1 //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===//
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 file contains the implementation of the scalar evolution expander,
11 // which is used to generate the code corresponding to a given scalar evolution
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Analysis/ScalarEvolutionExpander.h"
17 #include "llvm/Analysis/LoopInfo.h"
18 #include "llvm/IntrinsicInst.h"
19 #include "llvm/LLVMContext.h"
20 #include "llvm/Support/Debug.h"
21 #include "llvm/Target/TargetData.h"
22 #include "llvm/Target/TargetLowering.h"
23 #include "llvm/ADT/STLExtras.h"
27 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
28 /// reusing an existing cast if a suitable one exists, moving an existing
29 /// cast if a suitable one exists but isn't in the right place, or
30 /// creating a new one.
31 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
32 Instruction::CastOps Op,
33 BasicBlock::iterator IP) {
34 // This function must be called with the builder having a valid insertion
35 // point. It doesn't need to be the actual IP where the uses of the returned
36 // cast will be added, but it must dominate such IP.
37 // We use this precondition to assert that we can produce a cast that will
38 // dominate all its uses. In particular, this is crucial for the case
39 // where the builder's insertion point *is* the point where we were asked
41 // Since we don't know the the builder's insertion point is actually
42 // where the uses will be added (only that it dominates it), we are
43 // not allowed to move it.
44 BasicBlock::iterator BIP = Builder.GetInsertPoint();
46 assert(BIP == IP || SE.DT->dominates(IP, BIP));
48 // Check to see if there is already a cast!
49 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
52 if (U->getType() == Ty)
53 if (CastInst *CI = dyn_cast<CastInst>(U))
54 if (CI->getOpcode() == Op) {
55 // If the cast isn't where we want it, create a new cast at IP.
56 // Likewise, do not reuse a cast at BIP because it must dominate
57 // instructions that might be inserted before BIP.
58 if (BasicBlock::iterator(CI) != IP || BIP == IP) {
59 // Create a new cast, and leave the old cast in place in case
60 // it is being used as an insert point. Clear its operand
61 // so that it doesn't hold anything live.
62 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", IP);
64 CI->replaceAllUsesWith(NewCI);
65 CI->setOperand(0, UndefValue::get(V->getType()));
66 rememberInstruction(NewCI);
69 rememberInstruction(CI);
75 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), IP);
76 rememberInstruction(I);
80 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
81 /// which must be possible with a noop cast, doing what we can to share
83 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
84 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
85 assert((Op == Instruction::BitCast ||
86 Op == Instruction::PtrToInt ||
87 Op == Instruction::IntToPtr) &&
88 "InsertNoopCastOfTo cannot perform non-noop casts!");
89 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
90 "InsertNoopCastOfTo cannot change sizes!");
92 // Short-circuit unnecessary bitcasts.
93 if (Op == Instruction::BitCast) {
94 if (V->getType() == Ty)
96 if (CastInst *CI = dyn_cast<CastInst>(V)) {
97 if (CI->getOperand(0)->getType() == Ty)
98 return CI->getOperand(0);
101 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
102 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
103 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
104 if (CastInst *CI = dyn_cast<CastInst>(V))
105 if ((CI->getOpcode() == Instruction::PtrToInt ||
106 CI->getOpcode() == Instruction::IntToPtr) &&
107 SE.getTypeSizeInBits(CI->getType()) ==
108 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
109 return CI->getOperand(0);
110 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
111 if ((CE->getOpcode() == Instruction::PtrToInt ||
112 CE->getOpcode() == Instruction::IntToPtr) &&
113 SE.getTypeSizeInBits(CE->getType()) ==
114 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
115 return CE->getOperand(0);
118 // Fold a cast of a constant.
119 if (Constant *C = dyn_cast<Constant>(V))
120 return ConstantExpr::getCast(Op, C, Ty);
122 // Cast the argument at the beginning of the entry block, after
123 // any bitcasts of other arguments.
124 if (Argument *A = dyn_cast<Argument>(V)) {
125 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
126 while ((isa<BitCastInst>(IP) &&
127 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
128 cast<BitCastInst>(IP)->getOperand(0) != A) ||
129 isa<DbgInfoIntrinsic>(IP) ||
130 isa<LandingPadInst>(IP))
132 return ReuseOrCreateCast(A, Ty, Op, IP);
135 // Cast the instruction immediately after the instruction.
136 Instruction *I = cast<Instruction>(V);
137 BasicBlock::iterator IP = I; ++IP;
138 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
139 IP = II->getNormalDest()->begin();
140 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
142 return ReuseOrCreateCast(I, Ty, Op, IP);
145 /// InsertBinop - Insert the specified binary operator, doing a small amount
146 /// of work to avoid inserting an obviously redundant operation.
147 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
148 Value *LHS, Value *RHS) {
149 // Fold a binop with constant operands.
150 if (Constant *CLHS = dyn_cast<Constant>(LHS))
151 if (Constant *CRHS = dyn_cast<Constant>(RHS))
152 return ConstantExpr::get(Opcode, CLHS, CRHS);
154 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
155 unsigned ScanLimit = 6;
156 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
157 // Scanning starts from the last instruction before the insertion point.
158 BasicBlock::iterator IP = Builder.GetInsertPoint();
159 if (IP != BlockBegin) {
161 for (; ScanLimit; --IP, --ScanLimit) {
162 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
164 if (isa<DbgInfoIntrinsic>(IP))
166 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
167 IP->getOperand(1) == RHS)
169 if (IP == BlockBegin) break;
173 // Save the original insertion point so we can restore it when we're done.
174 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
175 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
177 // Move the insertion point out of as many loops as we can.
178 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
179 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
180 BasicBlock *Preheader = L->getLoopPreheader();
181 if (!Preheader) break;
183 // Ok, move up a level.
184 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
187 // If we haven't found this binop, insert it.
