1 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass is a simple loop invariant code motion pass. An interesting aspect
11 // of this pass is that it uses alias analysis for two purposes:
13 // 1. Moving loop invariant loads out of loops. If we can determine that a
14 // load inside of a loop never aliases anything stored to, we can hoist it
15 // like any other instruction.
16 // 2. Scalar Promotion of Memory - If there is a store instruction inside of
17 // the loop, we try to move the store to happen AFTER the loop instead of
18 // inside of the loop. This can only happen if a few conditions are true:
19 // A. The pointer stored through is loop invariant
20 // B. There are no stores or loads in the loop which _may_ alias the
21 // pointer. There are no calls in the loop which mod/ref the pointer.
22 // If these conditions are true, we can promote the loads and stores in the
23 // loop of the pointer to use a temporary alloca'd variable. We then use
24 // the mem2reg functionality to construct the appropriate SSA form for the
27 //===----------------------------------------------------------------------===//
29 #include "llvm/Transforms/Scalar.h"
30 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/AliasAnalysis.h"
34 #include "llvm/Analysis/AliasSetTracker.h"
35 #include "llvm/Analysis/Dominators.h"
36 #include "llvm/Instructions.h"
37 #include "llvm/DerivedTypes.h"
38 #include "llvm/Target/TargetData.h"
39 #include "llvm/Support/InstVisitor.h"
40 #include "llvm/Support/CFG.h"
41 #include "Support/CommandLine.h"
42 #include "Support/Debug.h"
43 #include "Support/Statistic.h"
44 #include "llvm/Assembly/Writer.h"
49 DisablePromotion("disable-licm-promotion", cl::Hidden,
50 cl::desc("Disable memory promotion in LICM pass"));
52 Statistic<> NumHoisted("licm", "Number of instructions hoisted out of loop");
53 Statistic<> NumHoistedLoads("licm", "Number of load insts hoisted");
54 Statistic<> NumPromoted("licm",
55 "Number of memory locations promoted to registers");
57 struct LICM : public FunctionPass, public InstVisitor<LICM> {
58 virtual bool runOnFunction(Function &F);
60 /// This transformation requires natural loop information & requires that
61 /// loop preheaders be inserted into the CFG...
63 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
65 AU.addRequiredID(LoopSimplifyID);
66 AU.addRequired<LoopInfo>();
67 AU.addRequired<DominatorTree>();
68 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
69 AU.addRequired<AliasAnalysis>();
73 LoopInfo *LI; // Current LoopInfo
74 AliasAnalysis *AA; // Current AliasAnalysis information
75 DominanceFrontier *DF; // Current Dominance Frontier
76 bool Changed; // Set to true when we change anything.
77 BasicBlock *Preheader; // The preheader block of the current loop...
78 Loop *CurLoop; // The current loop we are working on...
79 AliasSetTracker *CurAST; // AliasSet information for the current loop...
80 DominatorTree *DT; // Dominator Tree for the current Loop...
82 /// visitLoop - Hoist expressions out of the specified loop...
84 void visitLoop(Loop *L, AliasSetTracker &AST);
86 /// HoistRegion - Walk the specified region of the CFG (defined by all
87 /// blocks dominated by the specified block, and that are in the current
88 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
89 /// visit definitions before uses, allowing us to hoist a loop body in one
90 /// pass without iteration.
92 void HoistRegion(DominatorTree::Node *N);
94 /// inSubLoop - Little predicate that returns true if the specified basic
95 /// block is in a subloop of the current one, not the current one itself.
97 bool inSubLoop(BasicBlock *BB) {
98 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
99 for (unsigned i = 0, e = CurLoop->getSubLoops().size(); i != e; ++i)
100 if (CurLoop->getSubLoops()[i]->contains(BB))
101 return true; // A subloop actually contains this block!
105 /// hoist - When an instruction is found to only use loop invariant operands
106 /// that is safe to hoist, this instruction is called to do the dirty work.
