1 //===- ScalarEvolution.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 analysis
11 // engine, which is used primarily to analyze expressions involving induction
12 // variables in loops.
14 // There are several aspects to this library. First is the representation of
15 // scalar expressions, which are represented as subclasses of the SCEV class.
16 // These classes are used to represent certain types of subexpressions that we
17 // can handle. We only create one SCEV of a particular shape, so
18 // pointer-comparisons for equality are legal.
20 // One important aspect of the SCEV objects is that they are never cyclic, even
21 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
22 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
23 // recurrence) then we represent it directly as a recurrence node, otherwise we
24 // represent it as a SCEVUnknown node.
26 // In addition to being able to represent expressions of various types, we also
27 // have folders that are used to build the *canonical* representation for a
28 // particular expression. These folders are capable of using a variety of
29 // rewrite rules to simplify the expressions.
31 // Once the folders are defined, we can implement the more interesting
32 // higher-level code, such as the code that recognizes PHI nodes of various
33 // types, computes the execution count of a loop, etc.
35 // TODO: We should use these routines and value representations to implement
36 // dependence analysis!
38 //===----------------------------------------------------------------------===//
40 // There are several good references for the techniques used in this analysis.
42 // Chains of recurrences -- a method to expedite the evaluation
43 // of closed-form functions
44 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
46 // On computational properties of chains of recurrences
49 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
50 // Robert A. van Engelen
52 // Efficient Symbolic Analysis for Optimizing Compilers
53 // Robert A. van Engelen
55 // Using the chains of recurrences algebra for data dependence testing and
56 // induction variable substitution
57 // MS Thesis, Johnie Birch
59 //===----------------------------------------------------------------------===//
61 #define DEBUG_TYPE "scalar-evolution"
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/Constants.h"
64 #include "llvm/DerivedTypes.h"
65 #include "llvm/GlobalVariable.h"
66 #include "llvm/GlobalAlias.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/LLVMContext.h"
69 #include "llvm/Operator.h"
70 #include "llvm/Analysis/ConstantFolding.h"
71 #include "llvm/Analysis/Dominators.h"
72 #include "llvm/Analysis/InstructionSimplify.h"
73 #include "llvm/Analysis/LoopInfo.h"
74 #include "llvm/Analysis/ValueTracking.h"
75 #include "llvm/Assembly/Writer.h"
76 #include "llvm/Target/TargetData.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/ConstantRange.h"
79 #include "llvm/Support/Debug.h"
80 #include "llvm/Support/ErrorHandling.h"
81 #include "llvm/Support/GetElementPtrTypeIterator.h"
82 #include "llvm/Support/InstIterator.h"
83 #include "llvm/Support/MathExtras.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/ADT/Statistic.h"
86 #include "llvm/ADT/STLExtras.h"
87 #include "llvm/ADT/SmallPtrSet.h"
91 STATISTIC(NumArrayLenItCounts,
92 "Number of trip counts computed with array length");
93 STATISTIC(NumTripCountsComputed,
94 "Number of loops with predictable loop counts");
95 STATISTIC(NumTripCountsNotComputed,
96 "Number of loops without predictable loop counts");
97 STATISTIC(NumBruteForceTripCountsComputed,
98 "Number of loops with trip counts computed by force");
100 static cl::opt<unsigned>
101 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
102 cl::desc("Maximum number of iterations SCEV will "
103 "symbolically execute a constant "
107 INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
108 "Scalar Evolution Analysis", false, true)
109 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
110 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
111 INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
112 "Scalar Evolution Analysis", false, true)
113 char ScalarEvolution::ID = 0;
115 //===----------------------------------------------------------------------===//
116 // SCEV class definitions
117 //===----------------------------------------------------------------------===//
119 //===----------------------------------------------------------------------===//
120 // Implementation of the SCEV class.
125 void SCEV::dump() const {
130 bool SCEV::isZero() const {
131 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
132 return SC->getValue()->isZero();
136 bool SCEV::isOne() const {
137 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
138 return SC->getValue()->isOne();
142 bool SCEV::isAllOnesValue() const {
143 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
144 return SC->getValue()->isAllOnesValue();
148 SCEVCouldNotCompute::SCEVCouldNotCompute() :
149 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
151 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
152 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
156 const Type *SCEVCouldNotCompute::getType() const {
157 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
161 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
162 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
166 bool SCEVCouldNotCompute::hasOperand(const SCEV *) const {
167 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
171 void SCEVCouldNotCompute::print(raw_ostream &OS) const {
172 OS << "***COULDNOTCOMPUTE***";
175 bool SCEVCouldNotCompute::classof(const SCEV *S) {
176 return S->getSCEVType() == scCouldNotCompute;
179 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
181 ID.AddInteger(scConstant);
184 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
185 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V);
186 UniqueSCEVs.InsertNode(S, IP);
190 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
191 return getConstant(ConstantInt::get(getContext(), Val));
195 ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
196 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
197 return getConstant(ConstantInt::get(ITy, V, isSigned));
200 const Type *SCEVConstant::getType() const { return V->getType(); }
202 void SCEVConstant::print(raw_ostream &OS) const {
203 WriteAsOperand(OS, V, false);
206 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
207 unsigned SCEVTy, const SCEV *op, const Type *ty)
208 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
210 bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
211 return Op->dominates(BB, DT);
214 bool SCEVCastExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
215 return Op->properlyDominates(BB, DT);
218 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
219 const SCEV *op, const Type *ty)
220 : SCEVCastExpr(ID, scTruncate, op, ty) {
221 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
222 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
223 "Cannot truncate non-integer value!");
226 void SCEVTruncateExpr::print(raw_ostream &OS) const {
227 OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
230 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
231 const SCEV *op, const Type *ty)
232 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
233 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
234 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
235 "Cannot zero extend non-integer value!");
238 void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
239 OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
242 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
243 const SCEV *op, const Type *ty)
244 : SCEVCastExpr(ID, scSignExtend, op, ty) {
245 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
246 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
247 "Cannot sign extend non-integer value!");
250 void SCEVSignExtendExpr::print(raw_ostream &OS) const {
251 OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
254 void SCEVCommutativeExpr::print(raw_ostream &OS) const {
255 const char *OpStr = getOperationStr();
257 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I) {
259 if (llvm::next(I) != E)
265 bool SCEVNAryExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
266 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
267 if (!(*I)->dominates(BB, DT))
272 bool SCEVNAryExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
273 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
274 if (!(*I)->properlyDominates(BB, DT))
279 bool SCEVNAryExpr::isLoopInvariant(const Loop *L) const {
280 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
281 if (!(*I)->isLoopInvariant(L))
286 // hasComputableLoopEvolution - N-ary expressions have computable loop
287 // evolutions iff they have at least one operand that varies with the loop,
288 // but that all varying operands are computable.
289 bool SCEVNAryExpr::hasComputableLoopEvolution(const Loop *L) const {
290 bool HasVarying = false;
291 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I) {
293 if (!S->isLoopInvariant(L)) {
294 if (S->hasComputableLoopEvolution(L))
303 bool SCEVNAryExpr::hasOperand(const SCEV *O) const {
304 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I) {
306 if (O == S || S->hasOperand(O))
312 bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
313 return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
316 bool SCEVUDivExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
317 return LHS->properlyDominates(BB, DT) && RHS->properlyDominates(BB, DT);
320 void SCEVUDivExpr::print(raw_ostream &OS) const {
321 OS << "(" << *LHS << " /u " << *RHS << ")";
324 const Type *SCEVUDivExpr::getType() const {
325 // In most cases the types of LHS and RHS will be the same, but in some
326 // crazy cases one or the other may be a pointer. ScalarEvolution doesn't
327 // depend on the type for correctness, but handling types carefully can
328 // avoid extra casts in the SCEVExpander. The LHS is more likely to be
329 // a pointer type than the RHS, so use the RHS' type here.
330 return RHS->getType();
333 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
334 // Add recurrences are never invariant in the function-body (null loop).
338 // This recurrence is variant w.r.t. QueryLoop if QueryLoop contains L.
339 if (QueryLoop->contains(L))
342 // This recurrence is invariant w.r.t. QueryLoop if L contains QueryLoop.
343 if (L->contains(QueryLoop))
346 // This recurrence is variant w.r.t. QueryLoop if any of its operands
348 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
349 if (!(*I)->isLoopInvariant(QueryLoop))
352 // Otherwise it's loop-invariant.
357 SCEVAddRecExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
358 return DT->dominates(L->getHeader(), BB) &&
359 SCEVNAryExpr::dominates(BB, DT);
363 SCEVAddRecExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
364 // This uses a "dominates" query instead of "properly dominates" query because
365 // the instruction which produces the addrec's value is a PHI, and a PHI
366 // effectively properly dominates its entire containing block.
367 return DT->dominates(L->getHeader(), BB) &&
368 SCEVNAryExpr::properlyDominates(BB, DT);
371 void SCEVAddRecExpr::print(raw_ostream &OS) const {
372 OS << "{" << *Operands[0];
373 for (unsigned i = 1, e = NumOperands; i != e; ++i)
374 OS << ",+," << *Operands[i];
376 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
380 void SCEVUnknown::deleted() {
381 // Clear this SCEVUnknown from various maps.
382 SE->ValuesAtScopes.erase(this);
383 SE->UnsignedRanges.erase(this);
384 SE->SignedRanges.erase(this);
386 // Remove this SCEVUnknown from the uniquing map.
387 SE->UniqueSCEVs.RemoveNode(this);
389 // Release the value.
393 void SCEVUnknown::allUsesReplacedWith(Value *New) {
394 // Clear this SCEVUnknown from various maps.
395 SE->ValuesAtScopes.erase(this);
396 SE->UnsignedRanges.erase(this);
397 SE->SignedRanges.erase(this);
399 // Remove this SCEVUnknown from the uniquing map.
400 SE->UniqueSCEVs.RemoveNode(this);
402 // Update this SCEVUnknown to point to the new value. This is needed
403 // because there may still be outstanding SCEVs which still point to
408 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
409 // All non-instruction values are loop invariant. All instructions are loop
410 // invariant if they are not contained in the specified loop.
411 // Instructions are never considered invariant in the function body
412 // (null loop) because they are defined within the "loop".
413 if (Instruction *I = dyn_cast<Instruction>(getValue()))
414 return L && !L->contains(I);
418 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
419 if (Instruction *I = dyn_cast<Instruction>(getValue()))
420 return DT->dominates(I->getParent(), BB);
424 bool SCEVUnknown::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
425 if (Instruction *I = dyn_cast<Instruction>(getValue()))
426 return DT->properlyDominates(I->getParent(), BB);
430 const Type *SCEVUnknown::getType() const {
431 return getValue()->getType();
434 bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
435 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
436 if (VCE->getOpcode() == Instruction::PtrToInt)
437 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
438 if (CE->getOpcode() == Instruction::GetElementPtr &&
439 CE->getOperand(0)->isNullValue() &&
440 CE->getNumOperands() == 2)
441 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
443 AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
451 bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
452 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
453 if (VCE->getOpcode() == Instruction::PtrToInt)
454 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
455 if (CE->getOpcode() == Instruction::GetElementPtr &&
456 CE->getOperand(0)->isNullValue()) {
458 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
459 if (const StructType *STy = dyn_cast<StructType>(Ty))
460 if (!STy->isPacked() &&
461 CE->getNumOperands() == 3 &&
462 CE->getOperand(1)->isNullValue()) {
463 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
465 STy->getNumElements() == 2 &&
466 STy->getElementType(0)->isIntegerTy(1)) {
467 AllocTy = STy->getElementType(1);
476 bool SCEVUnknown::isOffsetOf(const Type *&CTy, Constant *&FieldNo) const {
477 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
478 if (VCE->getOpcode() == Instruction::PtrToInt)
479 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
480 if (CE->getOpcode() == Instruction::GetElementPtr &&
481 CE->getNumOperands() == 3 &&
482 CE->getOperand(0)->isNullValue() &&
483 CE->getOperand(1)->isNullValue()) {
485 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
486 // Ignore vector types here so that ScalarEvolutionExpander doesn't
487 // emit getelementptrs that index into vectors.
488 if (Ty->isStructTy() || Ty->isArrayTy()) {
490 FieldNo = CE->getOperand(2);
498 void SCEVUnknown::print(raw_ostream &OS) const {
500 if (isSizeOf(AllocTy)) {
501 OS << "sizeof(" << *AllocTy << ")";
504 if (isAlignOf(AllocTy)) {
505 OS << "alignof(" << *AllocTy << ")";
511 if (isOffsetOf(CTy, FieldNo)) {
512 OS << "offsetof(" << *CTy << ", ";
513 WriteAsOperand(OS, FieldNo, false);
518 // Otherwise just print it normally.
519 WriteAsOperand(OS, getValue(), false);
522 //===----------------------------------------------------------------------===//
524 //===----------------------------------------------------------------------===//
527 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
528 /// than the complexity of the RHS. This comparator is used to canonicalize
530 class SCEVComplexityCompare {
531 const LoopInfo *const LI;
533 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
535 // Return true or false if LHS is less than, or at least RHS, respectively.
536 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
537 return compare(LHS, RHS) < 0;
540 // Return negative, zero, or positive, if LHS is less than, equal to, or
541 // greater than RHS, respectively. A three-way result allows recursive
542 // comparisons to be more efficient.
543 int compare(const SCEV *LHS, const SCEV *RHS) const {
544 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
548 // Primarily, sort the SCEVs by their getSCEVType().
549 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
551 return (int)LType - (int)RType;
553 // Aside from the getSCEVType() ordering, the particular ordering
554 // isn't very important except that it's beneficial to be consistent,
555 // so that (a + b) and (b + a) don't end up as different expressions.
558 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS);
559 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
561 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
562 // not as complete as it could be.
563 const Value *LV = LU->getValue(), *RV = RU->getValue();
565 // Order pointer values after integer values. This helps SCEVExpander
567 bool LIsPointer = LV->getType()->isPointerTy(),
568 RIsPointer = RV->getType()->isPointerTy();
569 if (LIsPointer != RIsPointer)
570 return (int)LIsPointer - (int)RIsPointer;
572 // Compare getValueID values.
573 unsigned LID = LV->getValueID(),
574 RID = RV->getValueID();
576 return (int)LID - (int)RID;
578 // Sort arguments by their position.
579 if (const Argument *LA = dyn_cast<Argument>(LV)) {
580 const Argument *RA = cast<Argument>(RV);
581 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo();
582 return (int)LArgNo - (int)RArgNo;
585 // For instructions, compare their loop depth, and their operand
586 // count. This is pretty loose.
587 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) {
588 const Instruction *RInst = cast<Instruction>(RV);
590 // Compare loop depths.
591 const BasicBlock *LParent = LInst->getParent(),
592 *RParent = RInst->getParent();
593 if (LParent != RParent) {
594 unsigned LDepth = LI->getLoopDepth(LParent),
595 RDepth = LI->getLoopDepth(RParent);
596 if (LDepth != RDepth)
597 return (int)LDepth - (int)RDepth;
600 // Compare the number of operands.
601 unsigned LNumOps = LInst->getNumOperands(),
602 RNumOps = RInst->getNumOperands();
603 return (int)LNumOps - (int)RNumOps;
610 const SCEVConstant *LC = cast<SCEVConstant>(LHS);
611 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
613 // Compare constant values.
614 const APInt &LA = LC->getValue()->getValue();
615 const APInt &RA = RC->getValue()->getValue();
616 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth();
617 if (LBitWidth != RBitWidth)
618 return (int)LBitWidth - (int)RBitWidth;
619 return LA.ult(RA) ? -1 : 1;
623 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS);
624 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
626 // Compare addrec loop depths.
627 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop();
628 if (LLoop != RLoop) {
629 unsigned LDepth = LLoop->getLoopDepth(),
630 RDepth = RLoop->getLoopDepth();
631 if (LDepth != RDepth)
632 return (int)LDepth - (int)RDepth;
635 // Addrec complexity grows with operand count.
636 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands();
637 if (LNumOps != RNumOps)
638 return (int)LNumOps - (int)RNumOps;
640 // Lexicographically compare.
641 for (unsigned i = 0; i != LNumOps; ++i) {
642 long X = compare(LA->getOperand(i), RA->getOperand(i));
654 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS);
655 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
657 // Lexicographically compare n-ary expressions.
658 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
659 for (unsigned i = 0; i != LNumOps; ++i) {
662 long X = compare(LC->getOperand(i), RC->getOperand(i));
666 return (int)LNumOps - (int)RNumOps;
670 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
671 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
673 // Lexicographically compare udiv expressions.
674 long X = compare(LC->getLHS(), RC->getLHS());
677 return compare(LC->getRHS(), RC->getRHS());
683 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
684 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
686 // Compare cast expressions by operand.
687 return compare(LC->getOperand(), RC->getOperand());
694 llvm_unreachable("Unknown SCEV kind!");
700 /// GroupByComplexity - Given a list of SCEV objects, order them by their
701 /// complexity, and group objects of the same complexity together by value.
702 /// When this routine is finished, we know that any duplicates in the vector are
703 /// consecutive and that complexity is monotonically increasing.
705 /// Note that we go take special precautions to ensure that we get deterministic
706 /// results from this routine. In other words, we don't want the results of
707 /// this to depend on where the addresses of various SCEV objects happened to
710 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
712 if (Ops.size() < 2) return; // Noop
713 if (Ops.size() == 2) {
714 // This is the common case, which also happens to be trivially simple.
716 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
717 if (SCEVComplexityCompare(LI)(RHS, LHS))
722 // Do the rough sort by complexity.
723 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
725 // Now that we are sorted by complexity, group elements of the same
726 // complexity. Note that this is, at worst, N^2, but the vector is likely to
727 // be extremely short in practice. Note that we take this approach because we
728 // do not want to depend on the addresses of the objects we are grouping.
729 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
730 const SCEV *S = Ops[i];
731 unsigned Complexity = S->getSCEVType();
733 // If there are any objects of the same complexity and same value as this
735 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
736 if (Ops[j] == S) { // Found a duplicate.
737 // Move it to immediately after i'th element.
738 std::swap(Ops[i+1], Ops[j]);
739 ++i; // no need to rescan it.
740 if (i == e-2) return; // Done!
748 //===----------------------------------------------------------------------===//
749 // Simple SCEV method implementations
750 //===----------------------------------------------------------------------===//
752 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
754 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
756 const Type* ResultTy) {
757 // Handle the simplest case efficiently.
759 return SE.getTruncateOrZeroExtend(It, ResultTy);
761 // We are using the following formula for BC(It, K):
763 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
765 // Suppose, W is the bitwidth of the return value. We must be prepared for
766 // overflow. Hence, we must assure that the result of our computation is
767 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
768 // safe in modular arithmetic.
770 // However, this code doesn't use exactly that formula; the formula it uses
771 // is something like the following, where T is the number of factors of 2 in
772 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
775 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
777 // This formula is trivially equivalent to the previous formula. However,
778 // this formula can be implemented much more efficiently. The trick is that
779 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
780 // arithmetic. To do exact division in modular arithmetic, all we have
781 // to do is multiply by the inverse. Therefore, this step can be done at
784 // The next issue is how to safely do the division by 2^T. The way this
785 // is done is by doing the multiplication step at a width of at least W + T
786 // bits. This way, the bottom W+T bits of the product are accurate. Then,
787 // when we perform the division by 2^T (which is equivalent to a right shift
788 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
789 // truncated out after the division by 2^T.
791 // In comparison to just directly using the first formula, this technique
792 // is much more efficient; using the first formula requires W * K bits,
793 // but this formula less than W + K bits. Also, the first formula requires
794 // a division step, whereas this formula only requires multiplies and shifts.
796 // It doesn't matter whether the subtraction step is done in the calculation
797 // width or the input iteration count's width; if the subtraction overflows,
798 // the result must be zero anyway. We prefer here to do it in the width of
799 // the induction variable because it helps a lot for certain cases; CodeGen
800 // isn't smart enough to ignore the overflow, which leads to much less
801 // efficient code if the width of the subtraction is wider than the native
804 // (It's possible to not widen at all by pulling out factors of 2 before
805 // the multiplication; for example, K=2 can be calculated as
806 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
807 // extra arithmetic, so it's not an obvious win, and it gets
808 // much more complicated for K > 3.)
