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/ADT/STLExtras.h"
18 #include "llvm/Analysis/LoopInfo.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/IR/DataLayout.h"
21 #include "llvm/IR/IntrinsicInst.h"
22 #include "llvm/IR/LLVMContext.h"
23 #include "llvm/Support/Debug.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 produce a cast that will dominate all its
38 // uses. In particular, this is crucial for the case where the builder's
39 // insertion point *is* the point where we were asked to put the cast.
40 // Since we don't know the builder's insertion point is actually
41 // where the uses will be added (only that it dominates it), we are
42 // not allowed to move it.
43 BasicBlock::iterator BIP = Builder.GetInsertPoint();
45 Instruction *Ret = NULL;
47 // Check to see if there is already a cast!
48 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
51 if (U->getType() == Ty)
52 if (CastInst *CI = dyn_cast<CastInst>(U))
53 if (CI->getOpcode() == Op) {
54 // If the cast isn't where we want it, create a new cast at IP.
55 // Likewise, do not reuse a cast at BIP because it must dominate
56 // instructions that might be inserted before BIP.
57 if (BasicBlock::iterator(CI) != IP || BIP == IP) {
58 // Create a new cast, and leave the old cast in place in case
59 // it is being used as an insert point. Clear its operand
60 // so that it doesn't hold anything live.
61 Ret = CastInst::Create(Op, V, Ty, "", IP);
63 CI->replaceAllUsesWith(Ret);
64 CI->setOperand(0, UndefValue::get(V->getType()));
74 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP);
76 // We assert at the end of the function since IP might point to an
77 // instruction with different dominance properties than a cast
78 // (an invoke for example) and not dominate BIP (but the cast does).
79 assert(SE.DT->dominates(Ret, BIP));
81 rememberInstruction(Ret);
85 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
86 /// which must be possible with a noop cast, doing what we can to share
88 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
89 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
90 assert((Op == Instruction::BitCast ||
91 Op == Instruction::PtrToInt ||
92 Op == Instruction::IntToPtr) &&
93 "InsertNoopCastOfTo cannot perform non-noop casts!");
94 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
95 "InsertNoopCastOfTo cannot change sizes!");
97 // Short-circuit unnecessary bitcasts.
98 if (Op == Instruction::BitCast) {
99 if (V->getType() == Ty)
101 if (CastInst *CI = dyn_cast<CastInst>(V)) {
102 if (CI->getOperand(0)->getType() == Ty)
103 return CI->getOperand(0);
106 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
107 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
108 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
109 if (CastInst *CI = dyn_cast<CastInst>(V))
110 if ((CI->getOpcode() == Instruction::PtrToInt ||
111 CI->getOpcode() == Instruction::IntToPtr) &&
112 SE.getTypeSizeInBits(CI->getType()) ==
113 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
114 return CI->getOperand(0);
115 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
116 if ((CE->getOpcode() == Instruction::PtrToInt ||
117 CE->getOpcode() == Instruction::IntToPtr) &&
118 SE.getTypeSizeInBits(CE->getType()) ==
119 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
120 return CE->getOperand(0);
123 // Fold a cast of a constant.
124 if (Constant *C = dyn_cast<Constant>(V))
125 return ConstantExpr::getCast(Op, C, Ty);
127 // Cast the argument at the beginning of the entry block, after
128 // any bitcasts of other arguments.
129 if (Argument *A = dyn_cast<Argument>(V)) {
130 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
131 while ((isa<BitCastInst>(IP) &&
132 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
133 cast<BitCastInst>(IP)->getOperand(0) != A) ||
134 isa<DbgInfoIntrinsic>(IP) ||
135 isa<LandingPadInst>(IP))
137 return ReuseOrCreateCast(A, Ty, Op, IP);
140 // Cast the instruction immediately after the instruction.
141 Instruction *I = cast<Instruction>(V);
142 BasicBlock::iterator IP = I; ++IP;
143 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
144 IP = II->getNormalDest()->begin();
145 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
147 return ReuseOrCreateCast(I, Ty, Op, IP);
150 /// InsertBinop - Insert the specified binary operator, doing a small amount
151 /// of work to avoid inserting an obviously redundant operation.
152 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
153 Value *LHS, Value *RHS) {
154 // Fold a binop with constant operands.
155 if (Constant *CLHS = dyn_cast<Constant>(LHS))
156 if (Constant *CRHS = dyn_cast<Constant>(RHS))
157 return ConstantExpr::get(Opcode, CLHS, CRHS);
159 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
160 unsigned ScanLimit = 6;
161 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
162 // Scanning starts from the last instruction before the insertion point.
163 BasicBlock::iterator IP = Builder.GetInsertPoint();
164 if (IP != BlockBegin) {
166 for (; ScanLimit; --IP, --ScanLimit) {
167 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
169 if (isa<DbgInfoIntrinsic>(IP))
171 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
172 IP->getOperand(1) == RHS)
174 if (IP == BlockBegin) break;
178 // Save the original insertion point so we can restore it when we're done.
179 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
180 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
182 // Move the insertion point out of as many loops as we can.
183 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
184 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
185 BasicBlock *Preheader = L->getLoopPreheader();
186 if (!Preheader) break;
188 // Ok, move up a level.
189 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
192 // If we haven't found this binop, insert it.
193 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
194 BO->setDebugLoc(SaveInsertPt->getDebugLoc());
195 rememberInstruction(BO);
197 // Restore the original insert point.
199 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
204 /// FactorOutConstant - Test if S is divisible by Factor, using signed
205 /// division. If so, update S with Factor divided out and return true.
206 /// S need not be evenly divisible if a reasonable remainder can be
208 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
209 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
210 /// check to see if the divide was folded.
211 static bool FactorOutConstant(const SCEV *&S,
212 const SCEV *&Remainder,
215 const DataLayout *TD) {
216 // Everything is divisible by one.
222 S = SE.getConstant(S->getType(), 1);
226 // For a Constant, check for a multiple of the given factor.
