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/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
57 STATISTIC(NumRemoved , "Number of aux indvars removed");
58 STATISTIC(NumPointer , "Number of pointer indvars promoted");
59 STATISTIC(NumInserted, "Number of canonical indvars added");
60 STATISTIC(NumReplaced, "Number of exit values replaced");
61 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
64 class IndVarSimplify : public FunctionPass {
69 virtual bool runOnFunction(Function &) {
70 LI = &getAnalysis<LoopInfo>();
71 SE = &getAnalysis<ScalarEvolution>();
74 // Induction Variables live in the header nodes of loops
75 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
80 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
81 AU.addRequiredID(LoopSimplifyID);
82 AU.addRequired<ScalarEvolution>();
83 AU.addRequired<LoopInfo>();
84 AU.addPreservedID(LoopSimplifyID);
85 AU.addPreservedID(LCSSAID);
89 void runOnLoop(Loop *L);
90 void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
91 std::set<Instruction*> &DeadInsts);
92 Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
94 void RewriteLoopExitValues(Loop *L);
96 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
98 RegisterPass<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
101 FunctionPass *llvm::createIndVarSimplifyPass() {
102 return new IndVarSimplify();
105 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
106 /// specified set are trivially dead, delete them and see if this makes any of
107 /// their operands subsequently dead.
108 void IndVarSimplify::
109 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
110 while (!Insts.empty()) {
111 Instruction *I = *Insts.begin();
112 Insts.erase(Insts.begin());
113 if (isInstructionTriviallyDead(I)) {
114 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
115 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
117 SE->deleteInstructionFromRecords(I);
118 DOUT << "INDVARS: Deleting: " << *I;
119 I->eraseFromParent();
126 /// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer
127 /// recurrence. If so, change it into an integer recurrence, permitting
128 /// analysis by the SCEV routines.
129 void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN,
130 BasicBlock *Preheader,
131 std::set<Instruction*> &DeadInsts) {
132 assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
133 unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
134 unsigned BackedgeIdx = PreheaderIdx^1;
135 if (GetElementPtrInst *GEPI =
136 dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
137 if (GEPI->getOperand(0) == PN) {
138 assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
139 DOUT << "INDVARS: Eliminating pointer recurrence: " << *GEPI;
141 // Okay, we found a pointer recurrence. Transform this pointer
142 // recurrence into an integer recurrence. Compute the value that gets
143 // added to the pointer at every iteration.
144 Value *AddedVal = GEPI->getOperand(1);
146 // Insert a new integer PHI node into the top of the block.
147 PHINode *NewPhi = new PHINode(AddedVal->getType(),
148 PN->getName()+".rec", PN);
149 NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
151 // Create the new add instruction.
152 Value *NewAdd = BinaryOperator::createAdd(NewPhi, AddedVal,
153 GEPI->getName()+".rec", GEPI);
154 NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx));
156 // Update the existing GEP to use the recurrence.
157 GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx));
159 // Update the GEP to use the new recurrence we just inserted.
160 GEPI->setOperand(1, NewAdd);
162 // If the incoming value is a constant expr GEP, try peeling out the array
163 // 0 index if possible to make things simpler.
164 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEPI->getOperand(0)))
165 if (CE->getOpcode() == Instruction::GetElementPtr) {
166 unsigned NumOps = CE->getNumOperands();
167 assert(NumOps > 1 && "CE folding didn't work!");
168 if (CE->getOperand(NumOps-1)->isNullValue()) {
169 // Check to make sure the last index really is an array index.
170 gep_type_iterator GTI = gep_type_begin(CE);
171 for (unsigned i = 1, e = CE->getNumOperands()-1;
174 if (isa<SequentialType>(*GTI)) {
175 // Pull the last index out of the constant expr GEP.
176 std::vector<Value*> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
177 Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
179 GetElementPtrInst *NGEPI =
180 new GetElementPtrInst(NCE, Constant::getNullValue(Type::Int32Ty),
181 NewAdd, GEPI->getName(), GEPI);
182 GEPI->replaceAllUsesWith(NGEPI);
183 GEPI->eraseFromParent();
190 // Finally, if there are any other users of the PHI node, we must
191 // insert a new GEP instruction that uses the pre-incremented version
192 // of the induction amount.
193 if (!PN->use_empty()) {
194 BasicBlock::iterator InsertPos = PN; ++InsertPos;
195 while (isa<PHINode>(InsertPos)) ++InsertPos;
196 std::string Name = PN->getName(); PN->setName("");
198 new GetElementPtrInst(PN->getIncomingValue(PreheaderIdx),
199 std::vector<Value*>(1, NewPhi), Name,
201 PN->replaceAllUsesWith(PreInc);
204 // Delete the old PHI for sure, and the GEP if its otherwise unused.