810 // Protection from insane SCEVs; this bound is conservative,
811 // but it probably doesn't matter.
813 return SE.getCouldNotCompute();
815 unsigned W = SE.getTypeSizeInBits(ResultTy);
817 // Calculate K! / 2^T and T; we divide out the factors of two before
818 // multiplying for calculating K! / 2^T to avoid overflow.
819 // Other overflow doesn't matter because we only care about the bottom
820 // W bits of the result.
821 APInt OddFactorial(W, 1);
823 for (unsigned i = 3; i <= K; ++i) {
825 unsigned TwoFactors = Mult.countTrailingZeros();
827 Mult = Mult.lshr(TwoFactors);
828 OddFactorial *= Mult;
831 // We need at least W + T bits for the multiplication step
832 unsigned CalculationBits = W + T;
834 // Calculate 2^T, at width T+W.
835 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
837 // Calculate the multiplicative inverse of K! / 2^T;
838 // this multiplication factor will perform the exact division by
840 APInt Mod = APInt::getSignedMinValue(W+1);
841 APInt MultiplyFactor = OddFactorial.zext(W+1);
842 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
843 MultiplyFactor = MultiplyFactor.trunc(W);
845 // Calculate the product, at width T+W
846 const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
848 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
849 for (unsigned i = 1; i != K; ++i) {
850 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
851 Dividend = SE.getMulExpr(Dividend,
852 SE.getTruncateOrZeroExtend(S, CalculationTy));
856 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
858 // Truncate the result, and divide by K! / 2^T.
860 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
861 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
864 /// evaluateAtIteration - Return the value of this chain of recurrences at
865 /// the specified iteration number. We can evaluate this recurrence by
866 /// multiplying each element in the chain by the binomial coefficient
867 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
869 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
871 /// where BC(It, k) stands for binomial coefficient.
873 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
874 ScalarEvolution &SE) const {
875 const SCEV *Result = getStart();
876 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
877 // The computation is correct in the face of overflow provided that the
878 // multiplication is performed _after_ the evaluation of the binomial
880 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
881 if (isa<SCEVCouldNotCompute>(Coeff))
884 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
889 //===----------------------------------------------------------------------===//
890 // SCEV Expression folder implementations
891 //===----------------------------------------------------------------------===//
893 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
895 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
896 "This is not a truncating conversion!");
897 assert(isSCEVable(Ty) &&
898 "This is not a conversion to a SCEVable type!");
899 Ty = getEffectiveSCEVType(Ty);
902 ID.AddInteger(scTruncate);
906 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
908 // Fold if the operand is constant.
909 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
911 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(),
912 getEffectiveSCEVType(Ty))));
914 // trunc(trunc(x)) --> trunc(x)
915 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
916 return getTruncateExpr(ST->getOperand(), Ty);
918 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
919 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
920 return getTruncateOrSignExtend(SS->getOperand(), Ty);
922 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
923 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
924 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
926 // If the input value is a chrec scev, truncate the chrec's operands.
927 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
928 SmallVector<const SCEV *, 4> Operands;
929 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
930 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
931 return getAddRecExpr(Operands, AddRec->getLoop());
934 // As a special case, fold trunc(undef) to undef. We don't want to
935 // know too much about SCEVUnknowns, but this special case is handy
937 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
938 if (isa<UndefValue>(U->getValue()))
939 return getSCEV(UndefValue::get(Ty));
941 // The cast wasn't folded; create an explicit cast node. We can reuse
942 // the existing insert position since if we get here, we won't have
943 // made any changes which would invalidate it.
944 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
946 UniqueSCEVs.InsertNode(S, IP);
950 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
952 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
953 "This is not an extending conversion!");
954 assert(isSCEVable(Ty) &&
955 "This is not a conversion to a SCEVable type!");
956 Ty = getEffectiveSCEVType(Ty);
958 // Fold if the operand is constant.
959 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
961 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(),
962 getEffectiveSCEVType(Ty))));
964 // zext(zext(x)) --> zext(x)
965 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
966 return getZeroExtendExpr(SZ->getOperand(), Ty);
968 // Before doing any expensive analysis, check to see if we've already
969 // computed a SCEV for this Op and Ty.
971 ID.AddInteger(scZeroExtend);
975 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
977 // If the input value is a chrec scev, and we can prove that the value
978 // did not overflow the old, smaller, value, we can zero extend all of the
979 // operands (often constants). This allows analysis of something like
980 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
981 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
982 if (AR->isAffine()) {
983 const SCEV *Start = AR->getStart();
984 const SCEV *Step = AR->getStepRecurrence(*this);
985 unsigned BitWidth = getTypeSizeInBits(AR->getType());
986 const Loop *L = AR->getLoop();
988 // If we have special knowledge that this addrec won't overflow,
989 // we don't need to do any further analysis.
990 if (AR->hasNoUnsignedWrap())
991 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
992 getZeroExtendExpr(Step, Ty),
995 // Check whether the backedge-taken count is SCEVCouldNotCompute.
996 // Note that this serves two purposes: It filters out loops that are
997 // simply not analyzable, and it covers the case where this code is
998 // being called from within backedge-taken count analysis, such that
999 // attempting to ask for the backedge-taken count would likely result
1000 // in infinite recursion. In the later case, the analysis code will
1001 // cope with a conservative value, and it will take care to purge
1002 // that value once it has finished.
1003 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1004 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1005 // Manually compute the final value for AR, checking for
1008 // Check whether the backedge-taken count can be losslessly casted to
1009 // the addrec's type. The count is always unsigned.
1010 const SCEV *CastedMaxBECount =
1011 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1012 const SCEV *RecastedMaxBECount =
1013 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1014 if (MaxBECount == RecastedMaxBECount) {
1015 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1016 // Check whether Start+Step*MaxBECount has no unsigned overflow.
1017 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
1018 const SCEV *Add = getAddExpr(Start, ZMul);
1019 const SCEV *OperandExtendedAdd =
1020 getAddExpr(getZeroExtendExpr(Start, WideTy),
1021 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1022 getZeroExtendExpr(Step, WideTy)));
1023 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
1024 // Return the expression with the addrec on the outside.
1025 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1026 getZeroExtendExpr(Step, Ty),
1029 // Similar to above, only this time treat the step value as signed.
1030 // This covers loops that count down.
1031 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1032 Add = getAddExpr(Start, SMul);
1033 OperandExtendedAdd =
1034 getAddExpr(getZeroExtendExpr(Start, WideTy),
1035 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1036 getSignExtendExpr(Step, WideTy)));
1037 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
1038 // Return the expression with the addrec on the outside.
1039 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1040 getSignExtendExpr(Step, Ty),
1044 // If the backedge is guarded by a comparison with the pre-inc value
1045 // the addrec is safe. Also, if the entry is guarded by a comparison
1046 // with the start value and the backedge is guarded by a comparison
1047 // with the post-inc value, the addrec is safe.
1048 if (isKnownPositive(Step)) {
1049 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
1050 getUnsignedRange(Step).getUnsignedMax());
1051 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
1052 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
1053 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
1054 AR->getPostIncExpr(*this), N)))
1055 // Return the expression with the addrec on the outside.
1056 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1057 getZeroExtendExpr(Step, Ty),
1059 } else if (isKnownNegative(Step)) {
1060 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
1061 getSignedRange(Step).getSignedMin());
1062 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
1063 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
1064 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
1065 AR->getPostIncExpr(*this), N)))
1066 // Return the expression with the addrec on the outside.
1067 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1068 getSignExtendExpr(Step, Ty),
1074 // The cast wasn't folded; create an explicit cast node.
1075 // Recompute the insert position, as it may have been invalidated.
1076 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1077 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1079 UniqueSCEVs.InsertNode(S, IP);
1083 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1085 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1086 "This is not an extending conversion!");
1087 assert(isSCEVable(Ty) &&
1088 "This is not a conversion to a SCEVable type!");
1089 Ty = getEffectiveSCEVType(Ty);
1091 // Fold if the operand is constant.
1092 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1094 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(),
1095 getEffectiveSCEVType(Ty))));
1097 // sext(sext(x)) --> sext(x)
1098 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1099 return getSignExtendExpr(SS->getOperand(), Ty);
1101 // Before doing any expensive analysis, check to see if we've already
1102 // computed a SCEV for this Op and Ty.
1103 FoldingSetNodeID ID;
1104 ID.AddInteger(scSignExtend);
1108 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1110 // If the input value is a chrec scev, and we can prove that the value
1111 // did not overflow the old, smaller, value, we can sign extend all of the
1112 // operands (often constants). This allows analysis of something like
1113 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1114 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1115 if (AR->isAffine()) {
1116 const SCEV *Start = AR->getStart();
1117 const SCEV *Step = AR->getStepRecurrence(*this);
1118 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1119 const Loop *L = AR->getLoop();
1121 // If we have special knowledge that this addrec won't overflow,
1122 // we don't need to do any further analysis.
1123 if (AR->hasNoSignedWrap())
1124 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1125 getSignExtendExpr(Step, Ty),
1128 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1129 // Note that this serves two purposes: It filters out loops that are
1130 // simply not analyzable, and it covers the case where this code is
1131 // being called from within backedge-taken count analysis, such that
1132 // attempting to ask for the backedge-taken count would likely result
1133 // in infinite recursion. In the later case, the analysis code will
1134 // cope with a conservative value, and it will take care to purge
1135 // that value once it has finished.
1136 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1137 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1138 // Manually compute the final value for AR, checking for
1141 // Check whether the backedge-taken count can be losslessly casted to
1142 // the addrec's type. The count is always unsigned.
1143 const SCEV *CastedMaxBECount =
1144 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1145 const SCEV *RecastedMaxBECount =
1146 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1147 if (MaxBECount == RecastedMaxBECount) {
1148 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1149 // Check whether Start+Step*MaxBECount has no signed overflow.
1150 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1151 const SCEV *Add = getAddExpr(Start, SMul);
1152 const SCEV *OperandExtendedAdd =
1153 getAddExpr(getSignExtendExpr(Start, WideTy),
1154 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1155 getSignExtendExpr(Step, WideTy)));
1156 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1157 // Return the expression with the addrec on the outside.
1158 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1159 getSignExtendExpr(Step, Ty),
1162 // Similar to above, only this time treat the step value as unsigned.
1163 // This covers loops that count up with an unsigned step.
1164 const SCEV *UMul = getMulExpr(CastedMaxBECount, Step);
1165 Add = getAddExpr(Start, UMul);
1166 OperandExtendedAdd =
1167 getAddExpr(getSignExtendExpr(Start, WideTy),
1168 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1169 getZeroExtendExpr(Step, WideTy)));
1170 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1171 // Return the expression with the addrec on the outside.
1172 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1173 getZeroExtendExpr(Step, Ty),
1177 // If the backedge is guarded by a comparison with the pre-inc value
1178 // the addrec is safe. Also, if the entry is guarded by a comparison
1179 // with the start value and the backedge is guarded by a comparison
1180 // with the post-inc value, the addrec is safe.
1181 if (isKnownPositive(Step)) {
1182 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1183 getSignedRange(Step).getSignedMax());
1184 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1185 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1186 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1187 AR->getPostIncExpr(*this), N)))
1188 // Return the expression with the addrec on the outside.
1189 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1190 getSignExtendExpr(Step, Ty),
1192 } else if (isKnownNegative(Step)) {
1193 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1194 getSignedRange(Step).getSignedMin());
1195 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1196 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1197 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1198 AR->getPostIncExpr(*this), N)))
1199 // Return the expression with the addrec on the outside.
1200 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1201 getSignExtendExpr(Step, Ty),
1207 // The cast wasn't folded; create an explicit cast node.
1208 // Recompute the insert position, as it may have been invalidated.
1209 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1210 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1212 UniqueSCEVs.InsertNode(S, IP);
1216 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1217 /// unspecified bits out to the given type.
1219 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1221 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1222 "This is not an extending conversion!");
1223 assert(isSCEVable(Ty) &&
1224 "This is not a conversion to a SCEVable type!");
1225 Ty = getEffectiveSCEVType(Ty);
1227 // Sign-extend negative constants.
1228 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1229 if (SC->getValue()->getValue().isNegative())
1230 return getSignExtendExpr(Op, Ty);
1232 // Peel off a truncate cast.
1233 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1234 const SCEV *NewOp = T->getOperand();
1235 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1236 return getAnyExtendExpr(NewOp, Ty);
1237 return getTruncateOrNoop(NewOp, Ty);
1240 // Next try a zext cast. If the cast is folded, use it.
1241 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1242 if (!isa<SCEVZeroExtendExpr>(ZExt))
1245 // Next try a sext cast. If the cast is folded, use it.
1246 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1247 if (!isa<SCEVSignExtendExpr>(SExt))
1250 // Force the cast to be folded into the operands of an addrec.
1251 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1252 SmallVector<const SCEV *, 4> Ops;
1253 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1255 Ops.push_back(getAnyExtendExpr(*I, Ty));
1256 return getAddRecExpr(Ops, AR->getLoop());
1259 // As a special case, fold anyext(undef) to undef. We don't want to
1260 // know too much about SCEVUnknowns, but this special case is handy
1262 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
1263 if (isa<UndefValue>(U->getValue()))
1264 return getSCEV(UndefValue::get(Ty));
1266 // If the expression is obviously signed, use the sext cast value.
1267 if (isa<SCEVSMaxExpr>(Op))
1270 // Absent any other information, use the zext cast value.
1274 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1275 /// a list of operands to be added under the given scale, update the given
1276 /// map. This is a helper function for getAddRecExpr. As an example of
1277 /// what it does, given a sequence of operands that would form an add
1278 /// expression like this:
1280 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1282 /// where A and B are constants, update the map with these values:
1284 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1286 /// and add 13 + A*B*29 to AccumulatedConstant.
1287 /// This will allow getAddRecExpr to produce this:
1289 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1291 /// This form often exposes folding opportunities that are hidden in
1292 /// the original operand list.
1294 /// Return true iff it appears that any interesting folding opportunities
1295 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1296 /// the common case where no interesting opportunities are present, and
1297 /// is also used as a check to avoid infinite recursion.
1300 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1301 SmallVector<const SCEV *, 8> &NewOps,
1302 APInt &AccumulatedConstant,
1303 const SCEV *const *Ops, size_t NumOperands,
1305 ScalarEvolution &SE) {
1306 bool Interesting = false;
1308 // Iterate over the add operands. They are sorted, with constants first.
1310 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1312 // Pull a buried constant out to the outside.
1313 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1315 AccumulatedConstant += Scale * C->getValue()->getValue();
1318 // Next comes everything else. We're especially interested in multiplies
1319 // here, but they're in the middle, so just visit the rest with one loop.
1320 for (; i != NumOperands; ++i) {
1321 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1322 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1324 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1325 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1326 // A multiplication of a constant with another add; recurse.
1327 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1329 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1330 Add->op_begin(), Add->getNumOperands(),
1333 // A multiplication of a constant with some other value. Update
1335 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1336 const SCEV *Key = SE.getMulExpr(MulOps);
1337 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1338 M.insert(std::make_pair(Key, NewScale));
1340 NewOps.push_back(Pair.first->first);
1342 Pair.first->second += NewScale;
1343 // The map already had an entry for this value, which may indicate
1344 // a folding opportunity.
1349 // An ordinary operand. Update the map.
1350 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1351 M.insert(std::make_pair(Ops[i], Scale));
1353 NewOps.push_back(Pair.first->first);
1355 Pair.first->second += Scale;
1356 // The map already had an entry for this value, which may indicate
1357 // a folding opportunity.
1367 struct APIntCompare {
1368 bool operator()(const APInt &LHS, const APInt &RHS) const {
1369 return LHS.ult(RHS);
1374 /// getAddExpr - Get a canonical add expression, or something simpler if
1376 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1377 bool HasNUW, bool HasNSW) {
1378 assert(!Ops.empty() && "Cannot get empty add!");
1379 if (Ops.size() == 1) return Ops[0];
1381 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1382 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1383 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1384 "SCEVAddExpr operand types don't match!");
1387 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1388 if (!HasNUW && HasNSW) {
1390 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1391 E = Ops.end(); I != E; ++I)
1392 if (!isKnownNonNegative(*I)) {
1396 if (All) HasNUW = true;
1399 // Sort by complexity, this groups all similar expression types together.
1400 GroupByComplexity(Ops, LI);
1402 // If there are any constants, fold them together.
1404 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1406 assert(Idx < Ops.size());
1407 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1408 // We found two constants, fold them together!
1409 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1410 RHSC->getValue()->getValue());
1411 if (Ops.size() == 2) return Ops[0];
1412 Ops.erase(Ops.begin()+1); // Erase the folded element
1413 LHSC = cast<SCEVConstant>(Ops[0]);
1416 // If we are left with a constant zero being added, strip it off.
1417 if (LHSC->getValue()->isZero()) {
1418 Ops.erase(Ops.begin());
1422 if (Ops.size() == 1) return Ops[0];
1425 // Okay, check to see if the same value occurs in the operand list more than
1426 // once. If so, merge them together into an multiply expression. Since we
1427 // sorted the list, these values are required to be adjacent.
1428 const Type *Ty = Ops[0]->getType();
1429 bool FoundMatch = false;
1430 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1431 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1432 // Scan ahead to count how many equal operands there are.
1434 while (i+Count != e && Ops[i+Count] == Ops[i])
1436 // Merge the values into a multiply.
1437 const SCEV *Scale = getConstant(Ty, Count);
1438 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1439 if (Ops.size() == Count)
1442 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1443 --i; e -= Count - 1;
1447 return getAddExpr(Ops, HasNUW, HasNSW);
1449 // Check for truncates. If all the operands are truncated from the same
1450 // type, see if factoring out the truncate would permit the result to be
1451 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1452 // if the contents of the resulting outer trunc fold to something simple.
1453 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1454 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1455 const Type *DstType = Trunc->getType();
1456 const Type *SrcType = Trunc->getOperand()->getType();
1457 SmallVector<const SCEV *, 8> LargeOps;
1459 // Check all the operands to see if they can be represented in the
1460 // source type of the truncate.
1461 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1462 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1463 if (T->getOperand()->getType() != SrcType) {
1467 LargeOps.push_back(T->getOperand());
1468 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1469 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1470 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1471 SmallVector<const SCEV *, 8> LargeMulOps;
1472 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1473 if (const SCEVTruncateExpr *T =
1474 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1475 if (T->getOperand()->getType() != SrcType) {
1479 LargeMulOps.push_back(T->getOperand());
1480 } else if (const SCEVConstant *C =
1481 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1482 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1489 LargeOps.push_back(getMulExpr(LargeMulOps));
1496 // Evaluate the expression in the larger type.
1497 const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
1498 // If it folds to something simple, use it. Otherwise, don't.
1499 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1500 return getTruncateExpr(Fold, DstType);
1504 // Skip past any other cast SCEVs.
1505 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1508 // If there are add operands they would be next.
1509 if (Idx < Ops.size()) {
1510 bool DeletedAdd = false;
1511 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1512 // If we have an add, expand the add operands onto the end of the operands
1514 Ops.erase(Ops.begin()+Idx);
1515 Ops.append(Add->op_begin(), Add->op_end());
1519 // If we deleted at least one add, we added operands to the end of the list,
1520 // and they are not necessarily sorted. Recurse to resort and resimplify
1521 // any operands we just acquired.
1523 return getAddExpr(Ops);
1526 // Skip over the add expression until we get to a multiply.
1527 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1530 // Check to see if there are any folding opportunities present with
1531 // operands multiplied by constant values.
1532 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1533 uint64_t BitWidth = getTypeSizeInBits(Ty);
1534 DenseMap<const SCEV *, APInt> M;
1535 SmallVector<const SCEV *, 8> NewOps;
1536 APInt AccumulatedConstant(BitWidth, 0);
1537 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1538 Ops.data(), Ops.size(),
1539 APInt(BitWidth, 1), *this)) {
1540 // Some interesting folding opportunity is present, so its worthwhile to
1541 // re-generate the operands list. Group the operands by constant scale,
1542 // to avoid multiplying by the same constant scale multiple times.
1543 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1544 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1545 E = NewOps.end(); I != E; ++I)
1546 MulOpLists[M.find(*I)->second].push_back(*I);
1547 // Re-generate the operands list.
