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/LLVMContext.h"
19 #include "llvm/Target/TargetData.h"
20 #include "llvm/ADT/STLExtras.h"
23 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
24 /// which must be possible with a noop cast, doing what we can to share
26 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
27 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
28 assert((Op == Instruction::BitCast ||
29 Op == Instruction::PtrToInt ||
30 Op == Instruction::IntToPtr) &&
31 "InsertNoopCastOfTo cannot perform non-noop casts!");
32 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
33 "InsertNoopCastOfTo cannot change sizes!");
35 // Short-circuit unnecessary bitcasts.
36 if (Op == Instruction::BitCast && V->getType() == Ty)
39 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
40 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
41 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
42 if (CastInst *CI = dyn_cast<CastInst>(V))
43 if ((CI->getOpcode() == Instruction::PtrToInt ||
44 CI->getOpcode() == Instruction::IntToPtr) &&
45 SE.getTypeSizeInBits(CI->getType()) ==
46 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
47 return CI->getOperand(0);
48 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
49 if ((CE->getOpcode() == Instruction::PtrToInt ||
50 CE->getOpcode() == Instruction::IntToPtr) &&
51 SE.getTypeSizeInBits(CE->getType()) ==
52 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
53 return CE->getOperand(0);
56 if (Constant *C = dyn_cast<Constant>(V))
57 return ConstantExpr::getCast(Op, C, Ty);
59 if (Argument *A = dyn_cast<Argument>(V)) {
60 // Check to see if there is already a cast!
61 for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
63 if ((*UI)->getType() == Ty)
64 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
65 if (CI->getOpcode() == Op) {
66 // If the cast isn't the first instruction of the function, move it.
67 if (BasicBlock::iterator(CI) !=
68 A->getParent()->getEntryBlock().begin()) {
69 // Recreate the cast at the beginning of the entry block.
70 // The old cast is left in place in case it is being used
71 // as an insert point.
73 CastInst::Create(Op, V, Ty, "",
74 A->getParent()->getEntryBlock().begin());
76 CI->replaceAllUsesWith(NewCI);
82 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(),
83 A->getParent()->getEntryBlock().begin());
84 rememberInstruction(I);
88 Instruction *I = cast<Instruction>(V);
90 // Check to see if there is already a cast. If there is, use it.
91 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
93 if ((*UI)->getType() == Ty)
94 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
95 if (CI->getOpcode() == Op) {
96 BasicBlock::iterator It = I; ++It;
97 if (isa<InvokeInst>(I))
98 It = cast<InvokeInst>(I)->getNormalDest()->begin();
99 while (isa<PHINode>(It)) ++It;
100 if (It != BasicBlock::iterator(CI)) {
101 // Recreate the cast after the user.
102 // The old cast is left in place in case it is being used
103 // as an insert point.
104 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It);
106 CI->replaceAllUsesWith(NewCI);
107 rememberInstruction(NewCI);
110 rememberInstruction(CI);
114 BasicBlock::iterator IP = I; ++IP;
115 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
116 IP = II->getNormalDest()->begin();
117 while (isa<PHINode>(IP)) ++IP;
118 Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
119 rememberInstruction(CI);
123 /// InsertBinop - Insert the specified binary operator, doing a small amount
124 /// of work to avoid inserting an obviously redundant operation.
125 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
126 Value *LHS, Value *RHS) {
127 // Fold a binop with constant operands.
128 if (Constant *CLHS = dyn_cast<Constant>(LHS))
129 if (Constant *CRHS = dyn_cast<Constant>(RHS))
130 return ConstantExpr::get(Opcode, CLHS, CRHS);
132 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
133 unsigned ScanLimit = 6;
134 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
135 // Scanning starts from the last instruction before the insertion point.
136 BasicBlock::iterator IP = Builder.GetInsertPoint();
137 if (IP != BlockBegin) {
139 for (; ScanLimit; --IP, --ScanLimit) {
140 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
141 IP->getOperand(1) == RHS)
143 if (IP == BlockBegin) break;
147 // If we haven't found this binop, insert it.
148 Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
149 rememberInstruction(BO);
153 /// FactorOutConstant - Test if S is divisible by Factor, using signed
154 /// division. If so, update S with Factor divided out and return true.
155 /// S need not be evenly divisble if a reasonable remainder can be
157 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
158 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
159 /// check to see if the divide was folded.
