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/Compiler.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/GetElementPtrTypeIterator.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/Statistic.h"
58 STATISTIC(NumRemoved , "Number of aux indvars removed");
59 STATISTIC(NumPointer , "Number of pointer indvars promoted");
60 STATISTIC(NumInserted, "Number of canonical indvars added");
61 STATISTIC(NumReplaced, "Number of exit values replaced");
62 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
65 class VISIBILITY_HIDDEN IndVarSimplify : public FunctionPass {
70 virtual bool runOnFunction(Function &) {
71 LI = &getAnalysis<LoopInfo>();
72 SE = &getAnalysis<ScalarEvolution>();
75 // Induction Variables live in the header nodes of loops
76 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
81 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
82 AU.addRequiredID(LoopSimplifyID);
83 AU.addRequired<ScalarEvolution>();
84 AU.addRequired<LoopInfo>();
85 AU.addPreservedID(LoopSimplifyID);
86 AU.addPreservedID(LCSSAID);
90 void runOnLoop(Loop *L);
91 void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
92 std::set<Instruction*> &DeadInsts);
93 Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
95 void RewriteLoopExitValues(Loop *L);
97 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
99 RegisterPass<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
102 FunctionPass *llvm::createIndVarSimplifyPass() {
103 return new IndVarSimplify();
106 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
107 /// specified set are trivially dead, delete them and see if this makes any of
108 /// their operands subsequently dead.
109 void IndVarSimplify::
110 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
111 while (!Insts.empty()) {
112 Instruction *I = *Insts.begin();
113 Insts.erase(Insts.begin());
114 if (isInstructionTriviallyDead(I)) {
115 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
116 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
118 SE->deleteInstructionFromRecords(I);
119 DOUT << "INDVARS: Deleting: " << *I;
120 I->eraseFromParent();
127 /// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer
128 /// recurrence. If so, change it into an integer recurrence, permitting
129 /// analysis by the SCEV routines.
130 void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN,
131 BasicBlock *Preheader,
132 std::set<Instruction*> &DeadInsts) {
133 assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
134 unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
135 unsigned BackedgeIdx = PreheaderIdx^1;
136 if (GetElementPtrInst *GEPI =
137 dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
138 if (GEPI->getOperand(0) == PN) {
139 assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
140 DOUT << "INDVARS: Eliminating pointer recurrence: " << *GEPI;
142 // Okay, we found a pointer recurrence. Transform this pointer
143 // recurrence into an integer recurrence. Compute the value that gets
144 // added to the pointer at every iteration.
145 Value *AddedVal = GEPI->getOperand(1);
147 // Insert a new integer PHI node into the top of the block.
148 PHINode *NewPhi = new PHINode(AddedVal->getType(),
149 PN->getName()+".rec", PN);
150 NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
152 // Create the new add instruction.
153 Value *NewAdd = BinaryOperator::createAdd(NewPhi, AddedVal,
154 GEPI->getName()+".rec", GEPI);
155 NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx));
157 // Update the existing GEP to use the recurrence.
158 GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx));
160 // Update the GEP to use the new recurrence we just inserted.
161 GEPI->setOperand(1, NewAdd);
163 // If the incoming value is a constant expr GEP, try peeling out the array
164 // 0 index if possible to make things simpler.
165 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEPI->getOperand(0)))
166 if (CE->getOpcode() == Instruction::GetElementPtr) {
167 unsigned NumOps = CE->getNumOperands();
168 assert(NumOps > 1 && "CE folding didn't work!");
169 if (CE->getOperand(NumOps-1)->isNullValue()) {
170 // Check to make sure the last index really is an array index.
171 gep_type_iterator GTI = gep_type_begin(CE);
172 for (unsigned i = 1, e = CE->getNumOperands()-1;
175 if (isa<SequentialType>(*GTI)) {
176 // Pull the last index out of the constant expr GEP.
177 SmallVector<Value*, 8> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
178 Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
181 GetElementPtrInst *NGEPI = new GetElementPtrInst(
182 NCE, Constant::getNullValue(Type::Int32Ty), NewAdd,
183 GEPI->getName(), GEPI);
184 GEPI->replaceAllUsesWith(NGEPI);
185 GEPI->eraseFromParent();
192 // Finally, if there are any other users of the PHI node, we must
193 // insert a new GEP instruction that uses the pre-incremented version
194 // of the induction amount.
195 if (!PN->use_empty()) {
196 BasicBlock::iterator InsertPos = PN; ++InsertPos;
197 while (isa<PHINode>(InsertPos)) ++InsertPos;
199 new GetElementPtrInst(PN->getIncomingValue(PreheaderIdx),
200 NewPhi, "", InsertPos);
201 PreInc->takeName(PN);
202 PN->replaceAllUsesWith(PreInc);
205 // Delete the old PHI for sure, and the GEP if its otherwise unused.
