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/Analysis/TargetLibraryInfo.h"
85 #include "llvm/IR/Metadata.h"
86 #include "llvm/IR/PatternMatch.h"
87 #include "llvm/Support/Allocator.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Debug.h"
90 #include "llvm/Support/raw_ostream.h"
91 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
92 #include "llvm/Transforms/Utils/SSAUpdater.h"
96 #define DEBUG_TYPE "mldst-motion"
98 //===----------------------------------------------------------------------===//
99 // MergedLoadStoreMotion Pass
100 //===----------------------------------------------------------------------===//
103 class MergedLoadStoreMotion : public FunctionPass {
105 MemoryDependenceAnalysis *MD;
108 static char ID; // Pass identification, replacement for typeid
109 explicit MergedLoadStoreMotion(void)
110 : FunctionPass(ID), MD(nullptr), MagicCompileTimeControl(250) {
111 initializeMergedLoadStoreMotionPass(*PassRegistry::getPassRegistry());
114 bool runOnFunction(Function &F) override;
117 // This transformation requires dominator postdominator info
118 void getAnalysisUsage(AnalysisUsage &AU) const override {
119 AU.addRequired<TargetLibraryInfoWrapperPass>();
120 AU.addRequired<AliasAnalysis>();
121 AU.addPreserved<MemoryDependenceAnalysis>();
122 AU.addPreserved<AliasAnalysis>();
128 /// \brief Remove instruction from parent and update memory dependence
131 void removeInstruction(Instruction *Inst);
132 BasicBlock *getDiamondTail(BasicBlock *BB);
133 bool isDiamondHead(BasicBlock *BB);
134 // Routines for hoisting loads
135 bool isLoadHoistBarrierInRange(const Instruction& Start,
136 const Instruction& End,
138 LoadInst *canHoistFromBlock(BasicBlock *BB, LoadInst *LI);
139 void hoistInstruction(BasicBlock *BB, Instruction *HoistCand,
140 Instruction *ElseInst);
141 bool isSafeToHoist(Instruction *I) const;
142 bool hoistLoad(BasicBlock *BB, LoadInst *HoistCand, LoadInst *ElseInst);
143 bool mergeLoads(BasicBlock *BB);
144 // Routines for sinking stores
145 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI);
146 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1);
147 bool isStoreSinkBarrierInRange(const Instruction &Start,
148 const Instruction &End, MemoryLocation Loc);
149 bool sinkStore(BasicBlock *BB, StoreInst *SinkCand, StoreInst *ElseInst);
150 bool mergeStores(BasicBlock *BB);
151 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity,
152 // where Size0 and Size1 are the #instructions on the two sides of
153 // the diamond. The constant chosen here is arbitrary. Compiler Time
154 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl.
155 const int MagicCompileTimeControl;
158 char MergedLoadStoreMotion::ID = 0;
162 /// \brief createMergedLoadStoreMotionPass - The public interface to this file.
164 FunctionPass *llvm::createMergedLoadStoreMotionPass() {
165 return new MergedLoadStoreMotion();
168 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotion, "mldst-motion",
169 "MergedLoadStoreMotion", false, false)
170 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
171 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
172 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
173 INITIALIZE_PASS_END(MergedLoadStoreMotion, "mldst-motion",
174 "MergedLoadStoreMotion", false, false)
177 /// \brief Remove instruction from parent and update memory dependence analysis.
179 void MergedLoadStoreMotion::removeInstruction(Instruction *Inst) {
180 // Notify the memory dependence analysis.
182 MD->removeInstruction(Inst);
183 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
184 MD->invalidateCachedPointerInfo(LI->getPointerOperand());
185 if (Inst->getType()->getScalarType()->isPointerTy()) {
186 MD->invalidateCachedPointerInfo(Inst);
189 Inst->eraseFromParent();
193 /// \brief Return tail block of a diamond.