227 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
231 // Check for divisibility.
232 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
234 ConstantInt::get(SE.getContext(),
235 C->getValue()->getValue().sdiv(
236 FC->getValue()->getValue()));
237 // If the quotient is zero and the remainder is non-zero, reject
238 // the value at this scale. It will be considered for subsequent
241 const SCEV *Div = SE.getConstant(CI);
244 SE.getAddExpr(Remainder,
245 SE.getConstant(C->getValue()->getValue().srem(
246 FC->getValue()->getValue())));
252 // In a Mul, check if there is a constant operand which is a multiple
253 // of the given factor.
254 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
256 // With DataLayout, the size is known. Check if there is a constant
257 // operand which is a multiple of the given factor. If so, we can
259 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
260 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
261 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
262 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
264 SE.getConstant(C->getValue()->getValue().sdiv(
265 FC->getValue()->getValue()));
266 S = SE.getMulExpr(NewMulOps);
270 // Without DataLayout, check if Factor can be factored out of any of the
271 // Mul's operands. If so, we can just remove it.
272 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
273 const SCEV *SOp = M->getOperand(i);
274 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
275 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
276 Remainder->isZero()) {
277 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
279 S = SE.getMulExpr(NewMulOps);
286 // In an AddRec, check if both start and step are divisible.
287 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
288 const SCEV *Step = A->getStepRecurrence(SE);
289 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
290 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
292 if (!StepRem->isZero())
294 const SCEV *Start = A->getStart();
295 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
297 S = SE.getAddRecExpr(Start, Step, A->getLoop(),
298 A->getNoWrapFlags(SCEV::FlagNW));
305 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
306 /// is the number of SCEVAddRecExprs present, which are kept at the end of
309 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
311 ScalarEvolution &SE) {
312 unsigned NumAddRecs = 0;
313 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
315 // Group Ops into non-addrecs and addrecs.
316 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
317 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
318 // Let ScalarEvolution sort and simplify the non-addrecs list.
319 const SCEV *Sum = NoAddRecs.empty() ?
320 SE.getConstant(Ty, 0) :
321 SE.getAddExpr(NoAddRecs);
322 // If it returned an add, use the operands. Otherwise it simplified
323 // the sum into a single value, so just use that.
325 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
326 Ops.append(Add->op_begin(), Add->op_end());
327 else if (!Sum->isZero())
329 // Then append the addrecs.
330 Ops.append(AddRecs.begin(), AddRecs.end());
333 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
334 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
335 /// This helps expose more opportunities for folding parts of the expressions
336 /// into GEP indices.
338 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
340 ScalarEvolution &SE) {
342 SmallVector<const SCEV *, 8> AddRecs;
343 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
344 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
345 const SCEV *Start = A->getStart();
346 if (Start->isZero()) break;
347 const SCEV *Zero = SE.getConstant(Ty, 0);
348 AddRecs.push_back(SE.getAddRecExpr(Zero,
349 A->getStepRecurrence(SE),
351 A->getNoWrapFlags(SCEV::FlagNW)));
352 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
354 Ops.append(Add->op_begin(), Add->op_end());
355 e += Add->getNumOperands();
360 if (!AddRecs.empty()) {
361 // Add the addrecs onto the end of the list.
362 Ops.append(AddRecs.begin(), AddRecs.end());
363 // Resort the operand list, moving any constants to the front.
364 SimplifyAddOperands(Ops, Ty, SE);
368 /// expandAddToGEP - Expand an addition expression with a pointer type into
369 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
370 /// BasicAliasAnalysis and other passes analyze the result. See the rules
371 /// for getelementptr vs. inttoptr in
372 /// http://llvm.org/docs/LangRef.html#pointeraliasing
375 /// Design note: The correctness of using getelementptr here depends on
376 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
377 /// they may introduce pointer arithmetic which may not be safely converted
378 /// into getelementptr.
380 /// Design note: It might seem desirable for this function to be more
381 /// loop-aware. If some of the indices are loop-invariant while others
382 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
383 /// loop-invariant portions of the overall computation outside the loop.
384 /// However, there are a few reasons this is not done here. Hoisting simple
385 /// arithmetic is a low-level optimization that often isn't very
386 /// important until late in the optimization process. In fact, passes
387 /// like InstructionCombining will combine GEPs, even if it means
388 /// pushing loop-invariant computation down into loops, so even if the
389 /// GEPs were split here, the work would quickly be undone. The
390 /// LoopStrengthReduction pass, which is usually run quite late (and
391 /// after the last InstructionCombining pass), takes care of hoisting
392 /// loop-invariant portions of expressions, after considering what
393 /// can be folded using target addressing modes.
395 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
396 const SCEV *const *op_end,
400 Type *ElTy = PTy->getElementType();
401 SmallVector<Value *, 4> GepIndices;
402 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
403 bool AnyNonZeroIndices = false;
405 // Split AddRecs up into parts as either of the parts may be usable
406 // without the other.
407 SplitAddRecs(Ops, Ty, SE);
409 // Descend down the pointer's type and attempt to convert the other
410 // operands into GEP indices, at each level. The first index in a GEP
411 // indexes into the array implied by the pointer operand; the rest of
412 // the indices index into the element or field type selected by the
415 // If the scale size is not 0, attempt to factor out a scale for
417 SmallVector<const SCEV *, 8> ScaledOps;
418 if (ElTy->isSized()) {
419 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
420 if (!ElSize->isZero()) {
421 SmallVector<const SCEV *, 8> NewOps;
422 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
423 const SCEV *Op = Ops[i];
424 const SCEV *Remainder = SE.getConstant(Ty, 0);
425 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
426 // Op now has ElSize factored out.
427 ScaledOps.push_back(Op);
428 if (!Remainder->isZero())
429 NewOps.push_back(Remainder);
430 AnyNonZeroIndices = true;
432 // The operand was not divisible, so add it to the list of operands
433 // we'll scan next iteration.
434 NewOps.push_back(Ops[i]);
437 // If we made any changes, update Ops.