1549 if (AccumulatedConstant != 0)
1550 Ops.push_back(getConstant(AccumulatedConstant));
1551 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1552 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1554 Ops.push_back(getMulExpr(getConstant(I->first),
1555 getAddExpr(I->second)));
1557 return getConstant(Ty, 0);
1558 if (Ops.size() == 1)
1560 return getAddExpr(Ops);
1564 // If we are adding something to a multiply expression, make sure the
1565 // something is not already an operand of the multiply. If so, merge it into
1567 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1568 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1569 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1570 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1571 if (isa<SCEVConstant>(MulOpSCEV))
1573 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1574 if (MulOpSCEV == Ops[AddOp]) {
1575 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1576 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1577 if (Mul->getNumOperands() != 2) {
1578 // If the multiply has more than two operands, we must get the
1580 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1581 Mul->op_begin()+MulOp);
1582 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1583 InnerMul = getMulExpr(MulOps);
1585 const SCEV *One = getConstant(Ty, 1);
1586 const SCEV *AddOne = getAddExpr(One, InnerMul);
1587 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1588 if (Ops.size() == 2) return OuterMul;
1590 Ops.erase(Ops.begin()+AddOp);
1591 Ops.erase(Ops.begin()+Idx-1);
1593 Ops.erase(Ops.begin()+Idx);
1594 Ops.erase(Ops.begin()+AddOp-1);
1596 Ops.push_back(OuterMul);
1597 return getAddExpr(Ops);
1600 // Check this multiply against other multiplies being added together.
1601 for (unsigned OtherMulIdx = Idx+1;
1602 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1604 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1605 // If MulOp occurs in OtherMul, we can fold the two multiplies
1607 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1608 OMulOp != e; ++OMulOp)
1609 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1610 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1611 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1612 if (Mul->getNumOperands() != 2) {
1613 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1614 Mul->op_begin()+MulOp);
1615 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1616 InnerMul1 = getMulExpr(MulOps);
1618 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1619 if (OtherMul->getNumOperands() != 2) {
1620 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1621 OtherMul->op_begin()+OMulOp);
1622 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1623 InnerMul2 = getMulExpr(MulOps);
1625 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1626 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1627 if (Ops.size() == 2) return OuterMul;
1628 Ops.erase(Ops.begin()+Idx);
1629 Ops.erase(Ops.begin()+OtherMulIdx-1);
1630 Ops.push_back(OuterMul);
1631 return getAddExpr(Ops);
1637 // If there are any add recurrences in the operands list, see if any other
1638 // added values are loop invariant. If so, we can fold them into the
1640 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1643 // Scan over all recurrences, trying to fold loop invariants into them.
1644 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1645 // Scan all of the other operands to this add and add them to the vector if
1646 // they are loop invariant w.r.t. the recurrence.
1647 SmallVector<const SCEV *, 8> LIOps;
1648 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1649 const Loop *AddRecLoop = AddRec->getLoop();
1650 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1651 if (Ops[i]->isLoopInvariant(AddRecLoop)) {
1652 LIOps.push_back(Ops[i]);
1653 Ops.erase(Ops.begin()+i);
1657 // If we found some loop invariants, fold them into the recurrence.
1658 if (!LIOps.empty()) {
1659 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1660 LIOps.push_back(AddRec->getStart());
1662 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1664 AddRecOps[0] = getAddExpr(LIOps);
1666 // Build the new addrec. Propagate the NUW and NSW flags if both the
1667 // outer add and the inner addrec are guaranteed to have no overflow.
1668 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop,
1669 HasNUW && AddRec->hasNoUnsignedWrap(),
1670 HasNSW && AddRec->hasNoSignedWrap());
1672 // If all of the other operands were loop invariant, we are done.
1673 if (Ops.size() == 1) return NewRec;
1675 // Otherwise, add the folded AddRec by the non-liv parts.
1676 for (unsigned i = 0;; ++i)
1677 if (Ops[i] == AddRec) {
1681 return getAddExpr(Ops);
1684 // Okay, if there weren't any loop invariants to be folded, check to see if
1685 // there are multiple AddRec's with the same loop induction variable being
1686 // added together. If so, we can fold them.
1687 for (unsigned OtherIdx = Idx+1;
1688 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1690 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1691 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1692 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1694 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1696 if (const SCEVAddRecExpr *OtherAddRec =
1697 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1698 if (OtherAddRec->getLoop() == AddRecLoop) {
1699 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1701 if (i >= AddRecOps.size()) {
1702 AddRecOps.append(OtherAddRec->op_begin()+i,
1703 OtherAddRec->op_end());
1706 AddRecOps[i] = getAddExpr(AddRecOps[i],
1707 OtherAddRec->getOperand(i));
1709 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1711 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop);
1712 return getAddExpr(Ops);
1715 // Otherwise couldn't fold anything into this recurrence. Move onto the
1719 // Okay, it looks like we really DO need an add expr. Check to see if we
1720 // already have one, otherwise create a new one.
1721 FoldingSetNodeID ID;
1722 ID.AddInteger(scAddExpr);
1723 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1724 ID.AddPointer(Ops[i]);
1727 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1729 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1730 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1731 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1733 UniqueSCEVs.InsertNode(S, IP);
1735 if (HasNUW) S->setHasNoUnsignedWrap(true);
1736 if (HasNSW) S->setHasNoSignedWrap(true);
1740 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1742 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1743 bool HasNUW, bool HasNSW) {
1744 assert(!Ops.empty() && "Cannot get empty mul!");
1745 if (Ops.size() == 1) return Ops[0];
1747 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1748 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1749 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1750 "SCEVMulExpr operand types don't match!");
1753 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1754 if (!HasNUW && HasNSW) {
1756 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1757 E = Ops.end(); I != E; ++I)
1758 if (!isKnownNonNegative(*I)) {
1762 if (All) HasNUW = true;
1765 // Sort by complexity, this groups all similar expression types together.
1766 GroupByComplexity(Ops, LI);
1768 // If there are any constants, fold them together.
1770 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1772 // C1*(C2+V) -> C1*C2 + C1*V
1773 if (Ops.size() == 2)
1774 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1775 if (Add->getNumOperands() == 2 &&
1776 isa<SCEVConstant>(Add->getOperand(0)))
1777 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1778 getMulExpr(LHSC, Add->getOperand(1)));
1781 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1782 // We found two constants, fold them together!
1783 ConstantInt *Fold = ConstantInt::get(getContext(),
1784 LHSC->getValue()->getValue() *
1785 RHSC->getValue()->getValue());
1786 Ops[0] = getConstant(Fold);
1787 Ops.erase(Ops.begin()+1); // Erase the folded element
1788 if (Ops.size() == 1) return Ops[0];
1789 LHSC = cast<SCEVConstant>(Ops[0]);
1792 // If we are left with a constant one being multiplied, strip it off.
1793 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1794 Ops.erase(Ops.begin());
1796 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1797 // If we have a multiply of zero, it will always be zero.
1799 } else if (Ops[0]->isAllOnesValue()) {
1800 // If we have a mul by -1 of an add, try distributing the -1 among the
1802 if (Ops.size() == 2)
1803 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1804 SmallVector<const SCEV *, 4> NewOps;
1805 bool AnyFolded = false;
1806 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1808 const SCEV *Mul = getMulExpr(Ops[0], *I);
1809 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1810 NewOps.push_back(Mul);
1813 return getAddExpr(NewOps);
1817 if (Ops.size() == 1)
1821 // Skip over the add expression until we get to a multiply.
1822 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1825 // If there are mul operands inline them all into this expression.
1826 if (Idx < Ops.size()) {
1827 bool DeletedMul = false;
1828 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1829 // If we have an mul, expand the mul operands onto the end of the operands
1831 Ops.erase(Ops.begin()+Idx);
1832 Ops.append(Mul->op_begin(), Mul->op_end());
1836 // If we deleted at least one mul, we added operands to the end of the list,
1837 // and they are not necessarily sorted. Recurse to resort and resimplify
1838 // any operands we just acquired.
1840 return getMulExpr(Ops);
1843 // If there are any add recurrences in the operands list, see if any other
1844 // added values are loop invariant. If so, we can fold them into the
1846 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1849 // Scan over all recurrences, trying to fold loop invariants into them.
1850 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1851 // Scan all of the other operands to this mul and add them to the vector if
1852 // they are loop invariant w.r.t. the recurrence.
1853 SmallVector<const SCEV *, 8> LIOps;
1854 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1855 const Loop *AddRecLoop = AddRec->getLoop();
1856 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1857 if (Ops[i]->isLoopInvariant(AddRecLoop)) {
1858 LIOps.push_back(Ops[i]);
1859 Ops.erase(Ops.begin()+i);
1863 // If we found some loop invariants, fold them into the recurrence.
1864 if (!LIOps.empty()) {
1865 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1866 SmallVector<const SCEV *, 4> NewOps;
1867 NewOps.reserve(AddRec->getNumOperands());
1868 const SCEV *Scale = getMulExpr(LIOps);
1869 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1870 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1872 // Build the new addrec. Propagate the NUW and NSW flags if both the
1873 // outer mul and the inner addrec are guaranteed to have no overflow.
1874 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop,
1875 HasNUW && AddRec->hasNoUnsignedWrap(),
1876 HasNSW && AddRec->hasNoSignedWrap());
1878 // If all of the other operands were loop invariant, we are done.
1879 if (Ops.size() == 1) return NewRec;
1881 // Otherwise, multiply the folded AddRec by the non-liv parts.
1882 for (unsigned i = 0;; ++i)
1883 if (Ops[i] == AddRec) {
1887 return getMulExpr(Ops);
1890 // Okay, if there weren't any loop invariants to be folded, check to see if
1891 // there are multiple AddRec's with the same loop induction variable being
1892 // multiplied together. If so, we can fold them.
1893 for (unsigned OtherIdx = Idx+1;
1894 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1896 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1897 // F * G, where F = {A,+,B}<L> and G = {C,+,D}<L> -->
1898 // {A*C,+,F*D + G*B + B*D}<L>
1899 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1901 if (const SCEVAddRecExpr *OtherAddRec =
1902 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1903 if (OtherAddRec->getLoop() == AddRecLoop) {
1904 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1905 const SCEV *NewStart = getMulExpr(F->getStart(), G->getStart());
1906 const SCEV *B = F->getStepRecurrence(*this);
1907 const SCEV *D = G->getStepRecurrence(*this);
1908 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1911 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1913 if (Ops.size() == 2) return NewAddRec;
1914 Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
1915 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1917 return getMulExpr(Ops);
1920 // Otherwise couldn't fold anything into this recurrence. Move onto the
1924 // Okay, it looks like we really DO need an mul expr. Check to see if we
1925 // already have one, otherwise create a new one.
1926 FoldingSetNodeID ID;
1927 ID.AddInteger(scMulExpr);
1928 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1929 ID.AddPointer(Ops[i]);
1932 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1934 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1935 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1936 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
1938 UniqueSCEVs.InsertNode(S, IP);
1940 if (HasNUW) S->setHasNoUnsignedWrap(true);
1941 if (HasNSW) S->setHasNoSignedWrap(true);
1945 /// getUDivExpr - Get a canonical unsigned division expression, or something
1946 /// simpler if possible.
1947 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1949 assert(getEffectiveSCEVType(LHS->getType()) ==
1950 getEffectiveSCEVType(RHS->getType()) &&
1951 "SCEVUDivExpr operand types don't match!");
1953 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1954 if (RHSC->getValue()->equalsInt(1))
1955 return LHS; // X udiv 1 --> x
1956 // If the denominator is zero, the result of the udiv is undefined. Don't
1957 // try to analyze it, because the resolution chosen here may differ from
1958 // the resolution chosen in other parts of the compiler.
1959 if (!RHSC->getValue()->isZero()) {
1960 // Determine if the division can be folded into the operands of
1962 // TODO: Generalize this to non-constants by using known-bits information.
1963 const Type *Ty = LHS->getType();
1964 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1965 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
1966 // For non-power-of-two values, effectively round the value up to the
1967 // nearest power of two.
1968 if (!RHSC->getValue()->getValue().isPowerOf2())
1970 const IntegerType *ExtTy =
1971 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1972 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1973 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1974 if (const SCEVConstant *Step =
1975 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1976 if (!Step->getValue()->getValue()
1977 .urem(RHSC->getValue()->getValue()) &&
1978 getZeroExtendExpr(AR, ExtTy) ==
1979 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1980 getZeroExtendExpr(Step, ExtTy),
1982 SmallVector<const SCEV *, 4> Operands;
1983 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1984 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1985 return getAddRecExpr(Operands, AR->getLoop());
1987 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1988 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1989 SmallVector<const SCEV *, 4> Operands;
1990 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1991 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1992 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1993 // Find an operand that's safely divisible.
1994 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1995 const SCEV *Op = M->getOperand(i);
1996 const SCEV *Div = getUDivExpr(Op, RHSC);
1997 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1998 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
2001 return getMulExpr(Operands);
2005 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
2006 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
2007 SmallVector<const SCEV *, 4> Operands;
2008 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
2009 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
2010 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
2012 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2013 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
2014 if (isa<SCEVUDivExpr>(Op) ||
2015 getMulExpr(Op, RHS) != A->getOperand(i))
2017 Operands.push_back(Op);
2019 if (Operands.size() == A->getNumOperands())
2020 return getAddExpr(Operands);
2024 // Fold if both operands are constant.
2025 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
2026 Constant *LHSCV = LHSC->getValue();
2027 Constant *RHSCV = RHSC->getValue();
2028 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
2034 FoldingSetNodeID ID;
2035 ID.AddInteger(scUDivExpr);
2039 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2040 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
2042 UniqueSCEVs.InsertNode(S, IP);
2047 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2048 /// Simplify the expression as much as possible.
2049 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
2050 const SCEV *Step, const Loop *L,
2051 bool HasNUW, bool HasNSW) {
2052 SmallVector<const SCEV *, 4> Operands;
2053 Operands.push_back(Start);
2054 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
2055 if (StepChrec->getLoop() == L) {
2056 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
2057 return getAddRecExpr(Operands, L);
2060 Operands.push_back(Step);
2061 return getAddRecExpr(Operands, L, HasNUW, HasNSW);
2064 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2065 /// Simplify the expression as much as possible.
2067 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2069 bool HasNUW, bool HasNSW) {
2070 if (Operands.size() == 1) return Operands[0];
2072 const Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2073 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2074 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2075 "SCEVAddRecExpr operand types don't match!");
2076 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2077 assert(Operands[i]->isLoopInvariant(L) &&
2078 "SCEVAddRecExpr operand is not loop-invariant!");
2081 if (Operands.back()->isZero()) {
2082 Operands.pop_back();
2083 return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
2086 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2087 // use that information to infer NUW and NSW flags. However, computing a
2088 // BE count requires calling getAddRecExpr, so we may not yet have a
2089 // meaningful BE count at this point (and if we don't, we'd be stuck
2090 // with a SCEVCouldNotCompute as the cached BE count).
2092 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
2093 if (!HasNUW && HasNSW) {
2095 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2096 E = Operands.end(); I != E; ++I)
2097 if (!isKnownNonNegative(*I)) {
2101 if (All) HasNUW = true;
2104 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2105 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2106 const Loop *NestedLoop = NestedAR->getLoop();
2107 if (L->contains(NestedLoop) ?
2108 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2109 (!NestedLoop->contains(L) &&
2110 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2111 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2112 NestedAR->op_end());
2113 Operands[0] = NestedAR->getStart();
2114 // AddRecs require their operands be loop-invariant with respect to their
2115 // loops. Don't perform this transformation if it would break this
2117 bool AllInvariant = true;
2118 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2119 if (!Operands[i]->isLoopInvariant(L)) {
2120 AllInvariant = false;
2124 NestedOperands[0] = getAddRecExpr(Operands, L);
2125 AllInvariant = true;
2126 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2127 if (!NestedOperands[i]->isLoopInvariant(NestedLoop)) {
2128 AllInvariant = false;
2132 // Ok, both add recurrences are valid after the transformation.
2133 return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
2135 // Reset Operands to its original state.
2136 Operands[0] = NestedAR;
2140 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2141 // already have one, otherwise create a new one.
2142 FoldingSetNodeID ID;
2143 ID.AddInteger(scAddRecExpr);
2144 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2145 ID.AddPointer(Operands[i]);
2149 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2151 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2152 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2153 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2154 O, Operands.size(), L);
2155 UniqueSCEVs.InsertNode(S, IP);
2157 if (HasNUW) S->setHasNoUnsignedWrap(true);
2158 if (HasNSW) S->setHasNoSignedWrap(true);
2162 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2164 SmallVector<const SCEV *, 2> Ops;
2167 return getSMaxExpr(Ops);
2171 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2172 assert(!Ops.empty() && "Cannot get empty smax!");
2173 if (Ops.size() == 1) return Ops[0];
2175 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2176 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2177 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2178 "SCEVSMaxExpr operand types don't match!");
2181 // Sort by complexity, this groups all similar expression types together.
2182 GroupByComplexity(Ops, LI);
2184 // If there are any constants, fold them together.
2186 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2188 assert(Idx < Ops.size());
2189 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2190 // We found two constants, fold them together!
2191 ConstantInt *Fold = ConstantInt::get(getContext(),
2192 APIntOps::smax(LHSC->getValue()->getValue(),
2193 RHSC->getValue()->getValue()));
2194 Ops[0] = getConstant(Fold);
2195 Ops.erase(Ops.begin()+1); // Erase the folded element
2196 if (Ops.size() == 1) return Ops[0];
2197 LHSC = cast<SCEVConstant>(Ops[0]);
2200 // If we are left with a constant minimum-int, strip it off.
2201 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2202 Ops.erase(Ops.begin());
2204 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2205 // If we have an smax with a constant maximum-int, it will always be
2210 if (Ops.size() == 1) return Ops[0];
2213 // Find the first SMax
2214 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2217 // Check to see if one of the operands is an SMax. If so, expand its operands
2218 // onto our operand list, and recurse to simplify.
2219 if (Idx < Ops.size()) {
2220 bool DeletedSMax = false;
2221 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2222 Ops.erase(Ops.begin()+Idx);
2223 Ops.append(SMax->op_begin(), SMax->op_end());
2228 return getSMaxExpr(Ops);
2231 // Okay, check to see if the same value occurs in the operand list twice. If
2232 // so, delete one. Since we sorted the list, these values are required to
2234 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2235 // X smax Y smax Y --> X smax Y
2236 // X smax Y --> X, if X is always greater than Y
2237 if (Ops[i] == Ops[i+1] ||
2238 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2239 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2241 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2242 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2246 if (Ops.size() == 1) return Ops[0];
2248 assert(!Ops.empty() && "Reduced smax down to nothing!");
2250 // Okay, it looks like we really DO need an smax expr. Check to see if we
2251 // already have one, otherwise create a new one.
2252 FoldingSetNodeID ID;
2253 ID.AddInteger(scSMaxExpr);
2254 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2255 ID.AddPointer(Ops[i]);
2257 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2258 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2259 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2260 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2262 UniqueSCEVs.InsertNode(S, IP);
2266 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2268 SmallVector<const SCEV *, 2> Ops;
2271 return getUMaxExpr(Ops);
2275 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2276 assert(!Ops.empty() && "Cannot get empty umax!");
2277 if (Ops.size() == 1) return Ops[0];
2279 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2280 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2281 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2282 "SCEVUMaxExpr operand types don't match!");
2285 // Sort by complexity, this groups all similar expression types together.
2286 GroupByComplexity(Ops, LI);
2288 // If there are any constants, fold them together.
2290 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2292 assert(Idx < Ops.size());
2293 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2294 // We found two constants, fold them together!
2295 ConstantInt *Fold = ConstantInt::get(getContext(),
2296 APIntOps::umax(LHSC->getValue()->getValue(),
2297 RHSC->getValue()->getValue()));
2298 Ops[0] = getConstant(Fold);
2299 Ops.erase(Ops.begin()+1); // Erase the folded element
2300 if (Ops.size() == 1) return Ops[0];
2301 LHSC = cast<SCEVConstant>(Ops[0]);
2304 // If we are left with a constant minimum-int, strip it off.
