1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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 // If the trip count of a loop is computable, this pass also makes the following
16 // 1. The exit condition for the loop is canonicalized to compare the
17 // induction value against the exit value. This turns loops like:
18 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
19 // 2. Any use outside of the loop of an expression derived from the indvar
20 // is changed to compute the derived value outside of the loop, eliminating
21 // the dependence on the exit value of the induction variable. If the only
22 // purpose of the loop is to compute the exit value of some derived
23 // expression, this transformation will make the loop dead.
25 //===----------------------------------------------------------------------===//
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/LoopPass.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/Analysis/TargetLibraryInfo.h"
35 #include "llvm/Analysis/TargetTransformInfo.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CFG.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
53 #define DEBUG_TYPE "indvars"
55 STATISTIC(NumWidened , "Number of indvars widened");
56 STATISTIC(NumReplaced , "Number of exit values replaced");
57 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
58 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
59 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
61 // Trip count verification can be enabled by default under NDEBUG if we
62 // implement a strong expression equivalence checker in SCEV. Until then, we
63 // use the verify-indvars flag, which may assert in some cases.
64 static cl::opt<bool> VerifyIndvars(
65 "verify-indvars", cl::Hidden,
66 cl::desc("Verify the ScalarEvolution result after running indvars"));
68 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
69 cl::desc("Reduce live induction variables."));
72 class IndVarSimplify : public LoopPass {
76 TargetLibraryInfo *TLI;
77 const TargetTransformInfo *TTI;
79 SmallVector<WeakVH, 16> DeadInsts;
83 static char ID; // Pass identification, replacement for typeid
85 : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
86 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
89 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
91 void getAnalysisUsage(AnalysisUsage &AU) const override {
92 AU.addRequired<DominatorTreeWrapperPass>();
93 AU.addRequired<LoopInfoWrapperPass>();
94 AU.addRequired<ScalarEvolution>();
95 AU.addRequiredID(LoopSimplifyID);
96 AU.addRequiredID(LCSSAID);
97 AU.addPreserved<ScalarEvolution>();
98 AU.addPreservedID(LoopSimplifyID);
99 AU.addPreservedID(LCSSAID);
100 AU.setPreservesCFG();
104 void releaseMemory() override {
108 bool isValidRewrite(Value *FromVal, Value *ToVal);
110 void HandleFloatingPointIV(Loop *L, PHINode *PH);
111 void RewriteNonIntegerIVs(Loop *L);
113 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
115 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
117 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
118 PHINode *IndVar, SCEVExpander &Rewriter);
120 void SinkUnusedInvariants(Loop *L);
124 char IndVarSimplify::ID = 0;
125 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
126 "Induction Variable Simplification", false, false)
127 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
128 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
129 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
130 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
131 INITIALIZE_PASS_DEPENDENCY(LCSSA)
132 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
133 "Induction Variable Simplification", false, false)
135 Pass *llvm::createIndVarSimplifyPass() {
136 return new IndVarSimplify();
139 /// isValidRewrite - Return true if the SCEV expansion generated by the
140 /// rewriter can replace the original value. SCEV guarantees that it
141 /// produces the same value, but the way it is produced may be illegal IR.
142 /// Ideally, this function will only be called for verification.
143 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
144 // If an SCEV expression subsumed multiple pointers, its expansion could
145 // reassociate the GEP changing the base pointer. This is illegal because the
146 // final address produced by a GEP chain must be inbounds relative to its
147 // underlying object. Otherwise basic alias analysis, among other things,
148 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
149 // producing an expression involving multiple pointers. Until then, we must
152 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
153 // because it understands lcssa phis while SCEV does not.
154 Value *FromPtr = FromVal;
155 Value *ToPtr = ToVal;
156 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
157 FromPtr = GEP->getPointerOperand();
159 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
160 ToPtr = GEP->getPointerOperand();
162 if (FromPtr != FromVal || ToPtr != ToVal) {
163 // Quickly check the common case
164 if (FromPtr == ToPtr)
167 // SCEV may have rewritten an expression that produces the GEP's pointer
168 // operand. That's ok as long as the pointer operand has the same base
169 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
170 // base of a recurrence. This handles the case in which SCEV expansion
171 // converts a pointer type recurrence into a nonrecurrent pointer base
172 // indexed by an integer recurrence.
174 // If the GEP base pointer is a vector of pointers, abort.
175 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
178 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
179 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
180 if (FromBase == ToBase)
183 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
184 << *FromBase << " != " << *ToBase << "\n");
191 /// Determine the insertion point for this user. By default, insert immediately
192 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
193 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
194 /// common dominator for the incoming blocks.
195 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
197 PHINode *PHI = dyn_cast<PHINode>(User);
201 Instruction *InsertPt = nullptr;
202 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
203 if (PHI->getIncomingValue(i) != Def)
206 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
208 InsertPt = InsertBB->getTerminator();
211 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
212 InsertPt = InsertBB->getTerminator();
214 assert(InsertPt && "Missing phi operand");
215 assert((!isa<Instruction>(Def) ||
216 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
217 "def does not dominate all uses");
221 //===----------------------------------------------------------------------===//
222 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
223 //===----------------------------------------------------------------------===//
225 /// ConvertToSInt - Convert APF to an integer, if possible.
226 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
227 bool isExact = false;
228 // See if we can convert this to an int64_t
230 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
231 &isExact) != APFloat::opOK || !isExact)
237 /// HandleFloatingPointIV - If the loop has floating induction variable
238 /// then insert corresponding integer induction variable if possible.
240 /// for(double i = 0; i < 10000; ++i)
242 /// is converted into
243 /// for(int i = 0; i < 10000; ++i)
246 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
247 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
248 unsigned BackEdge = IncomingEdge^1;
250 // Check incoming value.
251 ConstantFP *InitValueVal =
252 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
255 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
258 // Check IV increment. Reject this PN if increment operation is not
259 // an add or increment value can not be represented by an integer.
260 BinaryOperator *Incr =
261 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
262 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
264 // If this is not an add of the PHI with a constantfp, or if the constant fp
265 // is not an integer, bail out.
266 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
268 if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
269 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
272 // Check Incr uses. One user is PN and the other user is an exit condition
273 // used by the conditional terminator.
274 Value::user_iterator IncrUse = Incr->user_begin();
275 Instruction *U1 = cast<Instruction>(*IncrUse++);
276 if (IncrUse == Incr->user_end()) return;
277 Instruction *U2 = cast<Instruction>(*IncrUse++);
278 if (IncrUse != Incr->user_end()) return;
280 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
281 // only used by a branch, we can't transform it.
