1 //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the implementation of the scalar evolution expander,
11 // which is used to generate the code corresponding to a given scalar evolution
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Analysis/ScalarEvolutionExpander.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/Analysis/LoopInfo.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/LLVMContext.h"
24 #include "llvm/Support/Debug.h"
28 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
29 /// reusing an existing cast if a suitable one exists, moving an existing
30 /// cast if a suitable one exists but isn't in the right place, or
31 /// creating a new one.
32 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
33 Instruction::CastOps Op,
34 BasicBlock::iterator IP) {
35 // This function must be called with the builder having a valid insertion
36 // point. It doesn't need to be the actual IP where the uses of the returned
37 // cast will be added, but it must dominate such IP.
38 // We use this precondition to produce a cast that will dominate all its
39 // uses. In particular, this is crucial for the case where the builder's
40 // insertion point *is* the point where we were asked to put the cast.
41 // Since we don't know the builder's insertion point is actually
42 // where the uses will be added (only that it dominates it), we are
43 // not allowed to move it.
44 BasicBlock::iterator BIP = Builder.GetInsertPoint();
46 Instruction *Ret = NULL;
48 // Check to see if there is already a cast!
49 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
52 if (U->getType() == Ty)
53 if (CastInst *CI = dyn_cast<CastInst>(U))
54 if (CI->getOpcode() == Op) {
55 // If the cast isn't where we want it, create a new cast at IP.
56 // Likewise, do not reuse a cast at BIP because it must dominate
57 // instructions that might be inserted before BIP.
58 if (BasicBlock::iterator(CI) != IP || BIP == IP) {
59 // Create a new cast, and leave the old cast in place in case
60 // it is being used as an insert point. Clear its operand
61 // so that it doesn't hold anything live.
62 Ret = CastInst::Create(Op, V, Ty, "", IP);
64 CI->replaceAllUsesWith(Ret);
65 CI->setOperand(0, UndefValue::get(V->getType()));
75 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP);
77 // We assert at the end of the function since IP might point to an
78 // instruction with different dominance properties than a cast
79 // (an invoke for example) and not dominate BIP (but the cast does).
80 assert(SE.DT->dominates(Ret, BIP));
82 rememberInstruction(Ret);
86 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
87 /// which must be possible with a noop cast, doing what we can to share
89 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
90 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
91 assert((Op == Instruction::BitCast ||
92 Op == Instruction::PtrToInt ||
93 Op == Instruction::IntToPtr) &&
94 "InsertNoopCastOfTo cannot perform non-noop casts!");
95 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
96 "InsertNoopCastOfTo cannot change sizes!");
98 // Short-circuit unnecessary bitcasts.
99 if (Op == Instruction::BitCast) {
100 if (V->getType() == Ty)
102 if (CastInst *CI = dyn_cast<CastInst>(V)) {
103 if (CI->getOperand(0)->getType() == Ty)
104 return CI->getOperand(0);
107 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
108 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
109 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
110 if (CastInst *CI = dyn_cast<CastInst>(V))
111 if ((CI->getOpcode() == Instruction::PtrToInt ||
112 CI->getOpcode() == Instruction::IntToPtr) &&
113 SE.getTypeSizeInBits(CI->getType()) ==
114 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
115 return CI->getOperand(0);
116 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
117 if ((CE->getOpcode() == Instruction::PtrToInt ||
118 CE->getOpcode() == Instruction::IntToPtr) &&
119 SE.getTypeSizeInBits(CE->getType()) ==
120 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
121 return CE->getOperand(0);
124 // Fold a cast of a constant.
125 if (Constant *C = dyn_cast<Constant>(V))
126 return ConstantExpr::getCast(Op, C, Ty);
128 // Cast the argument at the beginning of the entry block, after
129 // any bitcasts of other arguments.
130 if (Argument *A = dyn_cast<Argument>(V)) {
131 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
132 while ((isa<BitCastInst>(IP) &&
133 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
134 cast<BitCastInst>(IP)->getOperand(0) != A) ||
135 isa<DbgInfoIntrinsic>(IP) ||
136 isa<LandingPadInst>(IP))
138 return ReuseOrCreateCast(A, Ty, Op, IP);
141 // Cast the instruction immediately after the instruction.
142 Instruction *I = cast<Instruction>(V);
143 BasicBlock::iterator IP = I; ++IP;
144 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
145 IP = II->getNormalDest()->begin();
146 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
148 return ReuseOrCreateCast(I, Ty, Op, IP);
151 /// InsertBinop - Insert the specified binary operator, doing a small amount
152 /// of work to avoid inserting an obviously redundant operation.
153 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
154 Value *LHS, Value *RHS) {
155 // Fold a binop with constant operands.
156 if (Constant *CLHS = dyn_cast<Constant>(LHS))
157 if (Constant *CRHS = dyn_cast<Constant>(RHS))
158 return ConstantExpr::get(Opcode, CLHS, CRHS);
160 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
161 unsigned ScanLimit = 6;
162 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
163 // Scanning starts from the last instruction before the insertion point.
164 BasicBlock::iterator IP = Builder.GetInsertPoint();
165 if (IP != BlockBegin) {
167 for (; ScanLimit; --IP, --ScanLimit) {
168 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
170 if (isa<DbgInfoIntrinsic>(IP))
172 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
173 IP->getOperand(1) == RHS)
175 if (IP == BlockBegin) break;
179 // Save the original insertion point so we can restore it when we're done.
180 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc();
181 BuilderType::InsertPointGuard Guard(Builder);
183 // Move the insertion point out of as many loops as we can.
184 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
185 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
186 BasicBlock *Preheader = L->getLoopPreheader();
187 if (!Preheader) break;
189 // Ok, move up a level.
190 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
193 // If we haven't found this binop, insert it.
194 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
195 BO->setDebugLoc(Loc);
196 rememberInstruction(BO);
201 /// FactorOutConstant - Test if S is divisible by Factor, using signed
202 /// division. If so, update S with Factor divided out and return true.
203 /// S need not be evenly divisible if a reasonable remainder can be
205 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
206 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
207 /// check to see if the divide was folded.