438 if (!ScaledOps.empty()) {
440 SimplifyAddOperands(Ops, Ty, SE);
445 // Record the scaled array index for this level of the type. If
446 // we didn't find any operands that could be factored, tentatively
447 // assume that element zero was selected (since the zero offset
448 // would obviously be folded away).
449 Value *Scaled = ScaledOps.empty() ?
450 Constant::getNullValue(Ty) :
451 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
452 GepIndices.push_back(Scaled);
454 // Collect struct field index operands.
455 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
456 bool FoundFieldNo = false;
457 // An empty struct has no fields.
458 if (STy->getNumElements() == 0) break;
460 // With DataLayout, field offsets are known. See if a constant offset
461 // falls within any of the struct fields.
462 if (Ops.empty()) break;
463 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
464 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
465 const StructLayout &SL = *SE.TD->getStructLayout(STy);
466 uint64_t FullOffset = C->getValue()->getZExtValue();
467 if (FullOffset < SL.getSizeInBytes()) {
468 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
469 GepIndices.push_back(
470 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
471 ElTy = STy->getTypeAtIndex(ElIdx);
473 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
474 AnyNonZeroIndices = true;
479 // Without DataLayout, just check for an offsetof expression of the
480 // appropriate struct type.
481 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
482 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
485 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
486 GepIndices.push_back(FieldNo);
488 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
489 Ops[i] = SE.getConstant(Ty, 0);
490 AnyNonZeroIndices = true;
496 // If no struct field offsets were found, tentatively assume that
497 // field zero was selected (since the zero offset would obviously
500 ElTy = STy->getTypeAtIndex(0u);
501 GepIndices.push_back(
502 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
506 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
507 ElTy = ATy->getElementType();
512 // If none of the operands were convertible to proper GEP indices, cast
513 // the base to i8* and do an ugly getelementptr with that. It's still
514 // better than ptrtoint+arithmetic+inttoptr at least.
515 if (!AnyNonZeroIndices) {
516 // Cast the base to i8*.
517 V = InsertNoopCastOfTo(V,
518 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
520 assert(!isa<Instruction>(V) ||
521 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint()));
523 // Expand the operands for a plain byte offset.
524 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
526 // Fold a GEP with constant operands.
527 if (Constant *CLHS = dyn_cast<Constant>(V))
528 if (Constant *CRHS = dyn_cast<Constant>(Idx))
529 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
531 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
532 unsigned ScanLimit = 6;
533 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
534 // Scanning starts from the last instruction before the insertion point.
535 BasicBlock::iterator IP = Builder.GetInsertPoint();
536 if (IP != BlockBegin) {
538 for (; ScanLimit; --IP, --ScanLimit) {
539 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
541 if (isa<DbgInfoIntrinsic>(IP))
543 if (IP->getOpcode() == Instruction::GetElementPtr &&
544 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
546 if (IP == BlockBegin) break;
550 // Save the original insertion point so we can restore it when we're done.
551 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
552 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
554 // Move the insertion point out of as many loops as we can.
555 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
556 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
557 BasicBlock *Preheader = L->getLoopPreheader();
558 if (!Preheader) break;
560 // Ok, move up a level.
561 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
565 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
566 rememberInstruction(GEP);
568 // Restore the original insert point.
570 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
575 // Save the original insertion point so we can restore it when we're done.
576 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
577 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
579 // Move the insertion point out of as many loops as we can.
580 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
581 if (!L->isLoopInvariant(V)) break;
583 bool AnyIndexNotLoopInvariant = false;
584 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
585 E = GepIndices.end(); I != E; ++I)
586 if (!L->isLoopInvariant(*I)) {
587 AnyIndexNotLoopInvariant = true;
590 if (AnyIndexNotLoopInvariant)
593 BasicBlock *Preheader = L->getLoopPreheader();
594 if (!Preheader) break;
596 // Ok, move up a level.
597 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
600 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
601 // because ScalarEvolution may have changed the address arithmetic to
602 // compute a value which is beyond the end of the allocated object.
604 if (V->getType() != PTy)
605 Casted = InsertNoopCastOfTo(Casted, PTy);
606 Value *GEP = Builder.CreateGEP(Casted,
609 Ops.push_back(SE.getUnknown(GEP));
610 rememberInstruction(GEP);
612 // Restore the original insert point.
614 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
616 return expand(SE.getAddExpr(Ops));
619 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
620 /// SCEV expansion. If they are nested, this is the most nested. If they are
621 /// neighboring, pick the later.
622 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
626 if (A->contains(B)) return B;
627 if (B->contains(A)) return A;
628 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
629 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
630 return A; // Arbitrarily break the tie.
633 /// getRelevantLoop - Get the most relevant loop associated with the given
634 /// expression, according to PickMostRelevantLoop.
635 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
636 // Test whether we've already computed the most relevant loop for this SCEV.
637 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
638 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
640 return Pair.first->second;
642 if (isa<SCEVConstant>(S))
643 // A constant has no relevant loops.
645 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
646 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
647 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
648 // A non-instruction has no relevant loops.
651 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
653 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
655 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
657 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
658 return RelevantLoops[N] = L;
660 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
661 const Loop *Result = getRelevantLoop(C->getOperand());
662 return RelevantLoops[C] = Result;
664 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
666 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
667 getRelevantLoop(D->getRHS()),
669 return RelevantLoops[D] = Result;
671 llvm_unreachable("Unexpected SCEV type!");
676 /// LoopCompare - Compare loops by PickMostRelevantLoop.
680 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
682 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
683 std::pair<const Loop *, const SCEV *> RHS) const {
684 // Keep pointer operands sorted at the end.
685 if (LHS.second->getType()->isPointerTy() !=
686 RHS.second->getType()->isPointerTy())
687 return LHS.second->getType()->isPointerTy();
689 // Compare loops with PickMostRelevantLoop.
690 if (LHS.first != RHS.first)
691 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
693 // If one operand is a non-constant negative and the other is not,
694 // put the non-constant negative on the right so that a sub can
695 // be used instead of a negate and add.
