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 at the beginning of the entry block.
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);
112 BasicBlock::iterator IP = I; ++IP;
113 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
114 IP = II->getNormalDest()->begin();
115 while (isa<PHINode>(IP)) ++IP;
116 Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
117 rememberInstruction(CI);
121 /// InsertBinop - Insert the specified binary operator, doing a small amount
122 /// of work to avoid inserting an obviously redundant operation.
123 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
124 Value *LHS, Value *RHS) {
125 // Fold a binop with constant operands.
126 if (Constant *CLHS = dyn_cast<Constant>(LHS))
127 if (Constant *CRHS = dyn_cast<Constant>(RHS))
128 return ConstantExpr::get(Opcode, CLHS, CRHS);
130 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
131 unsigned ScanLimit = 6;
132 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
133 // Scanning starts from the last instruction before the insertion point.
134 BasicBlock::iterator IP = Builder.GetInsertPoint();
135 if (IP != BlockBegin) {
137 for (; ScanLimit; --IP, --ScanLimit) {
138 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
139 IP->getOperand(1) == RHS)
141 if (IP == BlockBegin) break;
145 // If we haven't found this binop, insert it.
146 Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
147 rememberInstruction(BO);
151 /// FactorOutConstant - Test if S is divisible by Factor, using signed
152 /// division. If so, update S with Factor divided out and return true.
153 /// S need not be evenly divisble if a reasonable remainder can be
155 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
156 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
157 /// check to see if the divide was folded.
158 static bool FactorOutConstant(const SCEV *&S,
159 const SCEV *&Remainder,
162 const TargetData *TD) {
163 // Everything is divisible by one.
169 S = SE.getIntegerSCEV(1, S->getType());
173 // For a Constant, check for a multiple of the given factor.
174 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
178 // Check for divisibility.
179 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
181 ConstantInt::get(SE.getContext(),
182 C->getValue()->getValue().sdiv(
183 FC->getValue()->getValue()));
184 // If the quotient is zero and the remainder is non-zero, reject
185 // the value at this scale. It will be considered for subsequent
188 const SCEV *Div = SE.getConstant(CI);
191 SE.getAddExpr(Remainder,
192 SE.getConstant(C->getValue()->getValue().srem(
193 FC->getValue()->getValue())));
199 // In a Mul, check if there is a constant operand which is a multiple
200 // of the given factor.
201 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
203 // With TargetData, the size is known. Check if there is a constant
204 // operand which is a multiple of the given factor. If so, we can
206 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
207 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
208 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
209 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
210 SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
213 SE.getConstant(C->getValue()->getValue().sdiv(
214 FC->getValue()->getValue()));
215 S = SE.getMulExpr(NewMulOps);
219 // Without TargetData, check if Factor can be factored out of any of the
220 // Mul's operands. If so, we can just remove it.
221 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
222 const SCEV *SOp = M->getOperand(i);
223 const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType());
224 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
225 Remainder->isZero()) {
226 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
227 SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
230 S = SE.getMulExpr(NewMulOps);
237 // In an AddRec, check if both start and step are divisible.
238 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
239 const SCEV *Step = A->getStepRecurrence(SE);
240 const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
241 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
243 if (!StepRem->isZero())
245 const SCEV *Start = A->getStart();
246 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
248 S = SE.getAddRecExpr(Start, Step, A->getLoop());
255 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
256 /// is the number of SCEVAddRecExprs present, which are kept at the end of
259 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
261 ScalarEvolution &SE) {
262 unsigned NumAddRecs = 0;
263 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
265 // Group Ops into non-addrecs and addrecs.
266 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
267 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
268 // Let ScalarEvolution sort and simplify the non-addrecs list.
269 const SCEV *Sum = NoAddRecs.empty() ?
270 SE.getIntegerSCEV(0, Ty) :
271 SE.getAddExpr(NoAddRecs);
272 // If it returned an add, use the operands. Otherwise it simplified
273 // the sum into a single value, so just use that.