205 DeadInsts.insert(PN);
212 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
213 /// loop to be a canonical != comparison against the incremented loop induction
214 /// variable. This pass is able to rewrite the exit tests of any loop where the
215 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
216 /// is actually a much broader range than just linear tests.
218 /// This method returns a "potentially dead" instruction whose computation chain
219 /// should be deleted when convenient.
220 Instruction *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
221 SCEV *IterationCount,
223 // Find the exit block for the loop. We can currently only handle loops with
225 std::vector<BasicBlock*> ExitBlocks;
226 L->getExitBlocks(ExitBlocks);
227 if (ExitBlocks.size() != 1) return 0;
228 BasicBlock *ExitBlock = ExitBlocks[0];
230 // Make sure there is only one predecessor block in the loop.
231 BasicBlock *ExitingBlock = 0;
232 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
234 if (L->contains(*PI)) {
235 if (ExitingBlock == 0)
238 return 0; // Multiple exits from loop to this block.
240 assert(ExitingBlock && "Loop info is broken");
242 if (!isa<BranchInst>(ExitingBlock->getTerminator()))
243 return 0; // Can't rewrite non-branch yet
244 BranchInst *BI = cast<BranchInst>(ExitingBlock->getTerminator());
245 assert(BI->isConditional() && "Must be conditional to be part of loop!");
247 Instruction *PotentiallyDeadInst = dyn_cast<Instruction>(BI->getCondition());
249 // If the exiting block is not the same as the backedge block, we must compare
250 // against the preincremented value, otherwise we prefer to compare against
251 // the post-incremented value.
252 BasicBlock *Header = L->getHeader();
253 pred_iterator HPI = pred_begin(Header);
254 assert(HPI != pred_end(Header) && "Loop with zero preds???");
255 if (!L->contains(*HPI)) ++HPI;
256 assert(HPI != pred_end(Header) && L->contains(*HPI) &&
257 "No backedge in loop?");
259 SCEVHandle TripCount = IterationCount;
261 if (*HPI == ExitingBlock) {
262 // The IterationCount expression contains the number of times that the
263 // backedge actually branches to the loop header. This is one less than the
264 // number of times the loop executes, so add one to it.
265 Constant *OneC = ConstantInt::get(IterationCount->getType(), 1);
266 TripCount = SCEVAddExpr::get(IterationCount, SCEVUnknown::get(OneC));
267 IndVar = L->getCanonicalInductionVariableIncrement();
269 // We have to use the preincremented value...
270 IndVar = L->getCanonicalInductionVariable();
273 DOUT << "INDVARS: LFTR: TripCount = " << *TripCount
274 << " IndVar = " << *IndVar << "\n";
276 // Expand the code for the iteration count into the preheader of the loop.
277 BasicBlock *Preheader = L->getLoopPreheader();
278 Value *ExitCnt = RW.expandCodeFor(TripCount, Preheader->getTerminator(),
281 // Insert a new icmp_ne or icmp_eq instruction before the branch.
282 ICmpInst::Predicate Opcode;
283 if (L->contains(BI->getSuccessor(0)))
284 Opcode = ICmpInst::ICMP_NE;
286 Opcode = ICmpInst::ICMP_EQ;
288 Value *Cond = new ICmpInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
289 BI->setCondition(Cond);
292 return PotentiallyDeadInst;
296 /// RewriteLoopExitValues - Check to see if this loop has a computable
297 /// loop-invariant execution count. If so, this means that we can compute the
298 /// final value of any expressions that are recurrent in the loop, and
299 /// substitute the exit values from the loop into any instructions outside of
300 /// the loop that use the final values of the current expressions.
301 void IndVarSimplify::RewriteLoopExitValues(Loop *L) {
302 BasicBlock *Preheader = L->getLoopPreheader();
304 // Scan all of the instructions in the loop, looking at those that have
305 // extra-loop users and which are recurrences.
306 SCEVExpander Rewriter(*SE, *LI);
308 // We insert the code into the preheader of the loop if the loop contains
309 // multiple exit blocks, or in the exit block if there is exactly one.
310 BasicBlock *BlockToInsertInto;
311 std::vector<BasicBlock*> ExitBlocks;
312 L->getExitBlocks(ExitBlocks);
313 if (ExitBlocks.size() == 1)
314 BlockToInsertInto = ExitBlocks[0];
316 BlockToInsertInto = Preheader;
317 BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
318 while (isa<PHINode>(InsertPt)) ++InsertPt;
320 bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
322 std::set<Instruction*> InstructionsToDelete;
324 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
325 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
326 BasicBlock *BB = L->getBlocks()[i];
327 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
328 if (I->getType()->isInteger()) { // Is an integer instruction
329 SCEVHandle SH = SE->getSCEV(I);
330 if (SH->hasComputableLoopEvolution(L) || // Varies predictably
331 HasConstantItCount) {
332 // Find out if this predictably varying value is actually used
333 // outside of the loop. "extra" as opposed to "intra".