188 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
189 BO->setDebugLoc(SaveInsertPt->getDebugLoc());
190 rememberInstruction(BO);
192 // Restore the original insert point.
194 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
199 /// FactorOutConstant - Test if S is divisible by Factor, using signed
200 /// division. If so, update S with Factor divided out and return true.
201 /// S need not be evenly divisible if a reasonable remainder can be
203 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
204 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
205 /// check to see if the divide was folded.
206 static bool FactorOutConstant(const SCEV *&S,
207 const SCEV *&Remainder,
210 const TargetData *TD) {
211 // Everything is divisible by one.
217 S = SE.getConstant(S->getType(), 1);
221 // For a Constant, check for a multiple of the given factor.
222 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
226 // Check for divisibility.
227 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
229 ConstantInt::get(SE.getContext(),
230 C->getValue()->getValue().sdiv(
231 FC->getValue()->getValue()));
232 // If the quotient is zero and the remainder is non-zero, reject
233 // the value at this scale. It will be considered for subsequent
236 const SCEV *Div = SE.getConstant(CI);
239 SE.getAddExpr(Remainder,
240 SE.getConstant(C->getValue()->getValue().srem(
241 FC->getValue()->getValue())));
247 // In a Mul, check if there is a constant operand which is a multiple
248 // of the given factor.
249 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
251 // With TargetData, the size is known. Check if there is a constant
252 // operand which is a multiple of the given factor. If so, we can
254 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
255 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
256 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
257 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
259 SE.getConstant(C->getValue()->getValue().sdiv(
260 FC->getValue()->getValue()));
261 S = SE.getMulExpr(NewMulOps);
265 // Without TargetData, check if Factor can be factored out of any of the
266 // Mul's operands. If so, we can just remove it.
267 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
268 const SCEV *SOp = M->getOperand(i);
269 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
270 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
271 Remainder->isZero()) {
272 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
274 S = SE.getMulExpr(NewMulOps);
281 // In an AddRec, check if both start and step are divisible.
282 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
283 const SCEV *Step = A->getStepRecurrence(SE);
284 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
285 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
287 if (!StepRem->isZero())
289 const SCEV *Start = A->getStart();
290 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
292 // FIXME: can use A->getNoWrapFlags(FlagNW)
293 S = SE.getAddRecExpr(Start, Step, A->getLoop(), SCEV::FlagAnyWrap);
300 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
301 /// is the number of SCEVAddRecExprs present, which are kept at the end of
304 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
306 ScalarEvolution &SE) {
307 unsigned NumAddRecs = 0;
308 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
310 // Group Ops into non-addrecs and addrecs.
311 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
312 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
313 // Let ScalarEvolution sort and simplify the non-addrecs list.
314 const SCEV *Sum = NoAddRecs.empty() ?
315 SE.getConstant(Ty, 0) :
316 SE.getAddExpr(NoAddRecs);
317 // If it returned an add, use the operands. Otherwise it simplified
318 // the sum into a single value, so just use that.
320 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
321 Ops.append(Add->op_begin(), Add->op_end());
322 else if (!Sum->isZero())
324 // Then append the addrecs.
325 Ops.append(AddRecs.begin(), AddRecs.end());
328 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
329 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
330 /// This helps expose more opportunities for folding parts of the expressions
331 /// into GEP indices.
333 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
335 ScalarEvolution &SE) {
337 SmallVector<const SCEV *, 8> AddRecs;
338 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
339 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
340 const SCEV *Start = A->getStart();
341 if (Start->isZero()) break;
342 const SCEV *Zero = SE.getConstant(Ty, 0);
343 AddRecs.push_back(SE.getAddRecExpr(Zero,
344 A->getStepRecurrence(SE),
346 // FIXME: A->getNoWrapFlags(FlagNW)
348 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
350 Ops.append(Add->op_begin(), Add->op_end());
351 e += Add->getNumOperands();
356 if (!AddRecs.empty()) {
357 // Add the addrecs onto the end of the list.
358 Ops.append(AddRecs.begin(), AddRecs.end());
359 // Resort the operand list, moving any constants to the front.
360 SimplifyAddOperands(Ops, Ty, SE);
364 /// expandAddToGEP - Expand an addition expression with a pointer type into
365 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
366 /// BasicAliasAnalysis and other passes analyze the result. See the rules
367 /// for getelementptr vs. inttoptr in
368 /// http://llvm.org/docs/LangRef.html#pointeraliasing
371 /// Design note: The correctness of using getelementptr here depends on
372 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
373 /// they may introduce pointer arithmetic which may not be safely converted
374 /// into getelementptr.
376 /// Design note: It might seem desirable for this function to be more
377 /// loop-aware. If some of the indices are loop-invariant while others
378 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
379 /// loop-invariant portions of the overall computation outside the loop.
380 /// However, there are a few reasons this is not done here. Hoisting simple
381 /// arithmetic is a low-level optimization that often isn't very
382 /// important until late in the optimization process. In fact, passes
383 /// like InstructionCombining will combine GEPs, even if it means
384 /// pushing loop-invariant computation down into loops, so even if the
385 /// GEPs were split here, the work would quickly be undone. The
386 /// LoopStrengthReduction pass, which is usually run quite late (and
387 /// after the last InstructionCombining pass), takes care of hoisting
388 /// loop-invariant portions of expressions, after considering what
389 /// can be folded using target addressing modes.
391 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
392 const SCEV *const *op_end,
396 Type *ElTy = PTy->getElementType();
397 SmallVector<Value *, 4> GepIndices;
398 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
399 bool AnyNonZeroIndices = false;
401 // Split AddRecs up into parts as either of the parts may be usable
402 // without the other.
