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 performs loop invariant code motion, attempting to remove as much
11 // code from the body of a loop as possible. It does this by either hoisting
12 // code into the preheader block, or by sinking code to the exit blocks if it is
13 // safe. This pass also promotes must-aliased memory locations in the loop to
16 // This pass uses alias analysis for two purposes:
18 // 1. Moving loop invariant loads out of loops. If we can determine that a
19 // load inside of a loop never aliases anything stored to, we can hoist it
20 // or sink it like any other instruction.
21 // 2. Scalar Promotion of Memory - If there is a store instruction inside of
22 // the loop, we try to move the store to happen AFTER the loop instead of
23 // inside of the loop. This can only happen if a few conditions are true:
24 // A. The pointer stored through is loop invariant
25 // B. There are no stores or loads in the loop which _may_ alias the
26 // pointer. There are no calls in the loop which mod/ref the pointer.
27 // If these conditions are true, we can promote the loads and stores in the
28 // loop of the pointer to use a temporary alloca'd variable. We then use
29 // the mem2reg functionality to construct the appropriate SSA form for the
32 //===----------------------------------------------------------------------===//
34 #include "llvm/Transforms/Scalar.h"
35 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/Analysis/LoopInfo.h"
38 #include "llvm/Analysis/AliasAnalysis.h"
39 #include "llvm/Analysis/AliasSetTracker.h"
40 #include "llvm/Analysis/Dominators.h"
41 #include "llvm/Instructions.h"
42 #include "llvm/DerivedTypes.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Support/InstVisitor.h"
45 #include "llvm/Support/CFG.h"
46 #include "Support/CommandLine.h"
47 #include "Support/Debug.h"
48 #include "Support/Statistic.h"
49 #include "llvm/Assembly/Writer.h"
55 DisablePromotion("disable-licm-promotion", cl::Hidden,
56 cl::desc("Disable memory promotion in LICM pass"));
58 Statistic<> NumHoisted("licm", "Number of instructions hoisted out of loop");
59 Statistic<> NumHoistedLoads("licm", "Number of load insts hoisted");
60 Statistic<> NumPromoted("licm",
61 "Number of memory locations promoted to registers");
63 struct LICM : public FunctionPass, public InstVisitor<LICM> {
64 virtual bool runOnFunction(Function &F);
66 /// This transformation requires natural loop information & requires that
67 /// loop preheaders be inserted into the CFG...
69 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
71 AU.addRequiredID(LoopSimplifyID);
72 AU.addRequired<LoopInfo>();
73 AU.addRequired<DominatorTree>();
74 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
75 AU.addRequired<AliasAnalysis>();
79 // Various analyses that we use...
80 AliasAnalysis *AA; // Current AliasAnalysis information
81 LoopInfo *LI; // Current LoopInfo
82 DominatorTree *DT; // Dominator Tree for the current Loop...
83 DominanceFrontier *DF; // Current Dominance Frontier
85 // State that is updated as we process loops
86 bool Changed; // Set to true when we change anything.
87 BasicBlock *Preheader; // The preheader block of the current loop...
88 Loop *CurLoop; // The current loop we are working on...
89 AliasSetTracker *CurAST; // AliasSet information for the current loop...
91 /// visitLoop - Hoist expressions out of the specified loop...
93 void visitLoop(Loop *L, AliasSetTracker &AST);
95 /// HoistRegion - Walk the specified region of the CFG (defined by all
96 /// blocks dominated by the specified block, and that are in the current
97 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
98 /// visit definitions before uses, allowing us to hoist a loop body in one
99 /// pass without iteration.
101 void HoistRegion(DominatorTree::Node *N);
103 /// inSubLoop - Little predicate that returns true if the specified basic
104 /// block is in a subloop of the current one, not the current one itself.
106 bool inSubLoop(BasicBlock *BB) {
107 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
108 for (unsigned i = 0, e = CurLoop->getSubLoops().size(); i != e; ++i)
109 if (CurLoop->getSubLoops()[i]->contains(BB))
110 return true; // A subloop actually contains this block!
114 /// hoist - When an instruction is found to only use loop invariant operands
115 /// that is safe to hoist, this instruction is called to do the dirty work.