206 DeadInsts.insert(PN);
213 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
214 /// loop to be a canonical != comparison against the incremented loop induction
215 /// variable. This pass is able to rewrite the exit tests of any loop where the
216 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
217 /// is actually a much broader range than just linear tests.
219 /// This method returns a "potentially dead" instruction whose computation chain
220 /// should be deleted when convenient.
221 Instruction *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
222 SCEV *IterationCount,
224 // Find the exit block for the loop. We can currently only handle loops with
226 std::vector<BasicBlock*> ExitBlocks;
227 L->getExitBlocks(ExitBlocks);
228 if (ExitBlocks.size() != 1) return 0;
229 BasicBlock *ExitBlock = ExitBlocks[0];
231 // Make sure there is only one predecessor block in the loop.
232 BasicBlock *ExitingBlock = 0;
233 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
235 if (L->contains(*PI)) {
236 if (ExitingBlock == 0)
239 return 0; // Multiple exits from loop to this block.
241 assert(ExitingBlock && "Loop info is broken");
243 if (!isa<BranchInst>(ExitingBlock->getTerminator()))
244 return 0; // Can't rewrite non-branch yet
245 BranchInst *BI = cast<BranchInst>(ExitingBlock->getTerminator());
246 assert(BI->isConditional() && "Must be conditional to be part of loop!");
248 Instruction *PotentiallyDeadInst = dyn_cast<Instruction>(BI->getCondition());
250 // If the exiting block is not the same as the backedge block, we must compare
251 // against the preincremented value, otherwise we prefer to compare against
252 // the post-incremented value.
253 BasicBlock *Header = L->getHeader();
254 pred_iterator HPI = pred_begin(Header);
255 assert(HPI != pred_end(Header) && "Loop with zero preds???");
256 if (!L->contains(*HPI)) ++HPI;
257 assert(HPI != pred_end(Header) && L->contains(*HPI) &&
258 "No backedge in loop?");
260 SCEVHandle TripCount = IterationCount;
262 if (*HPI == ExitingBlock) {
263 // The IterationCount expression contains the number of times that the
264 // backedge actually branches to the loop header. This is one less than the
265 // number of times the loop executes, so add one to it.
266 Constant *OneC = ConstantInt::get(IterationCount->getType(), 1);
267 TripCount = SCEVAddExpr::get(IterationCount, SCEVUnknown::get(OneC));
268 IndVar = L->getCanonicalInductionVariableIncrement();
270 // We have to use the preincremented value...
271 IndVar = L->getCanonicalInductionVariable();
274 DOUT << "INDVARS: LFTR: TripCount = " << *TripCount
275 << " IndVar = " << *IndVar << "\n";
277 // Expand the code for the iteration count into the preheader of the loop.
278 BasicBlock *Preheader = L->getLoopPreheader();
279 Value *ExitCnt = RW.expandCodeFor(TripCount, Preheader->getTerminator(),
282 // Insert a new icmp_ne or icmp_eq instruction before the branch.
283 ICmpInst::Predicate Opcode;
284 if (L->contains(BI->getSuccessor(0)))
285 Opcode = ICmpInst::ICMP_NE;
287 Opcode = ICmpInst::ICMP_EQ;
289 Value *Cond = new ICmpInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
290 BI->setCondition(Cond);
293 return PotentiallyDeadInst;
297 /// RewriteLoopExitValues - Check to see if this loop has a computable
298 /// loop-invariant execution count. If so, this means that we can compute the
299 /// final value of any expressions that are recurrent in the loop, and
300 /// substitute the exit values from the loop into any instructions outside of
301 /// the loop that use the final values of the current expressions.
302 void IndVarSimplify::RewriteLoopExitValues(Loop *L) {
303 BasicBlock *Preheader = L->getLoopPreheader();
305 // Scan all of the instructions in the loop, looking at those that have
306 // extra-loop users and which are recurrences.
307 SCEVExpander Rewriter(*SE, *LI);
309 // We insert the code into the preheader of the loop if the loop contains
310 // multiple exit blocks, or in the exit block if there is exactly one.
311 BasicBlock *BlockToInsertInto;
312 std::vector<BasicBlock*> ExitBlocks;
313 L->getExitBlocks(ExitBlocks);
314 if (ExitBlocks.size() == 1)
315 BlockToInsertInto = ExitBlocks[0];
317 BlockToInsertInto = Preheader;
318 BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
319 while (isa<PHINode>(InsertPt)) ++InsertPt;
321 bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
323 std::set<Instruction*> InstructionsToDelete;
325 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
326 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
327 BasicBlock *BB = L->getBlocks()[i];
328 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
329 if (I->getType()->isInteger()) { // Is an integer instruction
330 SCEVHandle SH = SE->getSCEV(I);
331 if (SH->hasComputableLoopEvolution(L) || // Varies predictably
332 HasConstantItCount) {
333 // Find out if this predictably varying value is actually used
334 // outside of the loop. "extra" as opposed to "intra".
