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 // FIXME: enable once our implementation of dominates is fixed.
47 // assert(BIP == IP || SE.DT->dominates(IP, BIP));
49 // Check to see if there is already a cast!
50 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
53 if (U->getType() == Ty)
54 if (CastInst *CI = dyn_cast<CastInst>(U))
55 if (CI->getOpcode() == Op) {
56 // If the cast isn't where we want it, create a new cast at IP.
57 // Likewise, do not reuse a cast at BIP because it must dominate
58 // instructions that might be inserted before BIP.
59 if (BasicBlock::iterator(CI) != IP || BIP == IP) {
60 // Create a new cast, and leave the old cast in place in case
61 // it is being used as an insert point. Clear its operand
62 // so that it doesn't hold anything live.
63 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", IP);
65 CI->replaceAllUsesWith(NewCI);
66 CI->setOperand(0, UndefValue::get(V->getType()));
67 rememberInstruction(NewCI);
70 rememberInstruction(CI);
76 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), IP);
77 rememberInstruction(I);
81 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
82 /// which must be possible with a noop cast, doing what we can to share
84 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
85 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
86 assert((Op == Instruction::BitCast ||
87 Op == Instruction::PtrToInt ||
88 Op == Instruction::IntToPtr) &&
89 "InsertNoopCastOfTo cannot perform non-noop casts!");
90 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
91 "InsertNoopCastOfTo cannot change sizes!");
93 // Short-circuit unnecessary bitcasts.
94 if (Op == Instruction::BitCast) {
95 if (V->getType() == Ty)
97 if (CastInst *CI = dyn_cast<CastInst>(V)) {
98 if (CI->getOperand(0)->getType() == Ty)
99 return CI->getOperand(0);
102 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
103 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
104 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
105 if (CastInst *CI = dyn_cast<CastInst>(V))
106 if ((CI->getOpcode() == Instruction::PtrToInt ||
107 CI->getOpcode() == Instruction::IntToPtr) &&
108 SE.getTypeSizeInBits(CI->getType()) ==
109 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
110 return CI->getOperand(0);
111 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
112 if ((CE->getOpcode() == Instruction::PtrToInt ||
113 CE->getOpcode() == Instruction::IntToPtr) &&
114 SE.getTypeSizeInBits(CE->getType()) ==
115 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
116 return CE->getOperand(0);
119 // Fold a cast of a constant.
120 if (Constant *C = dyn_cast<Constant>(V))
121 return ConstantExpr::getCast(Op, C, Ty);
123 // Cast the argument at the beginning of the entry block, after
124 // any bitcasts of other arguments.
125 if (Argument *A = dyn_cast<Argument>(V)) {
126 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
127 while ((isa<BitCastInst>(IP) &&
128 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
129 cast<BitCastInst>(IP)->getOperand(0) != A) ||
130 isa<DbgInfoIntrinsic>(IP) ||
131 isa<LandingPadInst>(IP))
133 return ReuseOrCreateCast(A, Ty, Op, IP);
136 // Cast the instruction immediately after the instruction.
137 Instruction *I = cast<Instruction>(V);
138 BasicBlock::iterator IP = I; ++IP;
139 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
140 IP = II->getNormalDest()->begin();
141 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
143 return ReuseOrCreateCast(I, Ty, Op, IP);
146 /// InsertBinop - Insert the specified binary operator, doing a small amount
147 /// of work to avoid inserting an obviously redundant operation.
148 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
149 Value *LHS, Value *RHS) {
150 // Fold a binop with constant operands.
151 if (Constant *CLHS = dyn_cast<Constant>(LHS))
152 if (Constant *CRHS = dyn_cast<Constant>(RHS))
153 return ConstantExpr::get(Opcode, CLHS, CRHS);
155 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
156 unsigned ScanLimit = 6;
157 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
158 // Scanning starts from the last instruction before the insertion point.
159 BasicBlock::iterator IP = Builder.GetInsertPoint();
160 if (IP != BlockBegin) {
162 for (; ScanLimit; --IP, --ScanLimit) {
163 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
165 if (isa<DbgInfoIntrinsic>(IP))
167 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
168 IP->getOperand(1) == RHS)
170 if (IP == BlockBegin) break;
174 // Save the original insertion point so we can restore it when we're done.
175 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
176 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
178 // Move the insertion point out of as many loops as we can.
179 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
180 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
181 BasicBlock *Preheader = L->getLoopPreheader();
182 if (!Preheader) break;
184 // Ok, move up a level.
185 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
188 // If we haven't found this binop, insert it.
189 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
190 BO->setDebugLoc(SaveInsertPt->getDebugLoc());
191 rememberInstruction(BO);
193 // Restore the original insert point.
195 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
200 /// FactorOutConstant - Test if S is divisible by Factor, using signed
201 /// division. If so, update S with Factor divided out and return true.
202 /// S need not be evenly divisible if a reasonable remainder can be
204 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
205 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
206 /// check to see if the divide was folded.
207 static bool FactorOutConstant(const SCEV *&S,
208 const SCEV *&Remainder,
211 const TargetData *TD) {
212 // Everything is divisible by one.
218 S = SE.getConstant(S->getType(), 1);
222 // For a Constant, check for a multiple of the given factor.
223 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
227 // Check for divisibility.
228 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
230 ConstantInt::get(SE.getContext(),
231 C->getValue()->getValue().sdiv(
232 FC->getValue()->getValue()));
233 // If the quotient is zero and the remainder is non-zero, reject
234 // the value at this scale. It will be considered for subsequent
237 const SCEV *Div = SE.getConstant(CI);
240 SE.getAddExpr(Remainder,
241 SE.getConstant(C->getValue()->getValue().srem(
242 FC->getValue()->getValue())));
248 // In a Mul, check if there is a constant operand which is a multiple
249 // of the given factor.