160 static bool FactorOutConstant(const SCEV *&S,
161 const SCEV *&Remainder,
164 const TargetData *TD) {
165 // Everything is divisible by one.
171 S = SE.getIntegerSCEV(1, S->getType());
175 // For a Constant, check for a multiple of the given factor.
176 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
180 // Check for divisibility.
181 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
183 ConstantInt::get(SE.getContext(),
184 C->getValue()->getValue().sdiv(
185 FC->getValue()->getValue()));
186 // If the quotient is zero and the remainder is non-zero, reject
187 // the value at this scale. It will be considered for subsequent
190 const SCEV *Div = SE.getConstant(CI);
193 SE.getAddExpr(Remainder,
194 SE.getConstant(C->getValue()->getValue().srem(
195 FC->getValue()->getValue())));
201 // In a Mul, check if there is a constant operand which is a multiple
202 // of the given factor.
203 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
205 // With TargetData, the size is known. Check if there is a constant
206 // operand which is a multiple of the given factor. If so, we can
208 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
209 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
210 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
211 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
212 SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
215 SE.getConstant(C->getValue()->getValue().sdiv(
216 FC->getValue()->getValue()));
217 S = SE.getMulExpr(NewMulOps);
221 // Without TargetData, check if Factor can be factored out of any of the
222 // Mul's operands. If so, we can just remove it.
223 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
224 const SCEV *SOp = M->getOperand(i);
225 const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType());
226 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
227 Remainder->isZero()) {
228 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
229 SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
232 S = SE.getMulExpr(NewMulOps);
239 // In an AddRec, check if both start and step are divisible.
240 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
241 const SCEV *Step = A->getStepRecurrence(SE);
242 const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
243 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
245 if (!StepRem->isZero())
247 const SCEV *Start = A->getStart();
248 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
250 S = SE.getAddRecExpr(Start, Step, A->getLoop());
257 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
258 /// is the number of SCEVAddRecExprs present, which are kept at the end of
261 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
263 ScalarEvolution &SE) {
264 unsigned NumAddRecs = 0;
265 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
267 // Group Ops into non-addrecs and addrecs.
268 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
269 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
270 // Let ScalarEvolution sort and simplify the non-addrecs list.
271 const SCEV *Sum = NoAddRecs.empty() ?
272 SE.getIntegerSCEV(0, Ty) :
273 SE.getAddExpr(NoAddRecs);
274 // If it returned an add, use the operands. Otherwise it simplified
275 // the sum into a single value, so just use that.
276 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
277 Ops = Add->getOperands();
283 // Then append the addrecs.
284 Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
287 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
288 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
289 /// This helps expose more opportunities for folding parts of the expressions
290 /// into GEP indices.
292 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
294 ScalarEvolution &SE) {
296 SmallVector<const SCEV *, 8> AddRecs;
297 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
298 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
299 const SCEV *Start = A->getStart();
300 if (Start->isZero()) break;
301 const SCEV *Zero = SE.getIntegerSCEV(0, Ty);
302 AddRecs.push_back(SE.getAddRecExpr(Zero,
303 A->getStepRecurrence(SE),
305 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
307 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
308 e += Add->getNumOperands();
313 if (!AddRecs.empty()) {
314 // Add the addrecs onto the end of the list.
315 Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
316 // Resort the operand list, moving any constants to the front.
317 SimplifyAddOperands(Ops, Ty, SE);
321 /// expandAddToGEP - Expand an addition expression with a pointer type into
322 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
323 /// BasicAliasAnalysis and other passes analyze the result. See the rules
324 /// for getelementptr vs. inttoptr in
325 /// http://llvm.org/docs/LangRef.html#pointeraliasing
328 /// Design note: The correctness of using getelementptr here depends on
329 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
330 /// they may introduce pointer arithmetic which may not be safely converted
331 /// into getelementptr.
333 /// Design note: It might seem desirable for this function to be more
334 /// loop-aware. If some of the indices are loop-invariant while others
335 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
336 /// loop-invariant portions of the overall computation outside the loop.
