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 #include "llvm/Transforms/Scalar.h"
41 #include "llvm/BasicBlock.h"
42 #include "llvm/Constants.h"
43 #include "llvm/Instructions.h"
44 #include "llvm/Type.h"
45 #include "llvm/Analysis/ScalarEvolutionExpander.h"
46 #include "llvm/Analysis/LoopInfo.h"
47 #include "llvm/Support/CFG.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/ADT/Statistic.h"
55 Statistic<> NumRemoved ("indvars", "Number of aux indvars removed");
56 Statistic<> NumPointer ("indvars", "Number of pointer indvars promoted");
57 Statistic<> NumInserted("indvars", "Number of canonical indvars added");
58 Statistic<> NumReplaced("indvars", "Number of exit values replaced");
59 Statistic<> NumLFTR ("indvars", "Number of loop exit tests replaced");
61 class IndVarSimplify : public FunctionPass {
66 virtual bool runOnFunction(Function &) {
67 LI = &getAnalysis<LoopInfo>();
68 SE = &getAnalysis<ScalarEvolution>();
71 // Induction Variables live in the header nodes of loops
72 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
77 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
78 AU.addRequiredID(LoopSimplifyID);
79 AU.addRequired<ScalarEvolution>();
80 AU.addRequired<LoopInfo>();
81 AU.addPreservedID(LoopSimplifyID);
82 AU.addPreservedID(LCSSAID);
86 void runOnLoop(Loop *L);
87 void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
88 std::set<Instruction*> &DeadInsts);
89 void LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
91 void RewriteLoopExitValues(Loop *L);
93 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
95 RegisterPass<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
98 FunctionPass *llvm::createIndVarSimplifyPass() {
99 return new IndVarSimplify();
102 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
103 /// specified set are trivially dead, delete them and see if this makes any of
104 /// their operands subsequently dead.
105 void IndVarSimplify::
106 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
107 while (!Insts.empty()) {
108 Instruction *I = *Insts.begin();
109 Insts.erase(Insts.begin());
110 if (isInstructionTriviallyDead(I)) {
111 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
112 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
114 SE->deleteInstructionFromRecords(I);
115 I->eraseFromParent();
122 /// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer
123 /// recurrence. If so, change it into an integer recurrence, permitting
124 /// analysis by the SCEV routines.
125 void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN,
126 BasicBlock *Preheader,
127 std::set<Instruction*> &DeadInsts) {
128 assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
129 unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
130 unsigned BackedgeIdx = PreheaderIdx^1;
131 if (GetElementPtrInst *GEPI =
132 dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
133 if (GEPI->getOperand(0) == PN) {
134 assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
136 // Okay, we found a pointer recurrence. Transform this pointer
137 // recurrence into an integer recurrence. Compute the value that gets
138 // added to the pointer at every iteration.
139 Value *AddedVal = GEPI->getOperand(1);
141 // Insert a new integer PHI node into the top of the block.
142 PHINode *NewPhi = new PHINode(AddedVal->getType(),
143 PN->getName()+".rec", PN);
144 NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
146 // Create the new add instruction.
147 Value *NewAdd = BinaryOperator::createAdd(NewPhi, AddedVal,
148 GEPI->getName()+".rec", GEPI);
149 NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx));
151 // Update the existing GEP to use the recurrence.
152 GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx));
154 // Update the GEP to use the new recurrence we just inserted.
155 GEPI->setOperand(1, NewAdd);
157 // If the incoming value is a constant expr GEP, try peeling out the array
158 // 0 index if possible to make things simpler.
159 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEPI->getOperand(0)))
160 if (CE->getOpcode() == Instruction::GetElementPtr) {
161 unsigned NumOps = CE->getNumOperands();
162 assert(NumOps > 1 && "CE folding didn't work!");
163 if (CE->getOperand(NumOps-1)->isNullValue()) {
164 // Check to make sure the last index really is an array index.
165 gep_type_iterator GTI = gep_type_begin(CE);
166 for (unsigned i = 1, e = CE->getNumOperands()-1;
169 if (isa<SequentialType>(*GTI)) {
170 // Pull the last index out of the constant expr GEP.
171 std::vector<Value*> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
172 Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
174 GetElementPtrInst *NGEPI =
175 new GetElementPtrInst(NCE, Constant::getNullValue(Type::IntTy),
176 NewAdd, GEPI->getName(), GEPI);
177 GEPI->replaceAllUsesWith(NGEPI);
178 GEPI->eraseFromParent();
185 // Finally, if there are any other users of the PHI node, we must
186 // insert a new GEP instruction that uses the pre-incremented version
187 // of the induction amount.