117 void hoist(Instruction &I);
119 /// SafeToHoist - Only hoist an instruction if it is not a trapping
120 /// instruction or if it is a trapping instruction and is guaranteed to
123 bool SafeToHoist(Instruction &I);
125 /// pointerInvalidatedByLoop - Return true if the body of this loop may
126 /// store into the memory location pointed to by V.
128 bool pointerInvalidatedByLoop(Value *V) {
129 // Check to see if any of the basic blocks in CurLoop invalidate *V.
130 return CurAST->getAliasSetForPointer(V, 0).isMod();
133 /// isLoopInvariant - Return true if the specified value is loop invariant
135 inline bool isLoopInvariant(Value *V) {
136 if (Instruction *I = dyn_cast<Instruction>(V))
137 return !CurLoop->contains(I->getParent());
138 return true; // All non-instructions are loop invariant
141 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many
142 /// to scalars as we can.
144 void PromoteValuesInLoop();
146 /// findPromotableValuesInLoop - Check the current loop for stores to
147 /// definite pointers, which are not loaded and stored through may aliases.
148 /// If these are found, create an alloca for the value, add it to the
149 /// PromotedValues list, and keep track of the mapping from value to
152 void findPromotableValuesInLoop(
153 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
154 std::map<Value*, AllocaInst*> &Val2AlMap);
157 /// Instruction visitation handlers... these basically control whether or
158 /// not the specified instruction types are hoisted.
160 friend class InstVisitor<LICM>;
161 void visitBinaryOperator(Instruction &I) {
162 if (isLoopInvariant(I.getOperand(0)) &&
163 isLoopInvariant(I.getOperand(1)) && SafeToHoist(I))
166 void visitCastInst(CastInst &CI) {
167 Instruction &I = (Instruction&)CI;
168 if (isLoopInvariant(I.getOperand(0)) && SafeToHoist(CI)) hoist(I);
170 void visitShiftInst(ShiftInst &I) { visitBinaryOperator((Instruction&)I); }
172 void visitLoadInst(LoadInst &LI);
174 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
175 Instruction &I = (Instruction&)GEPI;
176 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
177 if (!isLoopInvariant(I.getOperand(i))) return;
178 if(SafeToHoist(GEPI))
183 RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion");
186 FunctionPass *llvm::createLICMPass() { return new LICM(); }
188 /// runOnFunction - For LICM, this simply traverses the loop structure of the
189 /// function, hoisting expressions out of loops if possible.
191 bool LICM::runOnFunction(Function &) {
194 // Get our Loop and Alias Analysis information...
195 LI = &getAnalysis<LoopInfo>();
196 AA = &getAnalysis<AliasAnalysis>();
197 DF = &getAnalysis<DominanceFrontier>();
198 DT = &getAnalysis<DominatorTree>();
200 // Hoist expressions out of all of the top-level loops.
201 const std::vector<Loop*> &TopLevelLoops = LI->getTopLevelLoops();
202 for (std::vector<Loop*>::const_iterator I = TopLevelLoops.begin(),
203 E = TopLevelLoops.end(); I != E; ++I) {
204 AliasSetTracker AST(*AA);
205 LICM::visitLoop(*I, AST);
211 /// visitLoop - Hoist expressions out of the specified loop...
213 void LICM::visitLoop(Loop *L, AliasSetTracker &AST) {
214 // Recurse through all subloops before we process this loop...
215 for (std::vector<Loop*>::const_iterator I = L->getSubLoops().begin(),
216 E = L->getSubLoops().end(); I != E; ++I) {
217 AliasSetTracker SubAST(*AA);
218 LICM::visitLoop(*I, SubAST);
220 // Incorporate information about the subloops into this loop...
226 // Get the preheader block to move instructions into...
227 Preheader = L->getLoopPreheader();
228 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!");
230 // Loop over the body of this loop, looking for calls, invokes, and stores.
231 // Because subloops have already been incorporated into AST, we skip blocks in
234 const std::vector<BasicBlock*> &LoopBBs = L->getBlocks();
235 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
236 E = LoopBBs.end(); I != E; ++I)
237 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops...