696 if (LHS.second->isNonConstantNegative()) {
697 if (!RHS.second->isNonConstantNegative())
699 } else if (RHS.second->isNonConstantNegative())
702 // Otherwise they are equivalent according to this comparison.
709 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
710 Type *Ty = SE.getEffectiveSCEVType(S->getType());
712 // Collect all the add operands in a loop, along with their associated loops.
713 // Iterate in reverse so that constants are emitted last, all else equal, and
714 // so that pointer operands are inserted first, which the code below relies on
715 // to form more involved GEPs.
716 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
717 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
718 E(S->op_begin()); I != E; ++I)
719 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
721 // Sort by loop. Use a stable sort so that constants follow non-constants and
722 // pointer operands precede non-pointer operands.
723 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
725 // Emit instructions to add all the operands. Hoist as much as possible
726 // out of loops, and form meaningful getelementptrs where possible.
728 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
729 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
730 const Loop *CurLoop = I->first;
731 const SCEV *Op = I->second;
733 // This is the first operand. Just expand it.
736 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
737 // The running sum expression is a pointer. Try to form a getelementptr
738 // at this level with that as the base.
739 SmallVector<const SCEV *, 4> NewOps;
740 for (; I != E && I->first == CurLoop; ++I) {
741 // If the operand is SCEVUnknown and not instructions, peek through
742 // it, to enable more of it to be folded into the GEP.
743 const SCEV *X = I->second;
744 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
745 if (!isa<Instruction>(U->getValue()))
746 X = SE.getSCEV(U->getValue());
749 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
750 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
751 // The running sum is an integer, and there's a pointer at this level.
752 // Try to form a getelementptr. If the running sum is instructions,
753 // use a SCEVUnknown to avoid re-analyzing them.
754 SmallVector<const SCEV *, 4> NewOps;
755 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
757 for (++I; I != E && I->first == CurLoop; ++I)
758 NewOps.push_back(I->second);
759 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
760 } else if (Op->isNonConstantNegative()) {
761 // Instead of doing a negate and add, just do a subtract.
762 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
763 Sum = InsertNoopCastOfTo(Sum, Ty);
764 Sum = InsertBinop(Instruction::Sub, Sum, W);
768 Value *W = expandCodeFor(Op, Ty);
769 Sum = InsertNoopCastOfTo(Sum, Ty);
770 // Canonicalize a constant to the RHS.
771 if (isa<Constant>(Sum)) std::swap(Sum, W);
772 Sum = InsertBinop(Instruction::Add, Sum, W);
780 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
781 Type *Ty = SE.getEffectiveSCEVType(S->getType());
783 // Collect all the mul operands in a loop, along with their associated loops.
784 // Iterate in reverse so that constants are emitted last, all else equal.
785 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
786 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
787 E(S->op_begin()); I != E; ++I)
788 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
790 // Sort by loop. Use a stable sort so that constants follow non-constants.
791 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
793 // Emit instructions to mul all the operands. Hoist as much as possible
796 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
797 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
798 const SCEV *Op = I->second;
800 // This is the first operand. Just expand it.
803 } else if (Op->isAllOnesValue()) {
804 // Instead of doing a multiply by negative one, just do a negate.
805 Prod = InsertNoopCastOfTo(Prod, Ty);
806 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
810 Value *W = expandCodeFor(Op, Ty);
811 Prod = InsertNoopCastOfTo(Prod, Ty);
812 // Canonicalize a constant to the RHS.
813 if (isa<Constant>(Prod)) std::swap(Prod, W);
814 Prod = InsertBinop(Instruction::Mul, Prod, W);
822 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
823 Type *Ty = SE.getEffectiveSCEVType(S->getType());
825 Value *LHS = expandCodeFor(S->getLHS(), Ty);
826 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
827 const APInt &RHS = SC->getValue()->getValue();
828 if (RHS.isPowerOf2())
829 return InsertBinop(Instruction::LShr, LHS,
830 ConstantInt::get(Ty, RHS.logBase2()));
833 Value *RHS = expandCodeFor(S->getRHS(), Ty);
834 return InsertBinop(Instruction::UDiv, LHS, RHS);
837 /// Move parts of Base into Rest to leave Base with the minimal
838 /// expression that provides a pointer operand suitable for a
840 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
841 ScalarEvolution &SE) {
842 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
843 Base = A->getStart();
844 Rest = SE.getAddExpr(Rest,
845 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
846 A->getStepRecurrence(SE),
848 A->getNoWrapFlags(SCEV::FlagNW)));
850 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
851 Base = A->getOperand(A->getNumOperands()-1);
852 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
853 NewAddOps.back() = Rest;
854 Rest = SE.getAddExpr(NewAddOps);
855 ExposePointerBase(Base, Rest, SE);
859 /// Determine if this is a well-behaved chain of instructions leading back to
860 /// the PHI. If so, it may be reused by expanded expressions.
861 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
863 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
864 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
866 // If any of the operands don't dominate the insert position, bail.
867 // Addrec operands are always loop-invariant, so this can only happen
868 // if there are instructions which haven't been hoisted.
869 if (L == IVIncInsertLoop) {
870 for (User::op_iterator OI = IncV->op_begin()+1,
871 OE = IncV->op_end(); OI != OE; ++OI)
872 if (Instruction *OInst = dyn_cast<Instruction>(OI))
873 if (!SE.DT->dominates(OInst, IVIncInsertPos))
876 // Advance to the next instruction.
877 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
881 if (IncV->mayHaveSideEffects())
887 return isNormalAddRecExprPHI(PN, IncV, L);
890 /// getIVIncOperand returns an induction variable increment's induction
891 /// variable operand.
893 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
894 /// operands dominate InsertPos.
896 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
897 /// simple patterns generated by getAddRecExprPHILiterally and
898 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
899 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
900 Instruction *InsertPos,
902 if (IncV == InsertPos)
905 switch (IncV->getOpcode()) {
908 // Check for a simple Add/Sub or GEP of a loop invariant step.