108 void hoist(Instruction &I);
110 /// SafeToHoist - Only hoist an instruction if it is not a trapping
111 /// instruction or if it is a trapping instruction and is guaranteed to
114 bool SafeToHoist(Instruction &I);
116 /// pointerInvalidatedByLoop - Return true if the body of this loop may
117 /// store into the memory location pointed to by V.
119 bool pointerInvalidatedByLoop(Value *V) {
120 // Check to see if any of the basic blocks in CurLoop invalidate *V.
121 return CurAST->getAliasSetForPointer(V, 0).isMod();
124 /// isLoopInvariant - Return true if the specified value is loop invariant
126 inline bool isLoopInvariant(Value *V) {
127 if (Instruction *I = dyn_cast<Instruction>(V))
128 return !CurLoop->contains(I->getParent());
129 return true; // All non-instructions are loop invariant
132 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many
133 /// to scalars as we can.
135 void PromoteValuesInLoop();
137 /// findPromotableValuesInLoop - Check the current loop for stores to
138 /// definite pointers, which are not loaded and stored through may aliases.
139 /// If these are found, create an alloca for the value, add it to the
140 /// PromotedValues list, and keep track of the mapping from value to
143 void findPromotableValuesInLoop(
144 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
145 std::map<Value*, AllocaInst*> &Val2AlMap);
148 /// Instruction visitation handlers... these basically control whether or
149 /// not the specified instruction types are hoisted.
151 friend class InstVisitor<LICM>;
152 void visitBinaryOperator(Instruction &I) {
153 if (isLoopInvariant(I.getOperand(0)) &&
154 isLoopInvariant(I.getOperand(1)) && SafeToHoist(I))
157 void visitCastInst(CastInst &CI) {
158 Instruction &I = (Instruction&)CI;
159 if (isLoopInvariant(I.getOperand(0)) && SafeToHoist(CI)) hoist(I);
161 void visitShiftInst(ShiftInst &I) { visitBinaryOperator((Instruction&)I); }
163 void visitLoadInst(LoadInst &LI);
165 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
166 Instruction &I = (Instruction&)GEPI;
167 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
168 if (!isLoopInvariant(I.getOperand(i))) return;
169 if(SafeToHoist(GEPI))
174 RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion");
177 Pass *createLICMPass() { return new LICM(); }
179 /// runOnFunction - For LICM, this simply traverses the loop structure of the
180 /// function, hoisting expressions out of loops if possible.
182 bool LICM::runOnFunction(Function &) {
185 // Get our Loop and Alias Analysis information...
186 LI = &getAnalysis<LoopInfo>();
187 AA = &getAnalysis<AliasAnalysis>();
188 DF = &getAnalysis<DominanceFrontier>();
189 DT = &getAnalysis<DominatorTree>();
191 // Hoist expressions out of all of the top-level loops.
192 const std::vector<Loop*> &TopLevelLoops = LI->getTopLevelLoops();
193 for (std::vector<Loop*>::const_iterator I = TopLevelLoops.begin(),
194 E = TopLevelLoops.end(); I != E; ++I) {
195 AliasSetTracker AST(*AA);
196 LICM::visitLoop(*I, AST);
202 /// visitLoop - Hoist expressions out of the specified loop...
204 void LICM::visitLoop(Loop *L, AliasSetTracker &AST) {
205 // Recurse through all subloops before we process this loop...
206 for (std::vector<Loop*>::const_iterator I = L->getSubLoops().begin(),
207 E = L->getSubLoops().end(); I != E; ++I) {
208 AliasSetTracker SubAST(*AA);
209 LICM::visitLoop(*I, SubAST);
211 // Incorporate information about the subloops into this loop...
217 // Get the preheader block to move instructions into...
218 Preheader = L->getLoopPreheader();
219 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!");
221 // Loop over the body of this loop, looking for calls, invokes, and stores.
222 // Because subloops have already been incorporated into AST, we skip blocks in
225 const std::vector<BasicBlock*> &LoopBBs = L->getBlocks();
226 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
227 E = LoopBBs.end(); I != E; ++I)
228 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops...