403 SplitAddRecs(Ops, Ty, SE);
405 // Descend down the pointer's type and attempt to convert the other
406 // operands into GEP indices, at each level. The first index in a GEP
407 // indexes into the array implied by the pointer operand; the rest of
408 // the indices index into the element or field type selected by the
411 // If the scale size is not 0, attempt to factor out a scale for
413 SmallVector<const SCEV *, 8> ScaledOps;
414 if (ElTy->isSized()) {
415 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
416 if (!ElSize->isZero()) {
417 SmallVector<const SCEV *, 8> NewOps;
418 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
419 const SCEV *Op = Ops[i];
420 const SCEV *Remainder = SE.getConstant(Ty, 0);
421 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
422 // Op now has ElSize factored out.
423 ScaledOps.push_back(Op);
424 if (!Remainder->isZero())
425 NewOps.push_back(Remainder);
426 AnyNonZeroIndices = true;
428 // The operand was not divisible, so add it to the list of operands
429 // we'll scan next iteration.
430 NewOps.push_back(Ops[i]);
433 // If we made any changes, update Ops.
434 if (!ScaledOps.empty()) {
436 SimplifyAddOperands(Ops, Ty, SE);
441 // Record the scaled array index for this level of the type. If
442 // we didn't find any operands that could be factored, tentatively
443 // assume that element zero was selected (since the zero offset
444 // would obviously be folded away).
445 Value *Scaled = ScaledOps.empty() ?
446 Constant::getNullValue(Ty) :
447 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
448 GepIndices.push_back(Scaled);
450 // Collect struct field index operands.
451 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
452 bool FoundFieldNo = false;
453 // An empty struct has no fields.
454 if (STy->getNumElements() == 0) break;
456 // With TargetData, field offsets are known. See if a constant offset
457 // falls within any of the struct fields.
458 if (Ops.empty()) break;
459 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
460 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
461 const StructLayout &SL = *SE.TD->getStructLayout(STy);
462 uint64_t FullOffset = C->getValue()->getZExtValue();
463 if (FullOffset < SL.getSizeInBytes()) {
464 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
465 GepIndices.push_back(
466 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
467 ElTy = STy->getTypeAtIndex(ElIdx);
469 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
470 AnyNonZeroIndices = true;
475 // Without TargetData, just check for an offsetof expression of the
476 // appropriate struct type.
477 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
478 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
481 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
482 GepIndices.push_back(FieldNo);
484 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
485 Ops[i] = SE.getConstant(Ty, 0);
486 AnyNonZeroIndices = true;
492 // If no struct field offsets were found, tentatively assume that
493 // field zero was selected (since the zero offset would obviously
496 ElTy = STy->getTypeAtIndex(0u);
497 GepIndices.push_back(
498 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
502 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
503 ElTy = ATy->getElementType();
508 // If none of the operands were convertible to proper GEP indices, cast
509 // the base to i8* and do an ugly getelementptr with that. It's still
510 // better than ptrtoint+arithmetic+inttoptr at least.
511 if (!AnyNonZeroIndices) {
512 // Cast the base to i8*.
513 V = InsertNoopCastOfTo(V,
514 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
516 assert(!isa<Instruction>(V) ||
517 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint()));
519 // Expand the operands for a plain byte offset.
520 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
522 // Fold a GEP with constant operands.
523 if (Constant *CLHS = dyn_cast<Constant>(V))
524 if (Constant *CRHS = dyn_cast<Constant>(Idx))
525 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
527 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
528 unsigned ScanLimit = 6;
529 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
530 // Scanning starts from the last instruction before the insertion point.
531 BasicBlock::iterator IP = Builder.GetInsertPoint();
532 if (IP != BlockBegin) {
534 for (; ScanLimit; --IP, --ScanLimit) {
535 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
537 if (isa<DbgInfoIntrinsic>(IP))
539 if (IP->getOpcode() == Instruction::GetElementPtr &&
540 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
542 if (IP == BlockBegin) break;
546 // Save the original insertion point so we can restore it when we're done.
547 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
548 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
550 // Move the insertion point out of as many loops as we can.
551 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
552 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
553 BasicBlock *Preheader = L->getLoopPreheader();
554 if (!Preheader) break;
556 // Ok, move up a level.
557 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
561 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
562 rememberInstruction(GEP);
564 // Restore the original insert point.
566 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
571 // Save the original insertion point so we can restore it when we're done.
572 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
573 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
575 // Move the insertion point out of as many loops as we can.
576 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
577 if (!L->isLoopInvariant(V)) break;
579 bool AnyIndexNotLoopInvariant = false;
580 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
581 E = GepIndices.end(); I != E; ++I)
582 if (!L->isLoopInvariant(*I)) {
583 AnyIndexNotLoopInvariant = true;
586 if (AnyIndexNotLoopInvariant)
589 BasicBlock *Preheader = L->getLoopPreheader();
590 if (!Preheader) break;
592 // Ok, move up a level.
593 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
596 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
597 // because ScalarEvolution may have changed the address arithmetic to
598 // compute a value which is beyond the end of the allocated object.
600 if (V->getType() != PTy)
601 Casted = InsertNoopCastOfTo(Casted, PTy);
602 Value *GEP = Builder.CreateGEP(Casted,
605 Ops.push_back(SE.getUnknown(GEP));
606 rememberInstruction(GEP);
608 // Restore the original insert point.
610 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
612 return expand(SE.getAddExpr(Ops));
615 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
616 /// SCEV expansion. If they are nested, this is the most nested. If they are
617 /// neighboring, pick the later.
618 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
622 if (A->contains(B)) return B;
623 if (B->contains(A)) return A;
624 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
625 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
626 return A; // Arbitrarily break the tie.
629 /// getRelevantLoop - Get the most relevant loop associated with the given
630 /// expression, according to PickMostRelevantLoop.
631 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
632 // Test whether we've already computed the most relevant loop for this SCEV.
633 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
634 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
636 return Pair.first->second;
638 if (isa<SCEVConstant>(S))
639 // A constant has no relevant loops.
641 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
642 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
643 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
644 // A non-instruction has no relevant loops.