282 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
284 Compare = dyn_cast<FCmpInst>(U2);
285 if (!Compare || !Compare->hasOneUse() ||
286 !isa<BranchInst>(Compare->user_back()))
289 BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
291 // We need to verify that the branch actually controls the iteration count
292 // of the loop. If not, the new IV can overflow and no one will notice.
293 // The branch block must be in the loop and one of the successors must be out
295 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
296 if (!L->contains(TheBr->getParent()) ||
297 (L->contains(TheBr->getSuccessor(0)) &&
298 L->contains(TheBr->getSuccessor(1))))
302 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
304 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
306 if (ExitValueVal == nullptr ||
307 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
310 // Find new predicate for integer comparison.
311 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
312 switch (Compare->getPredicate()) {
313 default: return; // Unknown comparison.
314 case CmpInst::FCMP_OEQ:
315 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
316 case CmpInst::FCMP_ONE:
317 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
318 case CmpInst::FCMP_OGT:
319 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
320 case CmpInst::FCMP_OGE:
321 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
322 case CmpInst::FCMP_OLT:
323 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
324 case CmpInst::FCMP_OLE:
325 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
328 // We convert the floating point induction variable to a signed i32 value if
329 // we can. This is only safe if the comparison will not overflow in a way
330 // that won't be trapped by the integer equivalent operations. Check for this
332 // TODO: We could use i64 if it is native and the range requires it.
334 // The start/stride/exit values must all fit in signed i32.
335 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
338 // If not actually striding (add x, 0.0), avoid touching the code.
342 // Positive and negative strides have different safety conditions.
344 // If we have a positive stride, we require the init to be less than the
346 if (InitValue >= ExitValue)
349 uint32_t Range = uint32_t(ExitValue-InitValue);
350 // Check for infinite loop, either:
351 // while (i <= Exit) or until (i > Exit)
352 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
353 if (++Range == 0) return; // Range overflows.
356 unsigned Leftover = Range % uint32_t(IncValue);
358 // If this is an equality comparison, we require that the strided value
359 // exactly land on the exit value, otherwise the IV condition will wrap
360 // around and do things the fp IV wouldn't.
361 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
365 // If the stride would wrap around the i32 before exiting, we can't
367 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
371 // If we have a negative stride, we require the init to be greater than the
373 if (InitValue <= ExitValue)
376 uint32_t Range = uint32_t(InitValue-ExitValue);
377 // Check for infinite loop, either:
378 // while (i >= Exit) or until (i < Exit)
379 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
380 if (++Range == 0) return; // Range overflows.
383 unsigned Leftover = Range % uint32_t(-IncValue);
385 // If this is an equality comparison, we require that the strided value
386 // exactly land on the exit value, otherwise the IV condition will wrap
387 // around and do things the fp IV wouldn't.
388 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
392 // If the stride would wrap around the i32 before exiting, we can't
394 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
398 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
400 // Insert new integer induction variable.
401 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
402 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
403 PN->getIncomingBlock(IncomingEdge));
406 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
407 Incr->getName()+".int", Incr);
408 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
410 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
411 ConstantInt::get(Int32Ty, ExitValue),
414 // In the following deletions, PN may become dead and may be deleted.
415 // Use a WeakVH to observe whether this happens.
418 // Delete the old floating point exit comparison. The branch starts using the
420 NewCompare->takeName(Compare);
421 Compare->replaceAllUsesWith(NewCompare);
422 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
424 // Delete the old floating point increment.
425 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
426 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
428 // If the FP induction variable still has uses, this is because something else
429 // in the loop uses its value. In order to canonicalize the induction
430 // variable, we chose to eliminate the IV and rewrite it in terms of an
433 // We give preference to sitofp over uitofp because it is faster on most
436 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
437 PN->getParent()->getFirstInsertionPt());
438 PN->replaceAllUsesWith(Conv);
439 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
444 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
445 // First step. Check to see if there are any floating-point recurrences.
446 // If there are, change them into integer recurrences, permitting analysis by
447 // the SCEV routines.
449 BasicBlock *Header = L->getHeader();
451 SmallVector<WeakVH, 8> PHIs;
452 for (BasicBlock::iterator I = Header->begin();
453 PHINode *PN = dyn_cast<PHINode>(I); ++I)
456 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
457 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
458 HandleFloatingPointIV(L, PN);
460 // If the loop previously had floating-point IV, ScalarEvolution
461 // may not have been able to compute a trip count. Now that we've done some
462 // re-writing, the trip count may be computable.
467 //===----------------------------------------------------------------------===//
468 // RewriteLoopExitValues - Optimize IV users outside the loop.
469 // As a side effect, reduces the amount of IV processing within the loop.
470 //===----------------------------------------------------------------------===//
472 /// RewriteLoopExitValues - Check to see if this loop has a computable
473 /// loop-invariant execution count. If so, this means that we can compute the
474 /// final value of any expressions that are recurrent in the loop, and
475 /// substitute the exit values from the loop into any instructions outside of
476 /// the loop that use the final values of the current expressions.
478 /// This is mostly redundant with the regular IndVarSimplify activities that
479 /// happen later, except that it's more powerful in some cases, because it's
480 /// able to brute-force evaluate arbitrary instructions as long as they have
481 /// constant operands at the beginning of the loop.
482 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
483 // Verify the input to the pass in already in LCSSA form.
484 assert(L->isLCSSAForm(*DT));
486 SmallVector<BasicBlock*, 8> ExitBlocks;
487 L->getUniqueExitBlocks(ExitBlocks);
489 // Find all values that are computed inside the loop, but used outside of it.
490 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
491 // the exit blocks of the loop to find them.
492 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
493 BasicBlock *ExitBB = ExitBlocks[i];
495 // If there are no PHI nodes in this exit block, then no values defined
496 // inside the loop are used on this path, skip it.
497 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
500 unsigned NumPreds = PN->getNumIncomingValues();
502 // We would like to be able to RAUW single-incoming value PHI nodes. We
503 // have to be certain this is safe even when this is an LCSSA PHI node.
504 // While the computed exit value is no longer varying in *this* loop, the
505 // exit block may be an exit block for an outer containing loop as well,
506 // the exit value may be varying in the outer loop, and thus it may still
507 // require an LCSSA PHI node. The safe case is when this is
508 // single-predecessor PHI node (LCSSA) and the exit block containing it is
509 // part of the enclosing loop, or this is the outer most loop of the nest.
