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/GetElementPtrTypeIterator.h"
50 #include "llvm/Transforms/Utils/Local.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/ADT/Statistic.h"
55 STATISTIC(NumRemoved , "Number of aux indvars removed");
56 STATISTIC(NumPointer , "Number of pointer indvars promoted");
57 STATISTIC(NumInserted, "Number of canonical indvars added");
58 STATISTIC(NumReplaced, "Number of exit values replaced");
59 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
62 class IndVarSimplify : public FunctionPass {
67 virtual bool runOnFunction(Function &) {
68 LI = &getAnalysis<LoopInfo>();
69 SE = &getAnalysis<ScalarEvolution>();
72 // Induction Variables live in the header nodes of loops
73 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
78 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
79 AU.addRequiredID(LoopSimplifyID);
80 AU.addRequired<ScalarEvolution>();
81 AU.addRequired<LoopInfo>();
82 AU.addPreservedID(LoopSimplifyID);
83 AU.addPreservedID(LCSSAID);
87 void runOnLoop(Loop *L);
88 void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
89 std::set<Instruction*> &DeadInsts);
90 Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
92 void RewriteLoopExitValues(Loop *L);
94 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
96 RegisterPass<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
99 FunctionPass *llvm::createIndVarSimplifyPass() {
100 return new IndVarSimplify();
103 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
104 /// specified set are trivially dead, delete them and see if this makes any of
105 /// their operands subsequently dead.
106 void IndVarSimplify::
107 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
108 while (!Insts.empty()) {
109 Instruction *I = *Insts.begin();
110 Insts.erase(Insts.begin());
111 if (isInstructionTriviallyDead(I)) {
112 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
113 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
115 SE->deleteInstructionFromRecords(I);
116 I->eraseFromParent();
123 /// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer
124 /// recurrence. If so, change it into an integer recurrence, permitting
125 /// analysis by the SCEV routines.
126 void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN,
127 BasicBlock *Preheader,
128 std::set<Instruction*> &DeadInsts) {
129 assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
130 unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
131 unsigned BackedgeIdx = PreheaderIdx^1;
132 if (GetElementPtrInst *GEPI =
133 dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
134 if (GEPI->getOperand(0) == PN) {
135 assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
137 // Okay, we found a pointer recurrence. Transform this pointer
138 // recurrence into an integer recurrence. Compute the value that gets
139 // added to the pointer at every iteration.
140 Value *AddedVal = GEPI->getOperand(1);
142 // Insert a new integer PHI node into the top of the block.
143 PHINode *NewPhi = new PHINode(AddedVal->getType(),
144 PN->getName()+".rec", PN);
145 NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
147 // Create the new add instruction.
148 Value *NewAdd = BinaryOperator::createAdd(NewPhi, AddedVal,
149 GEPI->getName()+".rec", GEPI);
150 NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx));
152 // Update the existing GEP to use the recurrence.
153 GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx));
155 // Update the GEP to use the new recurrence we just inserted.
156 GEPI->setOperand(1, NewAdd);
158 // If the incoming value is a constant expr GEP, try peeling out the array
159 // 0 index if possible to make things simpler.
160 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEPI->getOperand(0)))
161 if (CE->getOpcode() == Instruction::GetElementPtr) {
162 unsigned NumOps = CE->getNumOperands();
163 assert(NumOps > 1 && "CE folding didn't work!");
164 if (CE->getOperand(NumOps-1)->isNullValue()) {
165 // Check to make sure the last index really is an array index.
166 gep_type_iterator GTI = gep_type_begin(CE);
167 for (unsigned i = 1, e = CE->getNumOperands()-1;
170 if (isa<SequentialType>(*GTI)) {
171 // Pull the last index out of the constant expr GEP.
172 std::vector<Value*> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
173 Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
175 GetElementPtrInst *NGEPI =
176 new GetElementPtrInst(NCE, Constant::getNullValue(Type::Int32Ty),
177 NewAdd, GEPI->getName(), GEPI);
178 GEPI->replaceAllUsesWith(NGEPI);
179 GEPI->eraseFromParent();
186 // Finally, if there are any other users of the PHI node, we must
187 // insert a new GEP instruction that uses the pre-incremented version
188 // of the induction amount.
189 if (!PN->use_empty()) {
190 BasicBlock::iterator InsertPos = PN; ++InsertPos;
191 while (isa<PHINode>(InsertPos)) ++InsertPos;
192 std::string Name = PN->getName(); PN->setName("");
194 new GetElementPtrInst(PN->getIncomingValue(PreheaderIdx),
195 std::vector<Value*>(1, NewPhi), Name,
197 PN->replaceAllUsesWith(PreInc);
200 // Delete the old PHI for sure, and the GEP if its otherwise unused.