909 case Instruction::Add:
910 case Instruction::Sub: {
911 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
912 if (!OInst || SE.DT->dominates(OInst, InsertPos))
913 return dyn_cast<Instruction>(IncV->getOperand(0));
916 case Instruction::BitCast:
917 return dyn_cast<Instruction>(IncV->getOperand(0));
918 case Instruction::GetElementPtr:
919 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
921 if (isa<Constant>(*I))
923 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
924 if (!SE.DT->dominates(OInst, InsertPos))
928 // allow any kind of GEP as long as it can be hoisted.
931 // This must be a pointer addition of constants (pretty), which is already
932 // handled, or some number of address-size elements (ugly). Ugly geps
933 // have 2 operands. i1* is used by the expander to represent an
934 // address-size element.
935 if (IncV->getNumOperands() != 2)
937 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
938 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
939 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
943 return dyn_cast<Instruction>(IncV->getOperand(0));
947 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
948 /// it available to other uses in this loop. Recursively hoist any operands,
949 /// until we reach a value that dominates InsertPos.
950 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
951 if (SE.DT->dominates(IncV, InsertPos))
954 // InsertPos must itself dominate IncV so that IncV's new position satisfies
955 // its existing users.
956 if (isa<PHINode>(InsertPos)
957 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
960 // Check that the chain of IV operands leading back to Phi can be hoisted.
961 SmallVector<Instruction*, 4> IVIncs;
963 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
966 // IncV is safe to hoist.
967 IVIncs.push_back(IncV);
969 if (SE.DT->dominates(IncV, InsertPos))
972 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
973 E = IVIncs.rend(); I != E; ++I) {
974 (*I)->moveBefore(InsertPos);
979 /// Determine if this cyclic phi is in a form that would have been generated by
980 /// LSR. We don't care if the phi was actually expanded in this pass, as long
981 /// as it is in a low-cost form, for example, no implied multiplication. This
982 /// should match any patterns generated by getAddRecExprPHILiterally and
984 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
986 for(Instruction *IVOper = IncV;
987 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
988 /*allowScale=*/false));) {
995 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
996 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
997 /// need to materialize IV increments elsewhere to handle difficult situations.
998 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
999 Type *ExpandTy, Type *IntTy,
1002 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
1003 if (ExpandTy->isPointerTy()) {
1004 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
1005 // If the step isn't constant, don't use an implicitly scaled GEP, because
1006 // that would require a multiply inside the loop.
1007 if (!isa<ConstantInt>(StepV))
1008 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
1009 GEPPtrTy->getAddressSpace());
1010 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
1011 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
1012 if (IncV->getType() != PN->getType()) {
1013 IncV = Builder.CreateBitCast(IncV, PN->getType());
1014 rememberInstruction(IncV);
1017 IncV = useSubtract ?
1018 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1019 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1020 rememberInstruction(IncV);
1025 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1026 /// the base addrec, which is the addrec without any non-loop-dominating
1027 /// values, and return the PHI.
1029 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1033 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1035 // Reuse a previously-inserted PHI, if present.
1036 BasicBlock *LatchBlock = L->getLoopLatch();
1038 for (BasicBlock::iterator I = L->getHeader()->begin();
1039 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1040 if (!SE.isSCEVable(PN->getType()) ||
1041 (SE.getEffectiveSCEVType(PN->getType()) !=
1042 SE.getEffectiveSCEVType(Normalized->getType())) ||
1043 SE.getSCEV(PN) != Normalized)
1047 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1050 if (!isExpandedAddRecExprPHI(PN, IncV, L))
1052 if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos))
1056 if (!isNormalAddRecExprPHI(PN, IncV, L))
1058 if (L == IVIncInsertLoop)
1060 if (SE.DT->dominates(IncV, IVIncInsertPos))
1062 // Make sure the increment is where we want it. But don't move it
1063 // down past a potential existing post-inc user.
1064 IncV->moveBefore(IVIncInsertPos);
1065 IVIncInsertPos = IncV;
1066 IncV = cast<Instruction>(IncV->getOperand(0));
1067 } while (IncV != PN);
1069 // Ok, the add recurrence looks usable.
1070 // Remember this PHI, even in post-inc mode.
1071 InsertedValues.insert(PN);
1072 // Remember the increment.
1073 rememberInstruction(IncV);
1078 // Save the original insertion point so we can restore it when we're done.
1079 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1080 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1082 // Another AddRec may need to be recursively expanded below. For example, if
1083 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1084 // loop. Remove this loop from the PostIncLoops set before expanding such
1085 // AddRecs. Otherwise, we cannot find a valid position for the step
1086 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1087 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1088 // so it's not worth implementing SmallPtrSet::swap.
1089 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1090 PostIncLoops.clear();
1092 // Expand code for the start value.
1093 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1094 L->getHeader()->begin());
1096 // StartV must be hoisted into L's preheader to dominate the new phi.
1097 assert(!isa<Instruction>(StartV) ||
1098 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1101 // Expand code for the step value. Do this before creating the PHI so that PHI
1102 // reuse code doesn't see an incomplete PHI.
1103 const SCEV *Step = Normalized->getStepRecurrence(SE);
1104 // If the stride is negative, insert a sub instead of an add for the increment
1105 // (unless it's a constant, because subtracts of constants are canonicalized
1107 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1109 Step = SE.getNegativeSCEV(Step);
1110 // Expand the step somewhere that dominates the loop header.
1111 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1114 BasicBlock *Header = L->getHeader();
1115 Builder.SetInsertPoint(Header, Header->begin());
1116 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1117 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1118 Twine(IVName) + ".iv");
1119 rememberInstruction(PN);
1121 // Create the step instructions and populate the PHI.
1122 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1123 BasicBlock *Pred = *HPI;
1125 // Add a start value.
1126 if (!L->contains(Pred)) {
1127 PN->addIncoming(StartV, Pred);
1131 // Create a step value and add it to the PHI.