274 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
275 Ops = Add->getOperands();
281 // Then append the addrecs.
282 Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
285 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
286 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
287 /// This helps expose more opportunities for folding parts of the expressions
288 /// into GEP indices.
290 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
292 ScalarEvolution &SE) {
294 SmallVector<const SCEV *, 8> AddRecs;
295 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
296 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
297 const SCEV *Start = A->getStart();
298 if (Start->isZero()) break;
299 const SCEV *Zero = SE.getIntegerSCEV(0, Ty);
300 AddRecs.push_back(SE.getAddRecExpr(Zero,
301 A->getStepRecurrence(SE),
303 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
305 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
306 e += Add->getNumOperands();
311 if (!AddRecs.empty()) {
312 // Add the addrecs onto the end of the list.
313 Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
314 // Resort the operand list, moving any constants to the front.
315 SimplifyAddOperands(Ops, Ty, SE);
319 /// expandAddToGEP - Expand an addition expression with a pointer type into
320 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
321 /// BasicAliasAnalysis and other passes analyze the result. See the rules
322 /// for getelementptr vs. inttoptr in
323 /// http://llvm.org/docs/LangRef.html#pointeraliasing
326 /// Design note: The correctness of using getelementptr here depends on
327 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
328 /// they may introduce pointer arithmetic which may not be safely converted
329 /// into getelementptr.
331 /// Design note: It might seem desirable for this function to be more
332 /// loop-aware. If some of the indices are loop-invariant while others
333 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
334 /// loop-invariant portions of the overall computation outside the loop.
335 /// However, there are a few reasons this is not done here. Hoisting simple
336 /// arithmetic is a low-level optimization that often isn't very
337 /// important until late in the optimization process. In fact, passes
338 /// like InstructionCombining will combine GEPs, even if it means
339 /// pushing loop-invariant computation down into loops, so even if the
340 /// GEPs were split here, the work would quickly be undone. The
341 /// LoopStrengthReduction pass, which is usually run quite late (and
342 /// after the last InstructionCombining pass), takes care of hoisting
343 /// loop-invariant portions of expressions, after considering what
344 /// can be folded using target addressing modes.
346 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
347 const SCEV *const *op_end,
348 const PointerType *PTy,
351 const Type *ElTy = PTy->getElementType();
352 SmallVector<Value *, 4> GepIndices;
353 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
354 bool AnyNonZeroIndices = false;
356 // Split AddRecs up into parts as either of the parts may be usable
357 // without the other.
358 SplitAddRecs(Ops, Ty, SE);
360 // Descend down the pointer's type and attempt to convert the other
361 // operands into GEP indices, at each level. The first index in a GEP
362 // indexes into the array implied by the pointer operand; the rest of
363 // the indices index into the element or field type selected by the
366 const SCEV *ElSize = SE.getAllocSizeExpr(ElTy);
367 // If the scale size is not 0, attempt to factor out a scale for
369 SmallVector<const SCEV *, 8> ScaledOps;
370 if (ElTy->isSized() && !ElSize->isZero()) {
371 SmallVector<const SCEV *, 8> NewOps;
372 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
373 const SCEV *Op = Ops[i];
374 const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
375 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
376 // Op now has ElSize factored out.
377 ScaledOps.push_back(Op);
378 if (!Remainder->isZero())
379 NewOps.push_back(Remainder);
380 AnyNonZeroIndices = true;
382 // The operand was not divisible, so add it to the list of operands
383 // we'll scan next iteration.
384 NewOps.push_back(Ops[i]);
387 // If we made any changes, update Ops.
388 if (!ScaledOps.empty()) {
390 SimplifyAddOperands(Ops, Ty, SE);
394 // Record the scaled array index for this level of the type. If
395 // we didn't find any operands that could be factored, tentatively
396 // assume that element zero was selected (since the zero offset
397 // would obviously be folded away).
398 Value *Scaled = ScaledOps.empty() ?
399 Constant::getNullValue(Ty) :
400 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
401 GepIndices.push_back(Scaled);
403 // Collect struct field index operands.