208 static bool FactorOutConstant(const SCEV *&S,
209 const SCEV *&Remainder,
212 const DataLayout *TD) {
213 // Everything is divisible by one.
219 S = SE.getConstant(S->getType(), 1);
223 // For a Constant, check for a multiple of the given factor.
224 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
228 // Check for divisibility.
229 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
231 ConstantInt::get(SE.getContext(),
232 C->getValue()->getValue().sdiv(
233 FC->getValue()->getValue()));
234 // If the quotient is zero and the remainder is non-zero, reject
235 // the value at this scale. It will be considered for subsequent
238 const SCEV *Div = SE.getConstant(CI);
241 SE.getAddExpr(Remainder,
242 SE.getConstant(C->getValue()->getValue().srem(
243 FC->getValue()->getValue())));
249 // In a Mul, check if there is a constant operand which is a multiple
250 // of the given factor.
251 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
253 // With DataLayout, the size is known. Check if there is a constant
254 // operand which is a multiple of the given factor. If so, we can
256 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
257 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
258 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
259 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
261 SE.getConstant(C->getValue()->getValue().sdiv(
262 FC->getValue()->getValue()));
263 S = SE.getMulExpr(NewMulOps);
267 // Without DataLayout, check if Factor can be factored out of any of the
268 // Mul's operands. If so, we can just remove it.
269 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
270 const SCEV *SOp = M->getOperand(i);
271 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
272 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
273 Remainder->isZero()) {
274 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
276 S = SE.getMulExpr(NewMulOps);
283 // In an AddRec, check if both start and step are divisible.
284 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
285 const SCEV *Step = A->getStepRecurrence(SE);
286 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
287 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
289 if (!StepRem->isZero())
291 const SCEV *Start = A->getStart();
292 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
294 S = SE.getAddRecExpr(Start, Step, A->getLoop(),
295 A->getNoWrapFlags(SCEV::FlagNW));
302 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
303 /// is the number of SCEVAddRecExprs present, which are kept at the end of
306 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
308 ScalarEvolution &SE) {
309 unsigned NumAddRecs = 0;
310 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
312 // Group Ops into non-addrecs and addrecs.
313 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
314 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
315 // Let ScalarEvolution sort and simplify the non-addrecs list.
316 const SCEV *Sum = NoAddRecs.empty() ?
317 SE.getConstant(Ty, 0) :
318 SE.getAddExpr(NoAddRecs);
319 // If it returned an add, use the operands. Otherwise it simplified
320 // the sum into a single value, so just use that.
322 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
323 Ops.append(Add->op_begin(), Add->op_end());
324 else if (!Sum->isZero())
326 // Then append the addrecs.
327 Ops.append(AddRecs.begin(), AddRecs.end());
330 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
331 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
332 /// This helps expose more opportunities for folding parts of the expressions
333 /// into GEP indices.
335 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
337 ScalarEvolution &SE) {
339 SmallVector<const SCEV *, 8> AddRecs;
340 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
341 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
342 const SCEV *Start = A->getStart();
343 if (Start->isZero()) break;
344 const SCEV *Zero = SE.getConstant(Ty, 0);
345 AddRecs.push_back(SE.getAddRecExpr(Zero,
346 A->getStepRecurrence(SE),
348 A->getNoWrapFlags(SCEV::FlagNW)));
349 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
351 Ops.append(Add->op_begin(), Add->op_end());
352 e += Add->getNumOperands();
357 if (!AddRecs.empty()) {
358 // Add the addrecs onto the end of the list.
359 Ops.append(AddRecs.begin(), AddRecs.end());
360 // Resort the operand list, moving any constants to the front.
361 SimplifyAddOperands(Ops, Ty, SE);
365 /// expandAddToGEP - Expand an addition expression with a pointer type into
366 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
367 /// BasicAliasAnalysis and other passes analyze the result. See the rules
368 /// for getelementptr vs. inttoptr in
369 /// http://llvm.org/docs/LangRef.html#pointeraliasing
372 /// Design note: The correctness of using getelementptr here depends on
373 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
374 /// they may introduce pointer arithmetic which may not be safely converted
375 /// into getelementptr.
377 /// Design note: It might seem desirable for this function to be more
378 /// loop-aware. If some of the indices are loop-invariant while others
379 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
380 /// loop-invariant portions of the overall computation outside the loop.
381 /// However, there are a few reasons this is not done here. Hoisting simple
382 /// arithmetic is a low-level optimization that often isn't very
383 /// important until late in the optimization process. In fact, passes
384 /// like InstructionCombining will combine GEPs, even if it means
385 /// pushing loop-invariant computation down into loops, so even if the
386 /// GEPs were split here, the work would quickly be undone. The
387 /// LoopStrengthReduction pass, which is usually run quite late (and
388 /// after the last InstructionCombining pass), takes care of hoisting
389 /// loop-invariant portions of expressions, after considering what
390 /// can be folded using target addressing modes.
392 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
393 const SCEV *const *op_end,
397 Type *ElTy = PTy->getElementType();
398 SmallVector<Value *, 4> GepIndices;
399 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
400 bool AnyNonZeroIndices = false;
402 // Split AddRecs up into parts as either of the parts may be usable
403 // without the other.
404 SplitAddRecs(Ops, Ty, SE);
406 Type *IntPtrTy = SE.TD
407 ? SE.TD->getIntPtrType(PTy)
408 : Type::getInt64Ty(PTy->getContext());
410 // Descend down the pointer's type and attempt to convert the other
411 // operands into GEP indices, at each level. The first index in a GEP
412 // indexes into the array implied by the pointer operand; the rest of
413 // the indices index into the element or field type selected by the
416 // If the scale size is not 0, attempt to factor out a scale for
418 SmallVector<const SCEV *, 8> ScaledOps;
419 if (ElTy->isSized()) {
420 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy);
421 if (!ElSize->isZero()) {
422 SmallVector<const SCEV *, 8> NewOps;
423 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
424 const SCEV *Op = Ops[i];
425 const SCEV *Remainder = SE.getConstant(Ty, 0);
426 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
427 // Op now has ElSize factored out.