510 // In either case the exit value could (at most) be varying in the same
511 // loop body as the phi node itself. Thus if it is in turn used outside of
512 // an enclosing loop it will only be via a separate LCSSA node.
513 bool LCSSASafePhiForRAUW =
515 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
517 // Iterate over all of the PHI nodes.
518 BasicBlock::iterator BBI = ExitBB->begin();
519 while ((PN = dyn_cast<PHINode>(BBI++))) {
521 continue; // dead use, don't replace it
523 // SCEV only supports integer expressions for now.
524 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
527 // It's necessary to tell ScalarEvolution about this explicitly so that
528 // it can walk the def-use list and forget all SCEVs, as it may not be
529 // watching the PHI itself. Once the new exit value is in place, there
530 // may not be a def-use connection between the loop and every instruction
531 // which got a SCEVAddRecExpr for that loop.
534 // Iterate over all of the values in all the PHI nodes.
535 for (unsigned i = 0; i != NumPreds; ++i) {
536 // If the value being merged in is not integer or is not defined
537 // in the loop, skip it.
538 Value *InVal = PN->getIncomingValue(i);
539 if (!isa<Instruction>(InVal))
542 // If this pred is for a subloop, not L itself, skip it.
543 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
544 continue; // The Block is in a subloop, skip it.
546 // Check that InVal is defined in the loop.
547 Instruction *Inst = cast<Instruction>(InVal);
548 if (!L->contains(Inst))
551 // Okay, this instruction has a user outside of the current loop
552 // and varies predictably *inside* the loop. Evaluate the value it
553 // contains when the loop exits, if possible.
554 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
555 if (!SE->isLoopInvariant(ExitValue, L) ||
556 !isSafeToExpand(ExitValue, *SE))
559 // Computing the value outside of the loop brings no benefit if :
560 // - it is definitely used inside the loop in a way which can not be
562 // - no use outside of the loop can take advantage of hoisting the
563 // computation out of the loop
564 if (ExitValue->getSCEVType()>=scMulExpr) {
565 unsigned NumHardInternalUses = 0;
566 unsigned NumSoftExternalUses = 0;
567 unsigned NumUses = 0;
568 for (auto IB = Inst->user_begin(), IE = Inst->user_end();
569 IB != IE && NumUses <= 6; ++IB) {
570 Instruction *UseInstr = cast<Instruction>(*IB);
571 unsigned Opc = UseInstr->getOpcode();
573 if (L->contains(UseInstr)) {
574 if (Opc == Instruction::Call || Opc == Instruction::Ret)
575 NumHardInternalUses++;
577 if (Opc == Instruction::PHI) {
578 // Do not count the Phi as a use. LCSSA may have inserted
579 // plenty of trivial ones.
581 for (auto PB = UseInstr->user_begin(),
582 PE = UseInstr->user_end();
583 PB != PE && NumUses <= 6; ++PB, ++NumUses) {
584 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
585 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
586 NumSoftExternalUses++;
590 if (Opc != Instruction::Call && Opc != Instruction::Ret)
591 NumSoftExternalUses++;
594 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
598 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
600 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
601 << " LoopVal = " << *Inst << "\n");
603 if (!isValidRewrite(Inst, ExitVal)) {
604 DeadInsts.push_back(ExitVal);
610 PN->setIncomingValue(i, ExitVal);
612 // If this instruction is dead now, delete it. Don't do it now to avoid
613 // invalidating iterators.
614 if (isInstructionTriviallyDead(Inst, TLI))
615 DeadInsts.push_back(Inst);
617 // If we determined that this PHI is safe to replace even if an LCSSA
619 if (LCSSASafePhiForRAUW) {
620 PN->replaceAllUsesWith(ExitVal);
621 PN->eraseFromParent();
627 // The insertion point instruction may have been deleted; clear it out
628 // so that the rewriter doesn't trip over it later.
629 Rewriter.clearInsertPoint();
632 //===----------------------------------------------------------------------===//
633 // IV Widening - Extend the width of an IV to cover its widest uses.
634 //===----------------------------------------------------------------------===//
637 // Collect information about induction variables that are used by sign/zero
638 // extend operations. This information is recorded by CollectExtend and
639 // provides the input to WidenIV.
642 Type *WidestNativeType; // Widest integer type created [sz]ext
643 bool IsSigned; // Was a sext user seen before a zext?
645 WideIVInfo() : NarrowIV(nullptr), WidestNativeType(nullptr),
650 /// visitCast - Update information about the induction variable that is
651 /// extended by this sign or zero extend operation. This is used to determine
652 /// the final width of the IV before actually widening it.
653 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
654 const TargetTransformInfo *TTI) {
655 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
656 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
659 Type *Ty = Cast->getType();
660 uint64_t Width = SE->getTypeSizeInBits(Ty);
661 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
664 // Cast is either an sext or zext up to this point.
665 // We should not widen an indvar if arithmetics on the wider indvar are more
666 // expensive than those on the narrower indvar. We check only the cost of ADD
667 // because at least an ADD is required to increment the induction variable. We
668 // could compute more comprehensively the cost of all instructions on the
669 // induction variable when necessary.
671 TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
672 TTI->getArithmeticInstrCost(Instruction::Add,
673 Cast->getOperand(0)->getType())) {
677 if (!WI.WidestNativeType) {
678 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
679 WI.IsSigned = IsSigned;
683 // We extend the IV to satisfy the sign of its first user, arbitrarily.
684 if (WI.IsSigned != IsSigned)
687 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
688 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
693 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
694 /// WideIV that computes the same value as the Narrow IV def. This avoids
695 /// caching Use* pointers.
696 struct NarrowIVDefUse {
697 Instruction *NarrowDef;
698 Instruction *NarrowUse;
699 Instruction *WideDef;
701 NarrowIVDefUse(): NarrowDef(nullptr), NarrowUse(nullptr), WideDef(nullptr) {}
703 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
704 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
707 /// WidenIV - The goal of this transform is to remove sign and zero extends
708 /// without creating any new induction variables. To do this, it creates a new
709 /// phi of the wider type and redirects all users, either removing extends or
710 /// inserting truncs whenever we stop propagating the type.