404 while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
405 bool FoundFieldNo = false;
406 // An empty struct has no fields.
407 if (STy->getNumElements() == 0) break;
409 // With TargetData, field offsets are known. See if a constant offset
410 // falls within any of the struct fields.
411 if (Ops.empty()) break;
412 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
413 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
414 const StructLayout &SL = *SE.TD->getStructLayout(STy);
415 uint64_t FullOffset = C->getValue()->getZExtValue();
416 if (FullOffset < SL.getSizeInBytes()) {
417 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
418 GepIndices.push_back(
419 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
420 ElTy = STy->getTypeAtIndex(ElIdx);
422 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
423 AnyNonZeroIndices = true;
428 // Without TargetData, just check for a SCEVFieldOffsetExpr of the
429 // appropriate struct type.
430 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
431 if (const SCEVFieldOffsetExpr *FO =
432 dyn_cast<SCEVFieldOffsetExpr>(Ops[i]))
433 if (FO->getStructType() == STy) {
434 unsigned FieldNo = FO->getFieldNo();
435 GepIndices.push_back(
436 ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
438 ElTy = STy->getTypeAtIndex(FieldNo);
439 Ops[i] = SE.getConstant(Ty, 0);
440 AnyNonZeroIndices = true;
445 // If no struct field offsets were found, tentatively assume that
446 // field zero was selected (since the zero offset would obviously
449 ElTy = STy->getTypeAtIndex(0u);
450 GepIndices.push_back(
451 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
455 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
456 ElTy = ATy->getElementType();
461 // If none of the operands were convertable to proper GEP indices, cast
462 // the base to i8* and do an ugly getelementptr with that. It's still
463 // better than ptrtoint+arithmetic+inttoptr at least.
464 if (!AnyNonZeroIndices) {
465 // Cast the base to i8*.
466 V = InsertNoopCastOfTo(V,
467 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
469 // Expand the operands for a plain byte offset.
470 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
472 // Fold a GEP with constant operands.
473 if (Constant *CLHS = dyn_cast<Constant>(V))
474 if (Constant *CRHS = dyn_cast<Constant>(Idx))
475 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
477 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
478 unsigned ScanLimit = 6;
479 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
480 // Scanning starts from the last instruction before the insertion point.
481 BasicBlock::iterator IP = Builder.GetInsertPoint();
482 if (IP != BlockBegin) {
484 for (; ScanLimit; --IP, --ScanLimit) {
485 if (IP->getOpcode() == Instruction::GetElementPtr &&
486 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
488 if (IP == BlockBegin) break;
493 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
494 rememberInstruction(GEP);
498 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
499 // because ScalarEvolution may have changed the address arithmetic to
500 // compute a value which is beyond the end of the allocated object.
502 if (V->getType() != PTy)
503 Casted = InsertNoopCastOfTo(Casted, PTy);
504 Value *GEP = Builder.CreateGEP(Casted,
508 Ops.push_back(SE.getUnknown(GEP));
509 rememberInstruction(GEP);
510 return expand(SE.getAddExpr(Ops));
513 /// isNonConstantNegative - Return true if the specified scev is negated, but
515 static bool isNonConstantNegative(const SCEV *F) {
516 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
517 if (!Mul) return false;
519 // If there is a constant factor, it will be first.
520 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
521 if (!SC) return false;
523 // Return true if the value is negative, this matches things like (-42 * V).
524 return SC->getValue()->getValue().isNegative();
527 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
528 int NumOperands = S->getNumOperands();
529 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
531 // Find the index of an operand to start with. Choose the operand with
532 // pointer type, if there is one, or the last operand otherwise.
534 for (; PIdx != NumOperands - 1; ++PIdx)
535 if (isa<PointerType>(S->getOperand(PIdx)->getType())) break;
537 // Expand code for the operand that we chose.
538 Value *V = expand(S->getOperand(PIdx));
540 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
541 // comments on expandAddToGEP for details.
542 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
543 // Take the operand at PIdx out of the list.