647 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
649 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
651 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
653 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
654 return RelevantLoops[N] = L;
656 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
657 const Loop *Result = getRelevantLoop(C->getOperand());
658 return RelevantLoops[C] = Result;
660 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
662 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
663 getRelevantLoop(D->getRHS()),
665 return RelevantLoops[D] = Result;
667 llvm_unreachable("Unexpected SCEV type!");
672 /// LoopCompare - Compare loops by PickMostRelevantLoop.
676 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
678 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
679 std::pair<const Loop *, const SCEV *> RHS) const {
680 // Keep pointer operands sorted at the end.
681 if (LHS.second->getType()->isPointerTy() !=
682 RHS.second->getType()->isPointerTy())
683 return LHS.second->getType()->isPointerTy();
685 // Compare loops with PickMostRelevantLoop.
686 if (LHS.first != RHS.first)
687 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
689 // If one operand is a non-constant negative and the other is not,
690 // put the non-constant negative on the right so that a sub can
691 // be used instead of a negate and add.
692 if (LHS.second->isNonConstantNegative()) {
693 if (!RHS.second->isNonConstantNegative())
695 } else if (RHS.second->isNonConstantNegative())
698 // Otherwise they are equivalent according to this comparison.
705 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
706 Type *Ty = SE.getEffectiveSCEVType(S->getType());
708 // Collect all the add operands in a loop, along with their associated loops.
709 // Iterate in reverse so that constants are emitted last, all else equal, and
710 // so that pointer operands are inserted first, which the code below relies on
711 // to form more involved GEPs.
712 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
713 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
714 E(S->op_begin()); I != E; ++I)
715 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
717 // Sort by loop. Use a stable sort so that constants follow non-constants and
718 // pointer operands precede non-pointer operands.
719 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
721 // Emit instructions to add all the operands. Hoist as much as possible
722 // out of loops, and form meaningful getelementptrs where possible.
724 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
725 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
726 const Loop *CurLoop = I->first;
727 const SCEV *Op = I->second;
729 // This is the first operand. Just expand it.
732 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
733 // The running sum expression is a pointer. Try to form a getelementptr
734 // at this level with that as the base.
735 SmallVector<const SCEV *, 4> NewOps;
736 for (; I != E && I->first == CurLoop; ++I) {
737 // If the operand is SCEVUnknown and not instructions, peek through
738 // it, to enable more of it to be folded into the GEP.
739 const SCEV *X = I->second;
740 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
741 if (!isa<Instruction>(U->getValue()))
742 X = SE.getSCEV(U->getValue());
745 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
746 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
747 // The running sum is an integer, and there's a pointer at this level.
748 // Try to form a getelementptr. If the running sum is instructions,
749 // use a SCEVUnknown to avoid re-analyzing them.
750 SmallVector<const SCEV *, 4> NewOps;
751 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
753 for (++I; I != E && I->first == CurLoop; ++I)
754 NewOps.push_back(I->second);
755 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
756 } else if (Op->isNonConstantNegative()) {
757 // Instead of doing a negate and add, just do a subtract.
758 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
759 Sum = InsertNoopCastOfTo(Sum, Ty);
760 Sum = InsertBinop(Instruction::Sub, Sum, W);
764 Value *W = expandCodeFor(Op, Ty);
765 Sum = InsertNoopCastOfTo(Sum, Ty);
766 // Canonicalize a constant to the RHS.
767 if (isa<Constant>(Sum)) std::swap(Sum, W);
768 Sum = InsertBinop(Instruction::Add, Sum, W);
776 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
777 Type *Ty = SE.getEffectiveSCEVType(S->getType());
779 // Collect all the mul operands in a loop, along with their associated loops.
780 // Iterate in reverse so that constants are emitted last, all else equal.
781 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
782 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
783 E(S->op_begin()); I != E; ++I)
784 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
786 // Sort by loop. Use a stable sort so that constants follow non-constants.
787 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
789 // Emit instructions to mul all the operands. Hoist as much as possible
792 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
793 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
794 const SCEV *Op = I->second;
796 // This is the first operand. Just expand it.
799 } else if (Op->isAllOnesValue()) {
800 // Instead of doing a multiply by negative one, just do a negate.
801 Prod = InsertNoopCastOfTo(Prod, Ty);
802 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
806 Value *W = expandCodeFor(Op, Ty);
807 Prod = InsertNoopCastOfTo(Prod, Ty);
808 // Canonicalize a constant to the RHS.
809 if (isa<Constant>(Prod)) std::swap(Prod, W);
810 Prod = InsertBinop(Instruction::Mul, Prod, W);
818 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
819 Type *Ty = SE.getEffectiveSCEVType(S->getType());
821 Value *LHS = expandCodeFor(S->getLHS(), Ty);
822 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
823 const APInt &RHS = SC->getValue()->getValue();
824 if (RHS.isPowerOf2())
825 return InsertBinop(Instruction::LShr, LHS,
826 ConstantInt::get(Ty, RHS.logBase2()));
829 Value *RHS = expandCodeFor(S->getRHS(), Ty);
830 return InsertBinop(Instruction::UDiv, LHS, RHS);
833 /// Move parts of Base into Rest to leave Base with the minimal
834 /// expression that provides a pointer operand suitable for a
836 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
837 ScalarEvolution &SE) {
838 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
839 Base = A->getStart();
840 Rest = SE.getAddExpr(Rest,
841 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
842 A->getStepRecurrence(SE),
844 // FIXME: A->getNoWrapFlags(FlagNW)
847 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
848 Base = A->getOperand(A->getNumOperands()-1);
849 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
850 NewAddOps.back() = Rest;
851 Rest = SE.getAddExpr(NewAddOps);
852 ExposePointerBase(Base, Rest, SE);
856 /// Determine if this is a well-behaved chain of instructions leading back to
857 /// the PHI. If so, it may be reused by expanded expressions.
858 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
860 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
861 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
863 // If any of the operands don't dominate the insert position, bail.
