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 const SCEV *ElSize = SE.getAllocSizeExpr(ElTy);
369 // If the scale size is not 0, attempt to factor out a scale for
371 SmallVector<const SCEV *, 8> ScaledOps;
372 if (ElTy->isSized() && !ElSize->isZero()) {
373 SmallVector<const SCEV *, 8> NewOps;
374 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
375 const SCEV *Op = Ops[i];
376 const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
377 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
378 // Op now has ElSize factored out.
379 ScaledOps.push_back(Op);
380 if (!Remainder->isZero())
381 NewOps.push_back(Remainder);
382 AnyNonZeroIndices = true;
384 // The operand was not divisible, so add it to the list of operands
385 // we'll scan next iteration.
386 NewOps.push_back(Ops[i]);
389 // If we made any changes, update Ops.
390 if (!ScaledOps.empty()) {
392 SimplifyAddOperands(Ops, Ty, SE);
396 // Record the scaled array index for this level of the type. If
397 // we didn't find any operands that could be factored, tentatively
398 // assume that element zero was selected (since the zero offset
399 // would obviously be folded away).
400 Value *Scaled = ScaledOps.empty() ?
401 Constant::getNullValue(Ty) :
402 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
403 GepIndices.push_back(Scaled);
405 // Collect struct field index operands.
406 while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
407 bool FoundFieldNo = false;
408 // An empty struct has no fields.
409 if (STy->getNumElements() == 0) break;
411 // With TargetData, field offsets are known. See if a constant offset
412 // falls within any of the struct fields.
413 if (Ops.empty()) break;
414 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
415 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
416 const StructLayout &SL = *SE.TD->getStructLayout(STy);
417 uint64_t FullOffset = C->getValue()->getZExtValue();
418 if (FullOffset < SL.getSizeInBytes()) {
419 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
420 GepIndices.push_back(
421 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
422 ElTy = STy->getTypeAtIndex(ElIdx);
424 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
425 AnyNonZeroIndices = true;
430 // Without TargetData, just check for an offsetof expression of the
431 // appropriate struct type.
432 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
433 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
434 const StructType *StructTy;
436 if (U->isOffsetOf(StructTy, FieldNo) && StructTy == STy) {
437 GepIndices.push_back(FieldNo);
439 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
440 Ops[i] = SE.getConstant(Ty, 0);
441 AnyNonZeroIndices = true;
447 // If no struct field offsets were found, tentatively assume that
448 // field zero was selected (since the zero offset would obviously
451 ElTy = STy->getTypeAtIndex(0u);
452 GepIndices.push_back(
453 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
457 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
458 ElTy = ATy->getElementType();
463 // If none of the operands were convertable to proper GEP indices, cast
464 // the base to i8* and do an ugly getelementptr with that. It's still
465 // better than ptrtoint+arithmetic+inttoptr at least.
466 if (!AnyNonZeroIndices) {
467 // Cast the base to i8*.
468 V = InsertNoopCastOfTo(V,
469 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
471 // Expand the operands for a plain byte offset.
472 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
474 // Fold a GEP with constant operands.
475 if (Constant *CLHS = dyn_cast<Constant>(V))
476 if (Constant *CRHS = dyn_cast<Constant>(Idx))
477 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
479 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
480 unsigned ScanLimit = 6;
481 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
482 // Scanning starts from the last instruction before the insertion point.
483 BasicBlock::iterator IP = Builder.GetInsertPoint();
484 if (IP != BlockBegin) {
486 for (; ScanLimit; --IP, --ScanLimit) {
487 if (IP->getOpcode() == Instruction::GetElementPtr &&
488 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
490 if (IP == BlockBegin) break;
495 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
496 rememberInstruction(GEP);
500 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
501 // because ScalarEvolution may have changed the address arithmetic to
502 // compute a value which is beyond the end of the allocated object.
504 if (V->getType() != PTy)
505 Casted = InsertNoopCastOfTo(Casted, PTy);
506 Value *GEP = Builder.CreateGEP(Casted,
510 Ops.push_back(SE.getUnknown(GEP));
511 rememberInstruction(GEP);
512 return expand(SE.getAddExpr(Ops));
515 /// isNonConstantNegative - Return true if the specified scev is negated, but
517 static bool isNonConstantNegative(const SCEV *F) {
518 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
519 if (!Mul) return false;
521 // If there is a constant factor, it will be first.