337 /// However, there are a few reasons this is not done here. Hoisting simple
338 /// arithmetic is a low-level optimization that often isn't very
339 /// important until late in the optimization process. In fact, passes
340 /// like InstructionCombining will combine GEPs, even if it means
341 /// pushing loop-invariant computation down into loops, so even if the
342 /// GEPs were split here, the work would quickly be undone. The
343 /// LoopStrengthReduction pass, which is usually run quite late (and
344 /// after the last InstructionCombining pass), takes care of hoisting
345 /// loop-invariant portions of expressions, after considering what
346 /// can be folded using target addressing modes.
348 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
349 const SCEV *const *op_end,
350 const PointerType *PTy,
353 const Type *ElTy = PTy->getElementType();
354 SmallVector<Value *, 4> GepIndices;
355 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
356 bool AnyNonZeroIndices = false;
358 // Split AddRecs up into parts as either of the parts may be usable
359 // without the other.
360 SplitAddRecs(Ops, Ty, SE);
362 // Descend down the pointer's type and attempt to convert the other
363 // operands into GEP indices, at each level. The first index in a GEP
364 // indexes into the array implied by the pointer operand; the rest of
365 // the indices index into the element or field type selected by the
368 // If the scale size is not 0, attempt to factor out a scale for
370 SmallVector<const SCEV *, 8> ScaledOps;
371 if (ElTy->isSized()) {
372 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
373 if (!ElSize->isZero()) {
374 SmallVector<const SCEV *, 8> NewOps;
375 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
376 const SCEV *Op = Ops[i];
377 const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
378 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
379 // Op now has ElSize factored out.
380 ScaledOps.push_back(Op);
381 if (!Remainder->isZero())
382 NewOps.push_back(Remainder);
383 AnyNonZeroIndices = true;
385 // The operand was not divisible, so add it to the list of operands
386 // we'll scan next iteration.
387 NewOps.push_back(Ops[i]);
390 // If we made any changes, update Ops.
391 if (!ScaledOps.empty()) {
393 SimplifyAddOperands(Ops, Ty, SE);
398 // Record the scaled array index for this level of the type. If
399 // we didn't find any operands that could be factored, tentatively
400 // assume that element zero was selected (since the zero offset
401 // would obviously be folded away).
402 Value *Scaled = ScaledOps.empty() ?
403 Constant::getNullValue(Ty) :
404 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
405 GepIndices.push_back(Scaled);
407 // Collect struct field index operands.
408 while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
409 bool FoundFieldNo = false;
410 // An empty struct has no fields.
411 if (STy->getNumElements() == 0) break;
413 // With TargetData, field offsets are known. See if a constant offset
414 // falls within any of the struct fields.
415 if (Ops.empty()) break;
416 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
417 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
418 const StructLayout &SL = *SE.TD->getStructLayout(STy);
419 uint64_t FullOffset = C->getValue()->getZExtValue();
420 if (FullOffset < SL.getSizeInBytes()) {
421 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
422 GepIndices.push_back(
423 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
424 ElTy = STy->getTypeAtIndex(ElIdx);
426 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
427 AnyNonZeroIndices = true;
432 // Without TargetData, just check for an offsetof expression of the
433 // appropriate struct type.
434 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
435 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
438 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
439 GepIndices.push_back(FieldNo);
441 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
442 Ops[i] = SE.getConstant(Ty, 0);
443 AnyNonZeroIndices = true;
449 // If no struct field offsets were found, tentatively assume that
450 // field zero was selected (since the zero offset would obviously
453 ElTy = STy->getTypeAtIndex(0u);
454 GepIndices.push_back(
455 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
459 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
460 ElTy = ATy->getElementType();
465 // If none of the operands were convertable to proper GEP indices, cast
466 // the base to i8* and do an ugly getelementptr with that. It's still
467 // better than ptrtoint+arithmetic+inttoptr at least.
468 if (!AnyNonZeroIndices) {
469 // Cast the base to i8*.
470 V = InsertNoopCastOfTo(V,
471 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
473 // Expand the operands for a plain byte offset.
474 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
476 // Fold a GEP with constant operands.
477 if (Constant *CLHS = dyn_cast<Constant>(V))
478 if (Constant *CRHS = dyn_cast<Constant>(Idx))
479 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
481 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
482 unsigned ScanLimit = 6;
483 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
484 // Scanning starts from the last instruction before the insertion point.
