1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
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 transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into simpler forms suitable for subsequent
12 // analysis and transformation.
14 // This transformation makes the following changes to each loop with an
15 // identifiable induction variable:
16 // 1. All loops are transformed to have a SINGLE canonical induction variable
17 // which starts at zero and steps by one.
18 // 2. The canonical induction variable is guaranteed to be the first PHI node
19 // in the loop header block.
20 // 3. Any pointer arithmetic recurrences are raised to use array subscripts.
22 // If the trip count of a loop is computable, this pass also makes the following
24 // 1. The exit condition for the loop is canonicalized to compare the
25 // induction value against the exit value. This turns loops like:
26 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
27 // 2. Any use outside of the loop of an expression derived from the indvar
28 // is changed to compute the derived value outside of the loop, eliminating
29 // the dependence on the exit value of the induction variable. If the only
30 // purpose of the loop is to compute the exit value of some derived
31 // expression, this transformation will make the loop dead.
33 // This transformation should be followed by strength reduction after all of the
34 // desired loop transformations have been performed. Additionally, on targets
35 // where it is profitable, the loop could be transformed to count down to zero
36 // (the "do loop" optimization).
38 //===----------------------------------------------------------------------===//
40 #define DEBUG_TYPE "indvars"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/BasicBlock.h"
43 #include "llvm/Constants.h"
44 #include "llvm/Instructions.h"
45 #include "llvm/Type.h"
46 #include "llvm/Analysis/ScalarEvolutionExpander.h"
47 #include "llvm/Analysis/LoopInfo.h"
48 #include "llvm/Support/CFG.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Transforms/Utils/Local.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/ADT/Statistic.h"
56 STATISTIC(NumRemoved , "Number of aux indvars removed");
57 STATISTIC(NumPointer , "Number of pointer indvars promoted");
58 STATISTIC(NumInserted, "Number of canonical indvars added");
59 STATISTIC(NumReplaced, "Number of exit values replaced");
60 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
63 class IndVarSimplify : public FunctionPass {
68 virtual bool runOnFunction(Function &) {
69 LI = &getAnalysis<LoopInfo>();
70 SE = &getAnalysis<ScalarEvolution>();
73 // Induction Variables live in the header nodes of loops
74 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
79 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
80 AU.addRequiredID(LoopSimplifyID);
81 AU.addRequired<ScalarEvolution>();
82 AU.addRequired<LoopInfo>();
83 AU.addPreservedID(LoopSimplifyID);
84 AU.addPreservedID(LCSSAID);
88 void runOnLoop(Loop *L);
89 void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
90 std::set<Instruction*> &DeadInsts);
91 Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
93 void RewriteLoopExitValues(Loop *L);
95 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
97 RegisterPass<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
100 FunctionPass *llvm::createIndVarSimplifyPass() {
101 return new IndVarSimplify();
104 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
105 /// specified set are trivially dead, delete them and see if this makes any of
106 /// their operands subsequently dead.
107 void IndVarSimplify::
108 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
109 while (!Insts.empty()) {
110 Instruction *I = *Insts.begin();
111 Insts.erase(Insts.begin());
112 if (isInstructionTriviallyDead(I)) {
113 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
114 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
116 SE->deleteInstructionFromRecords(I);
117 DOUT << "INDVARS: Deleting: " << *I;
118 I->eraseFromParent();
125 /// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer
126 /// recurrence. If so, change it into an integer recurrence, permitting
127 /// analysis by the SCEV routines.
128 void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN,
129 BasicBlock *Preheader,
130 std::set<Instruction*> &DeadInsts) {
131 assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
132 unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
133 unsigned BackedgeIdx = PreheaderIdx^1;
134 if (GetElementPtrInst *GEPI =
135 dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
136 if (GEPI->getOperand(0) == PN) {
137 assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
138 DOUT << "INDVARS: Eliminating pointer recurrence: " << *GEPI;
140 // Okay, we found a pointer recurrence. Transform this pointer
141 // recurrence into an integer recurrence. Compute the value that gets
142 // added to the pointer at every iteration.
143 Value *AddedVal = GEPI->getOperand(1);
145 // Insert a new integer PHI node into the top of the block.