522 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
523 if (!SC) return false;
525 // Return true if the value is negative, this matches things like (-42 * V).
526 return SC->getValue()->getValue().isNegative();
529 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
530 int NumOperands = S->getNumOperands();
531 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
533 // Find the index of an operand to start with. Choose the operand with
534 // pointer type, if there is one, or the last operand otherwise.
536 for (; PIdx != NumOperands - 1; ++PIdx)
537 if (isa<PointerType>(S->getOperand(PIdx)->getType())) break;
539 // Expand code for the operand that we chose.
540 Value *V = expand(S->getOperand(PIdx));
542 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
543 // comments on expandAddToGEP for details.
544 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
545 // Take the operand at PIdx out of the list.
546 const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
547 SmallVector<const SCEV *, 8> NewOps;
548 NewOps.insert(NewOps.end(), Ops.begin(), Ops.begin() + PIdx);
549 NewOps.insert(NewOps.end(), Ops.begin() + PIdx + 1, Ops.end());
551 return expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, V);
554 // Otherwise, we'll expand the rest of the SCEVAddExpr as plain integer
556 V = InsertNoopCastOfTo(V, Ty);
558 // Emit a bunch of add instructions
559 for (int i = NumOperands-1; i >= 0; --i) {
560 if (i == PIdx) continue;
561 const SCEV *Op = S->getOperand(i);
562 if (isNonConstantNegative(Op)) {
563 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
564 V = InsertBinop(Instruction::Sub, V, W);
566 Value *W = expandCodeFor(Op, Ty);
567 V = InsertBinop(Instruction::Add, V, W);
573 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
574 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
575 int FirstOp = 0; // Set if we should emit a subtract.
576 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
577 if (SC->getValue()->isAllOnesValue())
580 int i = S->getNumOperands()-2;
581 Value *V = expandCodeFor(S->getOperand(i+1), Ty);
583 // Emit a bunch of multiply instructions
584 for (; i >= FirstOp; --i) {
585 Value *W = expandCodeFor(S->getOperand(i), Ty);
586 V = InsertBinop(Instruction::Mul, V, W);
589 // -1 * ... ---> 0 - ...
591 V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
595 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
596 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
598 Value *LHS = expandCodeFor(S->getLHS(), Ty);
599 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
600 const APInt &RHS = SC->getValue()->getValue();
601 if (RHS.isPowerOf2())
602 return InsertBinop(Instruction::LShr, LHS,
603 ConstantInt::get(Ty, RHS.logBase2()));
606 Value *RHS = expandCodeFor(S->getRHS(), Ty);
607 return InsertBinop(Instruction::UDiv, LHS, RHS);
610 /// Move parts of Base into Rest to leave Base with the minimal
611 /// expression that provides a pointer operand suitable for a
613 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
614 ScalarEvolution &SE) {
615 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
616 Base = A->getStart();
617 Rest = SE.getAddExpr(Rest,
618 SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
619 A->getStepRecurrence(SE),
622 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
623 Base = A->getOperand(A->getNumOperands()-1);
624 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
625 NewAddOps.back() = Rest;
626 Rest = SE.getAddExpr(NewAddOps);
627 ExposePointerBase(Base, Rest, SE);
631 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
632 /// the base addrec, which is the addrec without any non-loop-dominating
633 /// values, and return the PHI.
635 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
637 const Type *ExpandTy,
639 // Reuse a previously-inserted PHI, if present.
640 for (BasicBlock::iterator I = L->getHeader()->begin();
641 PHINode *PN = dyn_cast<PHINode>(I); ++I)
642 if (isInsertedInstruction(PN) && SE.getSCEV(PN) == Normalized)
645 // Save the original insertion point so we can restore it when we're done.
646 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
647 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
649 // Expand code for the start value.