428 ScaledOps.push_back(Op);
429 if (!Remainder->isZero())
430 NewOps.push_back(Remainder);
431 AnyNonZeroIndices = true;
433 // The operand was not divisible, so add it to the list of operands
434 // we'll scan next iteration.
435 NewOps.push_back(Ops[i]);
438 // If we made any changes, update Ops.
439 if (!ScaledOps.empty()) {
441 SimplifyAddOperands(Ops, Ty, SE);
446 // Record the scaled array index for this level of the type. If
447 // we didn't find any operands that could be factored, tentatively
448 // assume that element zero was selected (since the zero offset
449 // would obviously be folded away).
450 Value *Scaled = ScaledOps.empty() ?
451 Constant::getNullValue(Ty) :
452 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
453 GepIndices.push_back(Scaled);
455 // Collect struct field index operands.
456 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
457 bool FoundFieldNo = false;
458 // An empty struct has no fields.
459 if (STy->getNumElements() == 0) break;
461 // With DataLayout, field offsets are known. See if a constant offset
462 // falls within any of the struct fields.
463 if (Ops.empty()) break;
464 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
465 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
466 const StructLayout &SL = *SE.TD->getStructLayout(STy);
467 uint64_t FullOffset = C->getValue()->getZExtValue();
468 if (FullOffset < SL.getSizeInBytes()) {
469 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
470 GepIndices.push_back(
471 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
472 ElTy = STy->getTypeAtIndex(ElIdx);
474 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
475 AnyNonZeroIndices = true;
480 // Without DataLayout, just check for an offsetof expression of the
481 // appropriate struct type.
482 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
483 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
486 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
487 GepIndices.push_back(FieldNo);
489 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
490 Ops[i] = SE.getConstant(Ty, 0);
491 AnyNonZeroIndices = true;
497 // If no struct field offsets were found, tentatively assume that
498 // field zero was selected (since the zero offset would obviously
501 ElTy = STy->getTypeAtIndex(0u);
502 GepIndices.push_back(
503 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
507 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
508 ElTy = ATy->getElementType();
513 // If none of the operands were convertible to proper GEP indices, cast
514 // the base to i8* and do an ugly getelementptr with that. It's still
515 // better than ptrtoint+arithmetic+inttoptr at least.
516 if (!AnyNonZeroIndices) {
517 // Cast the base to i8*.
518 V = InsertNoopCastOfTo(V,
519 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
521 assert(!isa<Instruction>(V) ||
522 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint()));
524 // Expand the operands for a plain byte offset.
525 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
527 // Fold a GEP with constant operands.
528 if (Constant *CLHS = dyn_cast<Constant>(V))
529 if (Constant *CRHS = dyn_cast<Constant>(Idx))
530 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
532 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
533 unsigned ScanLimit = 6;
534 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
535 // Scanning starts from the last instruction before the insertion point.
536 BasicBlock::iterator IP = Builder.GetInsertPoint();
537 if (IP != BlockBegin) {
539 for (; ScanLimit; --IP, --ScanLimit) {
540 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
542 if (isa<DbgInfoIntrinsic>(IP))
544 if (IP->getOpcode() == Instruction::GetElementPtr &&
545 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
547 if (IP == BlockBegin) break;
551 // Save the original insertion point so we can restore it when we're done.
552 BuilderType::InsertPointGuard Guard(Builder);
554 // Move the insertion point out of as many loops as we can.
555 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
556 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
557 BasicBlock *Preheader = L->getLoopPreheader();
558 if (!Preheader) break;
560 // Ok, move up a level.
561 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
565 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
566 rememberInstruction(GEP);
571 // Save the original insertion point so we can restore it when we're done.
572 BuilderType::InsertPointGuard Guard(Builder);
574 // Move the insertion point out of as many loops as we can.
575 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
576 if (!L->isLoopInvariant(V)) break;
578 bool AnyIndexNotLoopInvariant = false;
579 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
580 E = GepIndices.end(); I != E; ++I)
581 if (!L->isLoopInvariant(*I)) {
582 AnyIndexNotLoopInvariant = true;
585 if (AnyIndexNotLoopInvariant)
588 BasicBlock *Preheader = L->getLoopPreheader();
589 if (!Preheader) break;
591 // Ok, move up a level.
592 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
595 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
596 // because ScalarEvolution may have changed the address arithmetic to
597 // compute a value which is beyond the end of the allocated object.
599 if (V->getType() != PTy)
600 Casted = InsertNoopCastOfTo(Casted, PTy);
601 Value *GEP = Builder.CreateGEP(Casted,
604 Ops.push_back(SE.getUnknown(GEP));
605 rememberInstruction(GEP);
607 return expand(SE.getAddExpr(Ops));
610 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
611 /// SCEV expansion. If they are nested, this is the most nested. If they are
612 /// neighboring, pick the later.
613 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
617 if (A->contains(B)) return B;
618 if (B->contains(A)) return A;
619 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
620 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
621 return A; // Arbitrarily break the tie.
624 /// getRelevantLoop - Get the most relevant loop associated with the given
625 /// expression, according to PickMostRelevantLoop.
626 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
627 // Test whether we've already computed the most relevant loop for this SCEV.
628 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
629 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
631 return Pair.first->second;
633 if (isa<SCEVConstant>(S))
634 // A constant has no relevant loops.
636 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
637 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
638 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
639 // A non-instruction has no relevant loops.
642 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
644 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
646 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
648 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
649 return RelevantLoops[N] = L;
651 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
652 const Loop *Result = getRelevantLoop(C->getOperand());
653 return RelevantLoops[C] = Result;
655 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
657 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
658 getRelevantLoop(D->getRHS()),
660 return RelevantLoops[D] = Result;
662 llvm_unreachable("Unexpected SCEV type!");
667 /// LoopCompare - Compare loops by PickMostRelevantLoop.
671 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
673 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
674 std::pair<const Loop *, const SCEV *> RHS) const {
675 // Keep pointer operands sorted at the end.