195 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) {
196 assert(isDiamondHead(BB) && "Basic block is not head of a diamond");
197 BranchInst *BI = (BranchInst *)(BB->getTerminator());
198 BasicBlock *Succ0 = BI->getSuccessor(0);
199 BasicBlock *Tail = Succ0->getTerminator()->getSuccessor(0);
204 /// \brief True when BB is the head of a diamond (hammock)
206 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) {
209 if (!isa<BranchInst>(BB->getTerminator()))
211 if (BB->getTerminator()->getNumSuccessors() != 2)
214 BranchInst *BI = (BranchInst *)(BB->getTerminator());
215 BasicBlock *Succ0 = BI->getSuccessor(0);
216 BasicBlock *Succ1 = BI->getSuccessor(1);
218 if (!Succ0->getSinglePredecessor() ||
219 Succ0->getTerminator()->getNumSuccessors() != 1)
221 if (!Succ1->getSinglePredecessor() ||
222 Succ1->getTerminator()->getNumSuccessors() != 1)
225 BasicBlock *Tail = Succ0->getTerminator()->getSuccessor(0);
227 if (Succ1->getTerminator()->getSuccessor(0) != Tail)
233 /// \brief True when instruction is a hoist barrier for a load
235 /// Whenever an instruction could possibly modify the value
236 /// being loaded or protect against the load from happening
237 /// it is considered a hoist barrier.
240 bool MergedLoadStoreMotion::isLoadHoistBarrierInRange(const Instruction& Start,
241 const Instruction& End,
243 MemoryLocation Loc = MemoryLocation::get(LI);
244 return AA->canInstructionRangeModRef(Start, End, Loc, AliasAnalysis::Mod);
248 /// \brief Decide if a load can be hoisted
250 /// When there is a load in \p BB to the same address as \p LI
251 /// and it can be hoisted from \p BB, return that load.
252 /// Otherwise return Null.
254 LoadInst *MergedLoadStoreMotion::canHoistFromBlock(BasicBlock *BB1,
257 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); BBI != BBE;
259 Instruction *Inst = BBI;
261 // Only merge and hoist loads when their result in used only in BB
262 if (!isa<LoadInst>(Inst) || Inst->isUsedOutsideOfBlock(BB1))
265 LoadInst *Load1 = dyn_cast<LoadInst>(Inst);
266 BasicBlock *BB0 = Load0->getParent();
268 MemoryLocation Loc0 = MemoryLocation::get(Load0);
269 MemoryLocation Loc1 = MemoryLocation::get(Load1);
270 if (AA->isMustAlias(Loc0, Loc1) && Load0->isSameOperationAs(Load1) &&
271 !isLoadHoistBarrierInRange(BB1->front(), *Load1, Load1) &&
272 !isLoadHoistBarrierInRange(BB0->front(), *Load0, Load0)) {
280 /// \brief Merge two equivalent instructions \p HoistCand and \p ElseInst into
283 /// BB is the head of a diamond
285 void MergedLoadStoreMotion::hoistInstruction(BasicBlock *BB,
286 Instruction *HoistCand,
287 Instruction *ElseInst) {
288 DEBUG(dbgs() << " Hoist Instruction into BB \n"; BB->dump();
289 dbgs() << "Instruction Left\n"; HoistCand->dump(); dbgs() << "\n";
290 dbgs() << "Instruction Right\n"; ElseInst->dump(); dbgs() << "\n");
291 // Hoist the instruction.
292 assert(HoistCand->getParent() != BB);
294 // Intersect optional metadata.
295 HoistCand->intersectOptionalDataWith(ElseInst);
296 HoistCand->dropUnknownMetadata();
298 // Prepend point for instruction insert
299 Instruction *HoistPt = BB->getTerminator();
301 // Merged instruction
302 Instruction *HoistedInst = HoistCand->clone();
304 // Notify AA of the new value.
305 if (isa<LoadInst>(HoistCand))
306 AA->copyValue(HoistCand, HoistedInst);
308 // Hoist instruction.
309 HoistedInst->insertBefore(HoistPt);
311 HoistCand->replaceAllUsesWith(HoistedInst);
312 removeInstruction(HoistCand);
313 // Replace the else block instruction.
314 ElseInst->replaceAllUsesWith(HoistedInst);
315 removeInstruction(ElseInst);
319 /// \brief Return true if no operand of \p I is defined in I's parent block
321 bool MergedLoadStoreMotion::isSafeToHoist(Instruction *I) const {
322 BasicBlock *Parent = I->getParent();
323 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
324 Instruction *Instr = dyn_cast<Instruction>(I->getOperand(i));
325 if (Instr && Instr->getParent() == Parent)
332 /// \brief Merge two equivalent loads and GEPs and hoist into diamond head
334 bool MergedLoadStoreMotion::hoistLoad(BasicBlock *BB, LoadInst *L0,
336 // Only one definition?