238 AST.add(**I); // Incorporate the specified basic block
240 // We want to visit all of the instructions in this loop... that are not parts
241 // of our subloops (they have already had their invariants hoisted out of
242 // their loop, into this loop, so there is no need to process the BODIES of
245 // Traverse the body of the loop in depth first order on the dominator tree so
246 // that we are guaranteed to see definitions before we see uses. This allows
247 // us to perform the LICM transformation in one pass, without iteration.
249 HoistRegion(DT->getNode(L->getHeader()));
251 // Now that all loop invariants have been removed from the loop, promote any
252 // memory references to scalars that we can...
253 if (!DisablePromotion)
254 PromoteValuesInLoop();
256 // Clear out loops state information for the next iteration
261 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
262 /// dominated by the specified block, and that are in the current loop) in depth
263 /// first order w.r.t the DominatorTree. This allows us to visit definitions
264 /// before uses, allowing us to hoist a loop body in one pass without iteration.
266 void LICM::HoistRegion(DominatorTree::Node *N) {
267 assert(N != 0 && "Null dominator tree node?");
269 // If this subregion is not in the top level loop at all, exit.
270 if (!CurLoop->contains(N->getBlock())) return;
272 // Only need to hoist the contents of this block if it is not part of a
273 // subloop (which would already have been hoisted)
274 if (!inSubLoop(N->getBlock()))
275 visit(*N->getBlock());
277 const std::vector<DominatorTree::Node*> &Children = N->getChildren();
278 for (unsigned i = 0, e = Children.size(); i != e; ++i)
279 HoistRegion(Children[i]);
283 /// hoist - When an instruction is found to only use loop invariant operands
284 /// that is safe to hoist, this instruction is called to do the dirty work.
286 void LICM::hoist(Instruction &Inst) {
287 DEBUG(std::cerr << "LICM hoisting to";
288 WriteAsOperand(std::cerr, Preheader, false);
289 std::cerr << ": " << Inst);
291 // Remove the instruction from its current basic block... but don't delete the
293 Inst.getParent()->getInstList().remove(&Inst);
295 // Insert the new node in Preheader, before the terminator.
296 Preheader->getInstList().insert(Preheader->getTerminator(), &Inst);
302 /// SafeToHoist - Only hoist an instruction if it is not a trapping instruction
303 /// or if it is a trapping instruction and is guaranteed to execute
305 bool LICM::SafeToHoist(Instruction &Inst) {
306 // If it is not a trapping instruction, it is always safe to hoist.
307 if (!Inst.isTrapping()) return true;
309 // Otherwise we have to check to make sure that the instruction dominates all
310 // of the exit blocks. If it doesn't, then there is a path out of the loop
311 // which does not execute this instruction, so we can't hoist it.
313 // If the instruction is in the header block for the loop (which is very
314 // common), it is always guaranteed to dominate the exit blocks. Since this
315 // is a common case, and can save some work, check it now.
316 BasicBlock *LoopHeader = CurLoop->getHeader();
317 if (Inst.getParent() == LoopHeader)
320 // Get the Dominator Tree Node for the instruction's basic block.
321 DominatorTree::Node *InstDTNode = DT->getNode(Inst.getParent());
323 // Get the exit blocks for the current loop.
324 const std::vector<BasicBlock* > &ExitBlocks = CurLoop->getExitBlocks();
326 // For each exit block, get the DT node and walk up the DT until the
327 // instruction's basic block is found or we exit the loop.
328 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
329 DominatorTree::Node *IDom = DT->getNode(ExitBlocks[i]);
332 // Get next Immediate Dominator.
333 IDom = IDom->getIDom();
335 // If we have got to the header of the loop, then the instructions block
336 // did not dominate the exit node, so we can't hoist it.
337 if (IDom->getBlock() == LoopHeader)
340 } while(IDom != InstDTNode);
347 void LICM::visitLoadInst(LoadInst &LI) {
348 if (isLoopInvariant(LI.getOperand(0)) && !LI.isVolatile() &&
349 !pointerInvalidatedByLoop(LI.getOperand(0)) && SafeToHoist(LI)) {
355 /// PromoteValuesInLoop - Try to promote memory values to scalars by sinking
356 /// stores out of the loop and moving loads to before the loop. We do this by
357 /// looping over the stores in the loop, looking for stores to Must pointers
358 /// which are loop invariant. We promote these memory locations to use allocas
359 /// instead. These allocas can easily be raised to register values by the
360 /// PromoteMem2Reg functionality.