676 if (LHS.second->getType()->isPointerTy() !=
677 RHS.second->getType()->isPointerTy())
678 return LHS.second->getType()->isPointerTy();
680 // Compare loops with PickMostRelevantLoop.
681 if (LHS.first != RHS.first)
682 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
684 // If one operand is a non-constant negative and the other is not,
685 // put the non-constant negative on the right so that a sub can
686 // be used instead of a negate and add.
687 if (LHS.second->isNonConstantNegative()) {
688 if (!RHS.second->isNonConstantNegative())
690 } else if (RHS.second->isNonConstantNegative())
693 // Otherwise they are equivalent according to this comparison.
700 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
701 Type *Ty = SE.getEffectiveSCEVType(S->getType());
703 // Collect all the add operands in a loop, along with their associated loops.
704 // Iterate in reverse so that constants are emitted last, all else equal, and
705 // so that pointer operands are inserted first, which the code below relies on
706 // to form more involved GEPs.
707 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
708 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
709 E(S->op_begin()); I != E; ++I)
710 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
712 // Sort by loop. Use a stable sort so that constants follow non-constants and
713 // pointer operands precede non-pointer operands.
714 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
716 // Emit instructions to add all the operands. Hoist as much as possible
717 // out of loops, and form meaningful getelementptrs where possible.
719 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
720 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
721 const Loop *CurLoop = I->first;
722 const SCEV *Op = I->second;
724 // This is the first operand. Just expand it.
727 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
728 // The running sum expression is a pointer. Try to form a getelementptr
729 // at this level with that as the base.
730 SmallVector<const SCEV *, 4> NewOps;
731 for (; I != E && I->first == CurLoop; ++I) {
732 // If the operand is SCEVUnknown and not instructions, peek through
733 // it, to enable more of it to be folded into the GEP.
734 const SCEV *X = I->second;
735 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
736 if (!isa<Instruction>(U->getValue()))
737 X = SE.getSCEV(U->getValue());
740 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
741 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
742 // The running sum is an integer, and there's a pointer at this level.
743 // Try to form a getelementptr. If the running sum is instructions,
744 // use a SCEVUnknown to avoid re-analyzing them.
745 SmallVector<const SCEV *, 4> NewOps;
746 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
748 for (++I; I != E && I->first == CurLoop; ++I)
749 NewOps.push_back(I->second);
750 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
751 } else if (Op->isNonConstantNegative()) {
752 // Instead of doing a negate and add, just do a subtract.
753 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
754 Sum = InsertNoopCastOfTo(Sum, Ty);
755 Sum = InsertBinop(Instruction::Sub, Sum, W);
759 Value *W = expandCodeFor(Op, Ty);
760 Sum = InsertNoopCastOfTo(Sum, Ty);
761 // Canonicalize a constant to the RHS.
762 if (isa<Constant>(Sum)) std::swap(Sum, W);
763 Sum = InsertBinop(Instruction::Add, Sum, W);
771 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
772 Type *Ty = SE.getEffectiveSCEVType(S->getType());
774 // Collect all the mul operands in a loop, along with their associated loops.
775 // Iterate in reverse so that constants are emitted last, all else equal.
776 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
777 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
778 E(S->op_begin()); I != E; ++I)
779 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
781 // Sort by loop. Use a stable sort so that constants follow non-constants.
782 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
784 // Emit instructions to mul all the operands. Hoist as much as possible
787 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
788 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
789 const SCEV *Op = I->second;
791 // This is the first operand. Just expand it.
794 } else if (Op->isAllOnesValue()) {
795 // Instead of doing a multiply by negative one, just do a negate.
796 Prod = InsertNoopCastOfTo(Prod, Ty);
797 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
801 Value *W = expandCodeFor(Op, Ty);
802 Prod = InsertNoopCastOfTo(Prod, Ty);
803 // Canonicalize a constant to the RHS.
804 if (isa<Constant>(Prod)) std::swap(Prod, W);
805 Prod = InsertBinop(Instruction::Mul, Prod, W);
813 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
814 Type *Ty = SE.getEffectiveSCEVType(S->getType());
816 Value *LHS = expandCodeFor(S->getLHS(), Ty);
817 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
818 const APInt &RHS = SC->getValue()->getValue();
819 if (RHS.isPowerOf2())
820 return InsertBinop(Instruction::LShr, LHS,
821 ConstantInt::get(Ty, RHS.logBase2()));
824 Value *RHS = expandCodeFor(S->getRHS(), Ty);
825 return InsertBinop(Instruction::UDiv, LHS, RHS);
828 /// Move parts of Base into Rest to leave Base with the minimal
829 /// expression that provides a pointer operand suitable for a
831 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
832 ScalarEvolution &SE) {
833 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
834 Base = A->getStart();
835 Rest = SE.getAddExpr(Rest,
836 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
837 A->getStepRecurrence(SE),
839 A->getNoWrapFlags(SCEV::FlagNW)));
841 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
842 Base = A->getOperand(A->getNumOperands()-1);
843 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
844 NewAddOps.back() = Rest;
845 Rest = SE.getAddExpr(NewAddOps);
846 ExposePointerBase(Base, Rest, SE);
850 /// Determine if this is a well-behaved chain of instructions leading back to
851 /// the PHI. If so, it may be reused by expanded expressions.
852 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
854 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
855 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
857 // If any of the operands don't dominate the insert position, bail.
858 // Addrec operands are always loop-invariant, so this can only happen
859 // if there are instructions which haven't been hoisted.
860 if (L == IVIncInsertLoop) {
861 for (User::op_iterator OI = IncV->op_begin()+1,
862 OE = IncV->op_end(); OI != OE; ++OI)
863 if (Instruction *OInst = dyn_cast<Instruction>(OI))
864 if (!SE.DT->dominates(OInst, IVIncInsertPos))
867 // Advance to the next instruction.
868 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
872 if (IncV->mayHaveSideEffects())
878 return isNormalAddRecExprPHI(PN, IncV, L);
881 /// getIVIncOperand returns an induction variable increment's induction
882 /// variable operand.