250 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
252 // With TargetData, the size is known. Check if there is a constant
253 // operand which is a multiple of the given factor. If so, we can
255 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
256 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
257 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
258 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
260 SE.getConstant(C->getValue()->getValue().sdiv(
261 FC->getValue()->getValue()));
262 S = SE.getMulExpr(NewMulOps);
266 // Without TargetData, check if Factor can be factored out of any of the
267 // Mul's operands. If so, we can just remove it.
268 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
269 const SCEV *SOp = M->getOperand(i);
270 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
271 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
272 Remainder->isZero()) {
273 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
275 S = SE.getMulExpr(NewMulOps);
282 // In an AddRec, check if both start and step are divisible.
283 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
284 const SCEV *Step = A->getStepRecurrence(SE);
285 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
286 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
288 if (!StepRem->isZero())
290 const SCEV *Start = A->getStart();
291 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
293 // FIXME: can use A->getNoWrapFlags(FlagNW)
294 S = SE.getAddRecExpr(Start, Step, A->getLoop(), SCEV::FlagAnyWrap);
301 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
302 /// is the number of SCEVAddRecExprs present, which are kept at the end of
305 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
307 ScalarEvolution &SE) {
308 unsigned NumAddRecs = 0;
309 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
311 // Group Ops into non-addrecs and addrecs.
312 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
313 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
314 // Let ScalarEvolution sort and simplify the non-addrecs list.
315 const SCEV *Sum = NoAddRecs.empty() ?
316 SE.getConstant(Ty, 0) :
317 SE.getAddExpr(NoAddRecs);
318 // If it returned an add, use the operands. Otherwise it simplified
319 // the sum into a single value, so just use that.
321 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
322 Ops.append(Add->op_begin(), Add->op_end());
323 else if (!Sum->isZero())
325 // Then append the addrecs.
326 Ops.append(AddRecs.begin(), AddRecs.end());
329 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
330 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
331 /// This helps expose more opportunities for folding parts of the expressions
332 /// into GEP indices.
334 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
336 ScalarEvolution &SE) {
338 SmallVector<const SCEV *, 8> AddRecs;
339 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
340 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
341 const SCEV *Start = A->getStart();
342 if (Start->isZero()) break;
343 const SCEV *Zero = SE.getConstant(Ty, 0);
344 AddRecs.push_back(SE.getAddRecExpr(Zero,
345 A->getStepRecurrence(SE),
347 // FIXME: A->getNoWrapFlags(FlagNW)
349 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
351 Ops.append(Add->op_begin(), Add->op_end());
352 e += Add->getNumOperands();
357 if (!AddRecs.empty()) {
358 // Add the addrecs onto the end of the list.
359 Ops.append(AddRecs.begin(), AddRecs.end());
360 // Resort the operand list, moving any constants to the front.
361 SimplifyAddOperands(Ops, Ty, SE);
365 /// expandAddToGEP - Expand an addition expression with a pointer type into
366 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
367 /// BasicAliasAnalysis and other passes analyze the result. See the rules
368 /// for getelementptr vs. inttoptr in
369 /// http://llvm.org/docs/LangRef.html#pointeraliasing
372 /// Design note: The correctness of using getelementptr here depends on
373 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
374 /// they may introduce pointer arithmetic which may not be safely converted
375 /// into getelementptr.
377 /// Design note: It might seem desirable for this function to be more
378 /// loop-aware. If some of the indices are loop-invariant while others
379 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
380 /// loop-invariant portions of the overall computation outside the loop.
381 /// However, there are a few reasons this is not done here. Hoisting simple
382 /// arithmetic is a low-level optimization that often isn't very
383 /// important until late in the optimization process. In fact, passes
384 /// like InstructionCombining will combine GEPs, even if it means
385 /// pushing loop-invariant computation down into loops, so even if the
386 /// GEPs were split here, the work would quickly be undone. The
387 /// LoopStrengthReduction pass, which is usually run quite late (and
388 /// after the last InstructionCombining pass), takes care of hoisting
389 /// loop-invariant portions of expressions, after considering what
390 /// can be folded using target addressing modes.
392 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
393 const SCEV *const *op_end,
397 Type *ElTy = PTy->getElementType();
398 SmallVector<Value *, 4> GepIndices;
399 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
400 bool AnyNonZeroIndices = false;
402 // Split AddRecs up into parts as either of the parts may be usable
403 // without the other.
404 SplitAddRecs(Ops, Ty, SE);
406 // Descend down the pointer's type and attempt to convert the other
407 // operands into GEP indices, at each level. The first index in a GEP
408 // indexes into the array implied by the pointer operand; the rest of
409 // the indices index into the element or field type selected by the
412 // If the scale size is not 0, attempt to factor out a scale for
414 SmallVector<const SCEV *, 8> ScaledOps;
415 if (ElTy->isSized()) {
416 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
417 if (!ElSize->isZero()) {
418 SmallVector<const SCEV *, 8> NewOps;
419 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
420 const SCEV *Op = Ops[i];
421 const SCEV *Remainder = SE.getConstant(Ty, 0);
422 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
423 // Op now has ElSize factored out.
424 ScaledOps.push_back(Op);
425 if (!Remainder->isZero())
426 NewOps.push_back(Remainder);
427 AnyNonZeroIndices = true;
429 // The operand was not divisible, so add it to the list of operands
430 // we'll scan next iteration.
431 NewOps.push_back(Ops[i]);
434 // If we made any changes, update Ops.
435 if (!ScaledOps.empty()) {
437 SimplifyAddOperands(Ops, Ty, SE);
442 // Record the scaled array index for this level of the type. If
443 // we didn't find any operands that could be factored, tentatively
444 // assume that element zero was selected (since the zero offset
445 // would obviously be folded away).