485 BasicBlock::iterator IP = Builder.GetInsertPoint();
486 if (IP != BlockBegin) {
488 for (; ScanLimit; --IP, --ScanLimit) {
489 if (IP->getOpcode() == Instruction::GetElementPtr &&
490 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
492 if (IP == BlockBegin) break;
497 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
498 rememberInstruction(GEP);
502 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
503 // because ScalarEvolution may have changed the address arithmetic to
504 // compute a value which is beyond the end of the allocated object.
506 if (V->getType() != PTy)
507 Casted = InsertNoopCastOfTo(Casted, PTy);
508 Value *GEP = Builder.CreateGEP(Casted,
512 Ops.push_back(SE.getUnknown(GEP));
513 rememberInstruction(GEP);
514 return expand(SE.getAddExpr(Ops));
517 /// isNonConstantNegative - Return true if the specified scev is negated, but
519 static bool isNonConstantNegative(const SCEV *F) {
520 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
521 if (!Mul) return false;
523 // If there is a constant factor, it will be first.
524 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
525 if (!SC) return false;
527 // Return true if the value is negative, this matches things like (-42 * V).
528 return SC->getValue()->getValue().isNegative();
531 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
532 int NumOperands = S->getNumOperands();
533 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
535 // Find the index of an operand to start with. Choose the operand with
536 // pointer type, if there is one, or the last operand otherwise.
538 for (; PIdx != NumOperands - 1; ++PIdx)
539 if (isa<PointerType>(S->getOperand(PIdx)->getType())) break;
541 // Expand code for the operand that we chose.
542 Value *V = expand(S->getOperand(PIdx));
544 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
545 // comments on expandAddToGEP for details.
546 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
547 // Take the operand at PIdx out of the list.
548 const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
549 SmallVector<const SCEV *, 8> NewOps;
550 NewOps.insert(NewOps.end(), Ops.begin(), Ops.begin() + PIdx);
551 NewOps.insert(NewOps.end(), Ops.begin() + PIdx + 1, Ops.end());
553 return expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, V);
556 // Otherwise, we'll expand the rest of the SCEVAddExpr as plain integer
558 V = InsertNoopCastOfTo(V, Ty);
560 // Emit a bunch of add instructions
561 for (int i = NumOperands-1; i >= 0; --i) {
562 if (i == PIdx) continue;
563 const SCEV *Op = S->getOperand(i);
564 if (isNonConstantNegative(Op)) {
565 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
566 V = InsertBinop(Instruction::Sub, V, W);
568 Value *W = expandCodeFor(Op, Ty);
569 V = InsertBinop(Instruction::Add, V, W);
575 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
576 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
577 int FirstOp = 0; // Set if we should emit a subtract.
578 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
579 if (SC->getValue()->isAllOnesValue())
582 int i = S->getNumOperands()-2;
583 Value *V = expandCodeFor(S->getOperand(i+1), Ty);
585 // Emit a bunch of multiply instructions
586 for (; i >= FirstOp; --i) {
587 Value *W = expandCodeFor(S->getOperand(i), Ty);
588 V = InsertBinop(Instruction::Mul, V, W);
591 // -1 * ... ---> 0 - ...
593 V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
597 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
598 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
600 Value *LHS = expandCodeFor(S->getLHS(), Ty);
601 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
602 const APInt &RHS = SC->getValue()->getValue();
603 if (RHS.isPowerOf2())
604 return InsertBinop(Instruction::LShr, LHS,
605 ConstantInt::get(Ty, RHS.logBase2()));
608 Value *RHS = expandCodeFor(S->getRHS(), Ty);
609 return InsertBinop(Instruction::UDiv, LHS, RHS);
612 /// Move parts of Base into Rest to leave Base with the minimal
613 /// expression that provides a pointer operand suitable for a
615 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
616 ScalarEvolution &SE) {
617 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
618 Base = A->getStart();
619 Rest = SE.getAddExpr(Rest,
620 SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
621 A->getStepRecurrence(SE),
624 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
625 Base = A->getOperand(A->getNumOperands()-1);
626 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
627 NewAddOps.back() = Rest;
628 Rest = SE.getAddExpr(NewAddOps);
629 ExposePointerBase(Base, Rest, SE);
633 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
634 /// the base addrec, which is the addrec without any non-loop-dominating
635 /// values, and return the PHI.
637 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
639 const Type *ExpandTy,
641 // Reuse a previously-inserted PHI, if present.