650 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
651 L->getHeader()->begin());
653 // Expand code for the step value. Insert instructions right before the
654 // terminator corresponding to the back-edge. Do this before creating the PHI
655 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
656 // negative, insert a sub instead of an add for the increment (unless it's a
657 // constant, because subtracts of constants are canonicalized to adds).
658 const SCEV *Step = Normalized->getStepRecurrence(SE);
659 bool isPointer = isa<PointerType>(ExpandTy);
660 bool isNegative = !isPointer && isNonConstantNegative(Step);
662 Step = SE.getNegativeSCEV(Step);
663 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
666 Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin());
667 PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv");
668 rememberInstruction(PN);
670 // Create the step instructions and populate the PHI.
671 BasicBlock *Header = L->getHeader();
672 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
674 BasicBlock *Pred = *HPI;
676 // Add a start value.
677 if (!L->contains(Pred)) {
678 PN->addIncoming(StartV, Pred);
682 // Create a step value and add it to the PHI. If IVIncInsertLoop is
683 // non-null and equal to the addrec's loop, insert the instructions
684 // at IVIncInsertPos.
685 Instruction *InsertPos = L == IVIncInsertLoop ?
686 IVIncInsertPos : Pred->getTerminator();
687 Builder.SetInsertPoint(InsertPos->getParent(), InsertPos);
689 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
691 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
692 // If the step isn't constant, don't use an implicitly scaled GEP, because
693 // that would require a multiply inside the loop.
694 if (!isa<ConstantInt>(StepV))
695 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
696 GEPPtrTy->getAddressSpace());
697 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
698 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
699 if (IncV->getType() != PN->getType()) {
700 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
701 rememberInstruction(IncV);
705 Builder.CreateSub(PN, StepV, "lsr.iv.next") :
706 Builder.CreateAdd(PN, StepV, "lsr.iv.next");
707 rememberInstruction(IncV);
709 PN->addIncoming(IncV, Pred);
712 // Restore the original insert point.
714 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
716 // Remember this PHI, even in post-inc mode.
717 InsertedValues.insert(PN);
722 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
723 const Type *STy = S->getType();
724 const Type *IntTy = SE.getEffectiveSCEVType(STy);
725 const Loop *L = S->getLoop();
727 // Determine a normalized form of this expression, which is the expression
728 // before any post-inc adjustment is made.
729 const SCEVAddRecExpr *Normalized = S;
730 if (L == PostIncLoop) {
731 const SCEV *Step = S->getStepRecurrence(SE);
732 Normalized = cast<SCEVAddRecExpr>(SE.getMinusSCEV(S, Step));
735 // Strip off any non-loop-dominating component from the addrec start.
736 const SCEV *Start = Normalized->getStart();
737 const SCEV *PostLoopOffset = 0;
738 if (!Start->properlyDominates(L->getHeader(), SE.DT)) {
739 PostLoopOffset = Start;
740 Start = SE.getIntegerSCEV(0, Normalized->getType());
742 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start,
743 Normalized->getStepRecurrence(SE),
744 Normalized->getLoop()));
747 // Strip off any non-loop-dominating component from the addrec step.
748 const SCEV *Step = Normalized->getStepRecurrence(SE);
749 const SCEV *PostLoopScale = 0;
750 if (!Step->hasComputableLoopEvolution(L) &&
751 !Step->dominates(L->getHeader(), SE.DT)) {
752 PostLoopScale = Step;
753 Step = SE.getIntegerSCEV(1, Normalized->getType());
755 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
756 Normalized->getLoop()));
759 // Expand the core addrec. If we need post-loop scaling, force it to
760 // expand to an integer type to avoid the need for additional casting.
761 const Type *ExpandTy = PostLoopScale ? IntTy : STy;
762 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
764 // Accomodate post-inc mode, if necessary.
766 if (L != PostIncLoop)
769 // In PostInc mode, use the post-incremented value.
770 BasicBlock *LatchBlock = L->getLoopLatch();
771 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
772 Result = PN->getIncomingValueForBlock(LatchBlock);
775 // Re-apply any non-loop-dominating scale.
777 Result = Builder.CreateMul(Result,
778 expandCodeFor(PostLoopScale, IntTy));
779 rememberInstruction(Result);
782 // Re-apply any non-loop-dominating offset.