884 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
885 /// operands dominate InsertPos.
887 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
888 /// simple patterns generated by getAddRecExprPHILiterally and
889 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
890 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
891 Instruction *InsertPos,
893 if (IncV == InsertPos)
896 switch (IncV->getOpcode()) {
899 // Check for a simple Add/Sub or GEP of a loop invariant step.
900 case Instruction::Add:
901 case Instruction::Sub: {
902 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
903 if (!OInst || SE.DT->dominates(OInst, InsertPos))
904 return dyn_cast<Instruction>(IncV->getOperand(0));
907 case Instruction::BitCast:
908 return dyn_cast<Instruction>(IncV->getOperand(0));
909 case Instruction::GetElementPtr:
910 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
912 if (isa<Constant>(*I))
914 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
915 if (!SE.DT->dominates(OInst, InsertPos))
919 // allow any kind of GEP as long as it can be hoisted.
922 // This must be a pointer addition of constants (pretty), which is already
923 // handled, or some number of address-size elements (ugly). Ugly geps
924 // have 2 operands. i1* is used by the expander to represent an
925 // address-size element.
926 if (IncV->getNumOperands() != 2)
928 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
929 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
930 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
934 return dyn_cast<Instruction>(IncV->getOperand(0));
938 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
939 /// it available to other uses in this loop. Recursively hoist any operands,
940 /// until we reach a value that dominates InsertPos.
941 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
942 if (SE.DT->dominates(IncV, InsertPos))
945 // InsertPos must itself dominate IncV so that IncV's new position satisfies
946 // its existing users.
947 if (isa<PHINode>(InsertPos)
948 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
951 // Check that the chain of IV operands leading back to Phi can be hoisted.
952 SmallVector<Instruction*, 4> IVIncs;
954 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
957 // IncV is safe to hoist.
958 IVIncs.push_back(IncV);
960 if (SE.DT->dominates(IncV, InsertPos))
963 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
964 E = IVIncs.rend(); I != E; ++I) {
965 (*I)->moveBefore(InsertPos);
970 /// Determine if this cyclic phi is in a form that would have been generated by
971 /// LSR. We don't care if the phi was actually expanded in this pass, as long
972 /// as it is in a low-cost form, for example, no implied multiplication. This
973 /// should match any patterns generated by getAddRecExprPHILiterally and
975 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
977 for(Instruction *IVOper = IncV;
978 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
979 /*allowScale=*/false));) {
986 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
987 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
988 /// need to materialize IV increments elsewhere to handle difficult situations.
989 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
990 Type *ExpandTy, Type *IntTy,
993 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
994 if (ExpandTy->isPointerTy()) {
995 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
996 // If the step isn't constant, don't use an implicitly scaled GEP, because
997 // that would require a multiply inside the loop.
998 if (!isa<ConstantInt>(StepV))
999 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
1000 GEPPtrTy->getAddressSpace());
1001 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
1002 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
1003 if (IncV->getType() != PN->getType()) {
1004 IncV = Builder.CreateBitCast(IncV, PN->getType());
1005 rememberInstruction(IncV);
1008 IncV = useSubtract ?
1009 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1010 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1011 rememberInstruction(IncV);
1016 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1017 /// the base addrec, which is the addrec without any non-loop-dominating
1018 /// values, and return the PHI.
1020 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1024 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1026 // Reuse a previously-inserted PHI, if present.
1027 BasicBlock *LatchBlock = L->getLoopLatch();
1029 for (BasicBlock::iterator I = L->getHeader()->begin();
1030 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1031 if (!SE.isSCEVable(PN->getType()) ||
1032 (SE.getEffectiveSCEVType(PN->getType()) !=
1033 SE.getEffectiveSCEVType(Normalized->getType())) ||
1034 SE.getSCEV(PN) != Normalized)
1038 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1041 if (!isExpandedAddRecExprPHI(PN, IncV, L))
1043 if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos))
1047 if (!isNormalAddRecExprPHI(PN, IncV, L))
1049 if (L == IVIncInsertLoop)
1051 if (SE.DT->dominates(IncV, IVIncInsertPos))
1053 // Make sure the increment is where we want it. But don't move it
1054 // down past a potential existing post-inc user.
1055 IncV->moveBefore(IVIncInsertPos);
1056 IVIncInsertPos = IncV;
1057 IncV = cast<Instruction>(IncV->getOperand(0));
1058 } while (IncV != PN);
1060 // Ok, the add recurrence looks usable.
1061 // Remember this PHI, even in post-inc mode.
1062 InsertedValues.insert(PN);
1063 // Remember the increment.
1064 rememberInstruction(IncV);
1069 // Save the original insertion point so we can restore it when we're done.
1070 BuilderType::InsertPointGuard Guard(Builder);
1072 // Another AddRec may need to be recursively expanded below. For example, if
1073 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1074 // loop. Remove this loop from the PostIncLoops set before expanding such
1075 // AddRecs. Otherwise, we cannot find a valid position for the step
1076 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1077 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1078 // so it's not worth implementing SmallPtrSet::swap.
1079 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1080 PostIncLoops.clear();
1082 // Expand code for the start value.
1083 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1084 L->getHeader()->begin());
1086 // StartV must be hoisted into L's preheader to dominate the new phi.
1087 assert(!isa<Instruction>(StartV) ||
1088 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1091 // Expand code for the step value. Do this before creating the PHI so that PHI
1092 // reuse code doesn't see an incomplete PHI.
1093 const SCEV *Step = Normalized->getStepRecurrence(SE);
1094 // If the stride is negative, insert a sub instead of an add for the increment
1095 // (unless it's a constant, because subtracts of constants are canonicalized
1097 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1099 Step = SE.getNegativeSCEV(Step);
1100 // Expand the step somewhere that dominates the loop header.
1101 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1104 BasicBlock *Header = L->getHeader();
1105 Builder.SetInsertPoint(Header, Header->begin());
1106 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1107 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1108 Twine(IVName) + ".iv");
1109 rememberInstruction(PN);
1111 // Create the step instructions and populate the PHI.