446 Value *Scaled = ScaledOps.empty() ?
447 Constant::getNullValue(Ty) :
448 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
449 GepIndices.push_back(Scaled);
451 // Collect struct field index operands.
452 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
453 bool FoundFieldNo = false;
454 // An empty struct has no fields.
455 if (STy->getNumElements() == 0) break;
457 // With TargetData, field offsets are known. See if a constant offset
458 // falls within any of the struct fields.
459 if (Ops.empty()) break;
460 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
461 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
462 const StructLayout &SL = *SE.TD->getStructLayout(STy);
463 uint64_t FullOffset = C->getValue()->getZExtValue();
464 if (FullOffset < SL.getSizeInBytes()) {
465 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
466 GepIndices.push_back(
467 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
468 ElTy = STy->getTypeAtIndex(ElIdx);
470 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
471 AnyNonZeroIndices = true;
476 // Without TargetData, just check for an offsetof expression of the
477 // appropriate struct type.
478 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
479 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
482 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
483 GepIndices.push_back(FieldNo);
485 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
486 Ops[i] = SE.getConstant(Ty, 0);
487 AnyNonZeroIndices = true;
493 // If no struct field offsets were found, tentatively assume that
494 // field zero was selected (since the zero offset would obviously
497 ElTy = STy->getTypeAtIndex(0u);
498 GepIndices.push_back(
499 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
503 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
504 ElTy = ATy->getElementType();
509 // If none of the operands were convertible to proper GEP indices, cast
510 // the base to i8* and do an ugly getelementptr with that. It's still
511 // better than ptrtoint+arithmetic+inttoptr at least.
512 if (!AnyNonZeroIndices) {
513 // Cast the base to i8*.
514 V = InsertNoopCastOfTo(V,
515 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
517 assert(!isa<Instruction>(V) ||
518 SE.DT->properlyDominates(cast<Instruction>(V),
519 Builder.GetInsertPoint()));
521 // Expand the operands for a plain byte offset.
522 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
524 // Fold a GEP with constant operands.
525 if (Constant *CLHS = dyn_cast<Constant>(V))
526 if (Constant *CRHS = dyn_cast<Constant>(Idx))
527 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
529 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
530 unsigned ScanLimit = 6;
531 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
532 // Scanning starts from the last instruction before the insertion point.
533 BasicBlock::iterator IP = Builder.GetInsertPoint();
534 if (IP != BlockBegin) {
536 for (; ScanLimit; --IP, --ScanLimit) {
537 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
539 if (isa<DbgInfoIntrinsic>(IP))
541 if (IP->getOpcode() == Instruction::GetElementPtr &&
542 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
544 if (IP == BlockBegin) break;
548 // Save the original insertion point so we can restore it when we're done.
549 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
550 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
552 // Move the insertion point out of as many loops as we can.
553 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
554 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
555 BasicBlock *Preheader = L->getLoopPreheader();
556 if (!Preheader) break;
558 // Ok, move up a level.
559 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
563 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
564 rememberInstruction(GEP);
566 // Restore the original insert point.
568 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
573 // Save the original insertion point so we can restore it when we're done.
574 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
575 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
577 // Move the insertion point out of as many loops as we can.
578 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
579 if (!L->isLoopInvariant(V)) break;
581 bool AnyIndexNotLoopInvariant = false;
582 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
583 E = GepIndices.end(); I != E; ++I)
584 if (!L->isLoopInvariant(*I)) {
585 AnyIndexNotLoopInvariant = true;
588 if (AnyIndexNotLoopInvariant)
591 BasicBlock *Preheader = L->getLoopPreheader();
592 if (!Preheader) break;
594 // Ok, move up a level.
595 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
598 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
599 // because ScalarEvolution may have changed the address arithmetic to
600 // compute a value which is beyond the end of the allocated object.
602 if (V->getType() != PTy)
603 Casted = InsertNoopCastOfTo(Casted, PTy);
604 Value *GEP = Builder.CreateGEP(Casted,
607 Ops.push_back(SE.getUnknown(GEP));
608 rememberInstruction(GEP);
610 // Restore the original insert point.
612 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
614 return expand(SE.getAddExpr(Ops));
617 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
618 /// SCEV expansion. If they are nested, this is the most nested. If they are
619 /// neighboring, pick the later.
620 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
624 if (A->contains(B)) return B;
625 if (B->contains(A)) return A;
626 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
627 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
628 return A; // Arbitrarily break the tie.
631 /// getRelevantLoop - Get the most relevant loop associated with the given
632 /// expression, according to PickMostRelevantLoop.
633 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
634 // Test whether we've already computed the most relevant loop for this SCEV.
635 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
636 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
638 return Pair.first->second;
640 if (isa<SCEVConstant>(S))
641 // A constant has no relevant loops.
643 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
644 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
645 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
646 // A non-instruction has no relevant loops.
649 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
651 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
653 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
655 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
656 return RelevantLoops[N] = L;
658 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
659 const Loop *Result = getRelevantLoop(C->getOperand());
660 return RelevantLoops[C] = Result;
662 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
664 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
665 getRelevantLoop(D->getRHS()),
667 return RelevantLoops[D] = Result;
669 llvm_unreachable("Unexpected SCEV type!");
674 /// LoopCompare - Compare loops by PickMostRelevantLoop.
678 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
680 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
681 std::pair<const Loop *, const SCEV *> RHS) const {
682 // Keep pointer operands sorted at the end.
683 if (LHS.second->getType()->isPointerTy() !=
684 RHS.second->getType()->isPointerTy())
685 return LHS.second->getType()->isPointerTy();
687 // Compare loops with PickMostRelevantLoop.