642 for (BasicBlock::iterator I = L->getHeader()->begin();
643 PHINode *PN = dyn_cast<PHINode>(I); ++I)
644 if (isInsertedInstruction(PN) && SE.getSCEV(PN) == Normalized)
647 // Save the original insertion point so we can restore it when we're done.
648 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
649 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
651 // Expand code for the start value.
652 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
653 L->getHeader()->begin());
655 // Expand code for the step value. Insert instructions right before the
656 // terminator corresponding to the back-edge. Do this before creating the PHI
657 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
658 // negative, insert a sub instead of an add for the increment (unless it's a
659 // constant, because subtracts of constants are canonicalized to adds).
660 const SCEV *Step = Normalized->getStepRecurrence(SE);
661 bool isPointer = isa<PointerType>(ExpandTy);
662 bool isNegative = !isPointer && isNonConstantNegative(Step);
664 Step = SE.getNegativeSCEV(Step);
665 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
668 Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin());
669 PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv");
670 rememberInstruction(PN);
672 // Create the step instructions and populate the PHI.
673 BasicBlock *Header = L->getHeader();
674 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
676 BasicBlock *Pred = *HPI;
678 // Add a start value.
679 if (!L->contains(Pred)) {
680 PN->addIncoming(StartV, Pred);
684 // Create a step value and add it to the PHI. If IVIncInsertLoop is
685 // non-null and equal to the addrec's loop, insert the instructions
686 // at IVIncInsertPos.
687 Instruction *InsertPos = L == IVIncInsertLoop ?
688 IVIncInsertPos : Pred->getTerminator();
689 Builder.SetInsertPoint(InsertPos->getParent(), InsertPos);
691 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
693 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
694 // If the step isn't constant, don't use an implicitly scaled GEP, because
695 // that would require a multiply inside the loop.
696 if (!isa<ConstantInt>(StepV))
697 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
698 GEPPtrTy->getAddressSpace());
699 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
700 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
701 if (IncV->getType() != PN->getType()) {
702 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
703 rememberInstruction(IncV);
707 Builder.CreateSub(PN, StepV, "lsr.iv.next") :
708 Builder.CreateAdd(PN, StepV, "lsr.iv.next");
709 rememberInstruction(IncV);
711 PN->addIncoming(IncV, Pred);
714 // Restore the original insert point.
716 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
718 // Remember this PHI, even in post-inc mode.
719 InsertedValues.insert(PN);
724 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
725 const Type *STy = S->getType();
726 const Type *IntTy = SE.getEffectiveSCEVType(STy);
727 const Loop *L = S->getLoop();
729 // Determine a normalized form of this expression, which is the expression
730 // before any post-inc adjustment is made.
731 const SCEVAddRecExpr *Normalized = S;
732 if (L == PostIncLoop) {
733 const SCEV *Step = S->getStepRecurrence(SE);
734 Normalized = cast<SCEVAddRecExpr>(SE.getMinusSCEV(S, Step));
737 // Strip off any non-loop-dominating component from the addrec start.
738 const SCEV *Start = Normalized->getStart();
739 const SCEV *PostLoopOffset = 0;
740 if (!Start->properlyDominates(L->getHeader(), SE.DT)) {
741 PostLoopOffset = Start;
742 Start = SE.getIntegerSCEV(0, Normalized->getType());
744 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start,
745 Normalized->getStepRecurrence(SE),
746 Normalized->getLoop()));
749 // Strip off any non-loop-dominating component from the addrec step.
750 const SCEV *Step = Normalized->getStepRecurrence(SE);
751 const SCEV *PostLoopScale = 0;
752 if (!Step->hasComputableLoopEvolution(L) &&
753 !Step->dominates(L->getHeader(), SE.DT)) {
754 PostLoopScale = Step;
755 Step = SE.getIntegerSCEV(1, Normalized->getType());
757 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
758 Normalized->getLoop()));
761 // Expand the core addrec. If we need post-loop scaling, force it to
762 // expand to an integer type to avoid the need for additional casting.
763 const Type *ExpandTy = PostLoopScale ? IntTy : STy;
764 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
766 // Accomodate post-inc mode, if necessary.
768 if (L != PostIncLoop)
771 // In PostInc mode, use the post-incremented value.