188 if (!PN->use_empty()) {
189 BasicBlock::iterator InsertPos = PN; ++InsertPos;
190 while (isa<PHINode>(InsertPos)) ++InsertPos;
191 std::string Name = PN->getName(); PN->setName("");
193 new GetElementPtrInst(PN->getIncomingValue(PreheaderIdx),
194 std::vector<Value*>(1, NewPhi), Name,
196 PN->replaceAllUsesWith(PreInc);
199 // Delete the old PHI for sure, and the GEP if its otherwise unused.
200 DeadInsts.insert(PN);
207 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
208 /// loop to be a canonical != comparison against the incremented loop induction
209 /// variable. This pass is able to rewrite the exit tests of any loop where the
210 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
211 /// is actually a much broader range than just linear tests.
212 void IndVarSimplify::LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
214 // Find the exit block for the loop. We can currently only handle loops with
216 std::vector<BasicBlock*> ExitBlocks;
217 L->getExitBlocks(ExitBlocks);
218 if (ExitBlocks.size() != 1) return;
219 BasicBlock *ExitBlock = ExitBlocks[0];
221 // Make sure there is only one predecessor block in the loop.
222 BasicBlock *ExitingBlock = 0;
223 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
225 if (L->contains(*PI)) {
226 if (ExitingBlock == 0)
229 return; // Multiple exits from loop to this block.
231 assert(ExitingBlock && "Loop info is broken");
233 if (!isa<BranchInst>(ExitingBlock->getTerminator()))
234 return; // Can't rewrite non-branch yet
235 BranchInst *BI = cast<BranchInst>(ExitingBlock->getTerminator());
236 assert(BI->isConditional() && "Must be conditional to be part of loop!");
238 std::set<Instruction*> InstructionsToDelete;
239 if (Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()))
240 InstructionsToDelete.insert(Cond);
242 // If the exiting block is not the same as the backedge block, we must compare
243 // against the preincremented value, otherwise we prefer to compare against
244 // the post-incremented value.
245 BasicBlock *Header = L->getHeader();
246 pred_iterator HPI = pred_begin(Header);
247 assert(HPI != pred_end(Header) && "Loop with zero preds???");
248 if (!L->contains(*HPI)) ++HPI;
249 assert(HPI != pred_end(Header) && L->contains(*HPI) &&
250 "No backedge in loop?");
252 SCEVHandle TripCount = IterationCount;
254 if (*HPI == ExitingBlock) {
255 // The IterationCount expression contains the number of times that the
256 // backedge actually branches to the loop header. This is one less than the
257 // number of times the loop executes, so add one to it.
258 Constant *OneC = ConstantInt::get(IterationCount->getType(), 1);
259 TripCount = SCEVAddExpr::get(IterationCount, SCEVUnknown::get(OneC));
260 IndVar = L->getCanonicalInductionVariableIncrement();
262 // We have to use the preincremented value...
263 IndVar = L->getCanonicalInductionVariable();
266 // Expand the code for the iteration count into the preheader of the loop.
267 BasicBlock *Preheader = L->getLoopPreheader();
268 Value *ExitCnt = RW.expandCodeFor(TripCount, Preheader->getTerminator(),
271 // Insert a new setne or seteq instruction before the branch.
272 Instruction::BinaryOps Opcode;
273 if (L->contains(BI->getSuccessor(0)))
274 Opcode = Instruction::SetNE;
276 Opcode = Instruction::SetEQ;
278 Value *Cond = new SetCondInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
279 BI->setCondition(Cond);
283 DeleteTriviallyDeadInstructions(InstructionsToDelete);
287 /// RewriteLoopExitValues - Check to see if this loop has a computable
288 /// loop-invariant execution count. If so, this means that we can compute the
289 /// final value of any expressions that are recurrent in the loop, and
290 /// substitute the exit values from the loop into any instructions outside of
291 /// the loop that use the final values of the current expressions.
292 void IndVarSimplify::RewriteLoopExitValues(Loop *L) {
293 BasicBlock *Preheader = L->getLoopPreheader();
295 // Scan all of the instructions in the loop, looking at those that have
296 // extra-loop users and which are recurrences.
297 SCEVExpander Rewriter(*SE, *LI);
299 // We insert the code into the preheader of the loop if the loop contains
300 // multiple exit blocks, or in the exit block if there is exactly one.