726 Instruction *WideInc;
727 const SCEV *WideIncExpr;
728 SmallVectorImpl<WeakVH> &DeadInsts;
730 SmallPtrSet<Instruction*,16> Widened;
731 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
734 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
735 ScalarEvolution *SEv, DominatorTree *DTree,
736 SmallVectorImpl<WeakVH> &DI) :
737 OrigPhi(WI.NarrowIV),
738 WideType(WI.WidestNativeType),
739 IsSigned(WI.IsSigned),
741 L(LI->getLoopFor(OrigPhi->getParent())),
746 WideIncExpr(nullptr),
748 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
751 PHINode *CreateWideIV(SCEVExpander &Rewriter);
754 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
757 Instruction *CloneIVUser(NarrowIVDefUse DU);
759 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
761 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
763 const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
764 unsigned OpCode) const;
766 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
768 bool WidenLoopCompare(NarrowIVDefUse DU);
770 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
772 } // anonymous namespace
774 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
775 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
776 /// gratuitous for this purpose.
777 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
778 Instruction *Inst = dyn_cast<Instruction>(V);
782 return DT->properlyDominates(Inst->getParent(), L->getHeader());
785 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
787 // Set the debug location and conservative insertion point.
788 IRBuilder<> Builder(Use);
789 // Hoist the insertion point into loop preheaders as far as possible.
790 for (const Loop *L = LI->getLoopFor(Use->getParent());
791 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
792 L = L->getParentLoop())
793 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
795 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
796 Builder.CreateZExt(NarrowOper, WideType);
799 /// CloneIVUser - Instantiate a wide operation to replace a narrow
800 /// operation. This only needs to handle operations that can evaluation to
801 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
802 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
803 unsigned Opcode = DU.NarrowUse->getOpcode();
807 case Instruction::Add:
808 case Instruction::Mul:
809 case Instruction::UDiv:
810 case Instruction::Sub:
811 case Instruction::And:
812 case Instruction::Or:
813 case Instruction::Xor:
814 case Instruction::Shl:
815 case Instruction::LShr:
816 case Instruction::AShr:
817 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
819 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
820 // anything about the narrow operand yet so must insert a [sz]ext. It is
821 // probably loop invariant and will be folded or hoisted. If it actually
822 // comes from a widened IV, it should be removed during a future call to
824 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
825 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
826 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
827 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
829 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
830 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
832 NarrowBO->getName());
833 IRBuilder<> Builder(DU.NarrowUse);
834 Builder.Insert(WideBO);
835 if (const OverflowingBinaryOperator *OBO =
836 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
837 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
838 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
844 const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
845 unsigned OpCode) const {
846 if (OpCode == Instruction::Add)
847 return SE->getAddExpr(LHS, RHS);
848 if (OpCode == Instruction::Sub)
849 return SE->getMinusSCEV(LHS, RHS);
850 if (OpCode == Instruction::Mul)
851 return SE->getMulExpr(LHS, RHS);
853 llvm_unreachable("Unsupported opcode.");
856 /// No-wrap operations can transfer sign extension of their result to their
857 /// operands. Generate the SCEV value for the widened operation without
858 /// actually modifying the IR yet. If the expression after extending the
859 /// operands is an AddRec for this loop, return it.
860 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
862 // Handle the common case of add<nsw/nuw>
863 const unsigned OpCode = DU.NarrowUse->getOpcode();
864 // Only Add/Sub/Mul instructions supported yet.
865 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
866 OpCode != Instruction::Mul)
869 // One operand (NarrowDef) has already been extended to WideDef. Now determine
870 // if extending the other will lead to a recurrence.
871 const unsigned ExtendOperIdx =
872 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
873 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
875 const SCEV *ExtendOperExpr = nullptr;
876 const OverflowingBinaryOperator *OBO =
877 cast<OverflowingBinaryOperator>(DU.NarrowUse);
878 if (IsSigned && OBO->hasNoSignedWrap())
879 ExtendOperExpr = SE->getSignExtendExpr(
880 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
881 else if(!IsSigned && OBO->hasNoUnsignedWrap())
882 ExtendOperExpr = SE->getZeroExtendExpr(
883 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
887 // When creating this SCEV expr, don't apply the current operations NSW or NUW
888 // flags. This instruction may be guarded by control flow that the no-wrap
889 // behavior depends on. Non-control-equivalent instructions can be mapped to
890 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
891 // semantics to those operations.
892 const SCEV *lhs = SE->getSCEV(DU.WideDef);
893 const SCEV *rhs = ExtendOperExpr;
895 // Let's swap operands to the initial order for the case of non-commutative
896 // operations, like SUB. See PR21014.
897 if (ExtendOperIdx == 0)
899 const SCEVAddRecExpr *AddRec =
900 dyn_cast<SCEVAddRecExpr>(GetSCEVByOpCode(lhs, rhs, OpCode));
902 if (!AddRec || AddRec->getLoop() != L)
907 /// GetWideRecurrence - Is this instruction potentially interesting for further
908 /// simplification after widening it's type? In other words, can the
909 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
910 /// recurrence on the same loop. If so, return the sign or zero extended
911 /// recurrence. Otherwise return NULL.
912 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
913 if (!SE->isSCEVable(NarrowUse->getType()))
916 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
917 if (SE->getTypeSizeInBits(NarrowExpr->getType())
918 >= SE->getTypeSizeInBits(WideType)) {
919 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
920 // index. So don't follow this use.
924 const SCEV *WideExpr = IsSigned ?
925 SE->getSignExtendExpr(NarrowExpr, WideType) :
926 SE->getZeroExtendExpr(NarrowExpr, WideType);
927 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
928 if (!AddRec || AddRec->getLoop() != L)
933 /// This IV user cannot be widen. Replace this use of the original narrow IV
934 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
935 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
936 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
937 << " for user " << *DU.NarrowUse << "\n");
938 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
939 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
940 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
943 /// If the narrow use is a compare instruction, then widen the compare
944 // (and possibly the other operand). The extend operation is hoisted into the
945 // loop preheader as far as possible.
946 bool WidenIV::WidenLoopCompare(NarrowIVDefUse DU) {
947 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
951 // Sign of IV user and compare must match.
952 if (IsSigned != CmpInst::isSigned(Cmp->getPredicate()))
955 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
956 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
957 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
958 assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
960 // Widen the compare instruction.
961 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
962 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
964 // Widen the other operand of the compare, if necessary.
965 if (CastWidth < IVWidth) {
966 Value *ExtOp = getExtend(Op, WideType, IsSigned, Cmp);
967 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
972 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
973 /// widened. If so, return the wide clone of the user.
974 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
976 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
977 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
978 if (LI->getLoopFor(UsePhi->getParent()) != L) {
979 // For LCSSA phis, sink the truncate outside the loop.
980 // After SimplifyCFG most loop exit targets have a single predecessor.