864 // Addrec operands are always loop-invariant, so this can only happen
865 // if there are instructions which haven't been hoisted.
866 if (L == IVIncInsertLoop) {
867 for (User::op_iterator OI = IncV->op_begin()+1,
868 OE = IncV->op_end(); OI != OE; ++OI)
869 if (Instruction *OInst = dyn_cast<Instruction>(OI))
870 if (!SE.DT->dominates(OInst, IVIncInsertPos))
873 // Advance to the next instruction.
874 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
878 if (IncV->mayHaveSideEffects())
884 return isNormalAddRecExprPHI(PN, IncV, L);
887 /// getIVIncOperand returns an induction variable increment's induction
888 /// variable operand.
890 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
891 /// operands dominate InsertPos.
893 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
894 /// simple patterns generated by getAddRecExprPHILiterally and
895 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
896 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
897 Instruction *InsertPos,
899 if (IncV == InsertPos)
902 switch (IncV->getOpcode()) {
905 // Check for a simple Add/Sub or GEP of a loop invariant step.
906 case Instruction::Add:
907 case Instruction::Sub: {
908 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
909 if (!OInst || SE.DT->dominates(OInst, InsertPos))
910 return dyn_cast<Instruction>(IncV->getOperand(0));
913 case Instruction::BitCast:
914 return dyn_cast<Instruction>(IncV->getOperand(0));
915 case Instruction::GetElementPtr:
916 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
918 if (isa<Constant>(*I))
920 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
921 if (!SE.DT->dominates(OInst, InsertPos))
925 // allow any kind of GEP as long as it can be hoisted.
928 // This must be a pointer addition of constants (pretty), which is already
929 // handled, or some number of address-size elements (ugly). Ugly geps
930 // have 2 operands. i1* is used by the expander to represent an
931 // address-size element.
932 if (IncV->getNumOperands() != 2)
934 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
935 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
936 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
940 return dyn_cast<Instruction>(IncV->getOperand(0));
944 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
945 /// it available to other uses in this loop. Recursively hoist any operands,
946 /// until we reach a value that dominates InsertPos.
947 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
948 if (SE.DT->dominates(IncV, InsertPos))
951 // InsertPos must itself dominate IncV so that IncV's new position satisfies
952 // its existing users.
953 if (!SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
956 // Check that the chain of IV operands leading back to Phi can be hoisted.
957 SmallVector<Instruction*, 4> IVIncs;
959 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
962 // IncV is safe to hoist.
963 IVIncs.push_back(IncV);
965 if (SE.DT->dominates(IncV, InsertPos))
968 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
969 E = IVIncs.rend(); I != E; ++I) {
970 (*I)->moveBefore(InsertPos);
975 /// Determine if this cyclic phi is in a form that would have been generated by
976 /// LSR. We don't care if the phi was actually expanded in this pass, as long
977 /// as it is in a low-cost form, for example, no implied multiplication. This
978 /// should match any patterns generated by getAddRecExprPHILiterally and
980 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
982 for(Instruction *IVOper = IncV;
983 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
984 /*allowScale=*/false));) {
991 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
992 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
993 /// need to materialize IV increments elsewhere to handle difficult situations.
994 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
995 Type *ExpandTy, Type *IntTy,
998 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
999 if (ExpandTy->isPointerTy()) {
1000 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
1001 // If the step isn't constant, don't use an implicitly scaled GEP, because
1002 // that would require a multiply inside the loop.
1003 if (!isa<ConstantInt>(StepV))
1004 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
1005 GEPPtrTy->getAddressSpace());
1006 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
1007 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
1008 if (IncV->getType() != PN->getType()) {
1009 IncV = Builder.CreateBitCast(IncV, PN->getType());
1010 rememberInstruction(IncV);
1013 IncV = useSubtract ?
1014 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1015 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1016 rememberInstruction(IncV);
1021 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1022 /// the base addrec, which is the addrec without any non-loop-dominating
1023 /// values, and return the PHI.
1025 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1029 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1031 // Reuse a previously-inserted PHI, if present.
1032 BasicBlock *LatchBlock = L->getLoopLatch();
1034 for (BasicBlock::iterator I = L->getHeader()->begin();
1035 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1036 if (!SE.isSCEVable(PN->getType()) ||
1037 (SE.getEffectiveSCEVType(PN->getType()) !=
1038 SE.getEffectiveSCEVType(Normalized->getType())) ||
1039 SE.getSCEV(PN) != Normalized)
1043 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1046 if (!isExpandedAddRecExprPHI(PN, IncV, L))
1048 if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos))
1052 if (!isNormalAddRecExprPHI(PN, IncV, L))
1054 if (L == IVIncInsertLoop)
1056 if (SE.DT->dominates(IncV, IVIncInsertPos))
1058 // Make sure the increment is where we want it. But don't move it
1059 // down past a potential existing post-inc user.
1060 IncV->moveBefore(IVIncInsertPos);
1061 IVIncInsertPos = IncV;
1062 IncV = cast<Instruction>(IncV->getOperand(0));
1063 } while (IncV != PN);
1065 // Ok, the add recurrence looks usable.
1066 // Remember this PHI, even in post-inc mode.
1067 InsertedValues.insert(PN);
1068 // Remember the increment.
1069 rememberInstruction(IncV);
1074 // Save the original insertion point so we can restore it when we're done.
1075 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1076 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1078 // Another AddRec may need to be recursively expanded below. For example, if
1079 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1080 // loop. Remove this loop from the PostIncLoops set before expanding such
1081 // AddRecs. Otherwise, we cannot find a valid position for the step
1082 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1083 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1084 // so it's not worth implementing SmallPtrSet::swap.
1085 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1086 PostIncLoops.clear();
1088 // Expand code for the start value.
1089 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1090 L->getHeader()->begin());
1092 // StartV must be hoisted into L's preheader to dominate the new phi.