544 const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
545 SmallVector<const SCEV *, 8> NewOps;
546 NewOps.insert(NewOps.end(), Ops.begin(), Ops.begin() + PIdx);
547 NewOps.insert(NewOps.end(), Ops.begin() + PIdx + 1, Ops.end());
549 return expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, V);
552 // Otherwise, we'll expand the rest of the SCEVAddExpr as plain integer
554 V = InsertNoopCastOfTo(V, Ty);
556 // Emit a bunch of add instructions
557 for (int i = NumOperands-1; i >= 0; --i) {
558 if (i == PIdx) continue;
559 const SCEV *Op = S->getOperand(i);
560 if (isNonConstantNegative(Op)) {
561 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
562 V = InsertBinop(Instruction::Sub, V, W);
564 Value *W = expandCodeFor(Op, Ty);
565 V = InsertBinop(Instruction::Add, V, W);
571 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
572 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
573 int FirstOp = 0; // Set if we should emit a subtract.
574 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
575 if (SC->getValue()->isAllOnesValue())
578 int i = S->getNumOperands()-2;
579 Value *V = expandCodeFor(S->getOperand(i+1), Ty);
581 // Emit a bunch of multiply instructions
582 for (; i >= FirstOp; --i) {
583 Value *W = expandCodeFor(S->getOperand(i), Ty);
584 V = InsertBinop(Instruction::Mul, V, W);
587 // -1 * ... ---> 0 - ...
589 V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
593 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
594 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
596 Value *LHS = expandCodeFor(S->getLHS(), Ty);
597 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
598 const APInt &RHS = SC->getValue()->getValue();
599 if (RHS.isPowerOf2())
600 return InsertBinop(Instruction::LShr, LHS,
601 ConstantInt::get(Ty, RHS.logBase2()));
604 Value *RHS = expandCodeFor(S->getRHS(), Ty);
605 return InsertBinop(Instruction::UDiv, LHS, RHS);
608 /// Move parts of Base into Rest to leave Base with the minimal
609 /// expression that provides a pointer operand suitable for a
611 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
612 ScalarEvolution &SE) {
613 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
614 Base = A->getStart();
615 Rest = SE.getAddExpr(Rest,
616 SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
617 A->getStepRecurrence(SE),
620 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
621 Base = A->getOperand(A->getNumOperands()-1);
622 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
623 NewAddOps.back() = Rest;
624 Rest = SE.getAddExpr(NewAddOps);
625 ExposePointerBase(Base, Rest, SE);
629 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
630 /// the base addrec, which is the addrec without any non-loop-dominating
631 /// values, and return the PHI.
633 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
635 const Type *ExpandTy,
637 // Reuse a previously-inserted PHI, if present.
638 for (BasicBlock::iterator I = L->getHeader()->begin();
639 PHINode *PN = dyn_cast<PHINode>(I); ++I)
640 if (isInsertedInstruction(PN) && SE.getSCEV(PN) == Normalized)
643 // Save the original insertion point so we can restore it when we're done.
644 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
645 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
647 // Expand code for the start value.
648 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
649 L->getHeader()->begin());
651 // Expand code for the step value. Insert instructions right before the
652 // terminator corresponding to the back-edge. Do this before creating the PHI
653 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
654 // negative, insert a sub instead of an add for the increment (unless it's a
655 // constant, because subtracts of constants are canonicalized to adds).
656 const SCEV *Step = Normalized->getStepRecurrence(SE);
657 bool isPointer = isa<PointerType>(ExpandTy);
658 bool isNegative = !isPointer && isNonConstantNegative(Step);
660 Step = SE.getNegativeSCEV(Step);
661 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
664 Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin());
665 PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv");
666 rememberInstruction(PN);
668 // Create the step instructions and populate the PHI.
669 BasicBlock *Header = L->getHeader();
670 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
672 BasicBlock *Pred = *HPI;
674 // Add a start value.