2305 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2306 Ops.erase(Ops.begin());
2308 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2309 // If we have an umax with a constant maximum-int, it will always be
2314 if (Ops.size() == 1) return Ops[0];
2317 // Find the first UMax
2318 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2321 // Check to see if one of the operands is a UMax. If so, expand its operands
2322 // onto our operand list, and recurse to simplify.
2323 if (Idx < Ops.size()) {
2324 bool DeletedUMax = false;
2325 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2326 Ops.erase(Ops.begin()+Idx);
2327 Ops.append(UMax->op_begin(), UMax->op_end());
2332 return getUMaxExpr(Ops);
2335 // Okay, check to see if the same value occurs in the operand list twice. If
2336 // so, delete one. Since we sorted the list, these values are required to
2338 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2339 // X umax Y umax Y --> X umax Y
2340 // X umax Y --> X, if X is always greater than Y
2341 if (Ops[i] == Ops[i+1] ||
2342 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2343 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2345 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2346 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2350 if (Ops.size() == 1) return Ops[0];
2352 assert(!Ops.empty() && "Reduced umax down to nothing!");
2354 // Okay, it looks like we really DO need a umax expr. Check to see if we
2355 // already have one, otherwise create a new one.
2356 FoldingSetNodeID ID;
2357 ID.AddInteger(scUMaxExpr);
2358 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2359 ID.AddPointer(Ops[i]);
2361 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2362 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2363 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2364 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2366 UniqueSCEVs.InsertNode(S, IP);
2370 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2372 // ~smax(~x, ~y) == smin(x, y).
2373 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2376 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2378 // ~umax(~x, ~y) == umin(x, y)
2379 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2382 const SCEV *ScalarEvolution::getSizeOfExpr(const Type *AllocTy) {
2383 // If we have TargetData, we can bypass creating a target-independent
2384 // constant expression and then folding it back into a ConstantInt.
2385 // This is just a compile-time optimization.
2387 return getConstant(TD->getIntPtrType(getContext()),
2388 TD->getTypeAllocSize(AllocTy));
2390 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2391 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2392 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2394 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2395 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2398 const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
2399 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2400 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2401 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2403 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2404 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2407 const SCEV *ScalarEvolution::getOffsetOfExpr(const StructType *STy,
2409 // If we have TargetData, we can bypass creating a target-independent
2410 // constant expression and then folding it back into a ConstantInt.
2411 // This is just a compile-time optimization.
2413 return getConstant(TD->getIntPtrType(getContext()),
2414 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2416 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2417 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2418 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2420 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2421 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2424 const SCEV *ScalarEvolution::getOffsetOfExpr(const Type *CTy,
2425 Constant *FieldNo) {
2426 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2427 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2428 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2430 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2431 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2434 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2435 // Don't attempt to do anything other than create a SCEVUnknown object
2436 // here. createSCEV only calls getUnknown after checking for all other
2437 // interesting possibilities, and any other code that calls getUnknown
2438 // is doing so in order to hide a value from SCEV canonicalization.
2440 FoldingSetNodeID ID;
2441 ID.AddInteger(scUnknown);
2444 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2445 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2446 "Stale SCEVUnknown in uniquing map!");
2449 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2451 FirstUnknown = cast<SCEVUnknown>(S);
2452 UniqueSCEVs.InsertNode(S, IP);
2456 //===----------------------------------------------------------------------===//
2457 // Basic SCEV Analysis and PHI Idiom Recognition Code
2460 /// isSCEVable - Test if values of the given type are analyzable within
2461 /// the SCEV framework. This primarily includes integer types, and it
2462 /// can optionally include pointer types if the ScalarEvolution class
2463 /// has access to target-specific information.
2464 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2465 // Integers and pointers are always SCEVable.
2466 return Ty->isIntegerTy() || Ty->isPointerTy();
2469 /// getTypeSizeInBits - Return the size in bits of the specified type,
2470 /// for which isSCEVable must return true.
2471 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2472 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2474 // If we have a TargetData, use it!
2476 return TD->getTypeSizeInBits(Ty);
2478 // Integer types have fixed sizes.
2479 if (Ty->isIntegerTy())
2480 return Ty->getPrimitiveSizeInBits();
2482 // The only other support type is pointer. Without TargetData, conservatively
2483 // assume pointers are 64-bit.
2484 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2488 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2489 /// the given type and which represents how SCEV will treat the given
2490 /// type, for which isSCEVable must return true. For pointer types,
2491 /// this is the pointer-sized integer type.
2492 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2493 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2495 if (Ty->isIntegerTy())
2498 // The only other support type is pointer.
2499 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2500 if (TD) return TD->getIntPtrType(getContext());
2502 // Without TargetData, conservatively assume pointers are 64-bit.
2503 return Type::getInt64Ty(getContext());
2506 const SCEV *ScalarEvolution::getCouldNotCompute() {
2507 return &CouldNotCompute;
2510 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2511 /// expression and create a new one.
2512 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2513 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2515 ValueExprMapType::const_iterator I = ValueExprMap.find(V);
2516 if (I != ValueExprMap.end()) return I->second;
2517 const SCEV *S = createSCEV(V);
2519 // The process of creating a SCEV for V may have caused other SCEVs
2520 // to have been created, so it's necessary to insert the new entry
2521 // from scratch, rather than trying to remember the insert position
2523 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2527 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2529 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2530 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2532 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2534 const Type *Ty = V->getType();
2535 Ty = getEffectiveSCEVType(Ty);
2536 return getMulExpr(V,
2537 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2540 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2541 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2542 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2544 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2546 const Type *Ty = V->getType();
2547 Ty = getEffectiveSCEVType(Ty);
2548 const SCEV *AllOnes =
2549 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2550 return getMinusSCEV(AllOnes, V);
2553 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
2555 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS,
2557 // Fast path: X - X --> 0.
2559 return getConstant(LHS->getType(), 0);
2562 return getAddExpr(LHS, getNegativeSCEV(RHS));
2565 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2566 /// input value to the specified type. If the type must be extended, it is zero
2569 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V,
2571 const Type *SrcTy = V->getType();
2572 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2573 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2574 "Cannot truncate or zero extend with non-integer arguments!");
2575 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2576 return V; // No conversion
2577 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2578 return getTruncateExpr(V, Ty);
2579 return getZeroExtendExpr(V, Ty);
2582 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2583 /// input value to the specified type. If the type must be extended, it is sign
2586 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2588 const Type *SrcTy = V->getType();
2589 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2590 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2591 "Cannot truncate or zero extend with non-integer arguments!");
2592 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2593 return V; // No conversion
2594 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2595 return getTruncateExpr(V, Ty);
2596 return getSignExtendExpr(V, Ty);
2599 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2600 /// input value to the specified type. If the type must be extended, it is zero
2601 /// extended. The conversion must not be narrowing.
2603 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2604 const Type *SrcTy = V->getType();
2605 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2606 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2607 "Cannot noop or zero extend with non-integer arguments!");
2608 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2609 "getNoopOrZeroExtend cannot truncate!");
2610 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2611 return V; // No conversion
2612 return getZeroExtendExpr(V, Ty);
2615 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2616 /// input value to the specified type. If the type must be extended, it is sign
2617 /// extended. The conversion must not be narrowing.
2619 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
2620 const Type *SrcTy = V->getType();
2621 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2622 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2623 "Cannot noop or sign extend with non-integer arguments!");
2624 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2625 "getNoopOrSignExtend cannot truncate!");
2626 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2627 return V; // No conversion
2628 return getSignExtendExpr(V, Ty);
2631 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2632 /// the input value to the specified type. If the type must be extended,
2633 /// it is extended with unspecified bits. The conversion must not be
2636 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2637 const Type *SrcTy = V->getType();
2638 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2639 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2640 "Cannot noop or any extend with non-integer arguments!");
2641 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2642 "getNoopOrAnyExtend cannot truncate!");
2643 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2644 return V; // No conversion
2645 return getAnyExtendExpr(V, Ty);
2648 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2649 /// input value to the specified type. The conversion must not be widening.
2651 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2652 const Type *SrcTy = V->getType();
2653 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2654 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2655 "Cannot truncate or noop with non-integer arguments!");
2656 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2657 "getTruncateOrNoop cannot extend!");
2658 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2659 return V; // No conversion
2660 return getTruncateExpr(V, Ty);
2663 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2664 /// the types using zero-extension, and then perform a umax operation
2666 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2668 const SCEV *PromotedLHS = LHS;
2669 const SCEV *PromotedRHS = RHS;
2671 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2672 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2674 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2676 return getUMaxExpr(PromotedLHS, PromotedRHS);
2679 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2680 /// the types using zero-extension, and then perform a umin operation
2682 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2684 const SCEV *PromotedLHS = LHS;
2685 const SCEV *PromotedRHS = RHS;
2687 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2688 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2690 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2692 return getUMinExpr(PromotedLHS, PromotedRHS);
2695 /// PushDefUseChildren - Push users of the given Instruction
2696 /// onto the given Worklist.
2698 PushDefUseChildren(Instruction *I,
2699 SmallVectorImpl<Instruction *> &Worklist) {
2700 // Push the def-use children onto the Worklist stack.
2701 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2703 Worklist.push_back(cast<Instruction>(*UI));
2706 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2707 /// instructions that depend on the given instruction and removes them from
2708 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2711 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2712 SmallVector<Instruction *, 16> Worklist;
2713 PushDefUseChildren(PN, Worklist);
2715 SmallPtrSet<Instruction *, 8> Visited;
2717 while (!Worklist.empty()) {
2718 Instruction *I = Worklist.pop_back_val();
2719 if (!Visited.insert(I)) continue;
2721 ValueExprMapType::iterator It =
2722 ValueExprMap.find(static_cast<Value *>(I));
2723 if (It != ValueExprMap.end()) {
2724 const SCEV *Old = It->second;
2726 // Short-circuit the def-use traversal if the symbolic name
2727 // ceases to appear in expressions.
2728 if (Old != SymName && !Old->hasOperand(SymName))
2731 // SCEVUnknown for a PHI either means that it has an unrecognized
2732 // structure, it's a PHI that's in the progress of being computed
2733 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2734 // additional loop trip count information isn't going to change anything.
2735 // In the second case, createNodeForPHI will perform the necessary
2736 // updates on its own when it gets to that point. In the third, we do
2737 // want to forget the SCEVUnknown.
2738 if (!isa<PHINode>(I) ||
2739 !isa<SCEVUnknown>(Old) ||
2740 (I != PN && Old == SymName)) {
2741 ValuesAtScopes.erase(Old);
2742 UnsignedRanges.erase(Old);
2743 SignedRanges.erase(Old);
2744 ValueExprMap.erase(It);
2748 PushDefUseChildren(I, Worklist);
2752 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2753 /// a loop header, making it a potential recurrence, or it doesn't.
2755 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2756 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2757 if (L->getHeader() == PN->getParent()) {
2758 // The loop may have multiple entrances or multiple exits; we can analyze
2759 // this phi as an addrec if it has a unique entry value and a unique
2761 Value *BEValueV = 0, *StartValueV = 0;
2762 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2763 Value *V = PN->getIncomingValue(i);
2764 if (L->contains(PN->getIncomingBlock(i))) {
2767 } else if (BEValueV != V) {
2771 } else if (!StartValueV) {
2773 } else if (StartValueV != V) {
2778 if (BEValueV && StartValueV) {
2779 // While we are analyzing this PHI node, handle its value symbolically.
2780 const SCEV *SymbolicName = getUnknown(PN);
2781 assert(ValueExprMap.find(PN) == ValueExprMap.end() &&
2782 "PHI node already processed?");
2783 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2785 // Using this symbolic name for the PHI, analyze the value coming around
2787 const SCEV *BEValue = getSCEV(BEValueV);
2789 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2790 // has a special value for the first iteration of the loop.
2792 // If the value coming around the backedge is an add with the symbolic
2793 // value we just inserted, then we found a simple induction variable!
2794 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2795 // If there is a single occurrence of the symbolic value, replace it
2796 // with a recurrence.
2797 unsigned FoundIndex = Add->getNumOperands();
2798 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2799 if (Add->getOperand(i) == SymbolicName)
2800 if (FoundIndex == e) {
2805 if (FoundIndex != Add->getNumOperands()) {
2806 // Create an add with everything but the specified operand.
2807 SmallVector<const SCEV *, 8> Ops;
2808 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2809 if (i != FoundIndex)
2810 Ops.push_back(Add->getOperand(i));
2811 const SCEV *Accum = getAddExpr(Ops);
2813 // This is not a valid addrec if the step amount is varying each
2814 // loop iteration, but is not itself an addrec in this loop.
2815 if (Accum->isLoopInvariant(L) ||
2816 (isa<SCEVAddRecExpr>(Accum) &&
2817 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2818 bool HasNUW = false;
2819 bool HasNSW = false;
2821 // If the increment doesn't overflow, then neither the addrec nor
2822 // the post-increment will overflow.
2823 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
2824 if (OBO->hasNoUnsignedWrap())
2826 if (OBO->hasNoSignedWrap())
2830 const SCEV *StartVal = getSCEV(StartValueV);
2831 const SCEV *PHISCEV =
2832 getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
2834 // Since the no-wrap flags are on the increment, they apply to the
2835 // post-incremented value as well.
2836 if (Accum->isLoopInvariant(L))
2837 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
2838 Accum, L, HasNUW, HasNSW);
2840 // Okay, for the entire analysis of this edge we assumed the PHI
2841 // to be symbolic. We now need to go back and purge all of the
2842 // entries for the scalars that use the symbolic expression.
2843 ForgetSymbolicName(PN, SymbolicName);
2844 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2848 } else if (const SCEVAddRecExpr *AddRec =
2849 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2850 // Otherwise, this could be a loop like this:
2851 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2852 // In this case, j = {1,+,1} and BEValue is j.
2853 // Because the other in-value of i (0) fits the evolution of BEValue
2854 // i really is an addrec evolution.
2855 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2856 const SCEV *StartVal = getSCEV(StartValueV);
2858 // If StartVal = j.start - j.stride, we can use StartVal as the
2859 // initial step of the addrec evolution.
2860 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2861 AddRec->getOperand(1))) {
2862 const SCEV *PHISCEV =
2863 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2865 // Okay, for the entire analysis of this edge we assumed the PHI
2866 // to be symbolic. We now need to go back and purge all of the
2867 // entries for the scalars that use the symbolic expression.
2868 ForgetSymbolicName(PN, SymbolicName);
2869 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2877 // If the PHI has a single incoming value, follow that value, unless the
2878 // PHI's incoming blocks are in a different loop, in which case doing so
2879 // risks breaking LCSSA form. Instcombine would normally zap these, but
2880 // it doesn't have DominatorTree information, so it may miss cases.
2881 if (Value *V = SimplifyInstruction(PN, TD, DT)) {
2882 // TODO: The following check is suboptimal. For example, it is pointless
2883 // if V is a constant. Since the problematic case is if V is defined inside
2884 // a deeper loop, it would be better to check for that directly.
2885 bool AllSameLoop = true;
2886 Loop *PNLoop = LI->getLoopFor(PN->getParent());
2887 for (size_t i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2888 if (LI->getLoopFor(PN->getIncomingBlock(i)) != PNLoop) {
2889 AllSameLoop = false;
2896 // If it's not a loop phi, we can't handle it yet.
2897 return getUnknown(PN);
2900 /// createNodeForGEP - Expand GEP instructions into add and multiply
2901 /// operations. This allows them to be analyzed by regular SCEV code.
2903 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
2905 // Don't blindly transfer the inbounds flag from the GEP instruction to the
2906 // Add expression, because the Instruction may be guarded by control flow
2907 // and the no-overflow bits may not be valid for the expression in any
2910 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2911 Value *Base = GEP->getOperand(0);
2912 // Don't attempt to analyze GEPs over unsized objects.
2913 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2914 return getUnknown(GEP);
2915 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
2916 gep_type_iterator GTI = gep_type_begin(GEP);
2917 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
2921 // Compute the (potentially symbolic) offset in bytes for this index.
2922 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2923 // For a struct, add the member offset.
2924 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2925 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
2927 // Add the field offset to the running total offset.
2928 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
2930 // For an array, add the element offset, explicitly scaled.
2931 const SCEV *ElementSize = getSizeOfExpr(*GTI);
2932 const SCEV *IndexS = getSCEV(Index);
2933 // Getelementptr indices are signed.
2934 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
2936 // Multiply the index by the element size to compute the element offset.
2937 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize);
2939 // Add the element offset to the running total offset.
2940 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2944 // Get the SCEV for the GEP base.
2945 const SCEV *BaseS = getSCEV(Base);
2947 // Add the total offset from all the GEP indices to the base.
2948 return getAddExpr(BaseS, TotalOffset);
2951 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2952 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2953 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2954 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2956 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2957 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2958 return C->getValue()->getValue().countTrailingZeros();
2960 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2961 return std::min(GetMinTrailingZeros(T->getOperand()),
2962 (uint32_t)getTypeSizeInBits(T->getType()));
2964 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2965 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2966 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2967 getTypeSizeInBits(E->getType()) : OpRes;
2970 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2971 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2972 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2973 getTypeSizeInBits(E->getType()) : OpRes;
2976 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2977 // The result is the min of all operands results.
2978 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2979 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2980 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2984 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2985 // The result is the sum of all operands results.
2986 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2987 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2988 for (unsigned i = 1, e = M->getNumOperands();
2989 SumOpRes != BitWidth && i != e; ++i)
2990 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2995 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2996 // The result is the min of all operands results.
2997 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2998 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2999 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3003 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
3004 // The result is the min of all operands results.
3005 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3006 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3007 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3011 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
3012 // The result is the min of all operands results.
3013 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3014 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3015 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3019 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3020 // For a SCEVUnknown, ask ValueTracking.
3021 unsigned BitWidth = getTypeSizeInBits(U->getType());
3022 APInt Mask = APInt::getAllOnesValue(BitWidth);
3023 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3024 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
3025 return Zeros.countTrailingOnes();
3032 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
3035 ScalarEvolution::getUnsignedRange(const SCEV *S) {
3036 // See if we've computed this range already.
3037 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
3038 if (I != UnsignedRanges.end())
3041 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3042 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
3044 unsigned BitWidth = getTypeSizeInBits(S->getType());
3045 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3047 // If the value has known zeros, the maximum unsigned value will have those
3048 // known zeros as well.
3049 uint32_t TZ = GetMinTrailingZeros(S);
3051 ConservativeResult =
3052 ConstantRange(APInt::getMinValue(BitWidth),
3053 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
3055 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3056 ConstantRange X = getUnsignedRange(Add->getOperand(0));
3057 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3058 X = X.add(getUnsignedRange(Add->getOperand(i)));
3059 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
3062 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3063 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
3064 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3065 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
3066 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
3069 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3070 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
3071 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3072 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
3073 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
3076 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3077 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3078 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3079 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3080 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
3083 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3084 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3085 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3086 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3089 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3090 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3091 return setUnsignedRange(ZExt,
3092 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3095 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3096 ConstantRange X = getUnsignedRange(SExt->getOperand());
3097 return setUnsignedRange(SExt,
3098 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3101 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3102 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3103 return setUnsignedRange(Trunc,
3104 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3107 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3108 // If there's no unsigned wrap, the value will never be less than its
3110 if (AddRec->hasNoUnsignedWrap())
3111 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3112 if (!C->getValue()->isZero())
3113 ConservativeResult =
3114 ConservativeResult.intersectWith(
3115 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3117 // TODO: non-affine addrec
3118 if (AddRec->isAffine()) {
3119 const Type *Ty = AddRec->getType();
3120 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3121 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3122 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3123 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3125 const SCEV *Start = AddRec->getStart();
3126 const SCEV *Step = AddRec->getStepRecurrence(*this);
3128 ConstantRange StartRange = getUnsignedRange(Start);
3129 ConstantRange StepRange = getSignedRange(Step);
3130 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3131 ConstantRange EndRange =
3132 StartRange.add(MaxBECountRange.multiply(StepRange));
3134 // Check for overflow. This must be done with ConstantRange arithmetic
3135 // because we could be called from within the ScalarEvolution overflow
3137 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3138 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3139 ConstantRange ExtMaxBECountRange =
3140 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3141 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3142 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3144 return setUnsignedRange(AddRec, ConservativeResult);
3146 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3147 EndRange.getUnsignedMin());
3148 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3149 EndRange.getUnsignedMax());
3150 if (Min.isMinValue() && Max.isMaxValue())
3151 return setUnsignedRange(AddRec, ConservativeResult);
3152 return setUnsignedRange(AddRec,
3153 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3157 return setUnsignedRange(AddRec, ConservativeResult);
3160 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3161 // For a SCEVUnknown, ask ValueTracking.