334 std::vector<Instruction*> ExtraLoopUsers;
335 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
337 Instruction *User = cast<Instruction>(*UI);
338 if (!L->contains(User->getParent())) {
339 // If this is a PHI node in the exit block and we're inserting,
340 // into the exit block, it must have a single entry. In this
341 // case, we can't insert the code after the PHI and have the PHI
342 // still use it. Instead, don't insert the the PHI.
343 if (PHINode *PN = dyn_cast<PHINode>(User)) {
344 // FIXME: This is a case where LCSSA pessimizes code, this
345 // should be fixed better.
346 if (PN->getNumOperands() == 2 &&
347 PN->getParent() == BlockToInsertInto)
350 ExtraLoopUsers.push_back(User);
354 if (!ExtraLoopUsers.empty()) {
355 // Okay, this instruction has a user outside of the current loop
356 // and varies predictably in this loop. Evaluate the value it
357 // contains when the loop exits, and insert code for it.
358 SCEVHandle ExitValue = SE->getSCEVAtScope(I, L->getParentLoop());
359 if (!isa<SCEVCouldNotCompute>(ExitValue)) {
362 // Remember the next instruction. The rewriter can move code
363 // around in some cases.
364 BasicBlock::iterator NextI = I; ++NextI;
366 Value *NewVal = Rewriter.expandCodeFor(ExitValue, InsertPt,
369 DOUT << "INDVARS: RLEV: AfterLoopVal = " << *NewVal
370 << " LoopVal = " << *I << "\n";
372 // Rewrite any users of the computed value outside of the loop
373 // with the newly computed value.
374 for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i) {
375 PHINode* PN = dyn_cast<PHINode>(ExtraLoopUsers[i]);
376 if (PN && PN->getNumOperands() == 2 &&
377 !L->contains(PN->getParent())) {
378 // We're dealing with an LCSSA Phi. Handle it specially.
379 Instruction* LCSSAInsertPt = BlockToInsertInto->begin();
381 Instruction* NewInstr = dyn_cast<Instruction>(NewVal);
382 if (NewInstr && !isa<PHINode>(NewInstr) &&
383 !L->contains(NewInstr->getParent()))
384 for (unsigned j = 0; j < NewInstr->getNumOperands(); ++j){
386 dyn_cast<Instruction>(NewInstr->getOperand(j));
387 if (PredI && L->contains(PredI->getParent())) {
388 PHINode* NewLCSSA = new PHINode(PredI->getType(),
389 PredI->getName() + ".lcssa",
391 NewLCSSA->addIncoming(PredI,
392 BlockToInsertInto->getSinglePredecessor());
394 NewInstr->replaceUsesOfWith(PredI, NewLCSSA);
398 PN->replaceAllUsesWith(NewVal);
399 PN->eraseFromParent();
401 ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
405 // If this instruction is dead now, schedule it to be removed.
407 InstructionsToDelete.insert(I);
409 continue; // Skip the ++I
415 // Next instruction. Continue instruction skips this.
420 DeleteTriviallyDeadInstructions(InstructionsToDelete);
424 void IndVarSimplify::runOnLoop(Loop *L) {
425 // First step. Check to see if there are any trivial GEP pointer recurrences.
426 // If there are, change them into integer recurrences, permitting analysis by
427 // the SCEV routines.
429 BasicBlock *Header = L->getHeader();
430 BasicBlock *Preheader = L->getLoopPreheader();
432 std::set<Instruction*> DeadInsts;
433 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
434 PHINode *PN = cast<PHINode>(I);
435 if (isa<PointerType>(PN->getType()))
436 EliminatePointerRecurrence(PN, Preheader, DeadInsts);
439 if (!DeadInsts.empty())
440 DeleteTriviallyDeadInstructions(DeadInsts);
443 // Next, transform all loops nesting inside of this loop.
444 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
447 // Check to see if this loop has a computable loop-invariant execution count.
448 // If so, this means that we can compute the final value of any expressions
449 // that are recurrent in the loop, and substitute the exit values from the
450 // loop into any instructions outside of the loop that use the final values of
451 // the current expressions.
453 SCEVHandle IterationCount = SE->getIterationCount(L);
454 if (!isa<SCEVCouldNotCompute>(IterationCount))
455 RewriteLoopExitValues(L);
457 // Next, analyze all of the induction variables in the loop, canonicalizing
458 // auxillary induction variables.