675 if (!L->contains(Pred)) {
676 PN->addIncoming(StartV, Pred);
680 // Create a step value and add it to the PHI. If IVIncInsertLoop is
681 // non-null and equal to the addrec's loop, insert the instructions
682 // at IVIncInsertPos.
683 Instruction *InsertPos = L == IVIncInsertLoop ?
684 IVIncInsertPos : Pred->getTerminator();
685 Builder.SetInsertPoint(InsertPos->getParent(), InsertPos);
687 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
689 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
690 // If the step isn't constant, don't use an implicitly scaled GEP, because
691 // that would require a multiply inside the loop.
692 if (!isa<ConstantInt>(StepV))
693 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
694 GEPPtrTy->getAddressSpace());
695 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
696 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
697 if (IncV->getType() != PN->getType()) {
698 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
699 rememberInstruction(IncV);
703 Builder.CreateSub(PN, StepV, "lsr.iv.next") :
704 Builder.CreateAdd(PN, StepV, "lsr.iv.next");
705 rememberInstruction(IncV);
707 PN->addIncoming(IncV, Pred);
710 // Restore the original insert point.
712 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
714 // Remember this PHI, even in post-inc mode.
715 InsertedValues.insert(PN);
720 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
721 const Type *STy = S->getType();
722 const Type *IntTy = SE.getEffectiveSCEVType(STy);
723 const Loop *L = S->getLoop();
725 // Determine a normalized form of this expression, which is the expression
726 // before any post-inc adjustment is made.
727 const SCEVAddRecExpr *Normalized = S;
728 if (L == PostIncLoop) {
729 const SCEV *Step = S->getStepRecurrence(SE);
730 Normalized = cast<SCEVAddRecExpr>(SE.getMinusSCEV(S, Step));
733 // Strip off any non-loop-dominating component from the addrec start.
734 const SCEV *Start = Normalized->getStart();
735 const SCEV *PostLoopOffset = 0;
736 if (!Start->properlyDominates(L->getHeader(), SE.DT)) {
737 PostLoopOffset = Start;
738 Start = SE.getIntegerSCEV(0, Normalized->getType());
740 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start,
741 Normalized->getStepRecurrence(SE),
742 Normalized->getLoop()));
745 // Strip off any non-loop-dominating component from the addrec step.
746 const SCEV *Step = Normalized->getStepRecurrence(SE);
747 const SCEV *PostLoopScale = 0;
748 if (!Step->hasComputableLoopEvolution(L) &&
749 !Step->dominates(L->getHeader(), SE.DT)) {
750 PostLoopScale = Step;
751 Step = SE.getIntegerSCEV(1, Normalized->getType());
753 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
754 Normalized->getLoop()));
757 // Expand the core addrec. If we need post-loop scaling, force it to
758 // expand to an integer type to avoid the need for additional casting.
759 const Type *ExpandTy = PostLoopScale ? IntTy : STy;
760 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
762 // Accomodate post-inc mode, if necessary.
764 if (L != PostIncLoop)
767 // In PostInc mode, use the post-incremented value.
768 BasicBlock *LatchBlock = L->getLoopLatch();
769 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
770 Result = PN->getIncomingValueForBlock(LatchBlock);
773 // Re-apply any non-loop-dominating scale.
775 Result = Builder.CreateMul(Result,
776 expandCodeFor(PostLoopScale, IntTy));
777 rememberInstruction(Result);
780 // Re-apply any non-loop-dominating offset.
781 if (PostLoopOffset) {
782 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
783 const SCEV *const OffsetArray[1] = { PostLoopOffset };
784 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
786 Result = Builder.CreateAdd(Result,
787 expandCodeFor(PostLoopOffset, IntTy));
788 rememberInstruction(Result);
795 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
796 if (!CanonicalMode) return expandAddRecExprLiterally(S);
798 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
799 const Loop *L = S->getLoop();
801 // First check for an existing canonical IV in a suitable type.
802 PHINode *CanonicalIV = 0;
803 if (PHINode *PN = L->getCanonicalInductionVariable())
804 if (SE.isSCEVable(PN->getType()) &&
805 isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
806 SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
809 // Rewrite an AddRec in terms of the canonical induction variable, if
810 // its type is more narrow.