688 if (LHS.first != RHS.first)
689 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
691 // If one operand is a non-constant negative and the other is not,
692 // put the non-constant negative on the right so that a sub can
693 // be used instead of a negate and add.
694 if (LHS.second->isNonConstantNegative()) {
695 if (!RHS.second->isNonConstantNegative())
697 } else if (RHS.second->isNonConstantNegative())
700 // Otherwise they are equivalent according to this comparison.
707 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
708 Type *Ty = SE.getEffectiveSCEVType(S->getType());
710 // Collect all the add operands in a loop, along with their associated loops.
711 // Iterate in reverse so that constants are emitted last, all else equal, and
712 // so that pointer operands are inserted first, which the code below relies on
713 // to form more involved GEPs.
714 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
715 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
716 E(S->op_begin()); I != E; ++I)
717 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
719 // Sort by loop. Use a stable sort so that constants follow non-constants and
720 // pointer operands precede non-pointer operands.
721 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
723 // Emit instructions to add all the operands. Hoist as much as possible
724 // out of loops, and form meaningful getelementptrs where possible.
726 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
727 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
728 const Loop *CurLoop = I->first;
729 const SCEV *Op = I->second;
731 // This is the first operand. Just expand it.
734 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
735 // The running sum expression is a pointer. Try to form a getelementptr
736 // at this level with that as the base.
737 SmallVector<const SCEV *, 4> NewOps;
738 for (; I != E && I->first == CurLoop; ++I) {
739 // If the operand is SCEVUnknown and not instructions, peek through
740 // it, to enable more of it to be folded into the GEP.
741 const SCEV *X = I->second;
742 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
743 if (!isa<Instruction>(U->getValue()))
744 X = SE.getSCEV(U->getValue());
747 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
748 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
749 // The running sum is an integer, and there's a pointer at this level.
750 // Try to form a getelementptr. If the running sum is instructions,
751 // use a SCEVUnknown to avoid re-analyzing them.
752 SmallVector<const SCEV *, 4> NewOps;
753 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
755 for (++I; I != E && I->first == CurLoop; ++I)
756 NewOps.push_back(I->second);
757 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
758 } else if (Op->isNonConstantNegative()) {
759 // Instead of doing a negate and add, just do a subtract.
760 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
761 Sum = InsertNoopCastOfTo(Sum, Ty);
762 Sum = InsertBinop(Instruction::Sub, Sum, W);
766 Value *W = expandCodeFor(Op, Ty);
767 Sum = InsertNoopCastOfTo(Sum, Ty);
768 // Canonicalize a constant to the RHS.
769 if (isa<Constant>(Sum)) std::swap(Sum, W);
770 Sum = InsertBinop(Instruction::Add, Sum, W);
778 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
779 Type *Ty = SE.getEffectiveSCEVType(S->getType());
781 // Collect all the mul operands in a loop, along with their associated loops.
782 // Iterate in reverse so that constants are emitted last, all else equal.
783 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
784 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
785 E(S->op_begin()); I != E; ++I)
786 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
788 // Sort by loop. Use a stable sort so that constants follow non-constants.
789 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
791 // Emit instructions to mul all the operands. Hoist as much as possible
794 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
795 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
796 const SCEV *Op = I->second;
798 // This is the first operand. Just expand it.
801 } else if (Op->isAllOnesValue()) {
802 // Instead of doing a multiply by negative one, just do a negate.
803 Prod = InsertNoopCastOfTo(Prod, Ty);
804 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
808 Value *W = expandCodeFor(Op, Ty);
809 Prod = InsertNoopCastOfTo(Prod, Ty);
810 // Canonicalize a constant to the RHS.
811 if (isa<Constant>(Prod)) std::swap(Prod, W);
812 Prod = InsertBinop(Instruction::Mul, Prod, W);
820 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
821 Type *Ty = SE.getEffectiveSCEVType(S->getType());
823 Value *LHS = expandCodeFor(S->getLHS(), Ty);
824 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
825 const APInt &RHS = SC->getValue()->getValue();
826 if (RHS.isPowerOf2())
827 return InsertBinop(Instruction::LShr, LHS,
828 ConstantInt::get(Ty, RHS.logBase2()));
831 Value *RHS = expandCodeFor(S->getRHS(), Ty);
832 return InsertBinop(Instruction::UDiv, LHS, RHS);
835 /// Move parts of Base into Rest to leave Base with the minimal
836 /// expression that provides a pointer operand suitable for a
838 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
839 ScalarEvolution &SE) {
840 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
841 Base = A->getStart();
842 Rest = SE.getAddExpr(Rest,
843 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
844 A->getStepRecurrence(SE),
846 // FIXME: A->getNoWrapFlags(FlagNW)
849 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
850 Base = A->getOperand(A->getNumOperands()-1);
851 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
852 NewAddOps.back() = Rest;
853 Rest = SE.getAddExpr(NewAddOps);
854 ExposePointerBase(Base, Rest, SE);
858 /// Determine if this is a well-behaved chain of instructions leading back to
859 /// the PHI. If so, it may be reused by expanded expressions.
860 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
862 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
863 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
865 // If any of the operands don't dominate the insert position, bail.
866 // Addrec operands are always loop-invariant, so this can only happen
867 // if there are instructions which haven't been hoisted.
868 if (L == IVIncInsertLoop) {
869 for (User::op_iterator OI = IncV->op_begin()+1,
870 OE = IncV->op_end(); OI != OE; ++OI)
871 if (Instruction *OInst = dyn_cast<Instruction>(OI))
872 if (!SE.DT->dominates(OInst, IVIncInsertPos))
875 // Advance to the next instruction.
876 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
880 if (IncV->mayHaveSideEffects())
886 return isNormalAddRecExprPHI(PN, IncV, L);
889 /// getIVIncOperand returns an induction variable increment's induction
890 /// variable operand.