1093 assert(!isa<Instruction>(StartV) ||
1094 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1097 // Expand code for the step value. Do this before creating the PHI so that PHI
1098 // reuse code doesn't see an incomplete PHI.
1099 const SCEV *Step = Normalized->getStepRecurrence(SE);
1100 // If the stride is negative, insert a sub instead of an add for the increment
1101 // (unless it's a constant, because subtracts of constants are canonicalized
1103 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1105 Step = SE.getNegativeSCEV(Step);
1106 // Expand the step somewhere that dominates the loop header.
1107 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1110 BasicBlock *Header = L->getHeader();
1111 Builder.SetInsertPoint(Header, Header->begin());
1112 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1113 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1114 Twine(IVName) + ".iv");
1115 rememberInstruction(PN);
1117 // Create the step instructions and populate the PHI.
1118 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1119 BasicBlock *Pred = *HPI;
1121 // Add a start value.
1122 if (!L->contains(Pred)) {
1123 PN->addIncoming(StartV, Pred);
1127 // Create a step value and add it to the PHI.
1128 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1129 // instructions at IVIncInsertPos.
1130 Instruction *InsertPos = L == IVIncInsertLoop ?
1131 IVIncInsertPos : Pred->getTerminator();
1132 Builder.SetInsertPoint(InsertPos);
1133 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1135 PN->addIncoming(IncV, Pred);
1138 // Restore the original insert point.
1140 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1142 // After expanding subexpressions, restore the PostIncLoops set so the caller
1143 // can ensure that IVIncrement dominates the current uses.
1144 PostIncLoops = SavedPostIncLoops;
1146 // Remember this PHI, even in post-inc mode.
1147 InsertedValues.insert(PN);
1152 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1153 Type *STy = S->getType();
1154 Type *IntTy = SE.getEffectiveSCEVType(STy);
1155 const Loop *L = S->getLoop();
1157 // Determine a normalized form of this expression, which is the expression
1158 // before any post-inc adjustment is made.
1159 const SCEVAddRecExpr *Normalized = S;
1160 if (PostIncLoops.count(L)) {
1161 PostIncLoopSet Loops;
1164 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1165 Loops, SE, *SE.DT));
1168 // Strip off any non-loop-dominating component from the addrec start.
1169 const SCEV *Start = Normalized->getStart();
1170 const SCEV *PostLoopOffset = 0;
1171 if (!SE.properlyDominates(Start, L->getHeader())) {
1172 PostLoopOffset = Start;
1173 Start = SE.getConstant(Normalized->getType(), 0);
1174 Normalized = cast<SCEVAddRecExpr>(
1175 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1176 Normalized->getLoop(),
1177 // FIXME: Normalized->getNoWrapFlags(FlagNW)
1178 SCEV::FlagAnyWrap));
1181 // Strip off any non-loop-dominating component from the addrec step.
1182 const SCEV *Step = Normalized->getStepRecurrence(SE);
1183 const SCEV *PostLoopScale = 0;
1184 if (!SE.dominates(Step, L->getHeader())) {
1185 PostLoopScale = Step;
1186 Step = SE.getConstant(Normalized->getType(), 1);
1188 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
1189 Normalized->getLoop(),
1190 // FIXME: Normalized
1191 // ->getNoWrapFlags(FlagNW)
1192 SCEV::FlagAnyWrap));
1195 // Expand the core addrec. If we need post-loop scaling, force it to
1196 // expand to an integer type to avoid the need for additional casting.
1197 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1198 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1200 // Accommodate post-inc mode, if necessary.
1202 if (!PostIncLoops.count(L))
1205 // In PostInc mode, use the post-incremented value.
1206 BasicBlock *LatchBlock = L->getLoopLatch();
1207 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1208 Result = PN->getIncomingValueForBlock(LatchBlock);
1210 // For an expansion to use the postinc form, the client must call
1211 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1212 // or dominated by IVIncInsertPos.
1213 if (isa<Instruction>(Result)
1214 && !SE.DT->dominates(cast<Instruction>(Result),
1215 Builder.GetInsertPoint())) {
1216 // The induction variable's postinc expansion does not dominate this use.
1217 // IVUsers tries to prevent this case, so it is rare. However, it can
1218 // happen when an IVUser outside the loop is not dominated by the latch
1219 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1220 // all cases. Consider a phi outide whose operand is replaced during
1221 // expansion with the value of the postinc user. Without fundamentally
1222 // changing the way postinc users are tracked, the only remedy is
1223 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1224 // but hopefully expandCodeFor handles that.
1226 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1228 Step = SE.getNegativeSCEV(Step);
1229 // Expand the step somewhere that dominates the loop header.
1230 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1231 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1232 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1233 // Restore the insertion point to the place where the caller has
1234 // determined dominates all uses.
1235 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1236 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1240 // Re-apply any non-loop-dominating scale.
1241 if (PostLoopScale) {
1242 Result = InsertNoopCastOfTo(Result, IntTy);
1243 Result = Builder.CreateMul(Result,
1244 expandCodeFor(PostLoopScale, IntTy));
1245 rememberInstruction(Result);
1248 // Re-apply any non-loop-dominating offset.
1249 if (PostLoopOffset) {
1250 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1251 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1252 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1254 Result = InsertNoopCastOfTo(Result, IntTy);
1255 Result = Builder.CreateAdd(Result,
1256 expandCodeFor(PostLoopOffset, IntTy));
1257 rememberInstruction(Result);
1264 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1265 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1267 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1268 const Loop *L = S->getLoop();
1270 // First check for an existing canonical IV in a suitable type.
1271 PHINode *CanonicalIV = 0;
1272 if (PHINode *PN = L->getCanonicalInductionVariable())
1273 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1276 // Rewrite an AddRec in terms of the canonical induction variable, if
1277 // its type is more narrow.