1112 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1113 BasicBlock *Pred = *HPI;
1115 // Add a start value.
1116 if (!L->contains(Pred)) {
1117 PN->addIncoming(StartV, Pred);
1121 // Create a step value and add it to the PHI.
1122 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1123 // instructions at IVIncInsertPos.
1124 Instruction *InsertPos = L == IVIncInsertLoop ?
1125 IVIncInsertPos : Pred->getTerminator();
1126 Builder.SetInsertPoint(InsertPos);
1127 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1128 if (isa<OverflowingBinaryOperator>(IncV)) {
1129 if (Normalized->getNoWrapFlags(SCEV::FlagNUW))
1130 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1131 if (Normalized->getNoWrapFlags(SCEV::FlagNSW))
1132 cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1134 PN->addIncoming(IncV, Pred);
1137 // After expanding subexpressions, restore the PostIncLoops set so the caller
1138 // can ensure that IVIncrement dominates the current uses.
1139 PostIncLoops = SavedPostIncLoops;
1141 // Remember this PHI, even in post-inc mode.
1142 InsertedValues.insert(PN);
1147 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1148 Type *STy = S->getType();
1149 Type *IntTy = SE.getEffectiveSCEVType(STy);
1150 const Loop *L = S->getLoop();
1152 // Determine a normalized form of this expression, which is the expression
1153 // before any post-inc adjustment is made.
1154 const SCEVAddRecExpr *Normalized = S;
1155 if (PostIncLoops.count(L)) {
1156 PostIncLoopSet Loops;
1159 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1160 Loops, SE, *SE.DT));
1163 // Strip off any non-loop-dominating component from the addrec start.
1164 const SCEV *Start = Normalized->getStart();
1165 const SCEV *PostLoopOffset = 0;
1166 if (!SE.properlyDominates(Start, L->getHeader())) {
1167 PostLoopOffset = Start;
1168 Start = SE.getConstant(Normalized->getType(), 0);
1169 Normalized = cast<SCEVAddRecExpr>(
1170 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1171 Normalized->getLoop(),
1172 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1175 // Strip off any non-loop-dominating component from the addrec step.
1176 const SCEV *Step = Normalized->getStepRecurrence(SE);
1177 const SCEV *PostLoopScale = 0;
1178 if (!SE.dominates(Step, L->getHeader())) {
1179 PostLoopScale = Step;
1180 Step = SE.getConstant(Normalized->getType(), 1);
1182 cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1183 Start, Step, Normalized->getLoop(),
1184 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1187 // Expand the core addrec. If we need post-loop scaling, force it to
1188 // expand to an integer type to avoid the need for additional casting.
1189 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1190 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1192 // Accommodate post-inc mode, if necessary.
1194 if (!PostIncLoops.count(L))
1197 // In PostInc mode, use the post-incremented value.
1198 BasicBlock *LatchBlock = L->getLoopLatch();
1199 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1200 Result = PN->getIncomingValueForBlock(LatchBlock);
1202 // For an expansion to use the postinc form, the client must call
1203 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1204 // or dominated by IVIncInsertPos.
1205 if (isa<Instruction>(Result)
1206 && !SE.DT->dominates(cast<Instruction>(Result),
1207 Builder.GetInsertPoint())) {
1208 // The induction variable's postinc expansion does not dominate this use.
1209 // IVUsers tries to prevent this case, so it is rare. However, it can
1210 // happen when an IVUser outside the loop is not dominated by the latch
1211 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1212 // all cases. Consider a phi outide whose operand is replaced during
1213 // expansion with the value of the postinc user. Without fundamentally
1214 // changing the way postinc users are tracked, the only remedy is
1215 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1216 // but hopefully expandCodeFor handles that.
1218 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1220 Step = SE.getNegativeSCEV(Step);
1223 // Expand the step somewhere that dominates the loop header.
1224 BuilderType::InsertPointGuard Guard(Builder);
1225 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1227 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1231 // Re-apply any non-loop-dominating scale.
1232 if (PostLoopScale) {
1233 Result = InsertNoopCastOfTo(Result, IntTy);
1234 Result = Builder.CreateMul(Result,
1235 expandCodeFor(PostLoopScale, IntTy));
1236 rememberInstruction(Result);
1239 // Re-apply any non-loop-dominating offset.
1240 if (PostLoopOffset) {
1241 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1242 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1243 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1245 Result = InsertNoopCastOfTo(Result, IntTy);
1246 Result = Builder.CreateAdd(Result,
1247 expandCodeFor(PostLoopOffset, IntTy));
1248 rememberInstruction(Result);
1255 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1256 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1258 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1259 const Loop *L = S->getLoop();
1261 // First check for an existing canonical IV in a suitable type.
1262 PHINode *CanonicalIV = 0;
1263 if (PHINode *PN = L->getCanonicalInductionVariable())
1264 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1267 // Rewrite an AddRec in terms of the canonical induction variable, if
1268 // its type is more narrow.
1270 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1271 SE.getTypeSizeInBits(Ty)) {
1272 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1273 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1274 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1275 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1276 S->getNoWrapFlags(SCEV::FlagNW)));
1277 BasicBlock::iterator NewInsertPt =
1278 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1279 BuilderType::InsertPointGuard Guard(Builder);
1280 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1281 isa<LandingPadInst>(NewInsertPt))
1283 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1288 // {X,+,F} --> X + {0,+,F}
1289 if (!S->getStart()->isZero()) {
1290 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1291 NewOps[0] = SE.getConstant(Ty, 0);
1292 const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1293 S->getNoWrapFlags(SCEV::FlagNW));
1295 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1296 // comments on expandAddToGEP for details.
1297 const SCEV *Base = S->getStart();
1298 const SCEV *RestArray[1] = { Rest };
1299 // Dig into the expression to find the pointer base for a GEP.
1300 ExposePointerBase(Base, RestArray[0], SE);
1301 // If we found a pointer, expand the AddRec with a GEP.
1302 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1303 // Make sure the Base isn't something exotic, such as a multiplied
1304 // or divided pointer value. In those cases, the result type isn't
1305 // actually a pointer type.