301 BasicBlock *BlockToInsertInto;
302 std::vector<BasicBlock*> ExitBlocks;
303 L->getExitBlocks(ExitBlocks);
304 if (ExitBlocks.size() == 1)
305 BlockToInsertInto = ExitBlocks[0];
307 BlockToInsertInto = Preheader;
308 BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
309 while (isa<PHINode>(InsertPt)) ++InsertPt;
311 bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
313 std::set<Instruction*> InstructionsToDelete;
315 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
316 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
317 BasicBlock *BB = L->getBlocks()[i];
318 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
319 if (I->getType()->isInteger()) { // Is an integer instruction
320 SCEVHandle SH = SE->getSCEV(I);
321 if (SH->hasComputableLoopEvolution(L) || // Varies predictably
322 HasConstantItCount) {
323 // Find out if this predictably varying value is actually used
324 // outside of the loop. "extra" as opposed to "intra".
325 std::vector<Instruction*> ExtraLoopUsers;
326 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
328 Instruction *User = cast<Instruction>(*UI);
329 if (!L->contains(User->getParent())) {
330 // If this is a PHI node in the exit block and we're inserting,
331 // into the exit block, it must have a single entry. In this
332 // case, we can't insert the code after the PHI and have the PHI
333 // still use it. Instead, don't insert the the PHI.
334 if (PHINode *PN = dyn_cast<PHINode>(User)) {
335 // FIXME: This is a case where LCSSA pessimizes code, this
336 // should be fixed better.
337 if (PN->getNumOperands() == 2 &&
338 PN->getParent() == BlockToInsertInto)
341 ExtraLoopUsers.push_back(User);
345 if (!ExtraLoopUsers.empty()) {
346 // Okay, this instruction has a user outside of the current loop
347 // and varies predictably in this loop. Evaluate the value it
348 // contains when the loop exits, and insert code for it.
349 SCEVHandle ExitValue = SE->getSCEVAtScope(I, L->getParentLoop());
350 if (!isa<SCEVCouldNotCompute>(ExitValue)) {
353 // Remember the next instruction. The rewriter can move code
354 // around in some cases.
355 BasicBlock::iterator NextI = I; ++NextI;
357 Value *NewVal = Rewriter.expandCodeFor(ExitValue, InsertPt,
360 // Rewrite any users of the computed value outside of the loop
361 // with the newly computed value.
362 for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i) {
363 PHINode* PN = dyn_cast<PHINode>(ExtraLoopUsers[i]);
364 if (PN && PN->getNumOperands() == 2 &&
365 !L->contains(PN->getParent())) {
366 // We're dealing with an LCSSA Phi. Handle it specially.
367 Instruction* LCSSAInsertPt = BlockToInsertInto->begin();
369 Instruction* NewInstr = dyn_cast<Instruction>(NewVal);
370 if (NewInstr && !isa<PHINode>(NewInstr) &&
371 !L->contains(NewInstr->getParent()))
372 for (unsigned j = 0; j < NewInstr->getNumOperands(); ++j){
374 dyn_cast<Instruction>(NewInstr->getOperand(j));
375 if (PredI && L->contains(PredI->getParent())) {
376 PHINode* NewLCSSA = new PHINode(PredI->getType(),
377 PredI->getName() + ".lcssa",
379 NewLCSSA->addIncoming(PredI,
380 BlockToInsertInto->getSinglePredecessor());
382 NewInstr->replaceUsesOfWith(PredI, NewLCSSA);
386 PN->replaceAllUsesWith(NewVal);
387 PN->eraseFromParent();
389 ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
393 // If this instruction is dead now, schedule it to be removed.
395 InstructionsToDelete.insert(I);
397 continue; // Skip the ++I
403 // Next instruction. Continue instruction skips this.
408 DeleteTriviallyDeadInstructions(InstructionsToDelete);
412 void IndVarSimplify::runOnLoop(Loop *L) {
413 // First step. Check to see if there are any trivial GEP pointer recurrences.
414 // If there are, change them into integer recurrences, permitting analysis by
415 // the SCEV routines.
417 BasicBlock *Header = L->getHeader();
418 BasicBlock *Preheader = L->getLoopPreheader();
420 std::set<Instruction*> DeadInsts;
421 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
422 PHINode *PN = cast<PHINode>(I);
423 if (isa<PointerType>(PN->getType()))
424 EliminatePointerRecurrence(PN, Preheader, DeadInsts);
427 if (!DeadInsts.empty())
428 DeleteTriviallyDeadInstructions(DeadInsts);
431 // Next, transform all loops nesting inside of this loop.
432 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
435 // Check to see if this loop has a computable loop-invariant execution count.
436 // If so, this means that we can compute the final value of any expressions
437 // that are recurrent in the loop, and substitute the exit values from the
438 // loop into any instructions outside of the loop that use the final values of
439 // the current expressions.