3162 APInt Mask = APInt::getAllOnesValue(BitWidth);
3163 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3164 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
3165 if (Ones == ~Zeros + 1)
3166 return setUnsignedRange(U, ConservativeResult);
3167 return setUnsignedRange(U,
3168 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3171 return setUnsignedRange(S, ConservativeResult);
3174 /// getSignedRange - Determine the signed range for a particular SCEV.
3177 ScalarEvolution::getSignedRange(const SCEV *S) {
3178 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3179 if (I != SignedRanges.end())
3182 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3183 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3185 unsigned BitWidth = getTypeSizeInBits(S->getType());
3186 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3188 // If the value has known zeros, the maximum signed value will have those
3189 // known zeros as well.
3190 uint32_t TZ = GetMinTrailingZeros(S);
3192 ConservativeResult =
3193 ConstantRange(APInt::getSignedMinValue(BitWidth),
3194 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3196 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3197 ConstantRange X = getSignedRange(Add->getOperand(0));
3198 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3199 X = X.add(getSignedRange(Add->getOperand(i)));
3200 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3203 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3204 ConstantRange X = getSignedRange(Mul->getOperand(0));
3205 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3206 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3207 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3210 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3211 ConstantRange X = getSignedRange(SMax->getOperand(0));
3212 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3213 X = X.smax(getSignedRange(SMax->getOperand(i)));
3214 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3217 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3218 ConstantRange X = getSignedRange(UMax->getOperand(0));
3219 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3220 X = X.umax(getSignedRange(UMax->getOperand(i)));
3221 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3224 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3225 ConstantRange X = getSignedRange(UDiv->getLHS());
3226 ConstantRange Y = getSignedRange(UDiv->getRHS());
3227 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3230 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3231 ConstantRange X = getSignedRange(ZExt->getOperand());
3232 return setSignedRange(ZExt,
3233 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3236 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3237 ConstantRange X = getSignedRange(SExt->getOperand());
3238 return setSignedRange(SExt,
3239 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3242 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3243 ConstantRange X = getSignedRange(Trunc->getOperand());
3244 return setSignedRange(Trunc,
3245 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3248 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3249 // If there's no signed wrap, and all the operands have the same sign or
3250 // zero, the value won't ever change sign.
3251 if (AddRec->hasNoSignedWrap()) {
3252 bool AllNonNeg = true;
3253 bool AllNonPos = true;
3254 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3255 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3256 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3259 ConservativeResult = ConservativeResult.intersectWith(
3260 ConstantRange(APInt(BitWidth, 0),
3261 APInt::getSignedMinValue(BitWidth)));
3263 ConservativeResult = ConservativeResult.intersectWith(
3264 ConstantRange(APInt::getSignedMinValue(BitWidth),
3265 APInt(BitWidth, 1)));
3268 // TODO: non-affine addrec
3269 if (AddRec->isAffine()) {
3270 const Type *Ty = AddRec->getType();
3271 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3272 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3273 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3274 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3276 const SCEV *Start = AddRec->getStart();
3277 const SCEV *Step = AddRec->getStepRecurrence(*this);
3279 ConstantRange StartRange = getSignedRange(Start);
3280 ConstantRange StepRange = getSignedRange(Step);
3281 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3282 ConstantRange EndRange =
3283 StartRange.add(MaxBECountRange.multiply(StepRange));
3285 // Check for overflow. This must be done with ConstantRange arithmetic
3286 // because we could be called from within the ScalarEvolution overflow
3288 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3289 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3290 ConstantRange ExtMaxBECountRange =
3291 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3292 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3293 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3295 return setSignedRange(AddRec, ConservativeResult);
3297 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3298 EndRange.getSignedMin());
3299 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3300 EndRange.getSignedMax());
3301 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3302 return setSignedRange(AddRec, ConservativeResult);
3303 return setSignedRange(AddRec,
3304 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3308 return setSignedRange(AddRec, ConservativeResult);
3311 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3312 // For a SCEVUnknown, ask ValueTracking.
3313 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3314 return setSignedRange(U, ConservativeResult);
3315 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3317 return setSignedRange(U, ConservativeResult);
3318 return setSignedRange(U, ConservativeResult.intersectWith(
3319 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3320 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3323 return setSignedRange(S, ConservativeResult);
3326 /// createSCEV - We know that there is no SCEV for the specified value.
3327 /// Analyze the expression.
3329 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3330 if (!isSCEVable(V->getType()))
3331 return getUnknown(V);
3333 unsigned Opcode = Instruction::UserOp1;
3334 if (Instruction *I = dyn_cast<Instruction>(V)) {
3335 Opcode = I->getOpcode();
3337 // Don't attempt to analyze instructions in blocks that aren't
3338 // reachable. Such instructions don't matter, and they aren't required
3339 // to obey basic rules for definitions dominating uses which this
3340 // analysis depends on.
3341 if (!DT->isReachableFromEntry(I->getParent()))
3342 return getUnknown(V);
3343 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3344 Opcode = CE->getOpcode();
3345 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3346 return getConstant(CI);
3347 else if (isa<ConstantPointerNull>(V))
3348 return getConstant(V->getType(), 0);
3349 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3350 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3352 return getUnknown(V);
3354 Operator *U = cast<Operator>(V);
3356 case Instruction::Add: {
3357 // The simple thing to do would be to just call getSCEV on both operands
3358 // and call getAddExpr with the result. However if we're looking at a
3359 // bunch of things all added together, this can be quite inefficient,
3360 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3361 // Instead, gather up all the operands and make a single getAddExpr call.
3362 // LLVM IR canonical form means we need only traverse the left operands.
3363 SmallVector<const SCEV *, 4> AddOps;
3364 AddOps.push_back(getSCEV(U->getOperand(1)));
3365 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3366 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3367 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3369 U = cast<Operator>(Op);
3370 const SCEV *Op1 = getSCEV(U->getOperand(1));
3371 if (Opcode == Instruction::Sub)
3372 AddOps.push_back(getNegativeSCEV(Op1));
3374 AddOps.push_back(Op1);
3376 AddOps.push_back(getSCEV(U->getOperand(0)));
3377 return getAddExpr(AddOps);
3379 case Instruction::Mul: {
3380 // See the Add code above.
3381 SmallVector<const SCEV *, 4> MulOps;
3382 MulOps.push_back(getSCEV(U->getOperand(1)));
3383 for (Value *Op = U->getOperand(0);
3384 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3385 Op = U->getOperand(0)) {
3386 U = cast<Operator>(Op);
3387 MulOps.push_back(getSCEV(U->getOperand(1)));
3389 MulOps.push_back(getSCEV(U->getOperand(0)));
3390 return getMulExpr(MulOps);
3392 case Instruction::UDiv:
3393 return getUDivExpr(getSCEV(U->getOperand(0)),
3394 getSCEV(U->getOperand(1)));
3395 case Instruction::Sub:
3396 return getMinusSCEV(getSCEV(U->getOperand(0)),
3397 getSCEV(U->getOperand(1)));
3398 case Instruction::And:
3399 // For an expression like x&255 that merely masks off the high bits,
3400 // use zext(trunc(x)) as the SCEV expression.
3401 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3402 if (CI->isNullValue())
3403 return getSCEV(U->getOperand(1));
3404 if (CI->isAllOnesValue())
3405 return getSCEV(U->getOperand(0));
3406 const APInt &A = CI->getValue();
3408 // Instcombine's ShrinkDemandedConstant may strip bits out of
3409 // constants, obscuring what would otherwise be a low-bits mask.
3410 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3411 // knew about to reconstruct a low-bits mask value.
3412 unsigned LZ = A.countLeadingZeros();
3413 unsigned BitWidth = A.getBitWidth();
3414 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
3415 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3416 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
3418 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3420 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3422 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3423 IntegerType::get(getContext(), BitWidth - LZ)),
3428 case Instruction::Or:
3429 // If the RHS of the Or is a constant, we may have something like:
3430 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3431 // optimizations will transparently handle this case.
3433 // In order for this transformation to be safe, the LHS must be of the
3434 // form X*(2^n) and the Or constant must be less than 2^n.
3435 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3436 const SCEV *LHS = getSCEV(U->getOperand(0));
3437 const APInt &CIVal = CI->getValue();
3438 if (GetMinTrailingZeros(LHS) >=
3439 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3440 // Build a plain add SCEV.
3441 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3442 // If the LHS of the add was an addrec and it has no-wrap flags,
3443 // transfer the no-wrap flags, since an or won't introduce a wrap.
3444 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3445 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3446 if (OldAR->hasNoUnsignedWrap())
3447 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3448 if (OldAR->hasNoSignedWrap())
3449 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3455 case Instruction::Xor:
3456 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3457 // If the RHS of the xor is a signbit, then this is just an add.
3458 // Instcombine turns add of signbit into xor as a strength reduction step.
3459 if (CI->getValue().isSignBit())
3460 return getAddExpr(getSCEV(U->getOperand(0)),
3461 getSCEV(U->getOperand(1)));
3463 // If the RHS of xor is -1, then this is a not operation.
3464 if (CI->isAllOnesValue())
3465 return getNotSCEV(getSCEV(U->getOperand(0)));
3467 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3468 // This is a variant of the check for xor with -1, and it handles
3469 // the case where instcombine has trimmed non-demanded bits out
3470 // of an xor with -1.
3471 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3472 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3473 if (BO->getOpcode() == Instruction::And &&
3474 LCI->getValue() == CI->getValue())
3475 if (const SCEVZeroExtendExpr *Z =
3476 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3477 const Type *UTy = U->getType();
3478 const SCEV *Z0 = Z->getOperand();
3479 const Type *Z0Ty = Z0->getType();
3480 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3482 // If C is a low-bits mask, the zero extend is serving to
3483 // mask off the high bits. Complement the operand and
3484 // re-apply the zext.
3485 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3486 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3488 // If C is a single bit, it may be in the sign-bit position
3489 // before the zero-extend. In this case, represent the xor
3490 // using an add, which is equivalent, and re-apply the zext.
3491 APInt Trunc = APInt(CI->getValue()).trunc(Z0TySize);
3492 if (APInt(Trunc).zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3494 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3500 case Instruction::Shl:
3501 // Turn shift left of a constant amount into a multiply.
3502 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3503 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3505 // If the shift count is not less than the bitwidth, the result of
3506 // the shift is undefined. Don't try to analyze it, because the
3507 // resolution chosen here may differ from the resolution chosen in
3508 // other parts of the compiler.
3509 if (SA->getValue().uge(BitWidth))
3512 Constant *X = ConstantInt::get(getContext(),
3513 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3514 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3518 case Instruction::LShr:
3519 // Turn logical shift right of a constant into a unsigned divide.
3520 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3521 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3523 // If the shift count is not less than the bitwidth, the result of
3524 // the shift is undefined. Don't try to analyze it, because the
3525 // resolution chosen here may differ from the resolution chosen in
3526 // other parts of the compiler.
3527 if (SA->getValue().uge(BitWidth))
3530 Constant *X = ConstantInt::get(getContext(),
3531 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3532 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3536 case Instruction::AShr:
3537 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3538 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3539 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3540 if (L->getOpcode() == Instruction::Shl &&
3541 L->getOperand(1) == U->getOperand(1)) {
3542 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3544 // If the shift count is not less than the bitwidth, the result of
3545 // the shift is undefined. Don't try to analyze it, because the
3546 // resolution chosen here may differ from the resolution chosen in
3547 // other parts of the compiler.
3548 if (CI->getValue().uge(BitWidth))
3551 uint64_t Amt = BitWidth - CI->getZExtValue();
3552 if (Amt == BitWidth)
3553 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3555 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3556 IntegerType::get(getContext(),
3562 case Instruction::Trunc:
3563 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3565 case Instruction::ZExt:
3566 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3568 case Instruction::SExt:
3569 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3571 case Instruction::BitCast:
3572 // BitCasts are no-op casts so we just eliminate the cast.
3573 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3574 return getSCEV(U->getOperand(0));
3577 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3578 // lead to pointer expressions which cannot safely be expanded to GEPs,
3579 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3580 // simplifying integer expressions.
3582 case Instruction::GetElementPtr:
3583 return createNodeForGEP(cast<GEPOperator>(U));
3585 case Instruction::PHI:
3586 return createNodeForPHI(cast<PHINode>(U));
3588 case Instruction::Select:
3589 // This could be a smax or umax that was lowered earlier.
3590 // Try to recover it.
3591 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3592 Value *LHS = ICI->getOperand(0);
3593 Value *RHS = ICI->getOperand(1);
3594 switch (ICI->getPredicate()) {
3595 case ICmpInst::ICMP_SLT:
3596 case ICmpInst::ICMP_SLE:
3597 std::swap(LHS, RHS);
3599 case ICmpInst::ICMP_SGT:
3600 case ICmpInst::ICMP_SGE:
3601 // a >s b ? a+x : b+x -> smax(a, b)+x
3602 // a >s b ? b+x : a+x -> smin(a, b)+x
3603 if (LHS->getType() == U->getType()) {
3604 const SCEV *LS = getSCEV(LHS);
3605 const SCEV *RS = getSCEV(RHS);
3606 const SCEV *LA = getSCEV(U->getOperand(1));
3607 const SCEV *RA = getSCEV(U->getOperand(2));
3608 const SCEV *LDiff = getMinusSCEV(LA, LS);
3609 const SCEV *RDiff = getMinusSCEV(RA, RS);
3611 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3612 LDiff = getMinusSCEV(LA, RS);
3613 RDiff = getMinusSCEV(RA, LS);
3615 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3618 case ICmpInst::ICMP_ULT:
3619 case ICmpInst::ICMP_ULE:
3620 std::swap(LHS, RHS);
3622 case ICmpInst::ICMP_UGT:
3623 case ICmpInst::ICMP_UGE:
3624 // a >u b ? a+x : b+x -> umax(a, b)+x
3625 // a >u b ? b+x : a+x -> umin(a, b)+x
3626 if (LHS->getType() == U->getType()) {
3627 const SCEV *LS = getSCEV(LHS);
3628 const SCEV *RS = getSCEV(RHS);
3629 const SCEV *LA = getSCEV(U->getOperand(1));
3630 const SCEV *RA = getSCEV(U->getOperand(2));
3631 const SCEV *LDiff = getMinusSCEV(LA, LS);
3632 const SCEV *RDiff = getMinusSCEV(RA, RS);
3634 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3635 LDiff = getMinusSCEV(LA, RS);
3636 RDiff = getMinusSCEV(RA, LS);
3638 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3641 case ICmpInst::ICMP_NE:
3642 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3643 if (LHS->getType() == U->getType() &&
3644 isa<ConstantInt>(RHS) &&
3645 cast<ConstantInt>(RHS)->isZero()) {
3646 const SCEV *One = getConstant(LHS->getType(), 1);
3647 const SCEV *LS = getSCEV(LHS);
3648 const SCEV *LA = getSCEV(U->getOperand(1));
3649 const SCEV *RA = getSCEV(U->getOperand(2));
3650 const SCEV *LDiff = getMinusSCEV(LA, LS);
3651 const SCEV *RDiff = getMinusSCEV(RA, One);
3653 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3656 case ICmpInst::ICMP_EQ:
3657 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3658 if (LHS->getType() == U->getType() &&
3659 isa<ConstantInt>(RHS) &&
3660 cast<ConstantInt>(RHS)->isZero()) {
3661 const SCEV *One = getConstant(LHS->getType(), 1);
3662 const SCEV *LS = getSCEV(LHS);
3663 const SCEV *LA = getSCEV(U->getOperand(1));
3664 const SCEV *RA = getSCEV(U->getOperand(2));
3665 const SCEV *LDiff = getMinusSCEV(LA, One);
3666 const SCEV *RDiff = getMinusSCEV(RA, LS);
3668 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3676 default: // We cannot analyze this expression.
3680 return getUnknown(V);
3685 //===----------------------------------------------------------------------===//
3686 // Iteration Count Computation Code
3689 /// getBackedgeTakenCount - If the specified loop has a predictable
3690 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3691 /// object. The backedge-taken count is the number of times the loop header
3692 /// will be branched to from within the loop. This is one less than the
3693 /// trip count of the loop, since it doesn't count the first iteration,
3694 /// when the header is branched to from outside the loop.
3696 /// Note that it is not valid to call this method on a loop without a
3697 /// loop-invariant backedge-taken count (see
3698 /// hasLoopInvariantBackedgeTakenCount).
3700 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3701 return getBackedgeTakenInfo(L).Exact;
3704 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3705 /// return the least SCEV value that is known never to be less than the
3706 /// actual backedge taken count.
3707 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3708 return getBackedgeTakenInfo(L).Max;
3711 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3712 /// onto the given Worklist.
3714 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3715 BasicBlock *Header = L->getHeader();
3717 // Push all Loop-header PHIs onto the Worklist stack.
3718 for (BasicBlock::iterator I = Header->begin();
3719 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3720 Worklist.push_back(PN);
3723 const ScalarEvolution::BackedgeTakenInfo &
3724 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3725 // Initially insert a CouldNotCompute for this loop. If the insertion
3726 // succeeds, proceed to actually compute a backedge-taken count and
3727 // update the value. The temporary CouldNotCompute value tells SCEV
3728 // code elsewhere that it shouldn't attempt to request a new
3729 // backedge-taken count, which could result in infinite recursion.
3730 std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3731 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3733 BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
3734 if (BECount.Exact != getCouldNotCompute()) {
3735 assert(BECount.Exact->isLoopInvariant(L) &&
3736 BECount.Max->isLoopInvariant(L) &&
3737 "Computed backedge-taken count isn't loop invariant for loop!");
3738 ++NumTripCountsComputed;
3740 // Update the value in the map.
3741 Pair.first->second = BECount;
3743 if (BECount.Max != getCouldNotCompute())
3744 // Update the value in the map.
3745 Pair.first->second = BECount;
3746 if (isa<PHINode>(L->getHeader()->begin()))
3747 // Only count loops that have phi nodes as not being computable.
3748 ++NumTripCountsNotComputed;
3751 // Now that we know more about the trip count for this loop, forget any
3752 // existing SCEV values for PHI nodes in this loop since they are only
3753 // conservative estimates made without the benefit of trip count
3754 // information. This is similar to the code in forgetLoop, except that
3755 // it handles SCEVUnknown PHI nodes specially.
3756 if (BECount.hasAnyInfo()) {
3757 SmallVector<Instruction *, 16> Worklist;
3758 PushLoopPHIs(L, Worklist);
3760 SmallPtrSet<Instruction *, 8> Visited;
3761 while (!Worklist.empty()) {
3762 Instruction *I = Worklist.pop_back_val();
3763 if (!Visited.insert(I)) continue;
3765 ValueExprMapType::iterator It =
3766 ValueExprMap.find(static_cast<Value *>(I));
3767 if (It != ValueExprMap.end()) {
3768 const SCEV *Old = It->second;
3770 // SCEVUnknown for a PHI either means that it has an unrecognized
3771 // structure, or it's a PHI that's in the progress of being computed
3772 // by createNodeForPHI. In the former case, additional loop trip
3773 // count information isn't going to change anything. In the later
3774 // case, createNodeForPHI will perform the necessary updates on its
3775 // own when it gets to that point.
3776 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
3777 ValuesAtScopes.erase(Old);
3778 UnsignedRanges.erase(Old);
3779 SignedRanges.erase(Old);
3780 ValueExprMap.erase(It);
3782 if (PHINode *PN = dyn_cast<PHINode>(I))
3783 ConstantEvolutionLoopExitValue.erase(PN);
3786 PushDefUseChildren(I, Worklist);
3790 return Pair.first->second;
3793 /// forgetLoop - This method should be called by the client when it has
3794 /// changed a loop in a way that may effect ScalarEvolution's ability to
3795 /// compute a trip count, or if the loop is deleted.
3796 void ScalarEvolution::forgetLoop(const Loop *L) {
3797 // Drop any stored trip count value.