812 SE.getTypeSizeInBits(CanonicalIV->getType()) >
813 SE.getTypeSizeInBits(Ty)) {
814 const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
815 SmallVector<const SCEV *, 4> NewOps(Ops.size());
816 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
817 NewOps[i] = SE.getAnyExtendExpr(Ops[i], CanonicalIV->getType());
818 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop()));
819 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
820 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
821 BasicBlock::iterator NewInsertPt =
822 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
823 while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
824 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
826 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
830 // {X,+,F} --> X + {0,+,F}
831 if (!S->getStart()->isZero()) {
832 const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
833 SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
834 NewOps[0] = SE.getIntegerSCEV(0, Ty);
835 const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
837 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
838 // comments on expandAddToGEP for details.
839 const SCEV *Base = S->getStart();
840 const SCEV *RestArray[1] = { Rest };
841 // Dig into the expression to find the pointer base for a GEP.
842 ExposePointerBase(Base, RestArray[0], SE);
843 // If we found a pointer, expand the AddRec with a GEP.
844 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
845 // Make sure the Base isn't something exotic, such as a multiplied
846 // or divided pointer value. In those cases, the result type isn't
847 // actually a pointer type.
848 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
849 Value *StartV = expand(Base);
850 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
851 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
855 // Just do a normal add. Pre-expand the operands to suppress folding.
856 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
857 SE.getUnknown(expand(Rest))));
860 // {0,+,1} --> Insert a canonical induction variable into the loop!
862 S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
863 // If there's a canonical IV, just use it.
865 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
866 "IVs with types different from the canonical IV should "
867 "already have been handled!");
871 // Create and insert the PHI node for the induction variable in the
873 BasicBlock *Header = L->getHeader();
874 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
875 rememberInstruction(PN);
877 Constant *One = ConstantInt::get(Ty, 1);
878 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
880 if (L->contains(*HPI)) {
881 // Insert a unit add instruction right before the terminator
882 // corresponding to the back-edge.
883 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
884 (*HPI)->getTerminator());
885 rememberInstruction(Add);
886 PN->addIncoming(Add, *HPI);
888 PN->addIncoming(Constant::getNullValue(Ty), *HPI);
892 // {0,+,F} --> {0,+,1} * F
893 // Get the canonical induction variable I for this loop.
894 Value *I = CanonicalIV ?
896 getOrInsertCanonicalInductionVariable(L, Ty);
898 // If this is a simple linear addrec, emit it now as a special case.
899 if (S->isAffine()) // {0,+,F} --> i*F
901 expand(SE.getTruncateOrNoop(
902 SE.getMulExpr(SE.getUnknown(I),
903 SE.getNoopOrAnyExtend(S->getOperand(1),
907 // If this is a chain of recurrences, turn it into a closed form, using the
908 // folders, then expandCodeFor the closed form. This allows the folders to
909 // simplify the expression without having to build a bunch of special code
911 const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
913 // Promote S up to the canonical IV type, if the cast is foldable.
914 const SCEV *NewS = S;
915 const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
916 if (isa<SCEVAddRecExpr>(Ext))
919 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
920 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
922 // Truncate the result down to the original type, if needed.
923 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
927 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
928 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
929 Value *V = expandCodeFor(S->getOperand(),
930 SE.getEffectiveSCEVType(S->getOperand()->getType()));
931 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
932 rememberInstruction(I);
936 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
937 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
938 Value *V = expandCodeFor(S->getOperand(),
939 SE.getEffectiveSCEVType(S->getOperand()->getType()));
940 Value *I = Builder.CreateZExt(V, Ty, "tmp");
941 rememberInstruction(I);
945 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
946 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
947 Value *V = expandCodeFor(S->getOperand(),
948 SE.getEffectiveSCEVType(S->getOperand()->getType()));
949 Value *I = Builder.CreateSExt(V, Ty, "tmp");
950 rememberInstruction(I);
954 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
955 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
956 const Type *Ty = LHS->getType();
957 for (int i = S->getNumOperands()-2; i >= 0; --i) {
958 // In the case of mixed integer and pointer types, do the
959 // rest of the comparisons as integer.