981 // Otherwise fall back to a truncate within the loop.
982 if (UsePhi->getNumOperands() != 1)
983 truncateIVUse(DU, DT);
986 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
988 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
989 IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
990 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
991 UsePhi->replaceAllUsesWith(Trunc);
992 DeadInsts.push_back(UsePhi);
993 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
994 << " to " << *WidePhi << "\n");
999 // Our raison d'etre! Eliminate sign and zero extension.
1000 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1001 Value *NewDef = DU.WideDef;
1002 if (DU.NarrowUse->getType() != WideType) {
1003 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1004 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1005 if (CastWidth < IVWidth) {
1006 // The cast isn't as wide as the IV, so insert a Trunc.
1007 IRBuilder<> Builder(DU.NarrowUse);
1008 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1011 // A wider extend was hidden behind a narrower one. This may induce
1012 // another round of IV widening in which the intermediate IV becomes
1013 // dead. It should be very rare.
1014 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1015 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1016 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1017 NewDef = DU.NarrowUse;
1020 if (NewDef != DU.NarrowUse) {
1021 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1022 << " replaced by " << *DU.WideDef << "\n");
1024 DU.NarrowUse->replaceAllUsesWith(NewDef);
1025 DeadInsts.push_back(DU.NarrowUse);
1027 // Now that the extend is gone, we want to expose it's uses for potential
1028 // further simplification. We don't need to directly inform SimplifyIVUsers
1029 // of the new users, because their parent IV will be processed later as a
1030 // new loop phi. If we preserved IVUsers analysis, we would also want to
1031 // push the uses of WideDef here.
1033 // No further widening is needed. The deceased [sz]ext had done it for us.
1037 // Does this user itself evaluate to a recurrence after widening?
1038 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1040 WideAddRec = GetExtendedOperandRecurrence(DU);
1043 // If use is a loop condition, try to promote the condition instead of
1044 // truncating the IV first.
1045 if (WidenLoopCompare(DU))
1048 // This user does not evaluate to a recurence after widening, so don't
1049 // follow it. Instead insert a Trunc to kill off the original use,
1050 // eventually isolating the original narrow IV so it can be removed.
1051 truncateIVUse(DU, DT);
1054 // Assume block terminators cannot evaluate to a recurrence. We can't to
1055 // insert a Trunc after a terminator if there happens to be a critical edge.
1056 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1057 "SCEV is not expected to evaluate a block terminator");
1059 // Reuse the IV increment that SCEVExpander created as long as it dominates
1061 Instruction *WideUse = nullptr;
1062 if (WideAddRec == WideIncExpr
1063 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1066 WideUse = CloneIVUser(DU);
1070 // Evaluation of WideAddRec ensured that the narrow expression could be
1071 // extended outside the loop without overflow. This suggests that the wide use
1072 // evaluates to the same expression as the extended narrow use, but doesn't
1073 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1074 // where it fails, we simply throw away the newly created wide use.
1075 if (WideAddRec != SE->getSCEV(WideUse)) {
1076 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1077 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1078 DeadInsts.push_back(WideUse);
1082 // Returning WideUse pushes it on the worklist.
1086 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1088 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1089 for (User *U : NarrowDef->users()) {
1090 Instruction *NarrowUser = cast<Instruction>(U);
1092 // Handle data flow merges and bizarre phi cycles.
1093 if (!Widened.insert(NarrowUser).second)
1096 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef));
1100 /// CreateWideIV - Process a single induction variable. First use the
1101 /// SCEVExpander to create a wide induction variable that evaluates to the same
1102 /// recurrence as the original narrow IV. Then use a worklist to forward
1103 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1104 /// interesting IV users, the narrow IV will be isolated for removal by
1107 /// It would be simpler to delete uses as they are processed, but we must avoid
1108 /// invalidating SCEV expressions.
1110 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1111 // Is this phi an induction variable?
1112 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1116 // Widen the induction variable expression.
1117 const SCEV *WideIVExpr = IsSigned ?
1118 SE->getSignExtendExpr(AddRec, WideType) :
1119 SE->getZeroExtendExpr(AddRec, WideType);
1121 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1122 "Expect the new IV expression to preserve its type");
1124 // Can the IV be extended outside the loop without overflow?
1125 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1126 if (!AddRec || AddRec->getLoop() != L)
1129 // An AddRec must have loop-invariant operands. Since this AddRec is
1130 // materialized by a loop header phi, the expression cannot have any post-loop
1131 // operands, so they must dominate the loop header.
1132 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1133 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1134 && "Loop header phi recurrence inputs do not dominate the loop");
1136 // The rewriter provides a value for the desired IV expression. This may
1137 // either find an existing phi or materialize a new one. Either way, we
1138 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1139 // of the phi-SCC dominates the loop entry.
1140 Instruction *InsertPt = L->getHeader()->begin();
1141 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1143 // Remembering the WideIV increment generated by SCEVExpander allows
1144 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1145 // employ a general reuse mechanism because the call above is the only call to
1146 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1147 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1149 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1150 WideIncExpr = SE->getSCEV(WideInc);
1153 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1156 // Traverse the def-use chain using a worklist starting at the original IV.
1157 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1159 Widened.insert(OrigPhi);
1160 pushNarrowIVUsers(OrigPhi, WidePhi);
1162 while (!NarrowIVUsers.empty()) {
1163 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1165 // Process a def-use edge. This may replace the use, so don't hold a
1166 // use_iterator across it.
1167 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1169 // Follow all def-use edges from the previous narrow use.
1171 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1173 // WidenIVUse may have removed the def-use edge.
1174 if (DU.NarrowDef->use_empty())
1175 DeadInsts.push_back(DU.NarrowDef);
1180 //===----------------------------------------------------------------------===//
1181 // Live IV Reduction - Minimize IVs live across the loop.
1182 //===----------------------------------------------------------------------===//
1185 //===----------------------------------------------------------------------===//
1186 // Simplification of IV users based on SCEV evaluation.
1187 //===----------------------------------------------------------------------===//
1190 class IndVarSimplifyVisitor : public IVVisitor {
1191 ScalarEvolution *SE;
1192 const TargetTransformInfo *TTI;
1198 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1199 const TargetTransformInfo *TTI,
1200 const DominatorTree *DTree)
1201 : SE(SCEV), TTI(TTI), IVPhi(IV) {
1203 WI.NarrowIV = IVPhi;
1205 setSplitOverflowIntrinsics();
1208 // Implement the interface used by simplifyUsersOfIV.