892 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
893 /// operands dominate InsertPos.
895 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
896 /// simple patterns generated by getAddRecExprPHILiterally and
897 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
898 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
899 Instruction *InsertPos,
901 if (IncV == InsertPos)
904 switch (IncV->getOpcode()) {
907 // Check for a simple Add/Sub or GEP of a loop invariant step.
908 case Instruction::Add:
909 case Instruction::Sub: {
910 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
911 if (!OInst || SE.DT->properlyDominates(OInst, InsertPos))
912 return dyn_cast<Instruction>(IncV->getOperand(0));
915 case Instruction::BitCast:
916 return dyn_cast<Instruction>(IncV->getOperand(0));
917 case Instruction::GetElementPtr:
918 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
920 if (isa<Constant>(*I))
922 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
923 if (!SE.DT->properlyDominates(OInst, InsertPos))
927 // allow any kind of GEP as long as it can be hoisted.
930 // This must be a pointer addition of constants (pretty), which is already
931 // handled, or some number of address-size elements (ugly). Ugly geps
932 // have 2 operands. i1* is used by the expander to represent an
933 // address-size element.
934 if (IncV->getNumOperands() != 2)
936 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
937 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
938 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
942 return dyn_cast<Instruction>(IncV->getOperand(0));
946 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
947 /// it available to other uses in this loop. Recursively hoist any operands,
948 /// until we reach a value that dominates InsertPos.
949 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
950 if (SE.DT->properlyDominates(IncV, InsertPos))
953 // InsertPos must itself dominate IncV so that IncV's new position satisfies
954 // its existing users.
955 if (!SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
958 // Check that the chain of IV operands leading back to Phi can be hoisted.
959 SmallVector<Instruction*, 4> IVIncs;
961 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
964 // IncV is safe to hoist.
965 IVIncs.push_back(IncV);
967 if (SE.DT->properlyDominates(IncV, InsertPos))
970 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
971 E = IVIncs.rend(); I != E; ++I) {
972 (*I)->moveBefore(InsertPos);
977 /// Determine if this cyclic phi is in a form that would have been generated by
978 /// LSR. We don't care if the phi was actually expanded in this pass, as long
979 /// as it is in a low-cost form, for example, no implied multiplication. This
980 /// should match any patterns generated by getAddRecExprPHILiterally and
982 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
984 for(Instruction *IVOper = IncV;
985 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
986 /*allowScale=*/false));) {
993 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
994 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
995 /// need to materialize IV increments elsewhere to handle difficult situations.
996 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
997 Type *ExpandTy, Type *IntTy,
1000 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
1001 if (ExpandTy->isPointerTy()) {
1002 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
1003 // If the step isn't constant, don't use an implicitly scaled GEP, because
1004 // that would require a multiply inside the loop.
1005 if (!isa<ConstantInt>(StepV))
1006 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
1007 GEPPtrTy->getAddressSpace());
1008 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
1009 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
1010 if (IncV->getType() != PN->getType()) {
1011 IncV = Builder.CreateBitCast(IncV, PN->getType());
1012 rememberInstruction(IncV);
1015 IncV = useSubtract ?
1016 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1017 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1018 rememberInstruction(IncV);
1023 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1024 /// the base addrec, which is the addrec without any non-loop-dominating
1025 /// values, and return the PHI.
1027 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1031 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1033 // Reuse a previously-inserted PHI, if present.
1034 BasicBlock *LatchBlock = L->getLoopLatch();
1036 for (BasicBlock::iterator I = L->getHeader()->begin();
1037 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1038 if (!SE.isSCEVable(PN->getType()) ||
1039 (SE.getEffectiveSCEVType(PN->getType()) !=
1040 SE.getEffectiveSCEVType(Normalized->getType())) ||
1041 SE.getSCEV(PN) != Normalized)
1045 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1048 if (!isExpandedAddRecExprPHI(PN, IncV, L))
1050 if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos))
1054 if (!isNormalAddRecExprPHI(PN, IncV, L))
1056 if (L == IVIncInsertLoop)
1058 if (SE.DT->dominates(IncV, IVIncInsertPos))
1060 // Make sure the increment is where we want it. But don't move it
1061 // down past a potential existing post-inc user.
1062 IncV->moveBefore(IVIncInsertPos);
1063 IVIncInsertPos = IncV;
1064 IncV = cast<Instruction>(IncV->getOperand(0));
1065 } while (IncV != PN);
1067 // Ok, the add recurrence looks usable.
1068 // Remember this PHI, even in post-inc mode.
1069 InsertedValues.insert(PN);
1070 // Remember the increment.
1071 rememberInstruction(IncV);
1076 // Save the original insertion point so we can restore it when we're done.
1077 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1078 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1080 // Another AddRec may need to be recursively expanded below. For example, if
1081 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1082 // loop. Remove this loop from the PostIncLoops set before expanding such
1083 // AddRecs. Otherwise, we cannot find a valid position for the step
1084 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1085 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1086 // so it's not worth implementing SmallPtrSet::swap.
1087 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1088 PostIncLoops.clear();
1090 // Expand code for the start value.
1091 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1092 L->getHeader()->begin());
1094 // StartV must be hoisted into L's preheader to dominate the new phi.
1095 assert(!isa<Instruction>(StartV) ||
1096 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1099 // Expand code for the step value. Do this before creating the PHI so that PHI
1100 // reuse code doesn't see an incomplete PHI.
1101 const SCEV *Step = Normalized->getStepRecurrence(SE);
1102 // If the stride is negative, insert a sub instead of an add for the increment
1103 // (unless it's a constant, because subtracts of constants are canonicalized
1105 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1107 Step = SE.getNegativeSCEV(Step);
1108 // Expand the step somewhere that dominates the loop header.