783 if (PostLoopOffset) {
784 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
785 const SCEV *const OffsetArray[1] = { PostLoopOffset };
786 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
788 Result = Builder.CreateAdd(Result,
789 expandCodeFor(PostLoopOffset, IntTy));
790 rememberInstruction(Result);
797 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
798 if (!CanonicalMode) return expandAddRecExprLiterally(S);
800 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
801 const Loop *L = S->getLoop();
803 // First check for an existing canonical IV in a suitable type.
804 PHINode *CanonicalIV = 0;
805 if (PHINode *PN = L->getCanonicalInductionVariable())
806 if (SE.isSCEVable(PN->getType()) &&
807 isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
808 SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
811 // Rewrite an AddRec in terms of the canonical induction variable, if
812 // its type is more narrow.
814 SE.getTypeSizeInBits(CanonicalIV->getType()) >
815 SE.getTypeSizeInBits(Ty)) {
816 const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
817 SmallVector<const SCEV *, 4> NewOps(Ops.size());
818 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
819 NewOps[i] = SE.getAnyExtendExpr(Ops[i], CanonicalIV->getType());
820 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop()));
821 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
822 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
823 BasicBlock::iterator NewInsertPt =
824 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
825 while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
826 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
828 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
832 // {X,+,F} --> X + {0,+,F}
833 if (!S->getStart()->isZero()) {
834 const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
835 SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
836 NewOps[0] = SE.getIntegerSCEV(0, Ty);
837 const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
839 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
840 // comments on expandAddToGEP for details.
841 const SCEV *Base = S->getStart();
842 const SCEV *RestArray[1] = { Rest };
843 // Dig into the expression to find the pointer base for a GEP.
844 ExposePointerBase(Base, RestArray[0], SE);
845 // If we found a pointer, expand the AddRec with a GEP.
846 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
847 // Make sure the Base isn't something exotic, such as a multiplied
848 // or divided pointer value. In those cases, the result type isn't
849 // actually a pointer type.
850 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
851 Value *StartV = expand(Base);
852 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
853 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
857 // Just do a normal add. Pre-expand the operands to suppress folding.
858 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
859 SE.getUnknown(expand(Rest))));
862 // {0,+,1} --> Insert a canonical induction variable into the loop!
864 S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
865 // If there's a canonical IV, just use it.
867 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
868 "IVs with types different from the canonical IV should "
869 "already have been handled!");
873 // Create and insert the PHI node for the induction variable in the
875 BasicBlock *Header = L->getHeader();
876 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
877 rememberInstruction(PN);
879 Constant *One = ConstantInt::get(Ty, 1);
880 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
882 if (L->contains(*HPI)) {
883 // Insert a unit add instruction right before the terminator
884 // corresponding to the back-edge.
885 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
886 (*HPI)->getTerminator());
887 rememberInstruction(Add);
888 PN->addIncoming(Add, *HPI);
890 PN->addIncoming(Constant::getNullValue(Ty), *HPI);
894 // {0,+,F} --> {0,+,1} * F
895 // Get the canonical induction variable I for this loop.
896 Value *I = CanonicalIV ?
898 getOrInsertCanonicalInductionVariable(L, Ty);
900 // If this is a simple linear addrec, emit it now as a special case.
901 if (S->isAffine()) // {0,+,F} --> i*F
903 expand(SE.getTruncateOrNoop(
904 SE.getMulExpr(SE.getUnknown(I),
905 SE.getNoopOrAnyExtend(S->getOperand(1),
909 // If this is a chain of recurrences, turn it into a closed form, using the
910 // folders, then expandCodeFor the closed form. This allows the folders to
911 // simplify the expression without having to build a bunch of special code
913 const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
915 // Promote S up to the canonical IV type, if the cast is foldable.
916 const SCEV *NewS = S;
917 const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
918 if (isa<SCEVAddRecExpr>(Ext))
921 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
922 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
924 // Truncate the result down to the original type, if needed.