1306 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1307 Value *StartV = expand(Base);
1308 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1309 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1313 // Just do a normal add. Pre-expand the operands to suppress folding.
1314 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1315 SE.getUnknown(expand(Rest))));
1318 // If we don't yet have a canonical IV, create one.
1320 // Create and insert the PHI node for the induction variable in the
1322 BasicBlock *Header = L->getHeader();
1323 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1324 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1326 rememberInstruction(CanonicalIV);
1328 SmallSet<BasicBlock *, 4> PredSeen;
1329 Constant *One = ConstantInt::get(Ty, 1);
1330 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1331 BasicBlock *HP = *HPI;
1332 if (!PredSeen.insert(HP))
1335 if (L->contains(HP)) {
1336 // Insert a unit add instruction right before the terminator
1337 // corresponding to the back-edge.
1338 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1340 HP->getTerminator());
1341 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1342 rememberInstruction(Add);
1343 CanonicalIV->addIncoming(Add, HP);
1345 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1350 // {0,+,1} --> Insert a canonical induction variable into the loop!
1351 if (S->isAffine() && S->getOperand(1)->isOne()) {
1352 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1353 "IVs with types different from the canonical IV should "
1354 "already have been handled!");
1358 // {0,+,F} --> {0,+,1} * F
1360 // If this is a simple linear addrec, emit it now as a special case.
1361 if (S->isAffine()) // {0,+,F} --> i*F
1363 expand(SE.getTruncateOrNoop(
1364 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1365 SE.getNoopOrAnyExtend(S->getOperand(1),
1366 CanonicalIV->getType())),
1369 // If this is a chain of recurrences, turn it into a closed form, using the
1370 // folders, then expandCodeFor the closed form. This allows the folders to
1371 // simplify the expression without having to build a bunch of special code
1372 // into this folder.
1373 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1375 // Promote S up to the canonical IV type, if the cast is foldable.
1376 const SCEV *NewS = S;
1377 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1378 if (isa<SCEVAddRecExpr>(Ext))
1381 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1382 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1384 // Truncate the result down to the original type, if needed.
1385 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1389 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1390 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1391 Value *V = expandCodeFor(S->getOperand(),
1392 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1393 Value *I = Builder.CreateTrunc(V, Ty);
1394 rememberInstruction(I);
1398 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1399 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1400 Value *V = expandCodeFor(S->getOperand(),
1401 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1402 Value *I = Builder.CreateZExt(V, Ty);
1403 rememberInstruction(I);
1407 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1408 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1409 Value *V = expandCodeFor(S->getOperand(),
1410 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1411 Value *I = Builder.CreateSExt(V, Ty);
1412 rememberInstruction(I);
1416 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1417 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1418 Type *Ty = LHS->getType();
1419 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1420 // In the case of mixed integer and pointer types, do the
1421 // rest of the comparisons as integer.
1422 if (S->getOperand(i)->getType() != Ty) {
1423 Ty = SE.getEffectiveSCEVType(Ty);
1424 LHS = InsertNoopCastOfTo(LHS, Ty);
1426 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1427 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1428 rememberInstruction(ICmp);
1429 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1430 rememberInstruction(Sel);
1433 // In the case of mixed integer and pointer types, cast the
1434 // final result back to the pointer type.
1435 if (LHS->getType() != S->getType())
1436 LHS = InsertNoopCastOfTo(LHS, S->getType());
1440 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1441 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1442 Type *Ty = LHS->getType();
1443 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1444 // In the case of mixed integer and pointer types, do the
1445 // rest of the comparisons as integer.
1446 if (S->getOperand(i)->getType() != Ty) {
1447 Ty = SE.getEffectiveSCEVType(Ty);
1448 LHS = InsertNoopCastOfTo(LHS, Ty);
1450 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1451 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1452 rememberInstruction(ICmp);
1453 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1454 rememberInstruction(Sel);
1457 // In the case of mixed integer and pointer types, cast the
1458 // final result back to the pointer type.
1459 if (LHS->getType() != S->getType())
1460 LHS = InsertNoopCastOfTo(LHS, S->getType());
1464 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1466 Builder.SetInsertPoint(IP->getParent(), IP);
1467 return expandCodeFor(SH, Ty);
1470 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1471 // Expand the code for this SCEV.
1472 Value *V = expand(SH);
1474 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1475 "non-trivial casts should be done with the SCEVs directly!");
1476 V = InsertNoopCastOfTo(V, Ty);
1481 Value *SCEVExpander::expand(const SCEV *S) {
1482 // Compute an insertion point for this SCEV object. Hoist the instructions
1483 // as far out in the loop nest as possible.
1484 Instruction *InsertPt = Builder.GetInsertPoint();
1485 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1486 L = L->getParentLoop())
1487 if (SE.isLoopInvariant(S, L)) {
1489 if (BasicBlock *Preheader = L->getLoopPreheader())
1490 InsertPt = Preheader->getTerminator();
1492 // LSR sets the insertion point for AddRec start/step values to the
1493 // block start to simplify value reuse, even though it's an invalid
1494 // position. SCEVExpander must correct for this in all cases.
1495 InsertPt = L->getHeader()->getFirstInsertionPt();
1498 // If the SCEV is computable at this level, insert it into the header
1499 // after the PHIs (and after any other instructions that we've inserted
1500 // there) so that it is guaranteed to dominate any user inside the loop.
1501 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1502 InsertPt = L->getHeader()->getFirstInsertionPt();
1503 while (InsertPt != Builder.GetInsertPoint()
1504 && (isInsertedInstruction(InsertPt)
1505 || isa<DbgInfoIntrinsic>(InsertPt))) {
1506 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1511 // Check to see if we already expanded this here.
1512 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator
1513 I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1514 if (I != InsertedExpressions.end())
1517 BuilderType::InsertPointGuard Guard(Builder);
1518 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1520 // Expand the expression into instructions.
1521 Value *V = visit(S);
1523 // Remember the expanded value for this SCEV at this location.