337 Instruction *A0 = dyn_cast<Instruction>(L0->getPointerOperand());
338 Instruction *A1 = dyn_cast<Instruction>(L1->getPointerOperand());
339 if (A0 && A1 && A0->isIdenticalTo(A1) && isSafeToHoist(A0) &&
340 A0->hasOneUse() && (A0->getParent() == L0->getParent()) &&
341 A1->hasOneUse() && (A1->getParent() == L1->getParent()) &&
342 isa<GetElementPtrInst>(A0)) {
343 DEBUG(dbgs() << "Hoist Instruction into BB \n"; BB->dump();
344 dbgs() << "Instruction Left\n"; L0->dump(); dbgs() << "\n";
345 dbgs() << "Instruction Right\n"; L1->dump(); dbgs() << "\n");
346 hoistInstruction(BB, A0, A1);
347 hoistInstruction(BB, L0, L1);
354 /// \brief Try to hoist two loads to same address into diamond header
356 /// Starting from a diamond head block, iterate over the instructions in one
357 /// successor block and try to match a load in the second successor.
359 bool MergedLoadStoreMotion::mergeLoads(BasicBlock *BB) {
360 bool MergedLoads = false;
361 assert(isDiamondHead(BB));
362 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
363 BasicBlock *Succ0 = BI->getSuccessor(0);
364 BasicBlock *Succ1 = BI->getSuccessor(1);
365 // #Instructions in Succ1 for Compile Time Control
366 int Size1 = Succ1->size();
368 for (BasicBlock::iterator BBI = Succ0->begin(), BBE = Succ0->end();
371 Instruction *I = BBI;
374 // Only move non-simple (atomic, volatile) loads.
375 LoadInst *L0 = dyn_cast<LoadInst>(I);
376 if (!L0 || !L0->isSimple() || L0->isUsedOutsideOfBlock(Succ0))
380 if (NLoads * Size1 >= MagicCompileTimeControl)
382 if (LoadInst *L1 = canHoistFromBlock(Succ1, L0)) {
383 bool Res = hoistLoad(BB, L0, L1);
385 // Don't attempt to hoist above loads that had not been hoisted.
394 /// \brief True when instruction is a sink barrier for a store
397 /// Whenever an instruction could possibly read or modify the
398 /// value being stored or protect against the store from
399 /// happening it is considered a sink barrier.
402 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start,
403 const Instruction &End,
404 MemoryLocation Loc) {
405 return AA->canInstructionRangeModRef(Start, End, Loc, AliasAnalysis::ModRef);
409 /// \brief Check if \p BB contains a store to the same address as \p SI
411 /// \return The store in \p when it is safe to sink. Otherwise return Null.
413 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1,
415 DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n");
416 BasicBlock *BB0 = Store0->getParent();
417 for (BasicBlock::reverse_iterator RBI = BB1->rbegin(), RBE = BB1->rend();
419 Instruction *Inst = &*RBI;
421 if (!isa<StoreInst>(Inst))
424 StoreInst *Store1 = cast<StoreInst>(Inst);
426 MemoryLocation Loc0 = MemoryLocation::get(Store0);
427 MemoryLocation Loc1 = MemoryLocation::get(Store1);
428 if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) &&
429 !isStoreSinkBarrierInRange(*(std::next(BasicBlock::iterator(Store1))),
430 BB1->back(), Loc1) &&
431 !isStoreSinkBarrierInRange(*(std::next(BasicBlock::iterator(Store0))),
432 BB0->back(), Loc0)) {
440 /// \brief Create a PHI node in BB for the operands of S0 and S1
442 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0,
444 // Create a phi if the values mismatch.
446 Value *Opd1 = S0->getValueOperand();
447 Value *Opd2 = S1->getValueOperand();
449 NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink",
451 NewPN->addIncoming(Opd1, S0->getParent());
452 NewPN->addIncoming(Opd2, S1->getParent());
453 if (NewPN->getType()->getScalarType()->isPointerTy()) {
454 // Notify AA of the new value.