335 std::vector<Instruction*> ExtraLoopUsers;
336 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
338 Instruction *User = cast<Instruction>(*UI);
339 if (!L->contains(User->getParent())) {
340 // If this is a PHI node in the exit block and we're inserting,
341 // into the exit block, it must have a single entry. In this
342 // case, we can't insert the code after the PHI and have the PHI
343 // still use it. Instead, don't insert the the PHI.
344 if (PHINode *PN = dyn_cast<PHINode>(User)) {
345 // FIXME: This is a case where LCSSA pessimizes code, this
346 // should be fixed better.
347 if (PN->getNumOperands() == 2 &&
348 PN->getParent() == BlockToInsertInto)
351 ExtraLoopUsers.push_back(User);
355 if (!ExtraLoopUsers.empty()) {
356 // Okay, this instruction has a user outside of the current loop
357 // and varies predictably in this loop. Evaluate the value it
358 // contains when the loop exits, and insert code for it.
359 SCEVHandle ExitValue = SE->getSCEVAtScope(I, L->getParentLoop());
360 if (!isa<SCEVCouldNotCompute>(ExitValue)) {
363 // Remember the next instruction. The rewriter can move code
364 // around in some cases.
365 BasicBlock::iterator NextI = I; ++NextI;
367 Value *NewVal = Rewriter.expandCodeFor(ExitValue, InsertPt,
370 DOUT << "INDVARS: RLEV: AfterLoopVal = " << *NewVal
371 << " LoopVal = " << *I << "\n";
373 // Rewrite any users of the computed value outside of the loop
374 // with the newly computed value.
375 for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i) {
376 PHINode* PN = dyn_cast<PHINode>(ExtraLoopUsers[i]);
377 if (PN && PN->getNumOperands() == 2 &&
378 !L->contains(PN->getParent())) {
379 // We're dealing with an LCSSA Phi. Handle it specially.
380 Instruction* LCSSAInsertPt = BlockToInsertInto->begin();
382 Instruction* NewInstr = dyn_cast<Instruction>(NewVal);
383 if (NewInstr && !isa<PHINode>(NewInstr) &&
384 !L->contains(NewInstr->getParent()))
385 for (unsigned j = 0; j < NewInstr->getNumOperands(); ++j){
387 dyn_cast<Instruction>(NewInstr->getOperand(j));
388 if (PredI && L->contains(PredI->getParent())) {
389 PHINode* NewLCSSA = new PHINode(PredI->getType(),
390 PredI->getName() + ".lcssa",
392 NewLCSSA->addIncoming(PredI,
393 BlockToInsertInto->getSinglePredecessor());
395 NewInstr->replaceUsesOfWith(PredI, NewLCSSA);
399 PN->replaceAllUsesWith(NewVal);
400 PN->eraseFromParent();
402 ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
406 // If this instruction is dead now, schedule it to be removed.
408 InstructionsToDelete.insert(I);
410 continue; // Skip the ++I
416 // Next instruction. Continue instruction skips this.
421 DeleteTriviallyDeadInstructions(InstructionsToDelete);
425 void IndVarSimplify::runOnLoop(Loop *L) {
426 // First step. Check to see if there are any trivial GEP pointer recurrences.
427 // If there are, change them into integer recurrences, permitting analysis by
428 // the SCEV routines.
430 BasicBlock *Header = L->getHeader();
431 BasicBlock *Preheader = L->getLoopPreheader();
433 std::set<Instruction*> DeadInsts;
434 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
435 PHINode *PN = cast<PHINode>(I);
436 if (isa<PointerType>(PN->getType()))
437 EliminatePointerRecurrence(PN, Preheader, DeadInsts);
440 if (!DeadInsts.empty())
441 DeleteTriviallyDeadInstructions(DeadInsts);
444 // Next, transform all loops nesting inside of this loop.
445 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
448 // Check to see if this loop has a computable loop-invariant execution count.
449 // If so, this means that we can compute the final value of any expressions
450 // that are recurrent in the loop, and substitute the exit values from the
451 // loop into any instructions outside of the loop that use the final values of
452 // the current expressions.
454 SCEVHandle IterationCount = SE->getIterationCount(L);
455 if (!isa<SCEVCouldNotCompute>(IterationCount))
456 RewriteLoopExitValues(L);
458 // Next, analyze all of the induction variables in the loop, canonicalizing
459 // auxillary induction variables.
