd1 = S0->getValueOperand();
466 Value *Opd2 = S1->getValueOperand();
468 NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink",
470 NewPN->addIncoming(Opd1, S0->getParent());
471 NewPN->addIncoming(Opd2, S1->getParent());
472 if (NewPN->getType()->getScalarType()->isPointerTy()) {
473 // Notify AA of the new value.
474 AA->copyValue(Opd1, NewPN);
475 AA->copyValue(Opd2, NewPN);
476 // AA needs to be informed when a PHI-use of the pointer value is added
477 for (unsigned I = 0, E = NewPN->getNumIncomingValues(); I != E; ++I) {
478 unsigned J = PHINode::getOperandNumForIncomingValue(I);
479 AA->addEscapingUse(NewPN->getOperandUse(J));
482 MD->invalidateCachedPointerInfo(NewPN);
489 /// \brief Merge two stores to same address and sink into \p BB
491 /// Also sinks GEP instruction computing the store address
493 bool MergedLoadStoreMotion::sinkStore(BasicBlock *BB, StoreInst *S0,
495 // Only one definition?
496 Instruction *A0 = dyn_cast<Instruction>(S0->getPointerOperand());
497 Instruction *A1 = dyn_cast<Instruction>(S1->getPointerOperand());
498 if (A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() &&
499 (A0->getParent() == S0->getParent()) && A1->hasOneUse() &&
500 (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0)) {
501 DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump();
502 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n";
503 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n");
504 // Hoist the instruction.
505 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt();
506 // Intersect optional metadata.
507 S0->intersectOptionalDataWith(S1);
508 S0->dropUnknownMetadata();
510 // Create the new store to be inserted at the join point.
511 StoreInst *SNew = (StoreInst *)(S0->clone());
512 Instruction *ANew = A0->clone();
513 AA->copyValue(S0, SNew);
514 SNew->insertBefore(InsertPt);
515 ANew->insertBefore(SNew);
517 assert(S0->getParent() == A0->getParent());
518 assert(S1->getParent() == A1->getParent());
520 PHINode *NewPN = getPHIOperand(BB, S0, S1);
521 // New PHI operand? Use it.
523 SNew->setOperand(0, NewPN);
524 removeInstruction(S0);
525 removeInstruction(S1);
526 A0->replaceAllUsesWith(ANew);
527 removeInstruction(A0);
528 A1->replaceAllUsesWith(ANew);
529 removeInstruction(A1);
536 /// \brief True when two stores are equivalent and can sink into the footer
538 /// Starting from a diamond tail block, iterate over the instructions in one
539 /// predecessor block and try to match a store in the second predecessor.
541 bool MergedLoadStoreMotion::mergeStores(BasicBlock *T) {
543 bool MergedStores = false;
544 assert(T && "Footer of a diamond cannot be empty");
546 pred_iterator PI = pred_begin(T), E = pred_end(T);
548 BasicBlock *Pred0 = *PI;
550 BasicBlock *Pred1 = *PI;
552 // tail block of a diamond/hammock?
556 return false; // No. More than 2 predecessors.
558 // #Instructions in Succ1 for Compile Time Control
559 int Size1 = Pred1->size();
562 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend();
565 Instruction *I = &*RBI;
567 if (isStoreSinkBarrier(I))
569 // Sink move non-simple (atomic, volatile) stores
570 if (!isa<StoreInst>(I))
572 StoreInst *S0 = (StoreInst *)I;
577 if (NStores * Size1 >= MagicCompileTimeControl)
579 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) {
580 bool Res = sinkStore(T, S0, S1);
582 // Don't attempt to sink below stores that had to stick around
583 // But after removal of a store and some of its feeding
584 // instruction search again from the beginning since the iterator
585 // is likely stale at this point.
589 RBI = Pred0->rbegin();
591 DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump());
598 /// \brief Run the transformation for each function
600 bool MergedLoadStoreMotion::runOnFunction(Function &F) {
601 MD = &getAnalysis<MemoryDependenceAnalysis>();
602 AA = &getAnalysis<AliasAnalysis>();
604 bool Changed = false;
607 DEBUG(dbgs() << "Instruction Merger\n");
609 // Merge unconditional branches, allowing PRE to catch more
610 // optimization opportunities.
611 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) {
612 BasicBlock *BB = FI++;
614 // Hoist equivalent loads and sink stores
615 // outside diamonds when possible
616 // Run outside core GVN
617 if (isDiamondHead(BB)) {
618 Changed |= mergeLoads(BB);
619 Changed |= mergeStores(getDiamondTail(BB));