201 DeadInsts.insert(PN);
208 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
209 /// loop to be a canonical != comparison against the incremented loop induction
210 /// variable. This pass is able to rewrite the exit tests of any loop where the
211 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
212 /// is actually a much broader range than just linear tests.
214 /// This method returns a "potentially dead" instruction whose computation chain
215 /// should be deleted when convenient.
216 Instruction *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
217 SCEV *IterationCount,
219 // Find the exit block for the loop. We can currently only handle loops with
221 std::vector<BasicBlock*> ExitBlocks;
222 L->getExitBlocks(ExitBlocks);
223 if (ExitBlocks.size() != 1) return 0;
224 BasicBlock *ExitBlock = ExitBlocks[0];
226 // Make sure there is only one predecessor block in the loop.
227 BasicBlock *ExitingBlock = 0;
228 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
230 if (L->contains(*PI)) {
231 if (ExitingBlock == 0)
234 return 0; // Multiple exits from loop to this block.
236 assert(ExitingBlock && "Loop info is broken");
238 if (!isa<BranchInst>(ExitingBlock->getTerminator()))
239 return 0; // Can't rewrite non-branch yet
240 BranchInst *BI = cast<BranchInst>(ExitingBlock->getTerminator());
241 assert(BI->isConditional() && "Must be conditional to be part of loop!");
243 Instruction *PotentiallyDeadInst = dyn_cast<Instruction>(BI->getCondition());
245 // If the exiting block is not the same as the backedge block, we must compare
246 // against the preincremented value, otherwise we prefer to compare against
247 // the post-incremented value.
248 BasicBlock *Header = L->getHeader();
249 pred_iterator HPI = pred_begin(Header);
250 assert(HPI != pred_end(Header) && "Loop with zero preds???");
251 if (!L->contains(*HPI)) ++HPI;
252 assert(HPI != pred_end(Header) && L->contains(*HPI) &&
253 "No backedge in loop?");
255 SCEVHandle TripCount = IterationCount;
257 if (*HPI == ExitingBlock) {
258 // The IterationCount expression contains the number of times that the
259 // backedge actually branches to the loop header. This is one less than the
260 // number of times the loop executes, so add one to it.
261 Constant *OneC = ConstantInt::get(IterationCount->getType(), 1);
262 TripCount = SCEVAddExpr::get(IterationCount, SCEVUnknown::get(OneC));
263 IndVar = L->getCanonicalInductionVariableIncrement();
265 // We have to use the preincremented value...
266 IndVar = L->getCanonicalInductionVariable();
269 // Expand the code for the iteration count into the preheader of the loop.
270 BasicBlock *Preheader = L->getLoopPreheader();
271 Value *ExitCnt = RW.expandCodeFor(TripCount, Preheader->getTerminator(),
274 // Insert a new icmp_ne or icmp_eq instruction before the branch.
275 ICmpInst::Predicate Opcode;
276 if (L->contains(BI->getSuccessor(0)))
277 Opcode = ICmpInst::ICMP_NE;
279 Opcode = ICmpInst::ICMP_EQ;
281 Value *Cond = new ICmpInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
282 BI->setCondition(Cond);
285 return PotentiallyDeadInst;
289 /// RewriteLoopExitValues - Check to see if this loop has a computable
290 /// loop-invariant execution count. If so, this means that we can compute the
291 /// final value of any expressions that are recurrent in the loop, and
292 /// substitute the exit values from the loop into any instructions outside of
293 /// the loop that use the final values of the current expressions.
294 void IndVarSimplify::RewriteLoopExitValues(Loop *L) {
295 BasicBlock *Preheader = L->getLoopPreheader();
297 // Scan all of the instructions in the loop, looking at those that have
298 // extra-loop users and which are recurrences.
299 SCEVExpander Rewriter(*SE, *LI);
301 // We insert the code into the preheader of the loop if the loop contains
302 // multiple exit blocks, or in the exit block if there is exactly one.
303 BasicBlock *BlockToInsertInto;
304 std::vector<BasicBlock*> ExitBlocks;
305 L->getExitBlocks(ExitBlocks);
306 if (ExitBlocks.size() == 1)
307 BlockToInsertInto = ExitBlocks[0];
309 BlockToInsertInto = Preheader;
310 BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
311 while (isa<PHINode>(InsertPt)) ++InsertPt;
313 bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
315 std::set<Instruction*> InstructionsToDelete;
317 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
318 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
319 BasicBlock *BB = L->getBlocks()[i];
320 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
321 if (I->getType()->isInteger()) { // Is an integer instruction
322 SCEVHandle SH = SE->getSCEV(I);
323 if (SH->hasComputableLoopEvolution(L) || // Varies predictably
324 HasConstantItCount) {
325 // Find out if this predictably varying value is actually used
326 // outside of the loop. "extra" as opposed to "intra".