925 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
929 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
930 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
931 Value *V = expandCodeFor(S->getOperand(),
932 SE.getEffectiveSCEVType(S->getOperand()->getType()));
933 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
934 rememberInstruction(I);
938 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
939 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
940 Value *V = expandCodeFor(S->getOperand(),
941 SE.getEffectiveSCEVType(S->getOperand()->getType()));
942 Value *I = Builder.CreateZExt(V, Ty, "tmp");
943 rememberInstruction(I);
947 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
948 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
949 Value *V = expandCodeFor(S->getOperand(),
950 SE.getEffectiveSCEVType(S->getOperand()->getType()));
951 Value *I = Builder.CreateSExt(V, Ty, "tmp");
952 rememberInstruction(I);
956 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
957 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
958 const Type *Ty = LHS->getType();
959 for (int i = S->getNumOperands()-2; i >= 0; --i) {
960 // In the case of mixed integer and pointer types, do the
961 // rest of the comparisons as integer.
962 if (S->getOperand(i)->getType() != Ty) {
963 Ty = SE.getEffectiveSCEVType(Ty);
964 LHS = InsertNoopCastOfTo(LHS, Ty);
966 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
967 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
968 rememberInstruction(ICmp);
969 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
970 rememberInstruction(Sel);
973 // In the case of mixed integer and pointer types, cast the
974 // final result back to the pointer type.
975 if (LHS->getType() != S->getType())
976 LHS = InsertNoopCastOfTo(LHS, S->getType());
980 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
981 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
982 const Type *Ty = LHS->getType();
983 for (int i = S->getNumOperands()-2; i >= 0; --i) {
984 // In the case of mixed integer and pointer types, do the
985 // rest of the comparisons as integer.
986 if (S->getOperand(i)->getType() != Ty) {
987 Ty = SE.getEffectiveSCEVType(Ty);
988 LHS = InsertNoopCastOfTo(LHS, Ty);
990 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
991 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
992 rememberInstruction(ICmp);
993 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
994 rememberInstruction(Sel);
997 // In the case of mixed integer and pointer types, cast the
998 // final result back to the pointer type.
999 if (LHS->getType() != S->getType())
1000 LHS = InsertNoopCastOfTo(LHS, S->getType());
1004 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
1005 // Expand the code for this SCEV.
1006 Value *V = expand(SH);
1008 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1009 "non-trivial casts should be done with the SCEVs directly!");
1010 V = InsertNoopCastOfTo(V, Ty);
1015 Value *SCEVExpander::expand(const SCEV *S) {
1016 // Compute an insertion point for this SCEV object. Hoist the instructions
1017 // as far out in the loop nest as possible.
1018 Instruction *InsertPt = Builder.GetInsertPoint();
1019 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1020 L = L->getParentLoop())
1021 if (S->isLoopInvariant(L)) {
1023 if (BasicBlock *Preheader = L->getLoopPreheader())
1024 InsertPt = Preheader->getTerminator();
1026 // If the SCEV is computable at this level, insert it into the header
1027 // after the PHIs (and after any other instructions that we've inserted
1028 // there) so that it is guaranteed to dominate any user inside the loop.
1029 if (L && S->hasComputableLoopEvolution(L))
1030 InsertPt = L->getHeader()->getFirstNonPHI();
1031 while (isInsertedInstruction(InsertPt))
1032 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1036 // Check to see if we already expanded this here.
1037 std::map<std::pair<const SCEV *, Instruction *>,
1038 AssertingVH<Value> >::iterator I =
1039 InsertedExpressions.find(std::make_pair(S, InsertPt));
1040 if (I != InsertedExpressions.end())
1043 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1044 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1045 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1047 // Expand the expression into instructions.
1048 Value *V = visit(S);
1050 // Remember the expanded value for this SCEV at this location.
1052 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1054 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
1058 /// getOrInsertCanonicalInductionVariable - This method returns the
1059 /// canonical induction variable of the specified type for the specified
1060 /// loop (inserting one if there is none). A canonical induction variable
1061 /// starts at zero and steps by one on each iteration.
1063 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1065 assert(Ty->isInteger() && "Can only insert integer induction variables!");
1066 const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
1067 SE.getIntegerSCEV(1, Ty), L);
1068 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1069 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1070 Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
1072 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);