772 BasicBlock *LatchBlock = L->getLoopLatch();
773 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
774 Result = PN->getIncomingValueForBlock(LatchBlock);
777 // Re-apply any non-loop-dominating scale.
779 Result = Builder.CreateMul(Result,
780 expandCodeFor(PostLoopScale, IntTy));
781 rememberInstruction(Result);
784 // Re-apply any non-loop-dominating offset.
785 if (PostLoopOffset) {
786 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
787 const SCEV *const OffsetArray[1] = { PostLoopOffset };
788 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
790 Result = Builder.CreateAdd(Result,
791 expandCodeFor(PostLoopOffset, IntTy));
792 rememberInstruction(Result);
799 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
800 if (!CanonicalMode) return expandAddRecExprLiterally(S);
802 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
803 const Loop *L = S->getLoop();
805 // First check for an existing canonical IV in a suitable type.
806 PHINode *CanonicalIV = 0;
807 if (PHINode *PN = L->getCanonicalInductionVariable())
808 if (SE.isSCEVable(PN->getType()) &&
809 isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
810 SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
813 // Rewrite an AddRec in terms of the canonical induction variable, if
814 // its type is more narrow.
816 SE.getTypeSizeInBits(CanonicalIV->getType()) >
817 SE.getTypeSizeInBits(Ty)) {
818 const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
819 SmallVector<const SCEV *, 4> NewOps(Ops.size());
820 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
821 NewOps[i] = SE.getAnyExtendExpr(Ops[i], CanonicalIV->getType());
822 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop()));
823 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
824 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
825 BasicBlock::iterator NewInsertPt =
826 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
827 while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
828 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
830 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
834 // {X,+,F} --> X + {0,+,F}
835 if (!S->getStart()->isZero()) {
836 const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
837 SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
838 NewOps[0] = SE.getIntegerSCEV(0, Ty);
839 const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
841 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
842 // comments on expandAddToGEP for details.
843 const SCEV *Base = S->getStart();
844 const SCEV *RestArray[1] = { Rest };
845 // Dig into the expression to find the pointer base for a GEP.
846 ExposePointerBase(Base, RestArray[0], SE);
847 // If we found a pointer, expand the AddRec with a GEP.
848 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
849 // Make sure the Base isn't something exotic, such as a multiplied
850 // or divided pointer value. In those cases, the result type isn't
851 // actually a pointer type.
852 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
853 Value *StartV = expand(Base);
854 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
855 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
859 // Just do a normal add. Pre-expand the operands to suppress folding.
860 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
861 SE.getUnknown(expand(Rest))));
864 // {0,+,1} --> Insert a canonical induction variable into the loop!
866 S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
867 // If there's a canonical IV, just use it.
869 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
870 "IVs with types different from the canonical IV should "
871 "already have been handled!");
875 // Create and insert the PHI node for the induction variable in the
877 BasicBlock *Header = L->getHeader();
878 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
879 rememberInstruction(PN);
881 Constant *One = ConstantInt::get(Ty, 1);
882 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
884 if (L->contains(*HPI)) {
885 // Insert a unit add instruction right before the terminator
886 // corresponding to the back-edge.
887 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
888 (*HPI)->getTerminator());
889 rememberInstruction(Add);
890 PN->addIncoming(Add, *HPI);
892 PN->addIncoming(Constant::getNullValue(Ty), *HPI);
896 // {0,+,F} --> {0,+,1} * F
897 // Get the canonical induction variable I for this loop.
898 Value *I = CanonicalIV ?
900 getOrInsertCanonicalInductionVariable(L, Ty);
902 // If this is a simple linear addrec, emit it now as a special case.
903 if (S->isAffine()) // {0,+,F} --> i*F
905 expand(SE.getTruncateOrNoop(
906 SE.getMulExpr(SE.getUnknown(I),
907 SE.getNoopOrAnyExtend(S->getOperand(1),
911 // If this is a chain of recurrences, turn it into a closed form, using the
912 // folders, then expandCodeFor the closed form. This allows the folders to
913 // simplify the expression without having to build a bunch of special code
915 const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
917 // Promote S up to the canonical IV type, if the cast is foldable.
918 const SCEV *NewS = S;
919 const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
920 if (isa<SCEVAddRecExpr>(Ext))
923 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
924 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
926 // Truncate the result down to the original type, if needed.