327 std::vector<Instruction*> ExtraLoopUsers;
328 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
330 Instruction *User = cast<Instruction>(*UI);
331 if (!L->contains(User->getParent())) {
332 // If this is a PHI node in the exit block and we're inserting,
333 // into the exit block, it must have a single entry. In this
334 // case, we can't insert the code after the PHI and have the PHI
335 // still use it. Instead, don't insert the the PHI.
336 if (PHINode *PN = dyn_cast<PHINode>(User)) {
337 // FIXME: This is a case where LCSSA pessimizes code, this
338 // should be fixed better.
339 if (PN->getNumOperands() == 2 &&
340 PN->getParent() == BlockToInsertInto)
343 ExtraLoopUsers.push_back(User);
347 if (!ExtraLoopUsers.empty()) {
348 // Okay, this instruction has a user outside of the current loop
349 // and varies predictably in this loop. Evaluate the value it
350 // contains when the loop exits, and insert code for it.
351 SCEVHandle ExitValue = SE->getSCEVAtScope(I, L->getParentLoop());
352 if (!isa<SCEVCouldNotCompute>(ExitValue)) {
355 // Remember the next instruction. The rewriter can move code
356 // around in some cases.
357 BasicBlock::iterator NextI = I; ++NextI;
359 Value *NewVal = Rewriter.expandCodeFor(ExitValue, InsertPt,
362 // Rewrite any users of the computed value outside of the loop
363 // with the newly computed value.
364 for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i) {
365 PHINode* PN = dyn_cast<PHINode>(ExtraLoopUsers[i]);
366 if (PN && PN->getNumOperands() == 2 &&
367 !L->contains(PN->getParent())) {
368 // We're dealing with an LCSSA Phi. Handle it specially.
369 Instruction* LCSSAInsertPt = BlockToInsertInto->begin();
371 Instruction* NewInstr = dyn_cast<Instruction>(NewVal);
372 if (NewInstr && !isa<PHINode>(NewInstr) &&
373 !L->contains(NewInstr->getParent()))
374 for (unsigned j = 0; j < NewInstr->getNumOperands(); ++j){
376 dyn_cast<Instruction>(NewInstr->getOperand(j));
377 if (PredI && L->contains(PredI->getParent())) {
378 PHINode* NewLCSSA = new PHINode(PredI->getType(),
379 PredI->getName() + ".lcssa",
381 NewLCSSA->addIncoming(PredI,
382 BlockToInsertInto->getSinglePredecessor());
384 NewInstr->replaceUsesOfWith(PredI, NewLCSSA);
388 PN->replaceAllUsesWith(NewVal);
389 PN->eraseFromParent();
391 ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
395 // If this instruction is dead now, schedule it to be removed.
397 InstructionsToDelete.insert(I);
399 continue; // Skip the ++I
405 // Next instruction. Continue instruction skips this.
410 DeleteTriviallyDeadInstructions(InstructionsToDelete);
414 void IndVarSimplify::runOnLoop(Loop *L) {
415 // First step. Check to see if there are any trivial GEP pointer recurrences.
416 // If there are, change them into integer recurrences, permitting analysis by
417 // the SCEV routines.
419 BasicBlock *Header = L->getHeader();
420 BasicBlock *Preheader = L->getLoopPreheader();
422 std::set<Instruction*> DeadInsts;
423 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
424 PHINode *PN = cast<PHINode>(I);
425 if (isa<PointerType>(PN->getType()))
426 EliminatePointerRecurrence(PN, Preheader, DeadInsts);
429 if (!DeadInsts.empty())
430 DeleteTriviallyDeadInstructions(DeadInsts);
433 // Next, transform all loops nesting inside of this loop.
434 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
437 // Check to see if this loop has a computable loop-invariant execution count.
438 // If so, this means that we can compute the final value of any expressions
439 // that are recurrent in the loop, and substitute the exit values from the
440 // loop into any instructions outside of the loop that use the final values of
441 // the current expressions.
443 SCEVHandle IterationCount = SE->getIterationCount(L);
444 if (!isa<SCEVCouldNotCompute>(IterationCount))
445 RewriteLoopExitValues(L);
447 // Next, analyze all of the induction variables in the loop, canonicalizing
448 // auxillary induction variables.