460 std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
462 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
463 PHINode *PN = cast<PHINode>(I);
464 if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
465 SCEVHandle SCEV = SE->getSCEV(PN);
466 if (SCEV->hasComputableLoopEvolution(L))
467 // FIXME: It is an extremely bad idea to indvar substitute anything more
468 // complex than affine induction variables. Doing so will put expensive
469 // polynomial evaluations inside of the loop, and the str reduction pass
470 // currently can only reduce affine polynomials. For now just disable
471 // indvar subst on anything more complex than an affine addrec.
472 if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
474 IndVars.push_back(std::make_pair(PN, SCEV));
478 // If there are no induction variables in the loop, there is nothing more to
480 if (IndVars.empty()) {
481 // Actually, if we know how many times the loop iterates, lets insert a
482 // canonical induction variable to help subsequent passes.
483 if (!isa<SCEVCouldNotCompute>(IterationCount)) {
484 SCEVExpander Rewriter(*SE, *LI);
485 Rewriter.getOrInsertCanonicalInductionVariable(L,
486 IterationCount->getType());
487 if (Instruction *I = LinearFunctionTestReplace(L, IterationCount,
489 std::set<Instruction*> InstructionsToDelete;
490 InstructionsToDelete.insert(I);
491 DeleteTriviallyDeadInstructions(InstructionsToDelete);
497 // Compute the type of the largest recurrence expression.
499 const Type *LargestType = IndVars[0].first->getType();
500 bool DifferingSizes = false;
501 for (unsigned i = 1, e = IndVars.size(); i != e; ++i) {
502 const Type *Ty = IndVars[i].first->getType();
504 Ty->getPrimitiveSizeInBits() != LargestType->getPrimitiveSizeInBits();
505 if (Ty->getPrimitiveSizeInBits() > LargestType->getPrimitiveSizeInBits())
509 // Create a rewriter object which we'll use to transform the code with.
510 SCEVExpander Rewriter(*SE, *LI);
512 // Now that we know the largest of of the induction variables in this loop,
513 // insert a canonical induction variable of the largest size.
514 Value *IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
517 DOUT << "INDVARS: New CanIV: " << *IndVar;
519 if (!isa<SCEVCouldNotCompute>(IterationCount))
520 if (Instruction *DI = LinearFunctionTestReplace(L, IterationCount,Rewriter))
521 DeadInsts.insert(DI);
523 // Now that we have a canonical induction variable, we can rewrite any
524 // recurrences in terms of the induction variable. Start with the auxillary
525 // induction variables, and recursively rewrite any of their uses.
526 BasicBlock::iterator InsertPt = Header->begin();
527 while (isa<PHINode>(InsertPt)) ++InsertPt;
529 // If there were induction variables of other sizes, cast the primary
530 // induction variable to the right size for them, avoiding the need for the
531 // code evaluation methods to insert induction variables of different sizes.
532 if (DifferingSizes) {
533 SmallVector<unsigned,4> InsertedSizes;
534 InsertedSizes.push_back(LargestType->getPrimitiveSizeInBits());
535 for (unsigned i = 0, e = IndVars.size(); i != e; ++i) {
536 unsigned ithSize = IndVars[i].first->getType()->getPrimitiveSizeInBits();
537 if (std::find(InsertedSizes.begin(), InsertedSizes.end(), ithSize)
538 == InsertedSizes.end()) {
539 PHINode *PN = IndVars[i].first;
540 InsertedSizes.push_back(ithSize);
541 Instruction *New = new TruncInst(IndVar, PN->getType(), "indvar",
543 Rewriter.addInsertedValue(New, SE->getSCEV(New));
544 DOUT << "INDVARS: Made trunc IV for " << *PN
545 << " NewVal = " << *New << "\n";
550 // Rewrite all induction variables in terms of the canonical induction
552 std::map<unsigned, Value*> InsertedSizes;
553 while (!IndVars.empty()) {
554 PHINode *PN = IndVars.back().first;
555 Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt,
557 DOUT << "INDVARS: Rewrote IV '" << *IndVars.back().second << "' " << *PN
558 << " into = " << *NewVal << "\n";
559 NewVal->takeName(PN);
561 // Replace the old PHI Node with the inserted computation.
562 PN->replaceAllUsesWith(NewVal);
563 DeadInsts.insert(PN);
570 // Now replace all derived expressions in the loop body with simpler
572 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
573 if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
574 BasicBlock *BB = L->getBlocks()[i];
575 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
576 if (I->getType()->isInteger() && // Is an integer instruction
578 !Rewriter.isInsertedInstruction(I)) {
579 SCEVHandle SH = SE->getSCEV(I);
580 Value *V = Rewriter.expandCodeFor(SH, I, I->getType());
582 if (isa<Instruction>(V))
584 I->replaceAllUsesWith(V);
593 DeleteTriviallyDeadInstructions(DeadInsts);
595 if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());