1109 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1112 BasicBlock *Header = L->getHeader();
1113 Builder.SetInsertPoint(Header, Header->begin());
1114 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1115 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1116 Twine(IVName) + ".iv");
1117 rememberInstruction(PN);
1119 // Create the step instructions and populate the PHI.
1120 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1121 BasicBlock *Pred = *HPI;
1123 // Add a start value.
1124 if (!L->contains(Pred)) {
1125 PN->addIncoming(StartV, Pred);
1129 // Create a step value and add it to the PHI.
1130 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1131 // instructions at IVIncInsertPos.
1132 Instruction *InsertPos = L == IVIncInsertLoop ?
1133 IVIncInsertPos : Pred->getTerminator();
1134 Builder.SetInsertPoint(InsertPos);
1135 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1137 PN->addIncoming(IncV, Pred);
1140 // Restore the original insert point.
1142 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1144 // After expanding subexpressions, restore the PostIncLoops set so the caller
1145 // can ensure that IVIncrement dominates the current uses.
1146 PostIncLoops = SavedPostIncLoops;
1148 // Remember this PHI, even in post-inc mode.
1149 InsertedValues.insert(PN);
1154 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1155 Type *STy = S->getType();
1156 Type *IntTy = SE.getEffectiveSCEVType(STy);
1157 const Loop *L = S->getLoop();
1159 // Determine a normalized form of this expression, which is the expression
1160 // before any post-inc adjustment is made.
1161 const SCEVAddRecExpr *Normalized = S;
1162 if (PostIncLoops.count(L)) {
1163 PostIncLoopSet Loops;
1166 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1167 Loops, SE, *SE.DT));
1170 // Strip off any non-loop-dominating component from the addrec start.
1171 const SCEV *Start = Normalized->getStart();
1172 const SCEV *PostLoopOffset = 0;
1173 if (!SE.properlyDominates(Start, L->getHeader())) {
1174 PostLoopOffset = Start;
1175 Start = SE.getConstant(Normalized->getType(), 0);
1176 Normalized = cast<SCEVAddRecExpr>(
1177 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1178 Normalized->getLoop(),
1179 // FIXME: Normalized->getNoWrapFlags(FlagNW)
1180 SCEV::FlagAnyWrap));
1183 // Strip off any non-loop-dominating component from the addrec step.
1184 const SCEV *Step = Normalized->getStepRecurrence(SE);
1185 const SCEV *PostLoopScale = 0;
1186 if (!SE.dominates(Step, L->getHeader())) {
1187 PostLoopScale = Step;
1188 Step = SE.getConstant(Normalized->getType(), 1);
1190 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
1191 Normalized->getLoop(),
1192 // FIXME: Normalized
1193 // ->getNoWrapFlags(FlagNW)
1194 SCEV::FlagAnyWrap));
1197 // Expand the core addrec. If we need post-loop scaling, force it to
1198 // expand to an integer type to avoid the need for additional casting.
1199 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1200 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1202 // Accommodate post-inc mode, if necessary.
1204 if (!PostIncLoops.count(L))
1207 // In PostInc mode, use the post-incremented value.
1208 BasicBlock *LatchBlock = L->getLoopLatch();
1209 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1210 Result = PN->getIncomingValueForBlock(LatchBlock);
1212 // For an expansion to use the postinc form, the client must call
1213 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1214 // or dominated by IVIncInsertPos.
1215 if (isa<Instruction>(Result)
1216 && !SE.DT->dominates(cast<Instruction>(Result),
1217 Builder.GetInsertPoint())) {
1218 // The induction variable's postinc expansion does not dominate this use.
1219 // IVUsers tries to prevent this case, so it is rare. However, it can
1220 // happen when an IVUser outside the loop is not dominated by the latch
1221 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1222 // all cases. Consider a phi outide whose operand is replaced during
1223 // expansion with the value of the postinc user. Without fundamentally
1224 // changing the way postinc users are tracked, the only remedy is
1225 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1226 // but hopefully expandCodeFor handles that.
1228 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1230 Step = SE.getNegativeSCEV(Step);
1231 // Expand the step somewhere that dominates the loop header.
1232 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1233 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1234 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1235 // Restore the insertion point to the place where the caller has
1236 // determined dominates all uses.
1237 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1238 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1242 // Re-apply any non-loop-dominating scale.
1243 if (PostLoopScale) {
1244 Result = InsertNoopCastOfTo(Result, IntTy);
1245 Result = Builder.CreateMul(Result,
1246 expandCodeFor(PostLoopScale, IntTy));
1247 rememberInstruction(Result);
1250 // Re-apply any non-loop-dominating offset.
1251 if (PostLoopOffset) {
1252 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1253 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1254 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1256 Result = InsertNoopCastOfTo(Result, IntTy);
1257 Result = Builder.CreateAdd(Result,
1258 expandCodeFor(PostLoopOffset, IntTy));
1259 rememberInstruction(Result);
1266 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1267 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1269 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1270 const Loop *L = S->getLoop();
1272 // First check for an existing canonical IV in a suitable type.
1273 PHINode *CanonicalIV = 0;
1274 if (PHINode *PN = L->getCanonicalInductionVariable())
1275 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1278 // Rewrite an AddRec in terms of the canonical induction variable, if
1279 // its type is more narrow.
1281 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1282 SE.getTypeSizeInBits(Ty)) {
1283 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1284 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1285 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1286 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1287 // FIXME: S->getNoWrapFlags(FlagNW)
1288 SCEV::FlagAnyWrap));
1289 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1290 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1291 BasicBlock::iterator NewInsertPt =
1292 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1293 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1294 isa<LandingPadInst>(NewInsertPt))
1296 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1298 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1302 // {X,+,F} --> X + {0,+,F}
1303 if (!S->getStart()->isZero()) {
1304 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1305 NewOps[0] = SE.getConstant(Ty, 0);
1306 // FIXME: can use S->getNoWrapFlags()
1307 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, SCEV::FlagAnyWrap);
1309 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1310 // comments on expandAddToGEP for details.