1132 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1133 // instructions at IVIncInsertPos.
1134 Instruction *InsertPos = L == IVIncInsertLoop ?
1135 IVIncInsertPos : Pred->getTerminator();
1136 Builder.SetInsertPoint(InsertPos);
1137 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1138 if (isa<OverflowingBinaryOperator>(IncV)) {
1139 if (Normalized->getNoWrapFlags(SCEV::FlagNUW))
1140 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1141 if (Normalized->getNoWrapFlags(SCEV::FlagNSW))
1142 cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1144 PN->addIncoming(IncV, Pred);
1147 // Restore the original insert point.
1149 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1151 // After expanding subexpressions, restore the PostIncLoops set so the caller
1152 // can ensure that IVIncrement dominates the current uses.
1153 PostIncLoops = SavedPostIncLoops;
1155 // Remember this PHI, even in post-inc mode.
1156 InsertedValues.insert(PN);
1161 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1162 Type *STy = S->getType();
1163 Type *IntTy = SE.getEffectiveSCEVType(STy);
1164 const Loop *L = S->getLoop();
1166 // Determine a normalized form of this expression, which is the expression
1167 // before any post-inc adjustment is made.
1168 const SCEVAddRecExpr *Normalized = S;
1169 if (PostIncLoops.count(L)) {
1170 PostIncLoopSet Loops;
1173 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1174 Loops, SE, *SE.DT));
1177 // Strip off any non-loop-dominating component from the addrec start.
1178 const SCEV *Start = Normalized->getStart();
1179 const SCEV *PostLoopOffset = 0;
1180 if (!SE.properlyDominates(Start, L->getHeader())) {
1181 PostLoopOffset = Start;
1182 Start = SE.getConstant(Normalized->getType(), 0);
1183 Normalized = cast<SCEVAddRecExpr>(
1184 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1185 Normalized->getLoop(),
1186 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1189 // Strip off any non-loop-dominating component from the addrec step.
1190 const SCEV *Step = Normalized->getStepRecurrence(SE);
1191 const SCEV *PostLoopScale = 0;
1192 if (!SE.dominates(Step, L->getHeader())) {
1193 PostLoopScale = Step;
1194 Step = SE.getConstant(Normalized->getType(), 1);
1196 cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1197 Start, Step, Normalized->getLoop(),
1198 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1201 // Expand the core addrec. If we need post-loop scaling, force it to
1202 // expand to an integer type to avoid the need for additional casting.
1203 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1204 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1206 // Accommodate post-inc mode, if necessary.
1208 if (!PostIncLoops.count(L))
1211 // In PostInc mode, use the post-incremented value.
1212 BasicBlock *LatchBlock = L->getLoopLatch();
1213 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1214 Result = PN->getIncomingValueForBlock(LatchBlock);
1216 // For an expansion to use the postinc form, the client must call
1217 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1218 // or dominated by IVIncInsertPos.
1219 if (isa<Instruction>(Result)
1220 && !SE.DT->dominates(cast<Instruction>(Result),
1221 Builder.GetInsertPoint())) {
1222 // The induction variable's postinc expansion does not dominate this use.
1223 // IVUsers tries to prevent this case, so it is rare. However, it can
1224 // happen when an IVUser outside the loop is not dominated by the latch
1225 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1226 // all cases. Consider a phi outide whose operand is replaced during
1227 // expansion with the value of the postinc user. Without fundamentally
1228 // changing the way postinc users are tracked, the only remedy is
1229 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1230 // but hopefully expandCodeFor handles that.
1232 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1234 Step = SE.getNegativeSCEV(Step);
1235 // Expand the step somewhere that dominates the loop header.
1236 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1237 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1238 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1239 // Restore the insertion point to the place where the caller has
1240 // determined dominates all uses.
1241 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1242 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1246 // Re-apply any non-loop-dominating scale.
1247 if (PostLoopScale) {
1248 Result = InsertNoopCastOfTo(Result, IntTy);
1249 Result = Builder.CreateMul(Result,
1250 expandCodeFor(PostLoopScale, IntTy));
1251 rememberInstruction(Result);
1254 // Re-apply any non-loop-dominating offset.
1255 if (PostLoopOffset) {
1256 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1257 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1258 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1260 Result = InsertNoopCastOfTo(Result, IntTy);
1261 Result = Builder.CreateAdd(Result,
1262 expandCodeFor(PostLoopOffset, IntTy));
1263 rememberInstruction(Result);
1270 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1271 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1273 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1274 const Loop *L = S->getLoop();
1276 // First check for an existing canonical IV in a suitable type.
1277 PHINode *CanonicalIV = 0;
1278 if (PHINode *PN = L->getCanonicalInductionVariable())
1279 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1282 // Rewrite an AddRec in terms of the canonical induction variable, if
1283 // its type is more narrow.
1285 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1286 SE.getTypeSizeInBits(Ty)) {
1287 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1288 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1289 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1290 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1291 S->getNoWrapFlags(SCEV::FlagNW)));
1292 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1293 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1294 BasicBlock::iterator NewInsertPt =
1295 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1296 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1297 isa<LandingPadInst>(NewInsertPt))
1299 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1301 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1305 // {X,+,F} --> X + {0,+,F}
1306 if (!S->getStart()->isZero()) {
1307 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1308 NewOps[0] = SE.getConstant(Ty, 0);
1309 const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1310 S->getNoWrapFlags(SCEV::FlagNW));
1312 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1313 // comments on expandAddToGEP for details.
1314 const SCEV *Base = S->getStart();
1315 const SCEV *RestArray[1] = { Rest };
1316 // Dig into the expression to find the pointer base for a GEP.
1317 ExposePointerBase(Base, RestArray[0], SE);
1318 // If we found a pointer, expand the AddRec with a GEP.
1319 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1320 // Make sure the Base isn't something exotic, such as a multiplied
1321 // or divided pointer value. In those cases, the result type isn't
1322 // actually a pointer type.