1209 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1213 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1214 /// users. Each successive simplification may push more users which may
1215 /// themselves be candidates for simplification.
1217 /// Sign/Zero extend elimination is interleaved with IV simplification.
1219 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1220 SCEVExpander &Rewriter,
1221 LPPassManager &LPM) {
1222 SmallVector<WideIVInfo, 8> WideIVs;
1224 SmallVector<PHINode*, 8> LoopPhis;
1225 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1226 LoopPhis.push_back(cast<PHINode>(I));
1228 // Each round of simplification iterates through the SimplifyIVUsers worklist
1229 // for all current phis, then determines whether any IVs can be
1230 // widened. Widening adds new phis to LoopPhis, inducing another round of
1231 // simplification on the wide IVs.
1232 while (!LoopPhis.empty()) {
1233 // Evaluate as many IV expressions as possible before widening any IVs. This
1234 // forces SCEV to set no-wrap flags before evaluating sign/zero
1235 // extension. The first time SCEV attempts to normalize sign/zero extension,
1236 // the result becomes final. So for the most predictable results, we delay
1237 // evaluation of sign/zero extend evaluation until needed, and avoid running
1238 // other SCEV based analysis prior to SimplifyAndExtend.
1240 PHINode *CurrIV = LoopPhis.pop_back_val();
1242 // Information about sign/zero extensions of CurrIV.
1243 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1245 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor);
1247 if (Visitor.WI.WidestNativeType) {
1248 WideIVs.push_back(Visitor.WI);
1250 } while(!LoopPhis.empty());
1252 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1253 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1254 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1256 LoopPhis.push_back(WidePhi);
1262 //===----------------------------------------------------------------------===//
1263 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1264 //===----------------------------------------------------------------------===//
1266 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1267 /// count expression can be safely and cheaply expanded into an instruction
1268 /// sequence that can be used by LinearFunctionTestReplace.
1270 /// TODO: This fails for pointer-type loop counters with greater than one byte
1271 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1272 /// we could skip this check in the case that the LFTR loop counter (chosen by
1273 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1274 /// the loop test to an inequality test by checking the target data's alignment
1275 /// of element types (given that the initial pointer value originates from or is
1276 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1277 /// However, we don't yet have a strong motivation for converting loop tests
1278 /// into inequality tests.
1279 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1280 SCEVExpander &Rewriter) {
1281 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1282 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1283 BackedgeTakenCount->isZero())
1286 if (!L->getExitingBlock())
1289 // Can't rewrite non-branch yet.
1290 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1293 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1299 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1300 /// invariant value to the phi.
1301 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1302 Instruction *IncI = dyn_cast<Instruction>(IncV);
1306 switch (IncI->getOpcode()) {
1307 case Instruction::Add:
1308 case Instruction::Sub:
1310 case Instruction::GetElementPtr:
1311 // An IV counter must preserve its type.
1312 if (IncI->getNumOperands() == 2)
1318 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1319 if (Phi && Phi->getParent() == L->getHeader()) {
1320 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1324 if (IncI->getOpcode() == Instruction::GetElementPtr)
1327 // Allow add/sub to be commuted.
1328 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1329 if (Phi && Phi->getParent() == L->getHeader()) {
1330 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1336 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1337 static ICmpInst *getLoopTest(Loop *L) {
1338 assert(L->getExitingBlock() && "expected loop exit");
1340 BasicBlock *LatchBlock = L->getLoopLatch();
1341 // Don't bother with LFTR if the loop is not properly simplified.
1345 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1346 assert(BI && "expected exit branch");
1348 return dyn_cast<ICmpInst>(BI->getCondition());
1351 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1352 /// that the current exit test is already sufficiently canonical.
1353 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1354 // Do LFTR to simplify the exit condition to an ICMP.
1355 ICmpInst *Cond = getLoopTest(L);
1359 // Do LFTR to simplify the exit ICMP to EQ/NE
1360 ICmpInst::Predicate Pred = Cond->getPredicate();
1361 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1364 // Look for a loop invariant RHS
1365 Value *LHS = Cond->getOperand(0);
1366 Value *RHS = Cond->getOperand(1);
1367 if (!isLoopInvariant(RHS, L, DT)) {
1368 if (!isLoopInvariant(LHS, L, DT))
1370 std::swap(LHS, RHS);
1372 // Look for a simple IV counter LHS
1373 PHINode *Phi = dyn_cast<PHINode>(LHS);
1375 Phi = getLoopPhiForCounter(LHS, L, DT);
1380 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1381 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1385 // Do LFTR if the exit condition's IV is *not* a simple counter.
1386 Value *IncV = Phi->getIncomingValue(Idx);
1387 return Phi != getLoopPhiForCounter(IncV, L, DT);
1390 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1391 /// down to checking that all operands are constant and listing instructions
1392 /// that may hide undef.
1393 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1395 if (isa<Constant>(V))
1396 return !isa<UndefValue>(V);
1401 // Conservatively handle non-constant non-instructions. For example, Arguments
1403 Instruction *I = dyn_cast<Instruction>(V);
1407 // Load and return values may be undef.
1408 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1411 // Optimistically handle other instructions.
1412 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1413 if (!Visited.insert(*OI).second)
1415 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1421 /// Return true if the given value is concrete. We must prove that undef can
1424 /// TODO: If we decide that this is a good approach to checking for undef, we
1425 /// may factor it into a common location.
1426 static bool hasConcreteDef(Value *V) {
1427 SmallPtrSet<Value*, 8> Visited;
1429 return hasConcreteDefImpl(V, Visited, 0);
1432 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1433 /// be rewritten) loop exit test.
1434 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1435 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1436 Value *IncV = Phi->getIncomingValue(LatchIdx);
1438 for (User *U : Phi->users())
1439 if (U != Cond && U != IncV) return false;
1441 for (User *U : IncV->users())
1442 if (U != Cond && U != Phi) return false;
1446 /// FindLoopCounter - Find an affine IV in canonical form.
1448 /// BECount may be an i8* pointer type. The pointer difference is already
1449 /// valid count without scaling the address stride, so it remains a pointer
1450 /// expression as far as SCEV is concerned.
1452 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1454 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1456 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1457 /// This is difficult in general for SCEV because of potential overflow. But we
1458 /// could at least handle constant BECounts.
1459 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1460 ScalarEvolution *SE, DominatorTree *DT) {
1461 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1464 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1466 // Loop over all of the PHI nodes, looking for a simple counter.