146 PHINode *NewPhi = new PHINode(AddedVal->getType(),
147 PN->getName()+".rec", PN);
148 NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
150 // Create the new add instruction.
151 Value *NewAdd = BinaryOperator::createAdd(NewPhi, AddedVal,
152 GEPI->getName()+".rec", GEPI);
153 NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx));
155 // Update the existing GEP to use the recurrence.
156 GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx));
158 // Update the GEP to use the new recurrence we just inserted.
159 GEPI->setOperand(1, NewAdd);
161 // If the incoming value is a constant expr GEP, try peeling out the array
162 // 0 index if possible to make things simpler.
163 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEPI->getOperand(0)))
164 if (CE->getOpcode() == Instruction::GetElementPtr) {
165 unsigned NumOps = CE->getNumOperands();
166 assert(NumOps > 1 && "CE folding didn't work!");
167 if (CE->getOperand(NumOps-1)->isNullValue()) {
168 // Check to make sure the last index really is an array index.
169 gep_type_iterator GTI = gep_type_begin(CE);
170 for (unsigned i = 1, e = CE->getNumOperands()-1;
173 if (isa<SequentialType>(*GTI)) {
174 // Pull the last index out of the constant expr GEP.
175 std::vector<Value*> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
176 Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
178 GetElementPtrInst *NGEPI =
179 new GetElementPtrInst(NCE, Constant::getNullValue(Type::Int32Ty),
180 NewAdd, GEPI->getName(), GEPI);
181 GEPI->replaceAllUsesWith(NGEPI);
182 GEPI->eraseFromParent();
189 // Finally, if there are any other users of the PHI node, we must
190 // insert a new GEP instruction that uses the pre-incremented version
191 // of the induction amount.
192 if (!PN->use_empty()) {
193 BasicBlock::iterator InsertPos = PN; ++InsertPos;
194 while (isa<PHINode>(InsertPos)) ++InsertPos;
195 std::string Name = PN->getName(); PN->setName("");
197 new GetElementPtrInst(PN->getIncomingValue(PreheaderIdx),
198 std::vector<Value*>(1, NewPhi), Name,
200 PN->replaceAllUsesWith(PreInc);
203 // Delete the old PHI for sure, and the GEP if its otherwise unused.
204 DeadInsts.insert(PN);
211 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
212 /// loop to be a canonical != comparison against the incremented loop induction
213 /// variable. This pass is able to rewrite the exit tests of any loop where the
214 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
215 /// is actually a much broader range than just linear tests.
217 /// This method returns a "potentially dead" instruction whose computation chain
218 /// should be deleted when convenient.
219 Instruction *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
220 SCEV *IterationCount,
222 // Find the exit block for the loop. We can currently only handle loops with
224 std::vector<BasicBlock*> ExitBlocks;
225 L->getExitBlocks(ExitBlocks);
226 if (ExitBlocks.size() != 1) return 0;
227 BasicBlock *ExitBlock = ExitBlocks[0];
229 // Make sure there is only one predecessor block in the loop.
230 BasicBlock *ExitingBlock = 0;
231 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
233 if (L->contains(*PI)) {
234 if (ExitingBlock == 0)
237 return 0; // Multiple exits from loop to this block.
239 assert(ExitingBlock && "Loop info is broken");
241 if (!isa<BranchInst>(ExitingBlock->getTerminator()))
242 return 0; // Can't rewrite non-branch yet
243 BranchInst *BI = cast<BranchInst>(ExitingBlock->getTerminator());
244 assert(BI->isConditional() && "Must be conditional to be part of loop!");
246 Instruction *PotentiallyDeadInst = dyn_cast<Instruction>(BI->getCondition());
248 // If the exiting block is not the same as the backedge block, we must compare
249 // against the preincremented value, otherwise we prefer to compare against
250 // the post-incremented value.