229 AST.add(**I); // Incorporate the specified basic block
231 // We want to visit all of the instructions in this loop... that are not parts
232 // of our subloops (they have already had their invariants hoisted out of
233 // their loop, into this loop, so there is no need to process the BODIES of
236 // Traverse the body of the loop in depth first order on the dominator tree so
237 // that we are guaranteed to see definitions before we see uses. This allows
238 // us to perform the LICM transformation in one pass, without iteration.
240 HoistRegion(DT->getNode(L->getHeader()));
242 // Now that all loop invariants have been removed from the loop, promote any
243 // memory references to scalars that we can...
244 if (!DisablePromotion)
245 PromoteValuesInLoop();
247 // Clear out loops state information for the next iteration
252 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
253 /// dominated by the specified block, and that are in the current loop) in depth
254 /// first order w.r.t the DominatorTree. This allows us to visit definitions
255 /// before uses, allowing us to hoist a loop body in one pass without iteration.
257 void LICM::HoistRegion(DominatorTree::Node *N) {
258 assert(N != 0 && "Null dominator tree node?");
260 // If this subregion is not in the top level loop at all, exit.
261 if (!CurLoop->contains(N->getBlock())) return;
263 // Only need to hoist the contents of this block if it is not part of a
264 // subloop (which would already have been hoisted)
265 if (!inSubLoop(N->getBlock()))
266 visit(*N->getBlock());
268 const std::vector<DominatorTree::Node*> &Children = N->getChildren();
269 for (unsigned i = 0, e = Children.size(); i != e; ++i)
270 HoistRegion(Children[i]);
274 /// hoist - When an instruction is found to only use loop invariant operands
275 /// that is safe to hoist, this instruction is called to do the dirty work.
277 void LICM::hoist(Instruction &Inst) {
278 DEBUG(std::cerr << "LICM hoisting to";
279 WriteAsOperand(std::cerr, Preheader, false);
280 std::cerr << ": " << Inst);
282 // Remove the instruction from its current basic block... but don't delete the
284 Inst.getParent()->getInstList().remove(&Inst);
286 // Insert the new node in Preheader, before the terminator.
287 Preheader->getInstList().insert(Preheader->getTerminator(), &Inst);
293 /// SafeToHoist - Only hoist an instruction if it is not a trapping instruction
294 /// or if it is a trapping instruction and is guaranteed to execute
296 bool LICM::SafeToHoist(Instruction &Inst) {
298 //If it is a trapping instruction, then check if its guaranteed to execute.
299 if(Inst.isTrapping()) {
301 //Get the instruction's basic block.
302 BasicBlock *InstBB = Inst.getParent();
304 //Get the Dominator Tree Node for the instruction's basic block/
305 DominatorTree::Node *InstDTNode = DT->getNode(InstBB);
307 //Get the exit blocks for the current loop.
308 const std::vector<BasicBlock* > &ExitBlocks = CurLoop->getExitBlocks();
310 //For each exit block, get the DT node and walk up the DT until
311 //the instruction's basic block is found or we exit the loop.
312 for(unsigned i=0; i < ExitBlocks.size(); ++i) {
313 DominatorTree::Node *IDom = DT->getNode(ExitBlocks[i]);
315 while(IDom != InstDTNode) {
317 //Get next Immediate Dominator.
318 IDom = IDom->getIDom();
320 //See if we exited the loop.
321 if(!CurLoop->contains(IDom->getBlock()))
331 void LICM::visitLoadInst(LoadInst &LI) {
332 if (isLoopInvariant(LI.getOperand(0)) && !LI.isVolatile() &&
333 !pointerInvalidatedByLoop(LI.getOperand(0)) && SafeToHoist(LI)) {
339 /// PromoteValuesInLoop - Try to promote memory values to scalars by sinking
340 /// stores out of the loop and moving loads to before the loop. We do this by
341 /// looping over the stores in the loop, looking for stores to Must pointers
342 /// which are loop invariant. We promote these memory locations to use allocas
343 /// instead. These allocas can easily be raised to register values by the
344 /// PromoteMem2Reg functionality.