3798 BackedgeTakenCounts.erase(L);
3800 // Drop information about expressions based on loop-header PHIs.
3801 SmallVector<Instruction *, 16> Worklist;
3802 PushLoopPHIs(L, Worklist);
3804 SmallPtrSet<Instruction *, 8> Visited;
3805 while (!Worklist.empty()) {
3806 Instruction *I = Worklist.pop_back_val();
3807 if (!Visited.insert(I)) continue;
3809 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3810 if (It != ValueExprMap.end()) {
3811 const SCEV *Old = It->second;
3812 ValuesAtScopes.erase(Old);
3813 UnsignedRanges.erase(Old);
3814 SignedRanges.erase(Old);
3815 ValueExprMap.erase(It);
3816 if (PHINode *PN = dyn_cast<PHINode>(I))
3817 ConstantEvolutionLoopExitValue.erase(PN);
3820 PushDefUseChildren(I, Worklist);
3823 // Forget all contained loops too, to avoid dangling entries in the
3824 // ValuesAtScopes map.
3825 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
3829 /// forgetValue - This method should be called by the client when it has
3830 /// changed a value in a way that may effect its value, or which may
3831 /// disconnect it from a def-use chain linking it to a loop.
3832 void ScalarEvolution::forgetValue(Value *V) {
3833 Instruction *I = dyn_cast<Instruction>(V);
3836 // Drop information about expressions based on loop-header PHIs.
3837 SmallVector<Instruction *, 16> Worklist;
3838 Worklist.push_back(I);
3840 SmallPtrSet<Instruction *, 8> Visited;
3841 while (!Worklist.empty()) {
3842 I = Worklist.pop_back_val();
3843 if (!Visited.insert(I)) continue;
3845 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3846 if (It != ValueExprMap.end()) {
3847 const SCEV *Old = It->second;
3848 ValuesAtScopes.erase(Old);
3849 UnsignedRanges.erase(Old);
3850 SignedRanges.erase(Old);
3851 ValueExprMap.erase(It);
3852 if (PHINode *PN = dyn_cast<PHINode>(I))
3853 ConstantEvolutionLoopExitValue.erase(PN);
3856 PushDefUseChildren(I, Worklist);
3860 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3861 /// of the specified loop will execute.
3862 ScalarEvolution::BackedgeTakenInfo
3863 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3864 SmallVector<BasicBlock *, 8> ExitingBlocks;
3865 L->getExitingBlocks(ExitingBlocks);
3867 // Examine all exits and pick the most conservative values.
3868 const SCEV *BECount = getCouldNotCompute();
3869 const SCEV *MaxBECount = getCouldNotCompute();
3870 bool CouldNotComputeBECount = false;
3871 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3872 BackedgeTakenInfo NewBTI =
3873 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3875 if (NewBTI.Exact == getCouldNotCompute()) {
3876 // We couldn't compute an exact value for this exit, so
3877 // we won't be able to compute an exact value for the loop.
3878 CouldNotComputeBECount = true;
3879 BECount = getCouldNotCompute();
3880 } else if (!CouldNotComputeBECount) {
3881 if (BECount == getCouldNotCompute())
3882 BECount = NewBTI.Exact;
3884 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3886 if (MaxBECount == getCouldNotCompute())
3887 MaxBECount = NewBTI.Max;
3888 else if (NewBTI.Max != getCouldNotCompute())
3889 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3892 return BackedgeTakenInfo(BECount, MaxBECount);
3895 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3896 /// of the specified loop will execute if it exits via the specified block.
3897 ScalarEvolution::BackedgeTakenInfo
3898 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3899 BasicBlock *ExitingBlock) {
3901 // Okay, we've chosen an exiting block. See what condition causes us to
3902 // exit at this block.
3904 // FIXME: we should be able to handle switch instructions (with a single exit)
3905 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3906 if (ExitBr == 0) return getCouldNotCompute();
3907 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3909 // At this point, we know we have a conditional branch that determines whether
3910 // the loop is exited. However, we don't know if the branch is executed each
3911 // time through the loop. If not, then the execution count of the branch will
3912 // not be equal to the trip count of the loop.
3914 // Currently we check for this by checking to see if the Exit branch goes to
3915 // the loop header. If so, we know it will always execute the same number of
3916 // times as the loop. We also handle the case where the exit block *is* the
3917 // loop header. This is common for un-rotated loops.
3919 // If both of those tests fail, walk up the unique predecessor chain to the
3920 // header, stopping if there is an edge that doesn't exit the loop. If the
3921 // header is reached, the execution count of the branch will be equal to the
3922 // trip count of the loop.
3924 // More extensive analysis could be done to handle more cases here.
3926 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3927 ExitBr->getSuccessor(1) != L->getHeader() &&
3928 ExitBr->getParent() != L->getHeader()) {
3929 // The simple checks failed, try climbing the unique predecessor chain
3930 // up to the header.
3932 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3933 BasicBlock *Pred = BB->getUniquePredecessor();
3935 return getCouldNotCompute();
3936 TerminatorInst *PredTerm = Pred->getTerminator();
3937 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3938 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3941 // If the predecessor has a successor that isn't BB and isn't
3942 // outside the loop, assume the worst.
3943 if (L->contains(PredSucc))
3944 return getCouldNotCompute();
3946 if (Pred == L->getHeader()) {
3953 return getCouldNotCompute();
3956 // Proceed to the next level to examine the exit condition expression.
3957 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3958 ExitBr->getSuccessor(0),
3959 ExitBr->getSuccessor(1));
3962 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3963 /// backedge of the specified loop will execute if its exit condition
3964 /// were a conditional branch of ExitCond, TBB, and FBB.
3965 ScalarEvolution::BackedgeTakenInfo
3966 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3970 // Check if the controlling expression for this loop is an And or Or.
3971 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3972 if (BO->getOpcode() == Instruction::And) {
3973 // Recurse on the operands of the and.
3974 BackedgeTakenInfo BTI0 =
3975 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3976 BackedgeTakenInfo BTI1 =
3977 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3978 const SCEV *BECount = getCouldNotCompute();
3979 const SCEV *MaxBECount = getCouldNotCompute();
3980 if (L->contains(TBB)) {
3981 // Both conditions must be true for the loop to continue executing.
3982 // Choose the less conservative count.
3983 if (BTI0.Exact == getCouldNotCompute() ||
3984 BTI1.Exact == getCouldNotCompute())
3985 BECount = getCouldNotCompute();
3987 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3988 if (BTI0.Max == getCouldNotCompute())
3989 MaxBECount = BTI1.Max;
3990 else if (BTI1.Max == getCouldNotCompute())
3991 MaxBECount = BTI0.Max;
3993 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3995 // Both conditions must be true at the same time for the loop to exit.
3996 // For now, be conservative.
3997 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3998 if (BTI0.Max == BTI1.Max)
3999 MaxBECount = BTI0.Max;
4000 if (BTI0.Exact == BTI1.Exact)
4001 BECount = BTI0.Exact;
4004 return BackedgeTakenInfo(BECount, MaxBECount);
4006 if (BO->getOpcode() == Instruction::Or) {
4007 // Recurse on the operands of the or.
4008 BackedgeTakenInfo BTI0 =
4009 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
4010 BackedgeTakenInfo BTI1 =
4011 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
4012 const SCEV *BECount = getCouldNotCompute();
4013 const SCEV *MaxBECount = getCouldNotCompute();
4014 if (L->contains(FBB)) {
4015 // Both conditions must be false for the loop to continue executing.
4016 // Choose the less conservative count.
4017 if (BTI0.Exact == getCouldNotCompute() ||
4018 BTI1.Exact == getCouldNotCompute())
4019 BECount = getCouldNotCompute();
4021 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
4022 if (BTI0.Max == getCouldNotCompute())
4023 MaxBECount = BTI1.Max;
4024 else if (BTI1.Max == getCouldNotCompute())
4025 MaxBECount = BTI0.Max;
4027 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
4029 // Both conditions must be false at the same time for the loop to exit.
4030 // For now, be conservative.
4031 assert(L->contains(TBB) && "Loop block has no successor in loop!");
4032 if (BTI0.Max == BTI1.Max)
4033 MaxBECount = BTI0.Max;
4034 if (BTI0.Exact == BTI1.Exact)
4035 BECount = BTI0.Exact;
4038 return BackedgeTakenInfo(BECount, MaxBECount);
4042 // With an icmp, it may be feasible to compute an exact backedge-taken count.
4043 // Proceed to the next level to examine the icmp.
4044 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
4045 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
4047 // Check for a constant condition. These are normally stripped out by
4048 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
4049 // preserve the CFG and is temporarily leaving constant conditions
4051 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
4052 if (L->contains(FBB) == !CI->getZExtValue())
4053 // The backedge is always taken.
4054 return getCouldNotCompute();
4056 // The backedge is never taken.
4057 return getConstant(CI->getType(), 0);
4060 // If it's not an integer or pointer comparison then compute it the hard way.
4061 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
4064 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
4065 /// backedge of the specified loop will execute if its exit condition
4066 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4067 ScalarEvolution::BackedgeTakenInfo
4068 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
4073 // If the condition was exit on true, convert the condition to exit on false
4074 ICmpInst::Predicate Cond;
4075 if (!L->contains(FBB))
4076 Cond = ExitCond->getPredicate();
4078 Cond = ExitCond->getInversePredicate();
4080 // Handle common loops like: for (X = "string"; *X; ++X)
4081 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4082 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4083 BackedgeTakenInfo ItCnt =
4084 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
4085 if (ItCnt.hasAnyInfo())
4089 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4090 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4092 // Try to evaluate any dependencies out of the loop.
4093 LHS = getSCEVAtScope(LHS, L);
4094 RHS = getSCEVAtScope(RHS, L);
4096 // At this point, we would like to compute how many iterations of the
4097 // loop the predicate will return true for these inputs.
4098 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
4099 // If there is a loop-invariant, force it into the RHS.
4100 std::swap(LHS, RHS);
4101 Cond = ICmpInst::getSwappedPredicate(Cond);
4104 // Simplify the operands before analyzing them.
4105 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4107 // If we have a comparison of a chrec against a constant, try to use value
4108 // ranges to answer this query.
4109 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4110 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4111 if (AddRec->getLoop() == L) {
4112 // Form the constant range.
4113 ConstantRange CompRange(
4114 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4116 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4117 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4121 case ICmpInst::ICMP_NE: { // while (X != Y)
4122 // Convert to: while (X-Y != 0)
4123 BackedgeTakenInfo BTI = HowFarToZero(getMinusSCEV(LHS, RHS), L);
4124 if (BTI.hasAnyInfo()) return BTI;
4127 case ICmpInst::ICMP_EQ: { // while (X == Y)
4128 // Convert to: while (X-Y == 0)
4129 BackedgeTakenInfo BTI = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4130 if (BTI.hasAnyInfo()) return BTI;
4133 case ICmpInst::ICMP_SLT: {
4134 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
4135 if (BTI.hasAnyInfo()) return BTI;
4138 case ICmpInst::ICMP_SGT: {
4139 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4140 getNotSCEV(RHS), L, true);
4141 if (BTI.hasAnyInfo()) return BTI;
4144 case ICmpInst::ICMP_ULT: {
4145 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
4146 if (BTI.hasAnyInfo()) return BTI;
4149 case ICmpInst::ICMP_UGT: {
4150 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4151 getNotSCEV(RHS), L, false);
4152 if (BTI.hasAnyInfo()) return BTI;
4157 dbgs() << "ComputeBackedgeTakenCount ";
4158 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4159 dbgs() << "[unsigned] ";
4160 dbgs() << *LHS << " "
4161 << Instruction::getOpcodeName(Instruction::ICmp)
4162 << " " << *RHS << "\n";
4167 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
4170 static ConstantInt *
4171 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4172 ScalarEvolution &SE) {
4173 const SCEV *InVal = SE.getConstant(C);
4174 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4175 assert(isa<SCEVConstant>(Val) &&
4176 "Evaluation of SCEV at constant didn't fold correctly?");
4177 return cast<SCEVConstant>(Val)->getValue();
4180 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
4181 /// and a GEP expression (missing the pointer index) indexing into it, return
4182 /// the addressed element of the initializer or null if the index expression is
4185 GetAddressedElementFromGlobal(GlobalVariable *GV,
4186 const std::vector<ConstantInt*> &Indices) {
4187 Constant *Init = GV->getInitializer();
4188 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
4189 uint64_t Idx = Indices[i]->getZExtValue();
4190 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
4191 assert(Idx < CS->getNumOperands() && "Bad struct index!");
4192 Init = cast<Constant>(CS->getOperand(Idx));
4193 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
4194 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
4195 Init = cast<Constant>(CA->getOperand(Idx));
4196 } else if (isa<ConstantAggregateZero>(Init)) {
4197 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
4198 assert(Idx < STy->getNumElements() && "Bad struct index!");
4199 Init = Constant::getNullValue(STy->getElementType(Idx));
4200 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
4201 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
4202 Init = Constant::getNullValue(ATy->getElementType());
4204 llvm_unreachable("Unknown constant aggregate type!");
4208 return 0; // Unknown initializer type
4214 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
4215 /// 'icmp op load X, cst', try to see if we can compute the backedge
4216 /// execution count.
4217 ScalarEvolution::BackedgeTakenInfo
4218 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
4222 ICmpInst::Predicate predicate) {
4223 if (LI->isVolatile()) return getCouldNotCompute();
4225 // Check to see if the loaded pointer is a getelementptr of a global.
4226 // TODO: Use SCEV instead of manually grubbing with GEPs.
4227 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4228 if (!GEP) return getCouldNotCompute();
4230 // Make sure that it is really a constant global we are gepping, with an
4231 // initializer, and make sure the first IDX is really 0.
4232 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4233 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4234 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4235 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4236 return getCouldNotCompute();
4238 // Okay, we allow one non-constant index into the GEP instruction.
4240 std::vector<ConstantInt*> Indexes;
4241 unsigned VarIdxNum = 0;
4242 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4243 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4244 Indexes.push_back(CI);
4245 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4246 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4247 VarIdx = GEP->getOperand(i);
4249 Indexes.push_back(0);
4252 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4253 // Check to see if X is a loop variant variable value now.
4254 const SCEV *Idx = getSCEV(VarIdx);
4255 Idx = getSCEVAtScope(Idx, L);
4257 // We can only recognize very limited forms of loop index expressions, in
4258 // particular, only affine AddRec's like {C1,+,C2}.
4259 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4260 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
4261 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4262 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4263 return getCouldNotCompute();
4265 unsigned MaxSteps = MaxBruteForceIterations;
4266 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4267 ConstantInt *ItCst = ConstantInt::get(
4268 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4269 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4271 // Form the GEP offset.
4272 Indexes[VarIdxNum] = Val;
4274 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
4275 if (Result == 0) break; // Cannot compute!
4277 // Evaluate the condition for this iteration.
4278 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4279 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4280 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4282 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4283 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4286 ++NumArrayLenItCounts;
4287 return getConstant(ItCst); // Found terminating iteration!
4290 return getCouldNotCompute();
4294 /// CanConstantFold - Return true if we can constant fold an instruction of the
4295 /// specified type, assuming that all operands were constants.
4296 static bool CanConstantFold(const Instruction *I) {
4297 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4298 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
4301 if (const CallInst *CI = dyn_cast<CallInst>(I))
4302 if (const Function *F = CI->getCalledFunction())
4303 return canConstantFoldCallTo(F);
4307 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4308 /// in the loop that V is derived from. We allow arbitrary operations along the
4309 /// way, but the operands of an operation must either be constants or a value
4310 /// derived from a constant PHI. If this expression does not fit with these
4311 /// constraints, return null.
4312 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4313 // If this is not an instruction, or if this is an instruction outside of the
4314 // loop, it can't be derived from a loop PHI.
4315 Instruction *I = dyn_cast<Instruction>(V);
4316 if (I == 0 || !L->contains(I)) return 0;
4318 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4319 if (L->getHeader() == I->getParent())
4322 // We don't currently keep track of the control flow needed to evaluate
4323 // PHIs, so we cannot handle PHIs inside of loops.
4327 // If we won't be able to constant fold this expression even if the operands
4328 // are constants, return early.
4329 if (!CanConstantFold(I)) return 0;
4331 // Otherwise, we can evaluate this instruction if all of its operands are
4332 // constant or derived from a PHI node themselves.
4334 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
4335 if (!isa<Constant>(I->getOperand(Op))) {
4336 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
4337 if (P == 0) return 0; // Not evolving from PHI
4341 return 0; // Evolving from multiple different PHIs.
4344 // This is a expression evolving from a constant PHI!
4348 /// EvaluateExpression - Given an expression that passes the
4349 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4350 /// in the loop has the value PHIVal. If we can't fold this expression for some
4351 /// reason, return null.
4352 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
4353 const TargetData *TD) {
4354 if (isa<PHINode>(V)) return PHIVal;
4355 if (Constant *C = dyn_cast<Constant>(V)) return C;
4356 Instruction *I = cast<Instruction>(V);
4358 std::vector<Constant*> Operands(I->getNumOperands());
4360 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4361 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
4362 if (Operands[i] == 0) return 0;
4365 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4366 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4368 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4369 &Operands[0], Operands.size(), TD);
4372 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4373 /// in the header of its containing loop, we know the loop executes a
4374 /// constant number of times, and the PHI node is just a recurrence
4375 /// involving constants, fold it.
4377 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4380 std::map<PHINode*, Constant*>::const_iterator I =
4381 ConstantEvolutionLoopExitValue.find(PN);
4382 if (I != ConstantEvolutionLoopExitValue.end())
4385 if (BEs.ugt(MaxBruteForceIterations))
4386 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4388 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4390 // Since the loop is canonicalized, the PHI node must have two entries. One
4391 // entry must be a constant (coming in from outside of the loop), and the
4392 // second must be derived from the same PHI.
4393 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4394 Constant *StartCST =
4395 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4397 return RetVal = 0; // Must be a constant.
4399 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4400 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4401 !isa<Constant>(BEValue))
4402 return RetVal = 0; // Not derived from same PHI.
4404 // Execute the loop symbolically to determine the exit value.
4405 if (BEs.getActiveBits() >= 32)
4406 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4408 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4409 unsigned IterationNum = 0;
4410 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
4411 if (IterationNum == NumIterations)
4412 return RetVal = PHIVal; // Got exit value!
4414 // Compute the value of the PHI node for the next iteration.
4415 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4416 if (NextPHI == PHIVal)
4417 return RetVal = NextPHI; // Stopped evolving!
4419 return 0; // Couldn't evaluate!
4424 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
4425 /// constant number of times (the condition evolves only from constants),
4426 /// try to evaluate a few iterations of the loop until we get the exit
4427 /// condition gets a value of ExitWhen (true or false). If we cannot
4428 /// evaluate the trip count of the loop, return getCouldNotCompute().
4430 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
4433 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4434 if (PN == 0) return getCouldNotCompute();
4436 // If the loop is canonicalized, the PHI will have exactly two entries.
4437 // That's the only form we support here.
4438 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4440 // One entry must be a constant (coming in from outside of the loop), and the
4441 // second must be derived from the same PHI.
4442 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4443 Constant *StartCST =
4444 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4445 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
4447 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4448 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4449 !isa<Constant>(BEValue))
4450 return getCouldNotCompute(); // Not derived from same PHI.
4452 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4453 // the loop symbolically to determine when the condition gets a value of
4455 unsigned IterationNum = 0;
4456 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4457 for (Constant *PHIVal = StartCST;
4458 IterationNum != MaxIterations; ++IterationNum) {
4459 ConstantInt *CondVal =
4460 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
4462 // Couldn't symbolically evaluate.
4463 if (!CondVal) return getCouldNotCompute();
4465 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4466 ++NumBruteForceTripCountsComputed;
4467 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4470 // Compute the value of the PHI node for the next iteration.
4471 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4472 if (NextPHI == 0 || NextPHI == PHIVal)
4473 return getCouldNotCompute();// Couldn't evaluate or not making progress...
4477 // Too many iterations were needed to evaluate.
4478 return getCouldNotCompute();
4481 /// getSCEVAtScope - Return a SCEV expression for the specified value
4482 /// at the specified scope in the program. The L value specifies a loop
4483 /// nest to evaluate the expression at, where null is the top-level or a
4484 /// specified loop is immediately inside of the loop.