251 BasicBlock *Header = L->getHeader();
252 pred_iterator HPI = pred_begin(Header);
253 assert(HPI != pred_end(Header) && "Loop with zero preds???");
254 if (!L->contains(*HPI)) ++HPI;
255 assert(HPI != pred_end(Header) && L->contains(*HPI) &&
256 "No backedge in loop?");
258 SCEVHandle TripCount = IterationCount;
260 if (*HPI == ExitingBlock) {
261 // The IterationCount expression contains the number of times that the
262 // backedge actually branches to the loop header. This is one less than the
263 // number of times the loop executes, so add one to it.
264 Constant *OneC = ConstantInt::get(IterationCount->getType(), 1);
265 TripCount = SCEVAddExpr::get(IterationCount, SCEVUnknown::get(OneC));
266 IndVar = L->getCanonicalInductionVariableIncrement();
268 // We have to use the preincremented value...
269 IndVar = L->getCanonicalInductionVariable();
272 DOUT << "INDVARS: LFTR: TripCount = " << *TripCount
273 << " IndVar = " << *IndVar << "\n";
275 // Expand the code for the iteration count into the preheader of the loop.
276 BasicBlock *Preheader = L->getLoopPreheader();
277 Value *ExitCnt = RW.expandCodeFor(TripCount, Preheader->getTerminator(),
280 // Insert a new icmp_ne or icmp_eq instruction before the branch.
281 ICmpInst::Predicate Opcode;
282 if (L->contains(BI->getSuccessor(0)))
283 Opcode = ICmpInst::ICMP_NE;
285 Opcode = ICmpInst::ICMP_EQ;
287 Value *Cond = new ICmpInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
288 BI->setCondition(Cond);
291 return PotentiallyDeadInst;
295 /// RewriteLoopExitValues - Check to see if this loop has a computable
296 /// loop-invariant execution count. If so, this means that we can compute the
297 /// final value of any expressions that are recurrent in the loop, and
298 /// substitute the exit values from the loop into any instructions outside of
299 /// the loop that use the final values of the current expressions.
300 void IndVarSimplify::RewriteLoopExitValues(Loop *L) {
301 BasicBlock *Preheader = L->getLoopPreheader();
303 // Scan all of the instructions in the loop, looking at those that have
304 // extra-loop users and which are recurrences.
305 SCEVExpander Rewriter(*SE, *LI);
307 // We insert the code into the preheader of the loop if the loop contains
308 // multiple exit blocks, or in the exit block if there is exactly one.
309 BasicBlock *BlockToInsertInto;
310 std::vector<BasicBlock*> ExitBlocks;
311 L->getExitBlocks(ExitBlocks);
312 if (ExitBlocks.size() == 1)
313 BlockToInsertInto = ExitBlocks[0];
315 BlockToInsertInto = Preheader;
316 BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
317 while (isa<PHINode>(InsertPt)) ++InsertPt;
319 bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
321 std::set<Instruction*> InstructionsToDelete;
323 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
324 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
325 BasicBlock *BB = L->getBlocks()[i];
326 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
327 if (I->getType()->isInteger()) { // Is an integer instruction
328 SCEVHandle SH = SE->getSCEV(I);
329 if (SH->hasComputableLoopEvolution(L) || // Varies predictably
330 HasConstantItCount) {
331 // Find out if this predictably varying value is actually used
332 // outside of the loop. "extra" as opposed to "intra".
333 std::vector<Instruction*> ExtraLoopUsers;
334 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
336 Instruction *User = cast<Instruction>(*UI);
337 if (!L->contains(User->getParent())) {
338 // If this is a PHI node in the exit block and we're inserting,
339 // into the exit block, it must have a single entry. In this
340 // case, we can't insert the code after the PHI and have the PHI
341 // still use it. Instead, don't insert the the PHI.
342 if (PHINode *PN = dyn_cast<PHINode>(User)) {
343 // FIXME: This is a case where LCSSA pessimizes code, this
344 // should be fixed better.
345 if (PN->getNumOperands() == 2 &&
346 PN->getParent() == BlockToInsertInto)
349 ExtraLoopUsers.push_back(User);
353 if (!ExtraLoopUsers.empty()) {
354 // Okay, this instruction has a user outside of the current loop
355 // and varies predictably in this loop. Evaluate the value it
356 // contains when the loop exits, and insert code for it.