1467 PHINode *BestPhi = nullptr;
1468 const SCEV *BestInit = nullptr;
1469 BasicBlock *LatchBlock = L->getLoopLatch();
1470 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1472 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1473 PHINode *Phi = cast<PHINode>(I);
1474 if (!SE->isSCEVable(Phi->getType()))
1477 // Avoid comparing an integer IV against a pointer Limit.
1478 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1481 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1482 if (!AR || AR->getLoop() != L || !AR->isAffine())
1485 // AR may be a pointer type, while BECount is an integer type.
1486 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1487 // AR may not be a narrower type, or we may never exit.
1488 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1489 if (PhiWidth < BCWidth ||
1490 !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1493 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1494 if (!Step || !Step->isOne())
1497 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1498 Value *IncV = Phi->getIncomingValue(LatchIdx);
1499 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1502 // Avoid reusing a potentially undef value to compute other values that may
1503 // have originally had a concrete definition.
1504 if (!hasConcreteDef(Phi)) {
1505 // We explicitly allow unknown phis as long as they are already used by
1506 // the loop test. In this case we assume that performing LFTR could not
1507 // increase the number of undef users.
1508 if (ICmpInst *Cond = getLoopTest(L)) {
1509 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1510 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1515 const SCEV *Init = AR->getStart();
1517 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1518 // Don't force a live loop counter if another IV can be used.
1519 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1522 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1523 // also prefers integer to pointer IVs.
1524 if (BestInit->isZero() != Init->isZero()) {
1525 if (BestInit->isZero())
1528 // If two IVs both count from zero or both count from nonzero then the
1529 // narrower is likely a dead phi that has been widened. Use the wider phi
1530 // to allow the other to be eliminated.
1531 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1540 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1541 /// holds the RHS of the new loop test.
1542 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1543 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1544 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1545 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1546 const SCEV *IVInit = AR->getStart();
1548 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1549 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1550 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1551 // the existing GEPs whenever possible.
1552 if (IndVar->getType()->isPointerTy()
1553 && !IVCount->getType()->isPointerTy()) {
1555 // IVOffset will be the new GEP offset that is interpreted by GEP as a
1556 // signed value. IVCount on the other hand represents the loop trip count,
1557 // which is an unsigned value. FindLoopCounter only allows induction
1558 // variables that have a positive unit stride of one. This means we don't
1559 // have to handle the case of negative offsets (yet) and just need to zero
1561 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1562 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1564 // Expand the code for the iteration count.
1565 assert(SE->isLoopInvariant(IVOffset, L) &&
1566 "Computed iteration count is not loop invariant!");
1567 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1568 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1570 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1571 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1572 // We could handle pointer IVs other than i8*, but we need to compensate for
1573 // gep index scaling. See canExpandBackedgeTakenCount comments.
1574 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1575 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1576 && "unit stride pointer IV must be i8*");
1578 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1579 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1582 // In any other case, convert both IVInit and IVCount to integers before
1583 // comparing. This may result in SCEV expension of pointers, but in practice
1584 // SCEV will fold the pointer arithmetic away as such:
1585 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1587 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1588 // for simple memset-style loops.
1590 // IVInit integer and IVCount pointer would only occur if a canonical IV
1591 // were generated on top of case #2, which is not expected.
1593 const SCEV *IVLimit = nullptr;
1594 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1595 // For non-zero Start, compute IVCount here.
1596 if (AR->getStart()->isZero())
1599 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1600 const SCEV *IVInit = AR->getStart();
1602 // For integer IVs, truncate the IV before computing IVInit + BECount.
1603 if (SE->getTypeSizeInBits(IVInit->getType())
1604 > SE->getTypeSizeInBits(IVCount->getType()))
1605 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1607 IVLimit = SE->getAddExpr(IVInit, IVCount);
1609 // Expand the code for the iteration count.
1610 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1611 IRBuilder<> Builder(BI);
1612 assert(SE->isLoopInvariant(IVLimit, L) &&
1613 "Computed iteration count is not loop invariant!");
1614 // Ensure that we generate the same type as IndVar, or a smaller integer
1615 // type. In the presence of null pointer values, we have an integer type
1616 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1617 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1618 IndVar->getType() : IVCount->getType();
1619 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1623 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1624 /// loop to be a canonical != comparison against the incremented loop induction
1625 /// variable. This pass is able to rewrite the exit tests of any loop where the
1626 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1627 /// is actually a much broader range than just linear tests.
1628 Value *IndVarSimplify::
1629 LinearFunctionTestReplace(Loop *L,
1630 const SCEV *BackedgeTakenCount,
1632 SCEVExpander &Rewriter) {
1633 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1635 // Initialize CmpIndVar and IVCount to their preincremented values.
1636 Value *CmpIndVar = IndVar;
1637 const SCEV *IVCount = BackedgeTakenCount;
1639 // If the exiting block is the same as the backedge block, we prefer to
1640 // compare against the post-incremented value, otherwise we must compare
1641 // against the preincremented value.
1642 if (L->getExitingBlock() == L->getLoopLatch()) {
1643 // Add one to the "backedge-taken" count to get the trip count.
1644 // This addition may overflow, which is valid as long as the comparison is
1645 // truncated to BackedgeTakenCount->getType().
1646 IVCount = SE->getAddExpr(BackedgeTakenCount,
1647 SE->getConstant(BackedgeTakenCount->getType(), 1));
1648 // The BackedgeTaken expression contains the number of times that the
1649 // backedge branches to the loop header. This is one less than the
1650 // number of times the loop executes, so use the incremented indvar.
1651 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1654 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1655 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1656 && "genLoopLimit missed a cast");
1658 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1659 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1660 ICmpInst::Predicate P;
1661 if (L->contains(BI->getSuccessor(0)))
1662 P = ICmpInst::ICMP_NE;
1664 P = ICmpInst::ICMP_EQ;
1666 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1667 << " LHS:" << *CmpIndVar << '\n'
1669 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1670 << " RHS:\t" << *ExitCnt << "\n"
1671 << " IVCount:\t" << *IVCount << "\n");
1673 IRBuilder<> Builder(BI);
1675 // LFTR can ignore IV overflow and truncate to the width of
1676 // BECount. This avoids materializing the add(zext(add)) expression.
1677 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1678 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1679 if (CmpIndVarSize > ExitCntSize) {
1680 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1681 const SCEV *ARStart = AR->getStart();
1682 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1683 // For constant IVCount, avoid truncation.
1684 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1685 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1686 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1687 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1688 // above such that IVCount is now zero.