346 void LICM::PromoteValuesInLoop() {
347 // PromotedValues - List of values that are promoted out of the loop. Each
348 // value has an alloca instruction for it, and a canonical version of the
350 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues;
351 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca
353 findPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
354 if (ValueToAllocaMap.empty()) return; // If there are values to promote...
357 NumPromoted += PromotedValues.size();
359 // Emit a copy from the value into the alloca'd value in the loop preheader
360 TerminatorInst *LoopPredInst = Preheader->getTerminator();
361 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
362 // Load from the memory we are promoting...
363 LoadInst *LI = new LoadInst(PromotedValues[i].second,
364 PromotedValues[i].second->getName()+".promoted",
366 // Store into the temporary alloca...
367 new StoreInst(LI, PromotedValues[i].first, LoopPredInst);
370 // Scan the basic blocks in the loop, replacing uses of our pointers with
371 // uses of the allocas in question. If we find a branch that exits the
372 // loop, make sure to put reload code into all of the successors of the
375 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks();
376 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
377 E = LoopBBs.end(); I != E; ++I) {
378 // Rewrite all loads and stores in the block of the pointer...
379 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end();
381 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
382 std::map<Value*, AllocaInst*>::iterator
383 I = ValueToAllocaMap.find(L->getOperand(0));
384 if (I != ValueToAllocaMap.end())
385 L->setOperand(0, I->second); // Rewrite load instruction...
386 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) {
387 std::map<Value*, AllocaInst*>::iterator
388 I = ValueToAllocaMap.find(S->getOperand(1));
389 if (I != ValueToAllocaMap.end())
390 S->setOperand(1, I->second); // Rewrite store instruction...
394 // Check to see if any successors of this block are outside of the loop.
395 // If so, we need to copy the value from the alloca back into the memory
398 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
399 if (!CurLoop->contains(*SI)) {
400 // Copy all of the allocas into their memory locations...
401 BasicBlock::iterator BI = (*SI)->begin();
402 while (isa<PHINode>(*BI))
403 ++BI; // Skip over all of the phi nodes in the block...
404 Instruction *InsertPos = BI;
405 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
406 // Load from the alloca...
407 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
408 // Store into the memory we promoted...
409 new StoreInst(LI, PromotedValues[i].second, InsertPos);
414 // Now that we have done the deed, use the mem2reg functionality to promote
415 // all of the new allocas we just created into real SSA registers...
417 std::vector<AllocaInst*> PromotedAllocas;
418 PromotedAllocas.reserve(PromotedValues.size());
419 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
420 PromotedAllocas.push_back(PromotedValues[i].first);
421 PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData());
424 /// findPromotableValuesInLoop - Check the current loop for stores to definite
425 /// pointers, which are not loaded and stored through may aliases. If these are
426 /// found, create an alloca for the value, add it to the PromotedValues list,
427 /// and keep track of the mapping from value to alloca...
429 void LICM::findPromotableValuesInLoop(
430 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
431 std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
432 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin();
434 // Loop over all of the alias sets in the tracker object...
435 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
438 // We can promote this alias set if it has a store, if it is a "Must" alias
439 // set, and if the pointer is loop invariant.
440 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() &&
441 isLoopInvariant(AS.begin()->first)) {
442 assert(AS.begin() != AS.end() &&
443 "Must alias set should have at least one pointer element in it!");
444 Value *V = AS.begin()->first;
446 // Check that all of the pointers in the alias set have the same type. We
447 // cannot (yet) promote a memory location that is loaded and stored in
449 bool PointerOk = true;
450 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
451 if (V->getType() != I->first->getType()) {
457 const Type *Ty = cast<PointerType>(V->getType())->getElementType();
458 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
459 PromotedValues.push_back(std::make_pair(AI, V));
461 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
462 ValueToAllocaMap.insert(std::make_pair(I->first, AI));
464 DEBUG(std::cerr << "LICM: Promoting value: " << *V << "\n");