449 std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
451 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
452 PHINode *PN = cast<PHINode>(I);
453 if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
454 SCEVHandle SCEV = SE->getSCEV(PN);
455 if (SCEV->hasComputableLoopEvolution(L))
456 // FIXME: It is an extremely bad idea to indvar substitute anything more
457 // complex than affine induction variables. Doing so will put expensive
458 // polynomial evaluations inside of the loop, and the str reduction pass
459 // currently can only reduce affine polynomials. For now just disable
460 // indvar subst on anything more complex than an affine addrec.
461 if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
463 IndVars.push_back(std::make_pair(PN, SCEV));
467 // If there are no induction variables in the loop, there is nothing more to
469 if (IndVars.empty()) {
470 // Actually, if we know how many times the loop iterates, lets insert a
471 // canonical induction variable to help subsequent passes.
472 if (!isa<SCEVCouldNotCompute>(IterationCount)) {
473 SCEVExpander Rewriter(*SE, *LI);
474 Rewriter.getOrInsertCanonicalInductionVariable(L,
475 IterationCount->getType());
476 if (Instruction *I = LinearFunctionTestReplace(L, IterationCount,
478 std::set<Instruction*> InstructionsToDelete;
479 InstructionsToDelete.insert(I);
480 DeleteTriviallyDeadInstructions(InstructionsToDelete);
486 // Compute the type of the largest recurrence expression.
488 const Type *LargestType = IndVars[0].first->getType();
489 bool DifferingSizes = false;
490 for (unsigned i = 1, e = IndVars.size(); i != e; ++i) {
491 const Type *Ty = IndVars[i].first->getType();
492 DifferingSizes |= Ty->getPrimitiveSize() != LargestType->getPrimitiveSize();
493 if (Ty->getPrimitiveSize() > LargestType->getPrimitiveSize())
497 // Create a rewriter object which we'll use to transform the code with.
498 SCEVExpander Rewriter(*SE, *LI);
500 // Now that we know the largest of of the induction variables in this loop,
501 // insert a canonical induction variable of the largest size.
502 Value *IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
506 if (!isa<SCEVCouldNotCompute>(IterationCount))
507 if (Instruction *DI = LinearFunctionTestReplace(L, IterationCount,Rewriter))
508 DeadInsts.insert(DI);
510 // Now that we have a canonical induction variable, we can rewrite any
511 // recurrences in terms of the induction variable. Start with the auxillary
512 // induction variables, and recursively rewrite any of their uses.
513 BasicBlock::iterator InsertPt = Header->begin();
514 while (isa<PHINode>(InsertPt)) ++InsertPt;
516 // If there were induction variables of other sizes, cast the primary
517 // induction variable to the right size for them, avoiding the need for the
518 // code evaluation methods to insert induction variables of different sizes.
519 if (DifferingSizes) {
520 bool InsertedSizes[17] = { false };
521 InsertedSizes[LargestType->getPrimitiveSize()] = true;
522 for (unsigned i = 0, e = IndVars.size(); i != e; ++i)
523 if (!InsertedSizes[IndVars[i].first->getType()->getPrimitiveSize()]) {
524 PHINode *PN = IndVars[i].first;
525 InsertedSizes[PN->getType()->getPrimitiveSize()] = true;
526 Instruction *New = CastInst::create(Instruction::Trunc, IndVar,
527 PN->getType(), "indvar", InsertPt);
528 Rewriter.addInsertedValue(New, SE->getSCEV(New));
532 // If there were induction variables of other sizes, cast the primary
533 // induction variable to the right size for them, avoiding the need for the
534 // code evaluation methods to insert induction variables of different sizes.
535 std::map<unsigned, Value*> InsertedSizes;
536 while (!IndVars.empty()) {
537 PHINode *PN = IndVars.back().first;
538 Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt,
540 std::string Name = PN->getName();
542 NewVal->setName(Name);
544 // Replace the old PHI Node with the inserted computation.
545 PN->replaceAllUsesWith(NewVal);
546 DeadInsts.insert(PN);
553 // Now replace all derived expressions in the loop body with simpler
555 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
556 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
557 BasicBlock *BB = L->getBlocks()[i];
558 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
559 if (I->getType()->isInteger() && // Is an integer instruction
561 !Rewriter.isInsertedInstruction(I)) {
562 SCEVHandle SH = SE->getSCEV(I);
563 Value *V = Rewriter.expandCodeFor(SH, I, I->getType());
565 if (isa<Instruction>(V)) {
566 std::string Name = I->getName();
570 I->replaceAllUsesWith(V);
579 DeleteTriviallyDeadInstructions(DeadInsts);
581 if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());