455 AA->copyValue(Opd1, NewPN);
456 AA->copyValue(Opd2, NewPN);
457 // AA needs to be informed when a PHI-use of the pointer value is added
458 for (unsigned I = 0, E = NewPN->getNumIncomingValues(); I != E; ++I) {
459 unsigned J = PHINode::getOperandNumForIncomingValue(I);
460 AA->addEscapingUse(NewPN->getOperandUse(J));
463 MD->invalidateCachedPointerInfo(NewPN);
470 /// \brief Merge two stores to same address and sink into \p BB
472 /// Also sinks GEP instruction computing the store address
474 bool MergedLoadStoreMotion::sinkStore(BasicBlock *BB, StoreInst *S0,
476 // Only one definition?
477 Instruction *A0 = dyn_cast<Instruction>(S0->getPointerOperand());
478 Instruction *A1 = dyn_cast<Instruction>(S1->getPointerOperand());
479 if (A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() &&
480 (A0->getParent() == S0->getParent()) && A1->hasOneUse() &&
481 (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0)) {
482 DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump();
483 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n";
484 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n");
485 // Hoist the instruction.
486 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt();
487 // Intersect optional metadata.
488 S0->intersectOptionalDataWith(S1);
489 S0->dropUnknownMetadata();
491 // Create the new store to be inserted at the join point.
492 StoreInst *SNew = (StoreInst *)(S0->clone());
493 Instruction *ANew = A0->clone();
494 AA->copyValue(S0, SNew);
495 SNew->insertBefore(InsertPt);
496 ANew->insertBefore(SNew);
498 assert(S0->getParent() == A0->getParent());
499 assert(S1->getParent() == A1->getParent());
501 PHINode *NewPN = getPHIOperand(BB, S0, S1);
502 // New PHI operand? Use it.
504 SNew->setOperand(0, NewPN);
505 removeInstruction(S0);
506 removeInstruction(S1);
507 A0->replaceAllUsesWith(ANew);
508 removeInstruction(A0);
509 A1->replaceAllUsesWith(ANew);
510 removeInstruction(A1);
517 /// \brief True when two stores are equivalent and can sink into the footer
519 /// Starting from a diamond tail block, iterate over the instructions in one
520 /// predecessor block and try to match a store in the second predecessor.
522 bool MergedLoadStoreMotion::mergeStores(BasicBlock *T) {
524 bool MergedStores = false;
525 assert(T && "Footer of a diamond cannot be empty");
527 pred_iterator PI = pred_begin(T), E = pred_end(T);
529 BasicBlock *Pred0 = *PI;
531 BasicBlock *Pred1 = *PI;
533 // tail block of a diamond/hammock?
537 return false; // No. More than 2 predecessors.
539 // #Instructions in Succ1 for Compile Time Control
540 int Size1 = Pred1->size();
543 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend();
546 Instruction *I = &*RBI;
549 // Sink move non-simple (atomic, volatile) stores
550 if (!isa<StoreInst>(I))
552 StoreInst *S0 = (StoreInst *)I;
557 if (NStores * Size1 >= MagicCompileTimeControl)
559 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) {
560 bool Res = sinkStore(T, S0, S1);
562 // Don't attempt to sink below stores that had to stick around
563 // But after removal of a store and some of its feeding
564 // instruction search again from the beginning since the iterator
565 // is likely stale at this point.
569 RBI = Pred0->rbegin();
571 DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump());
578 /// \brief Run the transformation for each function
580 bool MergedLoadStoreMotion::runOnFunction(Function &F) {
581 MD = getAnalysisIfAvailable<MemoryDependenceAnalysis>();
582 AA = &getAnalysis<AliasAnalysis>();
584 bool Changed = false;
585 DEBUG(dbgs() << "Instruction Merger\n");
587 // Merge unconditional branches, allowing PRE to catch more
588 // optimization opportunities.
589 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) {
590 BasicBlock *BB = FI++;
592 // Hoist equivalent loads and sink stores
593 // outside diamonds when possible
594 if (isDiamondHead(BB)) {
595 Changed |= mergeLoads(BB);
596 Changed |= mergeStores(getDiamondTail(BB));