960 if (S->getOperand(i)->getType() != Ty) {
961 Ty = SE.getEffectiveSCEVType(Ty);
962 LHS = InsertNoopCastOfTo(LHS, Ty);
964 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
965 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
966 rememberInstruction(ICmp);
967 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
968 rememberInstruction(Sel);
971 // In the case of mixed integer and pointer types, cast the
972 // final result back to the pointer type.
973 if (LHS->getType() != S->getType())
974 LHS = InsertNoopCastOfTo(LHS, S->getType());
978 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
979 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
980 const Type *Ty = LHS->getType();
981 for (int i = S->getNumOperands()-2; i >= 0; --i) {
982 // In the case of mixed integer and pointer types, do the
983 // rest of the comparisons as integer.
984 if (S->getOperand(i)->getType() != Ty) {
985 Ty = SE.getEffectiveSCEVType(Ty);
986 LHS = InsertNoopCastOfTo(LHS, Ty);
988 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
989 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
990 rememberInstruction(ICmp);
991 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
992 rememberInstruction(Sel);
995 // In the case of mixed integer and pointer types, cast the
996 // final result back to the pointer type.
997 if (LHS->getType() != S->getType())
998 LHS = InsertNoopCastOfTo(LHS, S->getType());
1002 Value *SCEVExpander::visitFieldOffsetExpr(const SCEVFieldOffsetExpr *S) {
1003 return ConstantExpr::getOffsetOf(S->getStructType(), S->getFieldNo());
1006 Value *SCEVExpander::visitAllocSizeExpr(const SCEVAllocSizeExpr *S) {
1007 return ConstantExpr::getSizeOf(S->getAllocType());
1010 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
1011 // Expand the code for this SCEV.
1012 Value *V = expand(SH);
1014 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1015 "non-trivial casts should be done with the SCEVs directly!");
1016 V = InsertNoopCastOfTo(V, Ty);
1021 Value *SCEVExpander::expand(const SCEV *S) {
1022 // Compute an insertion point for this SCEV object. Hoist the instructions
1023 // as far out in the loop nest as possible.
1024 Instruction *InsertPt = Builder.GetInsertPoint();
1025 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1026 L = L->getParentLoop())
1027 if (S->isLoopInvariant(L)) {
1029 if (BasicBlock *Preheader = L->getLoopPreheader())
1030 InsertPt = Preheader->getTerminator();
1032 // If the SCEV is computable at this level, insert it into the header
1033 // after the PHIs (and after any other instructions that we've inserted
1034 // there) so that it is guaranteed to dominate any user inside the loop.
1035 if (L && S->hasComputableLoopEvolution(L))
1036 InsertPt = L->getHeader()->getFirstNonPHI();
1037 while (isInsertedInstruction(InsertPt))
1038 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1042 // Check to see if we already expanded this here.
1043 std::map<std::pair<const SCEV *, Instruction *>,
1044 AssertingVH<Value> >::iterator I =
1045 InsertedExpressions.find(std::make_pair(S, InsertPt));
1046 if (I != InsertedExpressions.end())
1049 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1050 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1051 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1053 // Expand the expression into instructions.
1054 Value *V = visit(S);
1056 // Remember the expanded value for this SCEV at this location.
1058 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1060 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
1064 /// getOrInsertCanonicalInductionVariable - This method returns the
1065 /// canonical induction variable of the specified type for the specified
1066 /// loop (inserting one if there is none). A canonical induction variable
1067 /// starts at zero and steps by one on each iteration.
1069 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1071 assert(Ty->isInteger() && "Can only insert integer induction variables!");
1072 const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
1073 SE.getIntegerSCEV(1, Ty), L);
1074 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1075 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1076 Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
1078 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);