357 SCEVHandle ExitValue = SE->getSCEVAtScope(I, L->getParentLoop());
358 if (!isa<SCEVCouldNotCompute>(ExitValue)) {
361 // Remember the next instruction. The rewriter can move code
362 // around in some cases.
363 BasicBlock::iterator NextI = I; ++NextI;
365 Value *NewVal = Rewriter.expandCodeFor(ExitValue, InsertPt,
368 DOUT << "INDVARS: RLEV: AfterLoopVal = " << *NewVal
369 << " LoopVal = " << *I << "\n";
371 // Rewrite any users of the computed value outside of the loop
372 // with the newly computed value.
373 for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i) {
374 PHINode* PN = dyn_cast<PHINode>(ExtraLoopUsers[i]);
375 if (PN && PN->getNumOperands() == 2 &&
376 !L->contains(PN->getParent())) {
377 // We're dealing with an LCSSA Phi. Handle it specially.
378 Instruction* LCSSAInsertPt = BlockToInsertInto->begin();
380 Instruction* NewInstr = dyn_cast<Instruction>(NewVal);
381 if (NewInstr && !isa<PHINode>(NewInstr) &&
382 !L->contains(NewInstr->getParent()))
383 for (unsigned j = 0; j < NewInstr->getNumOperands(); ++j){
385 dyn_cast<Instruction>(NewInstr->getOperand(j));
386 if (PredI && L->contains(PredI->getParent())) {
387 PHINode* NewLCSSA = new PHINode(PredI->getType(),
388 PredI->getName() + ".lcssa",
390 NewLCSSA->addIncoming(PredI,
391 BlockToInsertInto->getSinglePredecessor());
393 NewInstr->replaceUsesOfWith(PredI, NewLCSSA);
397 PN->replaceAllUsesWith(NewVal);
398 PN->eraseFromParent();
400 ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
404 // If this instruction is dead now, schedule it to be removed.
406 InstructionsToDelete.insert(I);
408 continue; // Skip the ++I
414 // Next instruction. Continue instruction skips this.
419 DeleteTriviallyDeadInstructions(InstructionsToDelete);
423 void IndVarSimplify::runOnLoop(Loop *L) {
424 // First step. Check to see if there are any trivial GEP pointer recurrences.
425 // If there are, change them into integer recurrences, permitting analysis by
426 // the SCEV routines.
428 BasicBlock *Header = L->getHeader();
429 BasicBlock *Preheader = L->getLoopPreheader();
431 std::set<Instruction*> DeadInsts;
432 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
433 PHINode *PN = cast<PHINode>(I);
434 if (isa<PointerType>(PN->getType()))
435 EliminatePointerRecurrence(PN, Preheader, DeadInsts);
438 if (!DeadInsts.empty())
439 DeleteTriviallyDeadInstructions(DeadInsts);
442 // Next, transform all loops nesting inside of this loop.
443 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
446 // Check to see if this loop has a computable loop-invariant execution count.
447 // If so, this means that we can compute the final value of any expressions
448 // that are recurrent in the loop, and substitute the exit values from the
449 // loop into any instructions outside of the loop that use the final values of
450 // the current expressions.
452 SCEVHandle IterationCount = SE->getIterationCount(L);
453 if (!isa<SCEVCouldNotCompute>(IterationCount))
454 RewriteLoopExitValues(L);
456 // Next, analyze all of the induction variables in the loop, canonicalizing
457 // auxillary induction variables.
458 std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
460 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
461 PHINode *PN = cast<PHINode>(I);
462 if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
463 SCEVHandle SCEV = SE->getSCEV(PN);
464 if (SCEV->hasComputableLoopEvolution(L))
465 // FIXME: It is an extremely bad idea to indvar substitute anything more
466 // complex than affine induction variables. Doing so will put expensive
467 // polynomial evaluations inside of the loop, and the str reduction pass
468 // currently can only reduce affine polynomials. For now just disable
469 // indvar subst on anything more complex than an affine addrec.
470 if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
472 IndVars.push_back(std::make_pair(PN, SCEV));
476 // If there are no induction variables in the loop, there is nothing more to
478 if (IndVars.empty()) {
479 // Actually, if we know how many times the loop iterates, lets insert a
480 // canonical induction variable to help subsequent passes.