1279 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1280 SE.getTypeSizeInBits(Ty)) {
1281 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1282 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1283 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1284 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1285 // FIXME: S->getNoWrapFlags(FlagNW)
1286 SCEV::FlagAnyWrap));
1287 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1288 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1289 BasicBlock::iterator NewInsertPt =
1290 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1291 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1292 isa<LandingPadInst>(NewInsertPt))
1294 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1296 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1300 // {X,+,F} --> X + {0,+,F}
1301 if (!S->getStart()->isZero()) {
1302 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1303 NewOps[0] = SE.getConstant(Ty, 0);
1304 // FIXME: can use S->getNoWrapFlags()
1305 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, SCEV::FlagAnyWrap);
1307 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1308 // comments on expandAddToGEP for details.
1309 const SCEV *Base = S->getStart();
1310 const SCEV *RestArray[1] = { Rest };
1311 // Dig into the expression to find the pointer base for a GEP.
1312 ExposePointerBase(Base, RestArray[0], SE);
1313 // If we found a pointer, expand the AddRec with a GEP.
1314 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1315 // Make sure the Base isn't something exotic, such as a multiplied
1316 // or divided pointer value. In those cases, the result type isn't
1317 // actually a pointer type.
1318 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1319 Value *StartV = expand(Base);
1320 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1321 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1325 // Just do a normal add. Pre-expand the operands to suppress folding.
1326 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1327 SE.getUnknown(expand(Rest))));
1330 // If we don't yet have a canonical IV, create one.
1332 // Create and insert the PHI node for the induction variable in the
1334 BasicBlock *Header = L->getHeader();
1335 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1336 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1338 rememberInstruction(CanonicalIV);
1340 Constant *One = ConstantInt::get(Ty, 1);
1341 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1342 BasicBlock *HP = *HPI;
1343 if (L->contains(HP)) {
1344 // Insert a unit add instruction right before the terminator
1345 // corresponding to the back-edge.
1346 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1348 HP->getTerminator());
1349 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1350 rememberInstruction(Add);
1351 CanonicalIV->addIncoming(Add, HP);
1353 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1358 // {0,+,1} --> Insert a canonical induction variable into the loop!
1359 if (S->isAffine() && S->getOperand(1)->isOne()) {
1360 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1361 "IVs with types different from the canonical IV should "
1362 "already have been handled!");
1366 // {0,+,F} --> {0,+,1} * F
1368 // If this is a simple linear addrec, emit it now as a special case.
1369 if (S->isAffine()) // {0,+,F} --> i*F
1371 expand(SE.getTruncateOrNoop(
1372 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1373 SE.getNoopOrAnyExtend(S->getOperand(1),
1374 CanonicalIV->getType())),
1377 // If this is a chain of recurrences, turn it into a closed form, using the
1378 // folders, then expandCodeFor the closed form. This allows the folders to
1379 // simplify the expression without having to build a bunch of special code
1380 // into this folder.
1381 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1383 // Promote S up to the canonical IV type, if the cast is foldable.
1384 const SCEV *NewS = S;
1385 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1386 if (isa<SCEVAddRecExpr>(Ext))
1389 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1390 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1392 // Truncate the result down to the original type, if needed.
1393 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1397 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1398 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1399 Value *V = expandCodeFor(S->getOperand(),
1400 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1401 Value *I = Builder.CreateTrunc(V, Ty);
1402 rememberInstruction(I);
1406 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1407 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1408 Value *V = expandCodeFor(S->getOperand(),
1409 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1410 Value *I = Builder.CreateZExt(V, Ty);
1411 rememberInstruction(I);
1415 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1416 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1417 Value *V = expandCodeFor(S->getOperand(),
1418 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1419 Value *I = Builder.CreateSExt(V, Ty);
1420 rememberInstruction(I);
1424 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1425 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1426 Type *Ty = LHS->getType();
1427 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1428 // In the case of mixed integer and pointer types, do the
1429 // rest of the comparisons as integer.
1430 if (S->getOperand(i)->getType() != Ty) {
1431 Ty = SE.getEffectiveSCEVType(Ty);
1432 LHS = InsertNoopCastOfTo(LHS, Ty);
1434 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1435 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1436 rememberInstruction(ICmp);
1437 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1438 rememberInstruction(Sel);
1441 // In the case of mixed integer and pointer types, cast the
1442 // final result back to the pointer type.
1443 if (LHS->getType() != S->getType())
1444 LHS = InsertNoopCastOfTo(LHS, S->getType());
1448 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1449 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1450 Type *Ty = LHS->getType();
1451 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1452 // In the case of mixed integer and pointer types, do the
1453 // rest of the comparisons as integer.
1454 if (S->getOperand(i)->getType() != Ty) {
1455 Ty = SE.getEffectiveSCEVType(Ty);
1456 LHS = InsertNoopCastOfTo(LHS, Ty);
1458 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1459 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1460 rememberInstruction(ICmp);
1461 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1462 rememberInstruction(Sel);
1465 // In the case of mixed integer and pointer types, cast the
1466 // final result back to the pointer type.
1467 if (LHS->getType() != S->getType())
1468 LHS = InsertNoopCastOfTo(LHS, S->getType());
1472 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1474 Builder.SetInsertPoint(IP->getParent(), IP);
1475 return expandCodeFor(SH, Ty);
1478 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1479 // Expand the code for this SCEV.
1480 Value *V = expand(SH);
1482 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1483 "non-trivial casts should be done with the SCEVs directly!");
1484 V = InsertNoopCastOfTo(V, Ty);
1489 Value *SCEVExpander::expand(const SCEV *S) {
1490 // Compute an insertion point for this SCEV object. Hoist the instructions
1491 // as far out in the loop nest as possible.
1492 Instruction *InsertPt = Builder.GetInsertPoint();
1493 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1494 L = L->getParentLoop())
1495 if (SE.isLoopInvariant(S, L)) {
1497 if (BasicBlock *Preheader = L->getLoopPreheader())
1498 InsertPt = Preheader->getTerminator();
1500 // LSR sets the insertion point for AddRec start/step values to the
1501 // block start to simplify value reuse, even though it's an invalid
1502 // position. SCEVExpander must correct for this in all cases.