1525 // This is independent of PostIncLoops. The mapped value simply materializes
1526 // the expression at this insertion point. If the mapped value happened to be
1527 // a postinc expansion, it could be reused by a non postinc user, but only if
1528 // its insertion point was already at the head of the loop.
1529 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1533 void SCEVExpander::rememberInstruction(Value *I) {
1534 if (!PostIncLoops.empty())
1535 InsertedPostIncValues.insert(I);
1537 InsertedValues.insert(I);
1540 /// getOrInsertCanonicalInductionVariable - This method returns the
1541 /// canonical induction variable of the specified type for the specified
1542 /// loop (inserting one if there is none). A canonical induction variable
1543 /// starts at zero and steps by one on each iteration.
1545 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1547 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1549 // Build a SCEV for {0,+,1}<L>.
1550 // Conservatively use FlagAnyWrap for now.
1551 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1552 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1554 // Emit code for it.
1555 BuilderType::InsertPointGuard Guard(Builder);
1556 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1561 /// Sort values by integer width for replaceCongruentIVs.
1562 static bool width_descending(Value *lhs, Value *rhs) {
1563 // Put pointers at the back and make sure pointer < pointer = false.
1564 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy())
1565 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy();
1566 return rhs->getType()->getPrimitiveSizeInBits()
1567 < lhs->getType()->getPrimitiveSizeInBits();
1570 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1571 /// replace them with their most canonical representative. Return the number of
1572 /// phis eliminated.
1574 /// This does not depend on any SCEVExpander state but should be used in
1575 /// the same context that SCEVExpander is used.
1576 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1577 SmallVectorImpl<WeakVH> &DeadInsts,
1578 const TargetTransformInfo *TTI) {
1579 // Find integer phis in order of increasing width.
1580 SmallVector<PHINode*, 8> Phis;
1581 for (BasicBlock::iterator I = L->getHeader()->begin();
1582 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1583 Phis.push_back(Phi);
1586 std::sort(Phis.begin(), Phis.end(), width_descending);
1588 unsigned NumElim = 0;
1589 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1590 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1591 // so narrow phis can reuse them.
1592 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1593 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1594 PHINode *Phi = *PIter;
1596 // Fold constant phis. They may be congruent to other constant phis and
1597 // would confuse the logic below that expects proper IVs.
1598 if (Value *V = Phi->hasConstantValue()) {
1599 Phi->replaceAllUsesWith(V);
1600 DeadInsts.push_back(Phi);
1602 DEBUG_WITH_TYPE(DebugType, dbgs()
1603 << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1607 if (!SE.isSCEVable(Phi->getType()))
1610 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1613 if (Phi->getType()->isIntegerTy() && TTI
1614 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1615 // This phi can be freely truncated to the narrowest phi type. Map the
1616 // truncated expression to it so it will be reused for narrow types.
1617 const SCEV *TruncExpr =
1618 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1619 ExprToIVMap[TruncExpr] = Phi;
1624 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1626 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1629 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1630 Instruction *OrigInc =
1631 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1632 Instruction *IsomorphicInc =
1633 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1635 // If this phi has the same width but is more canonical, replace the
1636 // original with it. As part of the "more canonical" determination,
1637 // respect a prior decision to use an IV chain.
1638 if (OrigPhiRef->getType() == Phi->getType()
1639 && !(ChainedPhis.count(Phi)
1640 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1641 && (ChainedPhis.count(Phi)
1642 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1643 std::swap(OrigPhiRef, Phi);
1644 std::swap(OrigInc, IsomorphicInc);
1646 // Replacing the congruent phi is sufficient because acyclic redundancy
1647 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1648 // that a phi is congruent, it's often the head of an IV user cycle that
1649 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1650 // common case of a single IV increment so that DeleteDeadPHIs can remove
1651 // cycles that had postinc uses.
1652 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1653 IsomorphicInc->getType());
1654 if (OrigInc != IsomorphicInc
1655 && TruncExpr == SE.getSCEV(IsomorphicInc)
1656 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1657 || hoistIVInc(OrigInc, IsomorphicInc))) {
1658 DEBUG_WITH_TYPE(DebugType, dbgs()
1659 << "INDVARS: Eliminated congruent iv.inc: "
1660 << *IsomorphicInc << '\n');
1661 Value *NewInc = OrigInc;
1662 if (OrigInc->getType() != IsomorphicInc->getType()) {
1663 Instruction *IP = isa<PHINode>(OrigInc)
1664 ? (Instruction*)L->getHeader()->getFirstInsertionPt()
1665 : OrigInc->getNextNode();
1666 IRBuilder<> Builder(IP);
1667 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1669 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1671 IsomorphicInc->replaceAllUsesWith(NewInc);
1672 DeadInsts.push_back(IsomorphicInc);
1675 DEBUG_WITH_TYPE(DebugType, dbgs()
1676 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1678 Value *NewIV = OrigPhiRef;
1679 if (OrigPhiRef->getType() != Phi->getType()) {
1680 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1681 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1682 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1684 Phi->replaceAllUsesWith(NewIV);
1685 DeadInsts.push_back(Phi);
1691 // Search for a SCEV subexpression that is not safe to expand. Any expression
1692 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
1693 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
1694 // instruction, but the important thing is that we prove the denominator is
1695 // nonzero before expansion.
1697 // IVUsers already checks that IV-derived expressions are safe. So this check is
1698 // only needed when the expression includes some subexpression that is not IV
1701 // Currently, we only allow division by a nonzero constant here. If this is
1702 // inadequate, we could easily allow division by SCEVUnknown by using
1703 // ValueTracking to check isKnownNonZero().
1704 struct SCEVFindUnsafe {
1707 SCEVFindUnsafe(): IsUnsafe(false) {}
1709 bool follow(const SCEV *S) {
1710 const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S);
1713 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
1714 if (SC && !SC->getValue()->isZero())
1719 bool isDone() const { return IsUnsafe; }
1724 bool isSafeToExpand(const SCEV *S) {
1725 SCEVFindUnsafe Search;
1726 visitAll(S, Search);
1727 return !Search.IsUnsafe;