4486 /// This method can be used to compute the exit value for a variable defined
4487 /// in a loop by querying what the value will hold in the parent loop.
4489 /// In the case that a relevant loop exit value cannot be computed, the
4490 /// original value V is returned.
4491 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4492 // Check to see if we've folded this expression at this loop before.
4493 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4494 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4495 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4497 return Pair.first->second ? Pair.first->second : V;
4499 // Otherwise compute it.
4500 const SCEV *C = computeSCEVAtScope(V, L);
4501 ValuesAtScopes[V][L] = C;
4505 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
4506 if (isa<SCEVConstant>(V)) return V;
4508 // If this instruction is evolved from a constant-evolving PHI, compute the
4509 // exit value from the loop without using SCEVs.
4510 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
4511 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
4512 const Loop *LI = (*this->LI)[I->getParent()];
4513 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
4514 if (PHINode *PN = dyn_cast<PHINode>(I))
4515 if (PN->getParent() == LI->getHeader()) {
4516 // Okay, there is no closed form solution for the PHI node. Check
4517 // to see if the loop that contains it has a known backedge-taken
4518 // count. If so, we may be able to force computation of the exit
4520 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
4521 if (const SCEVConstant *BTCC =
4522 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
4523 // Okay, we know how many times the containing loop executes. If
4524 // this is a constant evolving PHI node, get the final value at
4525 // the specified iteration number.
4526 Constant *RV = getConstantEvolutionLoopExitValue(PN,
4527 BTCC->getValue()->getValue(),
4529 if (RV) return getSCEV(RV);
4533 // Okay, this is an expression that we cannot symbolically evaluate
4534 // into a SCEV. Check to see if it's possible to symbolically evaluate
4535 // the arguments into constants, and if so, try to constant propagate the
4536 // result. This is particularly useful for computing loop exit values.
4537 if (CanConstantFold(I)) {
4538 SmallVector<Constant *, 4> Operands;
4539 bool MadeImprovement = false;
4540 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4541 Value *Op = I->getOperand(i);
4542 if (Constant *C = dyn_cast<Constant>(Op)) {
4543 Operands.push_back(C);
4547 // If any of the operands is non-constant and if they are
4548 // non-integer and non-pointer, don't even try to analyze them
4549 // with scev techniques.
4550 if (!isSCEVable(Op->getType()))
4553 const SCEV *OrigV = getSCEV(Op);
4554 const SCEV *OpV = getSCEVAtScope(OrigV, L);
4555 MadeImprovement |= OrigV != OpV;
4558 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
4560 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV))
4561 C = dyn_cast<Constant>(SU->getValue());
4563 if (C->getType() != Op->getType())
4564 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4568 Operands.push_back(C);
4571 // Check to see if getSCEVAtScope actually made an improvement.
4572 if (MadeImprovement) {
4574 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4575 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4576 Operands[0], Operands[1], TD);
4578 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4579 &Operands[0], Operands.size(), TD);
4586 // This is some other type of SCEVUnknown, just return it.
4590 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4591 // Avoid performing the look-up in the common case where the specified
4592 // expression has no loop-variant portions.
4593 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4594 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4595 if (OpAtScope != Comm->getOperand(i)) {
4596 // Okay, at least one of these operands is loop variant but might be
4597 // foldable. Build a new instance of the folded commutative expression.
4598 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4599 Comm->op_begin()+i);
4600 NewOps.push_back(OpAtScope);
4602 for (++i; i != e; ++i) {
4603 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4604 NewOps.push_back(OpAtScope);
4606 if (isa<SCEVAddExpr>(Comm))
4607 return getAddExpr(NewOps);
4608 if (isa<SCEVMulExpr>(Comm))
4609 return getMulExpr(NewOps);
4610 if (isa<SCEVSMaxExpr>(Comm))
4611 return getSMaxExpr(NewOps);
4612 if (isa<SCEVUMaxExpr>(Comm))
4613 return getUMaxExpr(NewOps);
4614 llvm_unreachable("Unknown commutative SCEV type!");
4617 // If we got here, all operands are loop invariant.
4621 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4622 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4623 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4624 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4625 return Div; // must be loop invariant
4626 return getUDivExpr(LHS, RHS);
4629 // If this is a loop recurrence for a loop that does not contain L, then we
4630 // are dealing with the final value computed by the loop.
4631 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4632 // First, attempt to evaluate each operand.
4633 // Avoid performing the look-up in the common case where the specified
4634 // expression has no loop-variant portions.
4635 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
4636 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
4637 if (OpAtScope == AddRec->getOperand(i))
4640 // Okay, at least one of these operands is loop variant but might be
4641 // foldable. Build a new instance of the folded commutative expression.
4642 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
4643 AddRec->op_begin()+i);
4644 NewOps.push_back(OpAtScope);
4645 for (++i; i != e; ++i)
4646 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
4648 AddRec = cast<SCEVAddRecExpr>(getAddRecExpr(NewOps, AddRec->getLoop()));
4652 // If the scope is outside the addrec's loop, evaluate it by using the
4653 // loop exit value of the addrec.
4654 if (!AddRec->getLoop()->contains(L)) {
4655 // To evaluate this recurrence, we need to know how many times the AddRec
4656 // loop iterates. Compute this now.
4657 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4658 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4660 // Then, evaluate the AddRec.
4661 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4667 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4668 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4669 if (Op == Cast->getOperand())
4670 return Cast; // must be loop invariant
4671 return getZeroExtendExpr(Op, Cast->getType());
4674 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4675 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4676 if (Op == Cast->getOperand())
4677 return Cast; // must be loop invariant
4678 return getSignExtendExpr(Op, Cast->getType());
4681 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4682 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4683 if (Op == Cast->getOperand())
4684 return Cast; // must be loop invariant
4685 return getTruncateExpr(Op, Cast->getType());
4688 llvm_unreachable("Unknown SCEV type!");
4692 /// getSCEVAtScope - This is a convenience function which does
4693 /// getSCEVAtScope(getSCEV(V), L).
4694 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4695 return getSCEVAtScope(getSCEV(V), L);
4698 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4699 /// following equation:
4701 /// A * X = B (mod N)
4703 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4704 /// A and B isn't important.
4706 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4707 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4708 ScalarEvolution &SE) {
4709 uint32_t BW = A.getBitWidth();
4710 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4711 assert(A != 0 && "A must be non-zero.");
4715 // The gcd of A and N may have only one prime factor: 2. The number of
4716 // trailing zeros in A is its multiplicity
4717 uint32_t Mult2 = A.countTrailingZeros();
4720 // 2. Check if B is divisible by D.
4722 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4723 // is not less than multiplicity of this prime factor for D.
4724 if (B.countTrailingZeros() < Mult2)
4725 return SE.getCouldNotCompute();
4727 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4730 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4731 // bit width during computations.
4732 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4733 APInt Mod(BW + 1, 0);
4734 Mod.set(BW - Mult2); // Mod = N / D
4735 APInt I = AD.multiplicativeInverse(Mod);
4737 // 4. Compute the minimum unsigned root of the equation:
4738 // I * (B / D) mod (N / D)
4739 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4741 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4743 return SE.getConstant(Result.trunc(BW));
4746 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4747 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4748 /// might be the same) or two SCEVCouldNotCompute objects.
4750 static std::pair<const SCEV *,const SCEV *>
4751 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4752 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4753 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4754 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4755 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4757 // We currently can only solve this if the coefficients are constants.
4758 if (!LC || !MC || !NC) {
4759 const SCEV *CNC = SE.getCouldNotCompute();
4760 return std::make_pair(CNC, CNC);
4763 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4764 const APInt &L = LC->getValue()->getValue();
4765 const APInt &M = MC->getValue()->getValue();
4766 const APInt &N = NC->getValue()->getValue();
4767 APInt Two(BitWidth, 2);
4768 APInt Four(BitWidth, 4);
4771 using namespace APIntOps;
4773 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4774 // The B coefficient is M-N/2
4778 // The A coefficient is N/2
4779 APInt A(N.sdiv(Two));
4781 // Compute the B^2-4ac term.
4784 SqrtTerm -= Four * (A * C);
4786 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4787 // integer value or else APInt::sqrt() will assert.
4788 APInt SqrtVal(SqrtTerm.sqrt());
4790 // Compute the two solutions for the quadratic formula.
4791 // The divisions must be performed as signed divisions.
4793 APInt TwoA( A << 1 );
4794 if (TwoA.isMinValue()) {
4795 const SCEV *CNC = SE.getCouldNotCompute();
4796 return std::make_pair(CNC, CNC);
4799 LLVMContext &Context = SE.getContext();
4801 ConstantInt *Solution1 =
4802 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4803 ConstantInt *Solution2 =
4804 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4806 return std::make_pair(SE.getConstant(Solution1),
4807 SE.getConstant(Solution2));
4808 } // end APIntOps namespace
4811 /// HowFarToZero - Return the number of times a backedge comparing the specified
4812 /// value to zero will execute. If not computable, return CouldNotCompute.
4813 ScalarEvolution::BackedgeTakenInfo
4814 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4815 // If the value is a constant
4816 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4817 // If the value is already zero, the branch will execute zero times.
4818 if (C->getValue()->isZero()) return C;
4819 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4822 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4823 if (!AddRec || AddRec->getLoop() != L)
4824 return getCouldNotCompute();
4826 if (AddRec->isAffine()) {
4827 // If this is an affine expression, the execution count of this branch is
4828 // the minimum unsigned root of the following equation:
4830 // Start + Step*N = 0 (mod 2^BW)
4834 // Step*N = -Start (mod 2^BW)
4836 // where BW is the common bit width of Start and Step.
4838 // Get the initial value for the loop.
4839 const SCEV *Start = getSCEVAtScope(AddRec->getStart(),
4840 L->getParentLoop());
4841 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1),
4842 L->getParentLoop());
4844 if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
4845 // For now we handle only constant steps.
4847 // First, handle unitary steps.
4848 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4849 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4850 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4851 return Start; // N = Start (as unsigned)
4853 // Then, try to solve the above equation provided that Start is constant.
4854 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4855 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4856 -StartC->getValue()->getValue(),
4859 } else if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
4860 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4861 // the quadratic equation to solve it.
4862 std::pair<const SCEV *,const SCEV *> Roots = SolveQuadraticEquation(AddRec,
4864 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4865 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4868 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
4869 << " sol#2: " << *R2 << "\n";
4871 // Pick the smallest positive root value.
4872 if (ConstantInt *CB =
4873 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
4874 R1->getValue(), R2->getValue()))) {
4875 if (CB->getZExtValue() == false)
4876 std::swap(R1, R2); // R1 is the minimum root now.
4878 // We can only use this value if the chrec ends up with an exact zero
4879 // value at this index. When solving for "X*X != 5", for example, we
4880 // should not accept a root of 2.
4881 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4883 return R1; // We found a quadratic root!
4888 return getCouldNotCompute();
4891 /// HowFarToNonZero - Return the number of times a backedge checking the
4892 /// specified value for nonzero will execute. If not computable, return
4894 ScalarEvolution::BackedgeTakenInfo
4895 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4896 // Loops that look like: while (X == 0) are very strange indeed. We don't
4897 // handle them yet except for the trivial case. This could be expanded in the
4898 // future as needed.
4900 // If the value is a constant, check to see if it is known to be non-zero
4901 // already. If so, the backedge will execute zero times.
4902 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4903 if (!C->getValue()->isNullValue())
4904 return getConstant(C->getType(), 0);
4905 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4908 // We could implement others, but I really doubt anyone writes loops like
4909 // this, and if they did, they would already be constant folded.
4910 return getCouldNotCompute();
4913 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4914 /// (which may not be an immediate predecessor) which has exactly one
4915 /// successor from which BB is reachable, or null if no such block is
4918 std::pair<BasicBlock *, BasicBlock *>
4919 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4920 // If the block has a unique predecessor, then there is no path from the
4921 // predecessor to the block that does not go through the direct edge
4922 // from the predecessor to the block.
4923 if (BasicBlock *Pred = BB->getSinglePredecessor())
4924 return std::make_pair(Pred, BB);
4926 // A loop's header is defined to be a block that dominates the loop.
4927 // If the header has a unique predecessor outside the loop, it must be
4928 // a block that has exactly one successor that can reach the loop.
4929 if (Loop *L = LI->getLoopFor(BB))
4930 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
4932 return std::pair<BasicBlock *, BasicBlock *>();
4935 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4936 /// testing whether two expressions are equal, however for the purposes of
4937 /// looking for a condition guarding a loop, it can be useful to be a little
4938 /// more general, since a front-end may have replicated the controlling
4941 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4942 // Quick check to see if they are the same SCEV.
4943 if (A == B) return true;
4945 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4946 // two different instructions with the same value. Check for this case.
4947 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4948 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4949 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4950 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4951 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4954 // Otherwise assume they may have a different value.
4958 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
4959 /// predicate Pred. Return true iff any changes were made.
4961 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
4962 const SCEV *&LHS, const SCEV *&RHS) {
4963 bool Changed = false;
4965 // Canonicalize a constant to the right side.
4966 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
4967 // Check for both operands constant.
4968 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
4969 if (ConstantExpr::getICmp(Pred,
4971 RHSC->getValue())->isNullValue())
4972 goto trivially_false;
4974 goto trivially_true;
4976 // Otherwise swap the operands to put the constant on the right.
4977 std::swap(LHS, RHS);
4978 Pred = ICmpInst::getSwappedPredicate(Pred);
4982 // If we're comparing an addrec with a value which is loop-invariant in the
4983 // addrec's loop, put the addrec on the left. Also make a dominance check,
4984 // as both operands could be addrecs loop-invariant in each other's loop.
4985 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
4986 const Loop *L = AR->getLoop();
4987 if (LHS->isLoopInvariant(L) && LHS->properlyDominates(L->getHeader(), DT)) {
4988 std::swap(LHS, RHS);
4989 Pred = ICmpInst::getSwappedPredicate(Pred);
4994 // If there's a constant operand, canonicalize comparisons with boundary
4995 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
4996 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
4997 const APInt &RA = RC->getValue()->getValue();
4999 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5000 case ICmpInst::ICMP_EQ:
5001 case ICmpInst::ICMP_NE:
5003 case ICmpInst::ICMP_UGE:
5004 if ((RA - 1).isMinValue()) {
5005 Pred = ICmpInst::ICMP_NE;
5006 RHS = getConstant(RA - 1);
5010 if (RA.isMaxValue()) {
5011 Pred = ICmpInst::ICMP_EQ;
5015 if (RA.isMinValue()) goto trivially_true;
5017 Pred = ICmpInst::ICMP_UGT;
5018 RHS = getConstant(RA - 1);
5021 case ICmpInst::ICMP_ULE:
5022 if ((RA + 1).isMaxValue()) {
5023 Pred = ICmpInst::ICMP_NE;
5024 RHS = getConstant(RA + 1);
5028 if (RA.isMinValue()) {
5029 Pred = ICmpInst::ICMP_EQ;
5033 if (RA.isMaxValue()) goto trivially_true;
5035 Pred = ICmpInst::ICMP_ULT;
5036 RHS = getConstant(RA + 1);
5039 case ICmpInst::ICMP_SGE:
5040 if ((RA - 1).isMinSignedValue()) {
5041 Pred = ICmpInst::ICMP_NE;
5042 RHS = getConstant(RA - 1);
5046 if (RA.isMaxSignedValue()) {
5047 Pred = ICmpInst::ICMP_EQ;
5051 if (RA.isMinSignedValue()) goto trivially_true;
5053 Pred = ICmpInst::ICMP_SGT;
5054 RHS = getConstant(RA - 1);
5057 case ICmpInst::ICMP_SLE:
5058 if ((RA + 1).isMaxSignedValue()) {
5059 Pred = ICmpInst::ICMP_NE;
5060 RHS = getConstant(RA + 1);
5064 if (RA.isMinSignedValue()) {
5065 Pred = ICmpInst::ICMP_EQ;
5069 if (RA.isMaxSignedValue()) goto trivially_true;
5071 Pred = ICmpInst::ICMP_SLT;
5072 RHS = getConstant(RA + 1);
5075 case ICmpInst::ICMP_UGT:
5076 if (RA.isMinValue()) {
5077 Pred = ICmpInst::ICMP_NE;
5081 if ((RA + 1).isMaxValue()) {
5082 Pred = ICmpInst::ICMP_EQ;
5083 RHS = getConstant(RA + 1);
5087 if (RA.isMaxValue()) goto trivially_false;
5089 case ICmpInst::ICMP_ULT:
5090 if (RA.isMaxValue()) {
5091 Pred = ICmpInst::ICMP_NE;
5095 if ((RA - 1).isMinValue()) {
5096 Pred = ICmpInst::ICMP_EQ;
5097 RHS = getConstant(RA - 1);
5101 if (RA.isMinValue()) goto trivially_false;
5103 case ICmpInst::ICMP_SGT:
5104 if (RA.isMinSignedValue()) {
5105 Pred = ICmpInst::ICMP_NE;
5109 if ((RA + 1).isMaxSignedValue()) {
5110 Pred = ICmpInst::ICMP_EQ;
5111 RHS = getConstant(RA + 1);
5115 if (RA.isMaxSignedValue()) goto trivially_false;
5117 case ICmpInst::ICMP_SLT:
5118 if (RA.isMaxSignedValue()) {
5119 Pred = ICmpInst::ICMP_NE;
5123 if ((RA - 1).isMinSignedValue()) {
5124 Pred = ICmpInst::ICMP_EQ;
5125 RHS = getConstant(RA - 1);
5129 if (RA.isMinSignedValue()) goto trivially_false;
5134 // Check for obvious equality.
5135 if (HasSameValue(LHS, RHS)) {
5136 if (ICmpInst::isTrueWhenEqual(Pred))
5137 goto trivially_true;
5138 if (ICmpInst::isFalseWhenEqual(Pred))
5139 goto trivially_false;
5142 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5143 // adding or subtracting 1 from one of the operands.
5145 case ICmpInst::ICMP_SLE:
5146 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5147 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5148 /*HasNUW=*/false, /*HasNSW=*/true);
5149 Pred = ICmpInst::ICMP_SLT;
5151 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5152 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5153 /*HasNUW=*/false, /*HasNSW=*/true);
5154 Pred = ICmpInst::ICMP_SLT;
5158 case ICmpInst::ICMP_SGE:
5159 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5160 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5161 /*HasNUW=*/false, /*HasNSW=*/true);
5162 Pred = ICmpInst::ICMP_SGT;
5164 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5165 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5166 /*HasNUW=*/false, /*HasNSW=*/true);
5167 Pred = ICmpInst::ICMP_SGT;
5171 case ICmpInst::ICMP_ULE:
5172 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5173 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5174 /*HasNUW=*/true, /*HasNSW=*/false);
5175 Pred = ICmpInst::ICMP_ULT;
5177 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5178 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5179 /*HasNUW=*/true, /*HasNSW=*/false);
5180 Pred = ICmpInst::ICMP_ULT;
5184 case ICmpInst::ICMP_UGE:
5185 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5186 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5187 /*HasNUW=*/true, /*HasNSW=*/false);
5188 Pred = ICmpInst::ICMP_UGT;
5190 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5191 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5192 /*HasNUW=*/true, /*HasNSW=*/false);
5193 Pred = ICmpInst::ICMP_UGT;
5201 // TODO: More simplifications are possible here.
5207 LHS = RHS = getConstant(Type::getInt1Ty(getContext()), 0);
5208 Pred = ICmpInst::ICMP_EQ;
5213 LHS = RHS = getConstant(Type::getInt1Ty(getContext()), 0);
5214 Pred = ICmpInst::ICMP_NE;
5218 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5219 return getSignedRange(S).getSignedMax().isNegative();
5222 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5223 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5226 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5227 return !getSignedRange(S).getSignedMin().isNegative();
5230 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5231 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5234 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5235 return isKnownNegative(S) || isKnownPositive(S);
5238 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5239 const SCEV *LHS, const SCEV *RHS) {
5240 // Canonicalize the inputs first.
5241 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5243 // If LHS or RHS is an addrec, check to see if the condition is true in
5244 // every iteration of the loop.