1311 const SCEV *Base = S->getStart();
1312 const SCEV *RestArray[1] = { Rest };
1313 // Dig into the expression to find the pointer base for a GEP.
1314 ExposePointerBase(Base, RestArray[0], SE);
1315 // If we found a pointer, expand the AddRec with a GEP.
1316 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1317 // Make sure the Base isn't something exotic, such as a multiplied
1318 // or divided pointer value. In those cases, the result type isn't
1319 // actually a pointer type.
1320 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1321 Value *StartV = expand(Base);
1322 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1323 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1327 // Just do a normal add. Pre-expand the operands to suppress folding.
1328 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1329 SE.getUnknown(expand(Rest))));
1332 // If we don't yet have a canonical IV, create one.
1334 // Create and insert the PHI node for the induction variable in the
1336 BasicBlock *Header = L->getHeader();
1337 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1338 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1340 rememberInstruction(CanonicalIV);
1342 Constant *One = ConstantInt::get(Ty, 1);
1343 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1344 BasicBlock *HP = *HPI;
1345 if (L->contains(HP)) {
1346 // Insert a unit add instruction right before the terminator
1347 // corresponding to the back-edge.
1348 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1350 HP->getTerminator());
1351 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1352 rememberInstruction(Add);
1353 CanonicalIV->addIncoming(Add, HP);
1355 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1360 // {0,+,1} --> Insert a canonical induction variable into the loop!
1361 if (S->isAffine() && S->getOperand(1)->isOne()) {
1362 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1363 "IVs with types different from the canonical IV should "
1364 "already have been handled!");
1368 // {0,+,F} --> {0,+,1} * F
1370 // If this is a simple linear addrec, emit it now as a special case.
1371 if (S->isAffine()) // {0,+,F} --> i*F
1373 expand(SE.getTruncateOrNoop(
1374 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1375 SE.getNoopOrAnyExtend(S->getOperand(1),
1376 CanonicalIV->getType())),
1379 // If this is a chain of recurrences, turn it into a closed form, using the
1380 // folders, then expandCodeFor the closed form. This allows the folders to
1381 // simplify the expression without having to build a bunch of special code
1382 // into this folder.
1383 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1385 // Promote S up to the canonical IV type, if the cast is foldable.
1386 const SCEV *NewS = S;
1387 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1388 if (isa<SCEVAddRecExpr>(Ext))
1391 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1392 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1394 // Truncate the result down to the original type, if needed.
1395 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1399 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1400 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1401 Value *V = expandCodeFor(S->getOperand(),
1402 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1403 Value *I = Builder.CreateTrunc(V, Ty);
1404 rememberInstruction(I);
1408 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1409 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1410 Value *V = expandCodeFor(S->getOperand(),
1411 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1412 Value *I = Builder.CreateZExt(V, Ty);
1413 rememberInstruction(I);
1417 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1418 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1419 Value *V = expandCodeFor(S->getOperand(),
1420 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1421 Value *I = Builder.CreateSExt(V, Ty);
1422 rememberInstruction(I);
1426 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1427 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1428 Type *Ty = LHS->getType();
1429 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1430 // In the case of mixed integer and pointer types, do the
1431 // rest of the comparisons as integer.
1432 if (S->getOperand(i)->getType() != Ty) {
1433 Ty = SE.getEffectiveSCEVType(Ty);
1434 LHS = InsertNoopCastOfTo(LHS, Ty);
1436 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1437 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1438 rememberInstruction(ICmp);
1439 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1440 rememberInstruction(Sel);
1443 // In the case of mixed integer and pointer types, cast the
1444 // final result back to the pointer type.
1445 if (LHS->getType() != S->getType())
1446 LHS = InsertNoopCastOfTo(LHS, S->getType());
1450 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1451 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1452 Type *Ty = LHS->getType();
1453 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1454 // In the case of mixed integer and pointer types, do the
1455 // rest of the comparisons as integer.
1456 if (S->getOperand(i)->getType() != Ty) {
1457 Ty = SE.getEffectiveSCEVType(Ty);
1458 LHS = InsertNoopCastOfTo(LHS, Ty);
1460 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1461 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1462 rememberInstruction(ICmp);
1463 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1464 rememberInstruction(Sel);
1467 // In the case of mixed integer and pointer types, cast the
1468 // final result back to the pointer type.
1469 if (LHS->getType() != S->getType())
1470 LHS = InsertNoopCastOfTo(LHS, S->getType());
1474 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1476 Builder.SetInsertPoint(IP->getParent(), IP);
1477 return expandCodeFor(SH, Ty);
1480 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1481 // Expand the code for this SCEV.
1482 Value *V = expand(SH);
1484 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1485 "non-trivial casts should be done with the SCEVs directly!");
1486 V = InsertNoopCastOfTo(V, Ty);
1491 Value *SCEVExpander::expand(const SCEV *S) {
1492 // Compute an insertion point for this SCEV object. Hoist the instructions
1493 // as far out in the loop nest as possible.
1494 Instruction *InsertPt = Builder.GetInsertPoint();
1495 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1496 L = L->getParentLoop())
1497 if (SE.isLoopInvariant(S, L)) {
1499 if (BasicBlock *Preheader = L->getLoopPreheader())
1500 InsertPt = Preheader->getTerminator();
1502 // LSR sets the insertion point for AddRec start/step values to the
1503 // block start to simplify value reuse, even though it's an invalid
1504 // position. SCEVExpander must correct for this in all cases.