481 if (!isa<SCEVCouldNotCompute>(IterationCount)) {
482 SCEVExpander Rewriter(*SE, *LI);
483 Rewriter.getOrInsertCanonicalInductionVariable(L,
484 IterationCount->getType());
485 if (Instruction *I = LinearFunctionTestReplace(L, IterationCount,
487 std::set<Instruction*> InstructionsToDelete;
488 InstructionsToDelete.insert(I);
489 DeleteTriviallyDeadInstructions(InstructionsToDelete);
495 // Compute the type of the largest recurrence expression.
497 const Type *LargestType = IndVars[0].first->getType();
498 bool DifferingSizes = false;
499 for (unsigned i = 1, e = IndVars.size(); i != e; ++i) {
500 const Type *Ty = IndVars[i].first->getType();
502 Ty->getPrimitiveSizeInBits() != LargestType->getPrimitiveSizeInBits();
503 if (Ty->getPrimitiveSizeInBits() > LargestType->getPrimitiveSizeInBits())
507 // Create a rewriter object which we'll use to transform the code with.
508 SCEVExpander Rewriter(*SE, *LI);
510 // Now that we know the largest of of the induction variables in this loop,
511 // insert a canonical induction variable of the largest size.
512 Value *IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
515 DOUT << "INDVARS: New CanIV: " << *IndVar;
517 if (!isa<SCEVCouldNotCompute>(IterationCount))
518 if (Instruction *DI = LinearFunctionTestReplace(L, IterationCount,Rewriter))
519 DeadInsts.insert(DI);
521 // Now that we have a canonical induction variable, we can rewrite any
522 // recurrences in terms of the induction variable. Start with the auxillary
523 // induction variables, and recursively rewrite any of their uses.
524 BasicBlock::iterator InsertPt = Header->begin();
525 while (isa<PHINode>(InsertPt)) ++InsertPt;
527 // If there were induction variables of other sizes, cast the primary
528 // induction variable to the right size for them, avoiding the need for the
529 // code evaluation methods to insert induction variables of different sizes.
530 if (DifferingSizes) {
531 bool InsertedSizes[17] = { false };
532 InsertedSizes[LargestType->getPrimitiveSize()] = true;
533 for (unsigned i = 0, e = IndVars.size(); i != e; ++i)
534 if (!InsertedSizes[IndVars[i].first->getType()->getPrimitiveSize()]) {
535 PHINode *PN = IndVars[i].first;
536 InsertedSizes[PN->getType()->getPrimitiveSize()] = true;
537 Instruction *New = new TruncInst(IndVar, PN->getType(), "indvar",
539 Rewriter.addInsertedValue(New, SE->getSCEV(New));
540 DOUT << "INDVARS: Made trunc IV for " << *PN
541 << " NewVal = " << *New << "\n";
545 // Rewrite all induction variables in terms of the canonical induction
547 std::map<unsigned, Value*> InsertedSizes;
548 while (!IndVars.empty()) {
549 PHINode *PN = IndVars.back().first;
550 Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt,
552 DOUT << "INDVARS: Rewrote IV '" << *IndVars.back().second << "' " << *PN
553 << " into = " << *NewVal << "\n";
554 std::string Name = PN->getName();
556 NewVal->setName(Name);
558 // Replace the old PHI Node with the inserted computation.
559 PN->replaceAllUsesWith(NewVal);
560 DeadInsts.insert(PN);
567 // Now replace all derived expressions in the loop body with simpler
569 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
570 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
571 BasicBlock *BB = L->getBlocks()[i];
572 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
573 if (I->getType()->isInteger() && // Is an integer instruction
575 !Rewriter.isInsertedInstruction(I)) {
576 SCEVHandle SH = SE->getSCEV(I);
577 Value *V = Rewriter.expandCodeFor(SH, I, I->getType());
579 if (isa<Instruction>(V)) {
580 std::string Name = I->getName();
584 I->replaceAllUsesWith(V);
593 DeleteTriviallyDeadInstructions(DeadInsts);
595 if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());