459 std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
461 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
462 PHINode *PN = cast<PHINode>(I);
463 if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
464 SCEVHandle SCEV = SE->getSCEV(PN);
465 if (SCEV->hasComputableLoopEvolution(L))
466 // FIXME: It is an extremely bad idea to indvar substitute anything more
467 // complex than affine induction variables. Doing so will put expensive
468 // polynomial evaluations inside of the loop, and the str reduction pass
469 // currently can only reduce affine polynomials. For now just disable
470 // indvar subst on anything more complex than an affine addrec.
471 if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
473 IndVars.push_back(std::make_pair(PN, SCEV));
477 // If there are no induction variables in the loop, there is nothing more to
479 if (IndVars.empty()) {
480 // Actually, if we know how many times the loop iterates, lets insert a
481 // canonical induction variable to help subsequent passes.
482 if (!isa<SCEVCouldNotCompute>(IterationCount)) {
483 SCEVExpander Rewriter(*SE, *LI);
484 Rewriter.getOrInsertCanonicalInductionVariable(L,
485 IterationCount->getType());
486 if (Instruction *I = LinearFunctionTestReplace(L, IterationCount,
488 std::set<Instruction*> InstructionsToDelete;
489 InstructionsToDelete.insert(I);
490 DeleteTriviallyDeadInstructions(InstructionsToDelete);
496 // Compute the type of the largest recurrence expression.
498 const Type *LargestType = IndVars[0].first->getType();
499 bool DifferingSizes = false;
500 for (unsigned i = 1, e = IndVars.size(); i != e; ++i) {
501 const Type *Ty = IndVars[i].first->getType();
503 Ty->getPrimitiveSizeInBits() != LargestType->getPrimitiveSizeInBits();
504 if (Ty->getPrimitiveSizeInBits() > LargestType->getPrimitiveSizeInBits())
508 // Create a rewriter object which we'll use to transform the code with.
509 SCEVExpander Rewriter(*SE, *LI);
511 // Now that we know the largest of of the induction variables in this loop,
512 // insert a canonical induction variable of the largest size.
513 Value *IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
516 DOUT << "INDVARS: New CanIV: " << *IndVar;
518 if (!isa<SCEVCouldNotCompute>(IterationCount))
519 if (Instruction *DI = LinearFunctionTestReplace(L, IterationCount,Rewriter))
520 DeadInsts.insert(DI);
522 // Now that we have a canonical induction variable, we can rewrite any
523 // recurrences in terms of the induction variable. Start with the auxillary
524 // induction variables, and recursively rewrite any of their uses.
525 BasicBlock::iterator InsertPt = Header->begin();
526 while (isa<PHINode>(InsertPt)) ++InsertPt;
528 // If there were induction variables of other sizes, cast the primary
529 // induction variable to the right size for them, avoiding the need for the
530 // code evaluation methods to insert induction variables of different sizes.
531 if (DifferingSizes) {
532 SmallVector<unsigned,4> InsertedSizes;
533 InsertedSizes.push_back(LargestType->getPrimitiveSizeInBits());
534 for (unsigned i = 0, e = IndVars.size(); i != e; ++i) {
535 unsigned ithSize = IndVars[i].first->getType()->getPrimitiveSizeInBits();
536 if (std::find(InsertedSizes.begin(), InsertedSizes.end(), ithSize)
537 == InsertedSizes.end()) {
538 PHINode *PN = IndVars[i].first;
539 InsertedSizes.push_back(ithSize);
540 Instruction *New = new TruncInst(IndVar, PN->getType(), "indvar",
542 Rewriter.addInsertedValue(New, SE->getSCEV(New));
543 DOUT << "INDVARS: Made trunc IV for " << *PN
544 << " NewVal = " << *New << "\n";
549 // Rewrite all induction variables in terms of the canonical induction
551 std::map<unsigned, Value*> InsertedSizes;
552 while (!IndVars.empty()) {
553 PHINode *PN = IndVars.back().first;
554 Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt,
556 DOUT << "INDVARS: Rewrote IV '" << *IndVars.back().second << "' " << *PN
557 << " into = " << *NewVal << "\n";
558 std::string Name = PN->getName();
560 NewVal->setName(Name);
562 // Replace the old PHI Node with the inserted computation.
563 PN->replaceAllUsesWith(NewVal);
564 DeadInsts.insert(PN);
571 // Now replace all derived expressions in the loop body with simpler
573 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
574 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
575 BasicBlock *BB = L->getBlocks()[i];
576 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
577 if (I->getType()->isInteger() && // Is an integer instruction
579 !Rewriter.isInsertedInstruction(I)) {
580 SCEVHandle SH = SE->getSCEV(I);
581 Value *V = Rewriter.expandCodeFor(SH, I, I->getType());
583 if (isa<Instruction>(V)) {
584 std::string Name = I->getName();
588 I->replaceAllUsesWith(V);
597 DeleteTriviallyDeadInstructions(DeadInsts);
599 if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());