441 SCEVHandle IterationCount = SE->getIterationCount(L);
442 if (!isa<SCEVCouldNotCompute>(IterationCount))
443 RewriteLoopExitValues(L);
445 // Next, analyze all of the induction variables in the loop, canonicalizing
446 // auxillary induction variables.
447 std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
449 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
450 PHINode *PN = cast<PHINode>(I);
451 if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
452 SCEVHandle SCEV = SE->getSCEV(PN);
453 if (SCEV->hasComputableLoopEvolution(L))
454 // FIXME: It is an extremely bad idea to indvar substitute anything more
455 // complex than affine induction variables. Doing so will put expensive
456 // polynomial evaluations inside of the loop, and the str reduction pass
457 // currently can only reduce affine polynomials. For now just disable
458 // indvar subst on anything more complex than an affine addrec.
459 if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
461 IndVars.push_back(std::make_pair(PN, SCEV));
465 // If there are no induction variables in the loop, there is nothing more to
467 if (IndVars.empty()) {
468 // Actually, if we know how many times the loop iterates, lets insert a
469 // canonical induction variable to help subsequent passes.
470 if (!isa<SCEVCouldNotCompute>(IterationCount)) {
471 SCEVExpander Rewriter(*SE, *LI);
472 Rewriter.getOrInsertCanonicalInductionVariable(L,
473 IterationCount->getType());
474 LinearFunctionTestReplace(L, IterationCount, Rewriter);
479 // Compute the type of the largest recurrence expression.
481 const Type *LargestType = IndVars[0].first->getType();
482 bool DifferingSizes = false;
483 for (unsigned i = 1, e = IndVars.size(); i != e; ++i) {
484 const Type *Ty = IndVars[i].first->getType();
485 DifferingSizes |= Ty->getPrimitiveSize() != LargestType->getPrimitiveSize();
486 if (Ty->getPrimitiveSize() > LargestType->getPrimitiveSize())
490 // Create a rewriter object which we'll use to transform the code with.
491 SCEVExpander Rewriter(*SE, *LI);
493 // Now that we know the largest of of the induction variables in this loop,
494 // insert a canonical induction variable of the largest size.
495 LargestType = LargestType->getUnsignedVersion();
496 Value *IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
500 if (!isa<SCEVCouldNotCompute>(IterationCount))
501 LinearFunctionTestReplace(L, IterationCount, Rewriter);
503 // Now that we have a canonical induction variable, we can rewrite any
504 // recurrences in terms of the induction variable. Start with the auxillary
505 // induction variables, and recursively rewrite any of their uses.
506 BasicBlock::iterator InsertPt = Header->begin();
507 while (isa<PHINode>(InsertPt)) ++InsertPt;
509 // If there were induction variables of other sizes, cast the primary
510 // induction variable to the right size for them, avoiding the need for the
511 // code evaluation methods to insert induction variables of different sizes.
512 if (DifferingSizes) {
513 bool InsertedSizes[17] = { false };
514 InsertedSizes[LargestType->getPrimitiveSize()] = true;
515 for (unsigned i = 0, e = IndVars.size(); i != e; ++i)
516 if (!InsertedSizes[IndVars[i].first->getType()->getPrimitiveSize()]) {
517 PHINode *PN = IndVars[i].first;
518 InsertedSizes[PN->getType()->getPrimitiveSize()] = true;
519 Instruction *New = new CastInst(IndVar,
520 PN->getType()->getUnsignedVersion(),
522 Rewriter.addInsertedValue(New, SE->getSCEV(New));
526 // If there were induction variables of other sizes, cast the primary
527 // induction variable to the right size for them, avoiding the need for the
528 // code evaluation methods to insert induction variables of different sizes.
529 std::map<unsigned, Value*> InsertedSizes;
530 while (!IndVars.empty()) {
531 PHINode *PN = IndVars.back().first;
532 Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt,
534 std::string Name = PN->getName();
536 NewVal->setName(Name);
538 // Replace the old PHI Node with the inserted computation.
539 PN->replaceAllUsesWith(NewVal);
540 DeadInsts.insert(PN);
547 // Now replace all derived expressions in the loop body with simpler
549 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
550 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
551 BasicBlock *BB = L->getBlocks()[i];
552 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
553 if (I->getType()->isInteger() && // Is an integer instruction
555 !Rewriter.isInsertedInstruction(I)) {
556 SCEVHandle SH = SE->getSCEV(I);
557 Value *V = Rewriter.expandCodeFor(SH, I, I->getType());
559 if (isa<Instruction>(V)) {
560 std::string Name = I->getName();
564 I->replaceAllUsesWith(V);
573 DeleteTriviallyDeadInstructions(DeadInsts);
575 if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());