927 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
931 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
932 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
933 Value *V = expandCodeFor(S->getOperand(),
934 SE.getEffectiveSCEVType(S->getOperand()->getType()));
935 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
936 rememberInstruction(I);
940 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
941 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
942 Value *V = expandCodeFor(S->getOperand(),
943 SE.getEffectiveSCEVType(S->getOperand()->getType()));
944 Value *I = Builder.CreateZExt(V, Ty, "tmp");
945 rememberInstruction(I);
949 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
950 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
951 Value *V = expandCodeFor(S->getOperand(),
952 SE.getEffectiveSCEVType(S->getOperand()->getType()));
953 Value *I = Builder.CreateSExt(V, Ty, "tmp");
954 rememberInstruction(I);
958 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
959 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
960 const Type *Ty = LHS->getType();
961 for (int i = S->getNumOperands()-2; i >= 0; --i) {
962 // In the case of mixed integer and pointer types, do the
963 // rest of the comparisons as integer.
964 if (S->getOperand(i)->getType() != Ty) {
965 Ty = SE.getEffectiveSCEVType(Ty);
966 LHS = InsertNoopCastOfTo(LHS, Ty);
968 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
969 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
970 rememberInstruction(ICmp);
971 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
972 rememberInstruction(Sel);
975 // In the case of mixed integer and pointer types, cast the
976 // final result back to the pointer type.
977 if (LHS->getType() != S->getType())
978 LHS = InsertNoopCastOfTo(LHS, S->getType());
982 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
983 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
984 const Type *Ty = LHS->getType();
985 for (int i = S->getNumOperands()-2; i >= 0; --i) {
986 // In the case of mixed integer and pointer types, do the
987 // rest of the comparisons as integer.
988 if (S->getOperand(i)->getType() != Ty) {
989 Ty = SE.getEffectiveSCEVType(Ty);
990 LHS = InsertNoopCastOfTo(LHS, Ty);
992 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
993 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
994 rememberInstruction(ICmp);
995 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
996 rememberInstruction(Sel);
999 // In the case of mixed integer and pointer types, cast the
1000 // final result back to the pointer type.
1001 if (LHS->getType() != S->getType())
1002 LHS = InsertNoopCastOfTo(LHS, S->getType());
1006 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
1007 // Expand the code for this SCEV.
1008 Value *V = expand(SH);
1010 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1011 "non-trivial casts should be done with the SCEVs directly!");
1012 V = InsertNoopCastOfTo(V, Ty);
1017 Value *SCEVExpander::expand(const SCEV *S) {
1018 // Compute an insertion point for this SCEV object. Hoist the instructions
1019 // as far out in the loop nest as possible.
1020 Instruction *InsertPt = Builder.GetInsertPoint();
1021 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1022 L = L->getParentLoop())
1023 if (S->isLoopInvariant(L)) {
1025 if (BasicBlock *Preheader = L->getLoopPreheader())
1026 InsertPt = Preheader->getTerminator();
1028 // If the SCEV is computable at this level, insert it into the header
1029 // after the PHIs (and after any other instructions that we've inserted
1030 // there) so that it is guaranteed to dominate any user inside the loop.
1031 if (L && S->hasComputableLoopEvolution(L))
1032 InsertPt = L->getHeader()->getFirstNonPHI();
1033 while (isInsertedInstruction(InsertPt))
1034 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1038 // Check to see if we already expanded this here.
1039 std::map<std::pair<const SCEV *, Instruction *>,
1040 AssertingVH<Value> >::iterator I =
1041 InsertedExpressions.find(std::make_pair(S, InsertPt));
1042 if (I != InsertedExpressions.end())
1045 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1046 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1047 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1049 // Expand the expression into instructions.
1050 Value *V = visit(S);
1052 // Remember the expanded value for this SCEV at this location.
1054 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1056 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
1060 /// getOrInsertCanonicalInductionVariable - This method returns the
1061 /// canonical induction variable of the specified type for the specified
1062 /// loop (inserting one if there is none). A canonical induction variable
1063 /// starts at zero and steps by one on each iteration.
1065 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1067 assert(Ty->isInteger() && "Can only insert integer induction variables!");
1068 const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
1069 SE.getIntegerSCEV(1, Ty), L);
1070 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1071 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1072 Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
1074 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);