1503 InsertPt = L->getHeader()->getFirstInsertionPt();
1506 // If the SCEV is computable at this level, insert it into the header
1507 // after the PHIs (and after any other instructions that we've inserted
1508 // there) so that it is guaranteed to dominate any user inside the loop.
1509 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1510 InsertPt = L->getHeader()->getFirstInsertionPt();
1511 while (InsertPt != Builder.GetInsertPoint()
1512 && (isInsertedInstruction(InsertPt)
1513 || isa<DbgInfoIntrinsic>(InsertPt))) {
1514 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1519 // Check to see if we already expanded this here.
1520 std::map<std::pair<const SCEV *, Instruction *>,
1521 AssertingVH<Value> >::iterator I =
1522 InsertedExpressions.find(std::make_pair(S, InsertPt));
1523 if (I != InsertedExpressions.end())
1526 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1527 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1528 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1530 // Expand the expression into instructions.
1531 Value *V = visit(S);
1533 // Remember the expanded value for this SCEV at this location.
1535 // This is independent of PostIncLoops. The mapped value simply materializes
1536 // the expression at this insertion point. If the mapped value happened to be
1537 // a postinc expansion, it could be reused by a non postinc user, but only if
1538 // its insertion point was already at the head of the loop.
1539 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1541 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1545 void SCEVExpander::rememberInstruction(Value *I) {
1546 if (!PostIncLoops.empty())
1547 InsertedPostIncValues.insert(I);
1549 InsertedValues.insert(I);
1552 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1553 Builder.SetInsertPoint(BB, I);
1556 /// getOrInsertCanonicalInductionVariable - This method returns the
1557 /// canonical induction variable of the specified type for the specified
1558 /// loop (inserting one if there is none). A canonical induction variable
1559 /// starts at zero and steps by one on each iteration.
1561 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1563 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1565 // Build a SCEV for {0,+,1}<L>.
1566 // Conservatively use FlagAnyWrap for now.
1567 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1568 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1570 // Emit code for it.
1571 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1572 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1573 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1575 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1580 /// Sort values by integer width for replaceCongruentIVs.
1581 static bool width_descending(Value *lhs, Value *rhs) {
1582 // Put pointers at the back and make sure pointer < pointer = false.
1583 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy())
1584 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy();
1585 return rhs->getType()->getPrimitiveSizeInBits()
1586 < lhs->getType()->getPrimitiveSizeInBits();
1589 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1590 /// replace them with their most canonical representative. Return the number of
1591 /// phis eliminated.
1593 /// This does not depend on any SCEVExpander state but should be used in
1594 /// the same context that SCEVExpander is used.
1595 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1596 SmallVectorImpl<WeakVH> &DeadInsts,
1597 const TargetLowering *TLI) {
1598 // Find integer phis in order of increasing width.
1599 SmallVector<PHINode*, 8> Phis;
1600 for (BasicBlock::iterator I = L->getHeader()->begin();
1601 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1602 Phis.push_back(Phi);
1605 std::sort(Phis.begin(), Phis.end(), width_descending);
1607 unsigned NumElim = 0;
1608 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1609 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1610 // so narrow phis can reuse them.
1611 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1612 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1613 PHINode *Phi = *PIter;
1615 if (!SE.isSCEVable(Phi->getType()))
1618 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1621 if (Phi->getType()->isIntegerTy() && TLI
1622 && TLI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1623 // This phi can be freely truncated to the narrowest phi type. Map the
1624 // truncated expression to it so it will be reused for narrow types.
1625 const SCEV *TruncExpr =
1626 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1627 ExprToIVMap[TruncExpr] = Phi;
1632 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1634 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1637 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1638 Instruction *OrigInc =
1639 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1640 Instruction *IsomorphicInc =
1641 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1643 // If this phi has the same width but is more canonical, replace the
1644 // original with it. As part of the "more canonical" determination,
1645 // respect a prior decision to use an IV chain.
1646 if (OrigPhiRef->getType() == Phi->getType()
1647 && !(ChainedPhis.count(Phi)
1648 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1649 && (ChainedPhis.count(Phi)
1650 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1651 std::swap(OrigPhiRef, Phi);
1652 std::swap(OrigInc, IsomorphicInc);
1654 // Replacing the congruent phi is sufficient because acyclic redundancy
1655 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1656 // that a phi is congruent, it's often the head of an IV user cycle that
1657 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1658 // common case of a single IV increment so that DeleteDeadPHIs can remove
1659 // cycles that had postinc uses.
1660 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1661 IsomorphicInc->getType());
1662 if (OrigInc != IsomorphicInc
1663 && TruncExpr == SE.getSCEV(IsomorphicInc)
1664 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1665 || hoistIVInc(OrigInc, IsomorphicInc))) {
1666 DEBUG_WITH_TYPE(DebugType, dbgs()
1667 << "INDVARS: Eliminated congruent iv.inc: "
1668 << *IsomorphicInc << '\n');
1669 Value *NewInc = OrigInc;
1670 if (OrigInc->getType() != IsomorphicInc->getType()) {
1671 Instruction *IP = isa<PHINode>(OrigInc)
1672 ? (Instruction*)L->getHeader()->getFirstInsertionPt()
1673 : OrigInc->getNextNode();
1674 IRBuilder<> Builder(IP);
1675 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1677 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1679 IsomorphicInc->replaceAllUsesWith(NewInc);
1680 DeadInsts.push_back(IsomorphicInc);
1683 DEBUG_WITH_TYPE(DebugType, dbgs()
1684 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1686 Value *NewIV = OrigPhiRef;
1687 if (OrigPhiRef->getType() != Phi->getType()) {
1688 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1689 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1690 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1692 Phi->replaceAllUsesWith(NewIV);
1693 DeadInsts.push_back(Phi);