1505 InsertPt = L->getHeader()->getFirstInsertionPt();
1508 // If the SCEV is computable at this level, insert it into the header
1509 // after the PHIs (and after any other instructions that we've inserted
1510 // there) so that it is guaranteed to dominate any user inside the loop.
1511 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1512 InsertPt = L->getHeader()->getFirstInsertionPt();
1513 while (InsertPt != Builder.GetInsertPoint()
1514 && (isInsertedInstruction(InsertPt)
1515 || isa<DbgInfoIntrinsic>(InsertPt))) {
1516 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1521 // Check to see if we already expanded this here.
1522 std::map<std::pair<const SCEV *, Instruction *>,
1523 AssertingVH<Value> >::iterator I =
1524 InsertedExpressions.find(std::make_pair(S, InsertPt));
1525 if (I != InsertedExpressions.end())
1528 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1529 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1530 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1532 // Expand the expression into instructions.
1533 Value *V = visit(S);
1535 // Remember the expanded value for this SCEV at this location.
1537 // This is independent of PostIncLoops. The mapped value simply materializes
1538 // the expression at this insertion point. If the mapped value happened to be
1539 // a postinc expansion, it could be reused by a non postinc user, but only if
1540 // its insertion point was already at the head of the loop.
1541 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1543 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1547 void SCEVExpander::rememberInstruction(Value *I) {
1548 if (!PostIncLoops.empty())
1549 InsertedPostIncValues.insert(I);
1551 InsertedValues.insert(I);
1554 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1555 Builder.SetInsertPoint(BB, I);
1558 /// getOrInsertCanonicalInductionVariable - This method returns the
1559 /// canonical induction variable of the specified type for the specified
1560 /// loop (inserting one if there is none). A canonical induction variable
1561 /// starts at zero and steps by one on each iteration.
1563 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1565 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1567 // Build a SCEV for {0,+,1}<L>.
1568 // Conservatively use FlagAnyWrap for now.
1569 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1570 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1572 // Emit code for it.
1573 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1574 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1575 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1577 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1582 /// Sort values by integer width for replaceCongruentIVs.
1583 static bool width_descending(Value *lhs, Value *rhs) {
1584 // Put pointers at the back and make sure pointer < pointer = false.
1585 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy())
1586 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy();
1587 return rhs->getType()->getPrimitiveSizeInBits()
1588 < lhs->getType()->getPrimitiveSizeInBits();
1591 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1592 /// replace them with their most canonical representative. Return the number of
1593 /// phis eliminated.
1595 /// This does not depend on any SCEVExpander state but should be used in
1596 /// the same context that SCEVExpander is used.
1597 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1598 SmallVectorImpl<WeakVH> &DeadInsts,
1599 const TargetLowering *TLI) {
1600 // Find integer phis in order of increasing width.
1601 SmallVector<PHINode*, 8> Phis;
1602 for (BasicBlock::iterator I = L->getHeader()->begin();
1603 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1604 Phis.push_back(Phi);
1607 std::sort(Phis.begin(), Phis.end(), width_descending);
1609 unsigned NumElim = 0;
1610 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1611 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1612 // so narrow phis can reuse them.
1613 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1614 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1615 PHINode *Phi = *PIter;
1617 if (!SE.isSCEVable(Phi->getType()))
1620 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1623 if (Phi->getType()->isIntegerTy() && TLI
1624 && TLI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1625 // This phi can be freely truncated to the narrowest phi type. Map the
1626 // truncated expression to it so it will be reused for narrow types.
1627 const SCEV *TruncExpr =
1628 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1629 ExprToIVMap[TruncExpr] = Phi;
1634 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1636 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1639 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1640 Instruction *OrigInc =
1641 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1642 Instruction *IsomorphicInc =
1643 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1645 // If this phi has the same width but is more canonical, replace the
1646 // original with it. As part of the "more canonical" determination,
1647 // respect a prior decision to use an IV chain.
1648 if (OrigPhiRef->getType() == Phi->getType()
1649 && !(ChainedPhis.count(Phi)
1650 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1651 && (ChainedPhis.count(Phi)
1652 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1653 std::swap(OrigPhiRef, Phi);
1654 std::swap(OrigInc, IsomorphicInc);
1656 // Replacing the congruent phi is sufficient because acyclic redundancy
1657 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1658 // that a phi is congruent, it's often the head of an IV user cycle that
1659 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1660 // common case of a single IV increment so that DeleteDeadPHIs can remove
1661 // cycles that had postinc uses.
1662 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1663 IsomorphicInc->getType());
1664 if (OrigInc != IsomorphicInc
1665 && TruncExpr == SE.getSCEV(IsomorphicInc)
1666 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1667 || hoistIVInc(OrigInc, IsomorphicInc))) {
1668 DEBUG_WITH_TYPE(DebugType, dbgs()
1669 << "INDVARS: Eliminated congruent iv.inc: "
1670 << *IsomorphicInc << '\n');
1671 Value *NewInc = OrigInc;
1672 if (OrigInc->getType() != IsomorphicInc->getType()) {
1673 Instruction *IP = isa<PHINode>(OrigInc)
1674 ? (Instruction*)L->getHeader()->getFirstInsertionPt()
1675 : OrigInc->getNextNode();
1676 IRBuilder<> Builder(IP);
1677 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1679 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1681 IsomorphicInc->replaceAllUsesWith(NewInc);
1682 DeadInsts.push_back(IsomorphicInc);
1685 DEBUG_WITH_TYPE(DebugType, dbgs()
1686 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1688 Value *NewIV = OrigPhiRef;
1689 if (OrigPhiRef->getType() != Phi->getType()) {
1690 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1691 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1692 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1694 Phi->replaceAllUsesWith(NewIV);
1695 DeadInsts.push_back(Phi);