1689 if (IVCount != BackedgeTakenCount && Count == 0) {
1690 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1694 Count = Count.zext(CmpIndVarSize);
1696 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1697 NewLimit = Start - Count;
1699 NewLimit = Start + Count;
1700 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1702 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1704 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1708 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1709 Value *OrigCond = BI->getCondition();
1710 // It's tempting to use replaceAllUsesWith here to fully replace the old
1711 // comparison, but that's not immediately safe, since users of the old
1712 // comparison may not be dominated by the new comparison. Instead, just
1713 // update the branch to use the new comparison; in the common case this
1714 // will make old comparison dead.
1715 BI->setCondition(Cond);
1716 DeadInsts.push_back(OrigCond);
1723 //===----------------------------------------------------------------------===//
1724 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1725 //===----------------------------------------------------------------------===//
1727 /// If there's a single exit block, sink any loop-invariant values that
1728 /// were defined in the preheader but not used inside the loop into the
1729 /// exit block to reduce register pressure in the loop.
1730 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1731 BasicBlock *ExitBlock = L->getExitBlock();
1732 if (!ExitBlock) return;
1734 BasicBlock *Preheader = L->getLoopPreheader();
1735 if (!Preheader) return;
1737 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1738 BasicBlock::iterator I = Preheader->getTerminator();
1739 while (I != Preheader->begin()) {
1741 // New instructions were inserted at the end of the preheader.
1742 if (isa<PHINode>(I))
1745 // Don't move instructions which might have side effects, since the side
1746 // effects need to complete before instructions inside the loop. Also don't
1747 // move instructions which might read memory, since the loop may modify
1748 // memory. Note that it's okay if the instruction might have undefined
1749 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1751 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1754 // Skip debug info intrinsics.
1755 if (isa<DbgInfoIntrinsic>(I))
1758 // Skip landingpad instructions.
1759 if (isa<LandingPadInst>(I))
1762 // Don't sink alloca: we never want to sink static alloca's out of the
1763 // entry block, and correctly sinking dynamic alloca's requires
1764 // checks for stacksave/stackrestore intrinsics.
1765 // FIXME: Refactor this check somehow?
1766 if (isa<AllocaInst>(I))
1769 // Determine if there is a use in or before the loop (direct or
1771 bool UsedInLoop = false;
1772 for (Use &U : I->uses()) {
1773 Instruction *User = cast<Instruction>(U.getUser());
1774 BasicBlock *UseBB = User->getParent();
1775 if (PHINode *P = dyn_cast<PHINode>(User)) {
1777 PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1778 UseBB = P->getIncomingBlock(i);
1780 if (UseBB == Preheader || L->contains(UseBB)) {
1786 // If there is, the def must remain in the preheader.
1790 // Otherwise, sink it to the exit block.
1791 Instruction *ToMove = I;
1794 if (I != Preheader->begin()) {
1795 // Skip debug info intrinsics.
1798 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1800 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1806 ToMove->moveBefore(InsertPt);
1812 //===----------------------------------------------------------------------===//
1813 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1814 //===----------------------------------------------------------------------===//
1816 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1817 if (skipOptnoneFunction(L))
1820 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1821 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1822 // canonicalization can be a pessimization without LSR to "clean up"
1824 // - We depend on having a preheader; in particular,
1825 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1826 // and we're in trouble if we can't find the induction variable even when
1827 // we've manually inserted one.
1828 if (!L->isLoopSimplifyForm())
1831 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1832 SE = &getAnalysis<ScalarEvolution>();
1833 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1834 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1835 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1836 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
1837 TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
1838 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1843 // If there are any floating-point recurrences, attempt to
1844 // transform them to use integer recurrences.
1845 RewriteNonIntegerIVs(L);
1847 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1849 // Create a rewriter object which we'll use to transform the code with.
1850 SCEVExpander Rewriter(*SE, DL, "indvars");
1852 Rewriter.setDebugType(DEBUG_TYPE);
1855 // Eliminate redundant IV users.
1857 // Simplification works best when run before other consumers of SCEV. We
1858 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1859 // other expressions involving loop IVs have been evaluated. This helps SCEV
1860 // set no-wrap flags before normalizing sign/zero extension.
1861 Rewriter.disableCanonicalMode();
1862 SimplifyAndExtend(L, Rewriter, LPM);
1864 // Check to see if this loop has a computable loop-invariant execution count.
1865 // If so, this means that we can compute the final value of any expressions
1866 // that are recurrent in the loop, and substitute the exit values from the
1867 // loop into any instructions outside of the loop that use the final values of
1868 // the current expressions.
1870 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1871 RewriteLoopExitValues(L, Rewriter);
1873 // Eliminate redundant IV cycles.
1874 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
1876 // If we have a trip count expression, rewrite the loop's exit condition
1877 // using it. We can currently only handle loops with a single exit.
1878 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
1879 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
1881 // Check preconditions for proper SCEVExpander operation. SCEV does not
1882 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1883 // pass that uses the SCEVExpander must do it. This does not work well for
1884 // loop passes because SCEVExpander makes assumptions about all loops,
1885 // while LoopPassManager only forces the current loop to be simplified.
1887 // FIXME: SCEV expansion has no way to bail out, so the caller must
1888 // explicitly check any assumptions made by SCEV. Brittle.
1889 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
1890 if (!AR || AR->getLoop()->getLoopPreheader())
1891 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
1895 // Clear the rewriter cache, because values that are in the rewriter's cache
1896 // can be deleted in the loop below, causing the AssertingVH in the cache to
1900 // Now that we're done iterating through lists, clean up any instructions
1901 // which are now dead.
1902 while (!DeadInsts.empty())
1903 if (Instruction *Inst =
1904 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1905 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
1907 // The Rewriter may not be used from this point on.
1909 // Loop-invariant instructions in the preheader that aren't used in the
1910 // loop may be sunk below the loop to reduce register pressure.
1911 SinkUnusedInvariants(L);
1913 // Clean up dead instructions.
1914 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
1915 // Check a post-condition.
1916 assert(L->isLCSSAForm(*DT) &&
1917 "Indvars did not leave the loop in lcssa form!");
1919 // Verify that LFTR, and any other change have not interfered with SCEV's
1920 // ability to compute trip count.
1922 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1924 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
1925 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
1926 SE->getTypeSizeInBits(NewBECount->getType()))
1927 NewBECount = SE->getTruncateOrNoop(NewBECount,
1928 BackedgeTakenCount->getType());
1930 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
1931 NewBECount->getType());
1932 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");