362 void LICM::PromoteValuesInLoop() {
363 // PromotedValues - List of values that are promoted out of the loop. Each
364 // value has an alloca instruction for it, and a canonical version of the
366 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues;
367 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca
369 findPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
370 if (ValueToAllocaMap.empty()) return; // If there are values to promote...
373 NumPromoted += PromotedValues.size();
375 // Emit a copy from the value into the alloca'd value in the loop preheader
376 TerminatorInst *LoopPredInst = Preheader->getTerminator();
377 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
378 // Load from the memory we are promoting...
379 LoadInst *LI = new LoadInst(PromotedValues[i].second,
380 PromotedValues[i].second->getName()+".promoted",
382 // Store into the temporary alloca...
383 new StoreInst(LI, PromotedValues[i].first, LoopPredInst);
386 // Scan the basic blocks in the loop, replacing uses of our pointers with
387 // uses of the allocas in question. If we find a branch that exits the
388 // loop, make sure to put reload code into all of the successors of the
391 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks();
392 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
393 E = LoopBBs.end(); I != E; ++I) {
394 // Rewrite all loads and stores in the block of the pointer...
395 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end();
397 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
398 std::map<Value*, AllocaInst*>::iterator
399 I = ValueToAllocaMap.find(L->getOperand(0));
400 if (I != ValueToAllocaMap.end())
401 L->setOperand(0, I->second); // Rewrite load instruction...
402 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) {
403 std::map<Value*, AllocaInst*>::iterator
404 I = ValueToAllocaMap.find(S->getOperand(1));
405 if (I != ValueToAllocaMap.end())
406 S->setOperand(1, I->second); // Rewrite store instruction...
410 // Check to see if any successors of this block are outside of the loop.
411 // If so, we need to copy the value from the alloca back into the memory
414 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
415 if (!CurLoop->contains(*SI)) {
416 // Copy all of the allocas into their memory locations...
417 BasicBlock::iterator BI = (*SI)->begin();
418 while (isa<PHINode>(*BI))
419 ++BI; // Skip over all of the phi nodes in the block...
420 Instruction *InsertPos = BI;
421 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
422 // Load from the alloca...
423 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
424 // Store into the memory we promoted...
425 new StoreInst(LI, PromotedValues[i].second, InsertPos);
430 // Now that we have done the deed, use the mem2reg functionality to promote
431 // all of the new allocas we just created into real SSA registers...
433 std::vector<AllocaInst*> PromotedAllocas;
434 PromotedAllocas.reserve(PromotedValues.size());
435 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
436 PromotedAllocas.push_back(PromotedValues[i].first);
437 PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData());
440 /// findPromotableValuesInLoop - Check the current loop for stores to definite
441 /// pointers, which are not loaded and stored through may aliases. If these are
442 /// found, create an alloca for the value, add it to the PromotedValues list,
443 /// and keep track of the mapping from value to alloca...
445 void LICM::findPromotableValuesInLoop(
446 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
447 std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
448 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin();
450 // Loop over all of the alias sets in the tracker object...
451 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
454 // We can promote this alias set if it has a store, if it is a "Must" alias
455 // set, and if the pointer is loop invariant.
456 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() &&
457 isLoopInvariant(AS.begin()->first)) {
458 assert(AS.begin() != AS.end() &&
459 "Must alias set should have at least one pointer element in it!");
460 Value *V = AS.begin()->first;
462 // Check that all of the pointers in the alias set have the same type. We
463 // cannot (yet) promote a memory location that is loaded and stored in
465 bool PointerOk = true;
466 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
467 if (V->getType() != I->first->getType()) {
473 const Type *Ty = cast<PointerType>(V->getType())->getElementType();
474 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
475 PromotedValues.push_back(std::make_pair(AI, V));
477 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
478 ValueToAllocaMap.insert(std::make_pair(I->first, AI));
480 DEBUG(std::cerr << "LICM: Promoting value: " << *V << "\n");