1323 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1324 Value *StartV = expand(Base);
1325 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1326 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1330 // Just do a normal add. Pre-expand the operands to suppress folding.
1331 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1332 SE.getUnknown(expand(Rest))));
1335 // If we don't yet have a canonical IV, create one.
1337 // Create and insert the PHI node for the induction variable in the
1339 BasicBlock *Header = L->getHeader();
1340 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1341 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1343 rememberInstruction(CanonicalIV);
1345 Constant *One = ConstantInt::get(Ty, 1);
1346 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1347 BasicBlock *HP = *HPI;
1348 if (L->contains(HP)) {
1349 // Insert a unit add instruction right before the terminator
1350 // corresponding to the back-edge.
1351 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1353 HP->getTerminator());
1354 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1355 rememberInstruction(Add);
1356 CanonicalIV->addIncoming(Add, HP);
1358 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1363 // {0,+,1} --> Insert a canonical induction variable into the loop!
1364 if (S->isAffine() && S->getOperand(1)->isOne()) {
1365 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1366 "IVs with types different from the canonical IV should "
1367 "already have been handled!");
1371 // {0,+,F} --> {0,+,1} * F
1373 // If this is a simple linear addrec, emit it now as a special case.
1374 if (S->isAffine()) // {0,+,F} --> i*F
1376 expand(SE.getTruncateOrNoop(
1377 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1378 SE.getNoopOrAnyExtend(S->getOperand(1),
1379 CanonicalIV->getType())),
1382 // If this is a chain of recurrences, turn it into a closed form, using the
1383 // folders, then expandCodeFor the closed form. This allows the folders to
1384 // simplify the expression without having to build a bunch of special code
1385 // into this folder.
1386 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1388 // Promote S up to the canonical IV type, if the cast is foldable.
1389 const SCEV *NewS = S;
1390 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1391 if (isa<SCEVAddRecExpr>(Ext))
1394 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1395 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1397 // Truncate the result down to the original type, if needed.
1398 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1402 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1403 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1404 Value *V = expandCodeFor(S->getOperand(),
1405 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1406 Value *I = Builder.CreateTrunc(V, Ty);
1407 rememberInstruction(I);
1411 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1412 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1413 Value *V = expandCodeFor(S->getOperand(),
1414 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1415 Value *I = Builder.CreateZExt(V, Ty);
1416 rememberInstruction(I);
1420 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1421 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1422 Value *V = expandCodeFor(S->getOperand(),
1423 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1424 Value *I = Builder.CreateSExt(V, Ty);
1425 rememberInstruction(I);
1429 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1430 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1431 Type *Ty = LHS->getType();
1432 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1433 // In the case of mixed integer and pointer types, do the
1434 // rest of the comparisons as integer.
1435 if (S->getOperand(i)->getType() != Ty) {
1436 Ty = SE.getEffectiveSCEVType(Ty);
1437 LHS = InsertNoopCastOfTo(LHS, Ty);
1439 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1440 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1441 rememberInstruction(ICmp);
1442 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1443 rememberInstruction(Sel);
1446 // In the case of mixed integer and pointer types, cast the
1447 // final result back to the pointer type.
1448 if (LHS->getType() != S->getType())
1449 LHS = InsertNoopCastOfTo(LHS, S->getType());
1453 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1454 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1455 Type *Ty = LHS->getType();
1456 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1457 // In the case of mixed integer and pointer types, do the
1458 // rest of the comparisons as integer.
1459 if (S->getOperand(i)->getType() != Ty) {
1460 Ty = SE.getEffectiveSCEVType(Ty);
1461 LHS = InsertNoopCastOfTo(LHS, Ty);
1463 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1464 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1465 rememberInstruction(ICmp);
1466 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1467 rememberInstruction(Sel);
1470 // In the case of mixed integer and pointer types, cast the
1471 // final result back to the pointer type.
1472 if (LHS->getType() != S->getType())
1473 LHS = InsertNoopCastOfTo(LHS, S->getType());
1477 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1479 Builder.SetInsertPoint(IP->getParent(), IP);
1480 return expandCodeFor(SH, Ty);
1483 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1484 // Expand the code for this SCEV.
1485 Value *V = expand(SH);
1487 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1488 "non-trivial casts should be done with the SCEVs directly!");
1489 V = InsertNoopCastOfTo(V, Ty);
1494 Value *SCEVExpander::expand(const SCEV *S) {
1495 // Compute an insertion point for this SCEV object. Hoist the instructions
1496 // as far out in the loop nest as possible.
1497 Instruction *InsertPt = Builder.GetInsertPoint();
1498 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1499 L = L->getParentLoop())
1500 if (SE.isLoopInvariant(S, L)) {
1502 if (BasicBlock *Preheader = L->getLoopPreheader())
1503 InsertPt = Preheader->getTerminator();
1505 // LSR sets the insertion point for AddRec start/step values to the
1506 // block start to simplify value reuse, even though it's an invalid
1507 // position. SCEVExpander must correct for this in all cases.
1508 InsertPt = L->getHeader()->getFirstInsertionPt();
1511 // If the SCEV is computable at this level, insert it into the header
1512 // after the PHIs (and after any other instructions that we've inserted
1513 // there) so that it is guaranteed to dominate any user inside the loop.
1514 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1515 InsertPt = L->getHeader()->getFirstInsertionPt();
1516 while (InsertPt != Builder.GetInsertPoint()
1517 && (isInsertedInstruction(InsertPt)
1518 || isa<DbgInfoIntrinsic>(InsertPt))) {
1519 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1524 // Check to see if we already expanded this here.
1525 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator
1526 I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1527 if (I != InsertedExpressions.end())
1530 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1531 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1532 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1534 // Expand the expression into instructions.
1535 Value *V = visit(S);
1537 // Remember the expanded value for this SCEV at this location.
1539 // This is independent of PostIncLoops. The mapped value simply materializes
1540 // the expression at this insertion point. If the mapped value happened to be
1541 // a postinc expansion, it could be reused by a non postinc user, but only if
1542 // its insertion point was already at the head of the loop.
1543 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1545 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1549 void SCEVExpander::rememberInstruction(Value *I) {
1550 if (!PostIncLoops.empty())
1551 InsertedPostIncValues.insert(I);
1553 InsertedValues.insert(I);
1556 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1557 Builder.SetInsertPoint(BB, I);
1560 /// getOrInsertCanonicalInductionVariable - This method returns the
1561 /// canonical induction variable of the specified type for the specified
1562 /// loop (inserting one if there is none). A canonical induction variable
1563 /// starts at zero and steps by one on each iteration.
1565 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1567 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1569 // Build a SCEV for {0,+,1}<L>.
1570 // Conservatively use FlagAnyWrap for now.
1571 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1572 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1574 // Emit code for it.
1575 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1576 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1577 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1579 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1584 /// Sort values by integer width for replaceCongruentIVs.
1585 static bool width_descending(Value *lhs, Value *rhs) {
1586 // Put pointers at the back and make sure pointer < pointer = false.
1587 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy())
1588 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy();
1589 return rhs->getType()->getPrimitiveSizeInBits()
1590 < lhs->getType()->getPrimitiveSizeInBits();
1593 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1594 /// replace them with their most canonical representative. Return the number of
1595 /// phis eliminated.
1597 /// This does not depend on any SCEVExpander state but should be used in
1598 /// the same context that SCEVExpander is used.
1599 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1600 SmallVectorImpl<WeakVH> &DeadInsts,
1601 const TargetTransformInfo *TTI) {
1602 // Find integer phis in order of increasing width.
1603 SmallVector<PHINode*, 8> Phis;
1604 for (BasicBlock::iterator I = L->getHeader()->begin();
1605 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1606 Phis.push_back(Phi);
1609 std::sort(Phis.begin(), Phis.end(), width_descending);
1611 unsigned NumElim = 0;
1612 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1613 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1614 // so narrow phis can reuse them.
1615 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1616 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1617 PHINode *Phi = *PIter;
1619 // Fold constant phis. They may be congruent to other constant phis and
1620 // would confuse the logic below that expects proper IVs.
1621 if (Value *V = Phi->hasConstantValue()) {
1622 Phi->replaceAllUsesWith(V);
1623 DeadInsts.push_back(Phi);
1625 DEBUG_WITH_TYPE(DebugType, dbgs()
1626 << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1630 if (!SE.isSCEVable(Phi->getType()))
1633 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1636 if (Phi->getType()->isIntegerTy() && TTI
1637 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1638 // This phi can be freely truncated to the narrowest phi type. Map the
1639 // truncated expression to it so it will be reused for narrow types.
1640 const SCEV *TruncExpr =
1641 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1642 ExprToIVMap[TruncExpr] = Phi;
1647 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1649 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1652 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1653 Instruction *OrigInc =
1654 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1655 Instruction *IsomorphicInc =
1656 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1658 // If this phi has the same width but is more canonical, replace the
1659 // original with it. As part of the "more canonical" determination,
1660 // respect a prior decision to use an IV chain.
1661 if (OrigPhiRef->getType() == Phi->getType()
1662 && !(ChainedPhis.count(Phi)
1663 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1664 && (ChainedPhis.count(Phi)
1665 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1666 std::swap(OrigPhiRef, Phi);
1667 std::swap(OrigInc, IsomorphicInc);
1669 // Replacing the congruent phi is sufficient because acyclic redundancy
1670 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1671 // that a phi is congruent, it's often the head of an IV user cycle that
1672 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1673 // common case of a single IV increment so that DeleteDeadPHIs can remove
1674 // cycles that had postinc uses.
1675 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1676 IsomorphicInc->getType());
1677 if (OrigInc != IsomorphicInc
1678 && TruncExpr == SE.getSCEV(IsomorphicInc)
1679 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1680 || hoistIVInc(OrigInc, IsomorphicInc))) {
1681 DEBUG_WITH_TYPE(DebugType, dbgs()
1682 << "INDVARS: Eliminated congruent iv.inc: "
1683 << *IsomorphicInc << '\n');
1684 Value *NewInc = OrigInc;
1685 if (OrigInc->getType() != IsomorphicInc->getType()) {
1686 Instruction *IP = isa<PHINode>(OrigInc)
1687 ? (Instruction*)L->getHeader()->getFirstInsertionPt()
1688 : OrigInc->getNextNode();
1689 IRBuilder<> Builder(IP);
1690 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1692 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1694 IsomorphicInc->replaceAllUsesWith(NewInc);
1695 DeadInsts.push_back(IsomorphicInc);
1698 DEBUG_WITH_TYPE(DebugType, dbgs()
1699 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1701 Value *NewIV = OrigPhiRef;
1702 if (OrigPhiRef->getType() != Phi->getType()) {
1703 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1704 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1705 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1707 Phi->replaceAllUsesWith(NewIV);
1708 DeadInsts.push_back(Phi);
1714 // Search for a SCEV subexpression that is not safe to expand. Any expression
1715 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
1716 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
1717 // instruction, but the important thing is that we prove the denominator is
1718 // nonzero before expansion.
1720 // IVUsers already checks that IV-derived expressions are safe. So this check is
1721 // only needed when the expression includes some subexpression that is not IV
1724 // Currently, we only allow division by a nonzero constant here. If this is
1725 // inadequate, we could easily allow division by SCEVUnknown by using
1726 // ValueTracking to check isKnownNonZero().
1727 struct SCEVFindUnsafe {
1730 SCEVFindUnsafe(): IsUnsafe(false) {}
1732 bool follow(const SCEV *S) {
1733 const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S);
1736 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
1737 if (SC && !SC->getValue()->isZero())
1742 bool isDone() const { return IsUnsafe; }
1747 bool isSafeToExpand(const SCEV *S) {
1748 SCEVFindUnsafe Search;
1749 visitAll(S, Search);
1750 return !Search.IsUnsafe;