5245 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5246 if (isLoopEntryGuardedByCond(
5247 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5248 isLoopBackedgeGuardedByCond(
5249 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5251 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5252 if (isLoopEntryGuardedByCond(
5253 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5254 isLoopBackedgeGuardedByCond(
5255 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5258 // Otherwise see what can be done with known constant ranges.
5259 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5263 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5264 const SCEV *LHS, const SCEV *RHS) {
5265 if (HasSameValue(LHS, RHS))
5266 return ICmpInst::isTrueWhenEqual(Pred);
5268 // This code is split out from isKnownPredicate because it is called from
5269 // within isLoopEntryGuardedByCond.
5272 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5274 case ICmpInst::ICMP_SGT:
5275 Pred = ICmpInst::ICMP_SLT;
5276 std::swap(LHS, RHS);
5277 case ICmpInst::ICMP_SLT: {
5278 ConstantRange LHSRange = getSignedRange(LHS);
5279 ConstantRange RHSRange = getSignedRange(RHS);
5280 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5282 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5286 case ICmpInst::ICMP_SGE:
5287 Pred = ICmpInst::ICMP_SLE;
5288 std::swap(LHS, RHS);
5289 case ICmpInst::ICMP_SLE: {
5290 ConstantRange LHSRange = getSignedRange(LHS);
5291 ConstantRange RHSRange = getSignedRange(RHS);
5292 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5294 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
5298 case ICmpInst::ICMP_UGT:
5299 Pred = ICmpInst::ICMP_ULT;
5300 std::swap(LHS, RHS);
5301 case ICmpInst::ICMP_ULT: {
5302 ConstantRange LHSRange = getUnsignedRange(LHS);
5303 ConstantRange RHSRange = getUnsignedRange(RHS);
5304 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
5306 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
5310 case ICmpInst::ICMP_UGE:
5311 Pred = ICmpInst::ICMP_ULE;
5312 std::swap(LHS, RHS);
5313 case ICmpInst::ICMP_ULE: {
5314 ConstantRange LHSRange = getUnsignedRange(LHS);
5315 ConstantRange RHSRange = getUnsignedRange(RHS);
5316 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
5318 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
5322 case ICmpInst::ICMP_NE: {
5323 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
5325 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
5328 const SCEV *Diff = getMinusSCEV(LHS, RHS);
5329 if (isKnownNonZero(Diff))
5333 case ICmpInst::ICMP_EQ:
5334 // The check at the top of the function catches the case where
5335 // the values are known to be equal.
5341 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
5342 /// protected by a conditional between LHS and RHS. This is used to
5343 /// to eliminate casts.
5345 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
5346 ICmpInst::Predicate Pred,
5347 const SCEV *LHS, const SCEV *RHS) {
5348 // Interpret a null as meaning no loop, where there is obviously no guard
5349 // (interprocedural conditions notwithstanding).
5350 if (!L) return true;
5352 BasicBlock *Latch = L->getLoopLatch();
5356 BranchInst *LoopContinuePredicate =
5357 dyn_cast<BranchInst>(Latch->getTerminator());
5358 if (!LoopContinuePredicate ||
5359 LoopContinuePredicate->isUnconditional())
5362 return isImpliedCond(Pred, LHS, RHS,
5363 LoopContinuePredicate->getCondition(),
5364 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
5367 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
5368 /// by a conditional between LHS and RHS. This is used to help avoid max
5369 /// expressions in loop trip counts, and to eliminate casts.
5371 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
5372 ICmpInst::Predicate Pred,
5373 const SCEV *LHS, const SCEV *RHS) {
5374 // Interpret a null as meaning no loop, where there is obviously no guard
5375 // (interprocedural conditions notwithstanding).
5376 if (!L) return false;
5378 // Starting at the loop predecessor, climb up the predecessor chain, as long
5379 // as there are predecessors that can be found that have unique successors
5380 // leading to the original header.
5381 for (std::pair<BasicBlock *, BasicBlock *>
5382 Pair(L->getLoopPredecessor(), L->getHeader());
5384 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
5386 BranchInst *LoopEntryPredicate =
5387 dyn_cast<BranchInst>(Pair.first->getTerminator());
5388 if (!LoopEntryPredicate ||
5389 LoopEntryPredicate->isUnconditional())
5392 if (isImpliedCond(Pred, LHS, RHS,
5393 LoopEntryPredicate->getCondition(),
5394 LoopEntryPredicate->getSuccessor(0) != Pair.second))
5401 /// isImpliedCond - Test whether the condition described by Pred, LHS,
5402 /// and RHS is true whenever the given Cond value evaluates to true.
5403 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
5404 const SCEV *LHS, const SCEV *RHS,
5405 Value *FoundCondValue,
5407 // Recursively handle And and Or conditions.
5408 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
5409 if (BO->getOpcode() == Instruction::And) {
5411 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5412 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5413 } else if (BO->getOpcode() == Instruction::Or) {
5415 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5416 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5420 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
5421 if (!ICI) return false;
5423 // Bail if the ICmp's operands' types are wider than the needed type
5424 // before attempting to call getSCEV on them. This avoids infinite
5425 // recursion, since the analysis of widening casts can require loop
5426 // exit condition information for overflow checking, which would
5428 if (getTypeSizeInBits(LHS->getType()) <
5429 getTypeSizeInBits(ICI->getOperand(0)->getType()))
5432 // Now that we found a conditional branch that dominates the loop, check to
5433 // see if it is the comparison we are looking for.
5434 ICmpInst::Predicate FoundPred;
5436 FoundPred = ICI->getInversePredicate();
5438 FoundPred = ICI->getPredicate();
5440 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
5441 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
5443 // Balance the types. The case where FoundLHS' type is wider than
5444 // LHS' type is checked for above.
5445 if (getTypeSizeInBits(LHS->getType()) >
5446 getTypeSizeInBits(FoundLHS->getType())) {
5447 if (CmpInst::isSigned(Pred)) {
5448 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
5449 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
5451 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
5452 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
5456 // Canonicalize the query to match the way instcombine will have
5457 // canonicalized the comparison.
5458 if (SimplifyICmpOperands(Pred, LHS, RHS))
5460 return CmpInst::isTrueWhenEqual(Pred);
5461 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
5462 if (FoundLHS == FoundRHS)
5463 return CmpInst::isFalseWhenEqual(Pred);
5465 // Check to see if we can make the LHS or RHS match.
5466 if (LHS == FoundRHS || RHS == FoundLHS) {
5467 if (isa<SCEVConstant>(RHS)) {
5468 std::swap(FoundLHS, FoundRHS);
5469 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
5471 std::swap(LHS, RHS);
5472 Pred = ICmpInst::getSwappedPredicate(Pred);
5476 // Check whether the found predicate is the same as the desired predicate.
5477 if (FoundPred == Pred)
5478 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
5480 // Check whether swapping the found predicate makes it the same as the
5481 // desired predicate.
5482 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
5483 if (isa<SCEVConstant>(RHS))
5484 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
5486 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
5487 RHS, LHS, FoundLHS, FoundRHS);
5490 // Check whether the actual condition is beyond sufficient.
5491 if (FoundPred == ICmpInst::ICMP_EQ)
5492 if (ICmpInst::isTrueWhenEqual(Pred))
5493 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
5495 if (Pred == ICmpInst::ICMP_NE)
5496 if (!ICmpInst::isTrueWhenEqual(FoundPred))
5497 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
5500 // Otherwise assume the worst.
5504 /// isImpliedCondOperands - Test whether the condition described by Pred,
5505 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
5506 /// and FoundRHS is true.
5507 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
5508 const SCEV *LHS, const SCEV *RHS,
5509 const SCEV *FoundLHS,
5510 const SCEV *FoundRHS) {
5511 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
5512 FoundLHS, FoundRHS) ||
5513 // ~x < ~y --> x > y
5514 isImpliedCondOperandsHelper(Pred, LHS, RHS,
5515 getNotSCEV(FoundRHS),
5516 getNotSCEV(FoundLHS));
5519 /// isImpliedCondOperandsHelper - Test whether the condition described by
5520 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
5521 /// FoundLHS, and FoundRHS is true.
5523 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
5524 const SCEV *LHS, const SCEV *RHS,
5525 const SCEV *FoundLHS,
5526 const SCEV *FoundRHS) {
5528 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5529 case ICmpInst::ICMP_EQ:
5530 case ICmpInst::ICMP_NE:
5531 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
5534 case ICmpInst::ICMP_SLT:
5535 case ICmpInst::ICMP_SLE:
5536 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
5537 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
5540 case ICmpInst::ICMP_SGT:
5541 case ICmpInst::ICMP_SGE:
5542 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
5543 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
5546 case ICmpInst::ICMP_ULT:
5547 case ICmpInst::ICMP_ULE:
5548 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
5549 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
5552 case ICmpInst::ICMP_UGT:
5553 case ICmpInst::ICMP_UGE:
5554 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
5555 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
5563 /// getBECount - Subtract the end and start values and divide by the step,
5564 /// rounding up, to get the number of times the backedge is executed. Return
5565 /// CouldNotCompute if an intermediate computation overflows.
5566 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
5570 assert(!isKnownNegative(Step) &&
5571 "This code doesn't handle negative strides yet!");
5573 const Type *Ty = Start->getType();
5574 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
5575 const SCEV *Diff = getMinusSCEV(End, Start);
5576 const SCEV *RoundUp = getAddExpr(Step, NegOne);
5578 // Add an adjustment to the difference between End and Start so that
5579 // the division will effectively round up.
5580 const SCEV *Add = getAddExpr(Diff, RoundUp);
5583 // Check Add for unsigned overflow.
5584 // TODO: More sophisticated things could be done here.
5585 const Type *WideTy = IntegerType::get(getContext(),
5586 getTypeSizeInBits(Ty) + 1);
5587 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
5588 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
5589 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
5590 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
5591 return getCouldNotCompute();
5594 return getUDivExpr(Add, Step);
5597 /// HowManyLessThans - Return the number of times a backedge containing the
5598 /// specified less-than comparison will execute. If not computable, return
5599 /// CouldNotCompute.
5600 ScalarEvolution::BackedgeTakenInfo
5601 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
5602 const Loop *L, bool isSigned) {
5603 // Only handle: "ADDREC < LoopInvariant".
5604 if (!RHS->isLoopInvariant(L)) return getCouldNotCompute();
5606 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
5607 if (!AddRec || AddRec->getLoop() != L)
5608 return getCouldNotCompute();
5610 // Check to see if we have a flag which makes analysis easy.
5611 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
5612 AddRec->hasNoUnsignedWrap();
5614 if (AddRec->isAffine()) {
5615 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
5616 const SCEV *Step = AddRec->getStepRecurrence(*this);
5619 return getCouldNotCompute();
5620 if (Step->isOne()) {
5621 // With unit stride, the iteration never steps past the limit value.
5622 } else if (isKnownPositive(Step)) {
5623 // Test whether a positive iteration can step past the limit
5624 // value and past the maximum value for its type in a single step.
5625 // Note that it's not sufficient to check NoWrap here, because even
5626 // though the value after a wrap is undefined, it's not undefined
5627 // behavior, so if wrap does occur, the loop could either terminate or
5628 // loop infinitely, but in either case, the loop is guaranteed to
5629 // iterate at least until the iteration where the wrapping occurs.
5630 const SCEV *One = getConstant(Step->getType(), 1);
5632 APInt Max = APInt::getSignedMaxValue(BitWidth);
5633 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
5634 .slt(getSignedRange(RHS).getSignedMax()))
5635 return getCouldNotCompute();
5637 APInt Max = APInt::getMaxValue(BitWidth);
5638 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
5639 .ult(getUnsignedRange(RHS).getUnsignedMax()))
5640 return getCouldNotCompute();
5643 // TODO: Handle negative strides here and below.
5644 return getCouldNotCompute();
5646 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
5647 // m. So, we count the number of iterations in which {n,+,s} < m is true.
5648 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
5649 // treat m-n as signed nor unsigned due to overflow possibility.
5651 // First, we get the value of the LHS in the first iteration: n
5652 const SCEV *Start = AddRec->getOperand(0);
5654 // Determine the minimum constant start value.
5655 const SCEV *MinStart = getConstant(isSigned ?
5656 getSignedRange(Start).getSignedMin() :
5657 getUnsignedRange(Start).getUnsignedMin());
5659 // If we know that the condition is true in order to enter the loop,
5660 // then we know that it will run exactly (m-n)/s times. Otherwise, we
5661 // only know that it will execute (max(m,n)-n)/s times. In both cases,
5662 // the division must round up.
5663 const SCEV *End = RHS;
5664 if (!isLoopEntryGuardedByCond(L,
5665 isSigned ? ICmpInst::ICMP_SLT :
5667 getMinusSCEV(Start, Step), RHS))
5668 End = isSigned ? getSMaxExpr(RHS, Start)
5669 : getUMaxExpr(RHS, Start);
5671 // Determine the maximum constant end value.
5672 const SCEV *MaxEnd = getConstant(isSigned ?
5673 getSignedRange(End).getSignedMax() :
5674 getUnsignedRange(End).getUnsignedMax());
5676 // If MaxEnd is within a step of the maximum integer value in its type,
5677 // adjust it down to the minimum value which would produce the same effect.
5678 // This allows the subsequent ceiling division of (N+(step-1))/step to
5679 // compute the correct value.
5680 const SCEV *StepMinusOne = getMinusSCEV(Step,
5681 getConstant(Step->getType(), 1));
5684 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
5687 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
5690 // Finally, we subtract these two values and divide, rounding up, to get
5691 // the number of times the backedge is executed.
5692 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
5694 // The maximum backedge count is similar, except using the minimum start
5695 // value and the maximum end value.
5696 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
5698 return BackedgeTakenInfo(BECount, MaxBECount);
5701 return getCouldNotCompute();
5704 /// getNumIterationsInRange - Return the number of iterations of this loop that
5705 /// produce values in the specified constant range. Another way of looking at
5706 /// this is that it returns the first iteration number where the value is not in
5707 /// the condition, thus computing the exit count. If the iteration count can't
5708 /// be computed, an instance of SCEVCouldNotCompute is returned.
5709 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
5710 ScalarEvolution &SE) const {
5711 if (Range.isFullSet()) // Infinite loop.
5712 return SE.getCouldNotCompute();
5714 // If the start is a non-zero constant, shift the range to simplify things.
5715 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
5716 if (!SC->getValue()->isZero()) {
5717 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
5718 Operands[0] = SE.getConstant(SC->getType(), 0);
5719 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
5720 if (const SCEVAddRecExpr *ShiftedAddRec =
5721 dyn_cast<SCEVAddRecExpr>(Shifted))
5722 return ShiftedAddRec->getNumIterationsInRange(
5723 Range.subtract(SC->getValue()->getValue()), SE);
5724 // This is strange and shouldn't happen.
5725 return SE.getCouldNotCompute();
5728 // The only time we can solve this is when we have all constant indices.
5729 // Otherwise, we cannot determine the overflow conditions.
5730 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5731 if (!isa<SCEVConstant>(getOperand(i)))
5732 return SE.getCouldNotCompute();
5735 // Okay at this point we know that all elements of the chrec are constants and
5736 // that the start element is zero.
5738 // First check to see if the range contains zero. If not, the first
5740 unsigned BitWidth = SE.getTypeSizeInBits(getType());
5741 if (!Range.contains(APInt(BitWidth, 0)))
5742 return SE.getConstant(getType(), 0);
5745 // If this is an affine expression then we have this situation:
5746 // Solve {0,+,A} in Range === Ax in Range
5748 // We know that zero is in the range. If A is positive then we know that
5749 // the upper value of the range must be the first possible exit value.
5750 // If A is negative then the lower of the range is the last possible loop
5751 // value. Also note that we already checked for a full range.
5752 APInt One(BitWidth,1);
5753 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5754 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5756 // The exit value should be (End+A)/A.
5757 APInt ExitVal = (End + A).udiv(A);
5758 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5760 // Evaluate at the exit value. If we really did fall out of the valid
5761 // range, then we computed our trip count, otherwise wrap around or other
5762 // things must have happened.
5763 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5764 if (Range.contains(Val->getValue()))
5765 return SE.getCouldNotCompute(); // Something strange happened
5767 // Ensure that the previous value is in the range. This is a sanity check.
5768 assert(Range.contains(
5769 EvaluateConstantChrecAtConstant(this,
5770 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5771 "Linear scev computation is off in a bad way!");
5772 return SE.getConstant(ExitValue);
5773 } else if (isQuadratic()) {
5774 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5775 // quadratic equation to solve it. To do this, we must frame our problem in
5776 // terms of figuring out when zero is crossed, instead of when
5777 // Range.getUpper() is crossed.
5778 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5779 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5780 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5782 // Next, solve the constructed addrec
5783 std::pair<const SCEV *,const SCEV *> Roots =
5784 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5785 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5786 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5788 // Pick the smallest positive root value.
5789 if (ConstantInt *CB =
5790 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5791 R1->getValue(), R2->getValue()))) {
5792 if (CB->getZExtValue() == false)
5793 std::swap(R1, R2); // R1 is the minimum root now.
5795 // Make sure the root is not off by one. The returned iteration should
5796 // not be in the range, but the previous one should be. When solving
5797 // for "X*X < 5", for example, we should not return a root of 2.
5798 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5801 if (Range.contains(R1Val->getValue())) {
5802 // The next iteration must be out of the range...
5803 ConstantInt *NextVal =
5804 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5806 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5807 if (!Range.contains(R1Val->getValue()))
5808 return SE.getConstant(NextVal);
5809 return SE.getCouldNotCompute(); // Something strange happened
5812 // If R1 was not in the range, then it is a good return value. Make
5813 // sure that R1-1 WAS in the range though, just in case.
5814 ConstantInt *NextVal =
5815 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5816 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5817 if (Range.contains(R1Val->getValue()))
5819 return SE.getCouldNotCompute(); // Something strange happened
5824 return SE.getCouldNotCompute();
5829 //===----------------------------------------------------------------------===//
5830 // SCEVCallbackVH Class Implementation
5831 //===----------------------------------------------------------------------===//
5833 void ScalarEvolution::SCEVCallbackVH::deleted() {
5834 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5835 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5836 SE->ConstantEvolutionLoopExitValue.erase(PN);
5837 SE->ValueExprMap.erase(getValPtr());
5838 // this now dangles!
5841 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
5842 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5844 // Forget all the expressions associated with users of the old value,
5845 // so that future queries will recompute the expressions using the new
5847 Value *Old = getValPtr();
5848 SmallVector<User *, 16> Worklist;
5849 SmallPtrSet<User *, 8> Visited;
5850 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5852 Worklist.push_back(*UI);
5853 while (!Worklist.empty()) {
5854 User *U = Worklist.pop_back_val();
5855 // Deleting the Old value will cause this to dangle. Postpone
5856 // that until everything else is done.
5859 if (!Visited.insert(U))
5861 if (PHINode *PN = dyn_cast<PHINode>(U))
5862 SE->ConstantEvolutionLoopExitValue.erase(PN);
5863 SE->ValueExprMap.erase(U);
5864 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5866 Worklist.push_back(*UI);
5868 // Delete the Old value.
5869 if (PHINode *PN = dyn_cast<PHINode>(Old))
5870 SE->ConstantEvolutionLoopExitValue.erase(PN);
5871 SE->ValueExprMap.erase(Old);
5872 // this now dangles!
5875 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5876 : CallbackVH(V), SE(se) {}
5878 //===----------------------------------------------------------------------===//
5879 // ScalarEvolution Class Implementation
5880 //===----------------------------------------------------------------------===//
5882 ScalarEvolution::ScalarEvolution()
5883 : FunctionPass(ID), FirstUnknown(0) {
5884 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
5887 bool ScalarEvolution::runOnFunction(Function &F) {
5889 LI = &getAnalysis<LoopInfo>();
5890 TD = getAnalysisIfAvailable<TargetData>();
5891 DT = &getAnalysis<DominatorTree>();
5895 void ScalarEvolution::releaseMemory() {
5896 // Iterate through all the SCEVUnknown instances and call their
5897 // destructors, so that they release their references to their values.
5898 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
5902 ValueExprMap.clear();
5903 BackedgeTakenCounts.clear();
5904 ConstantEvolutionLoopExitValue.clear();
5905 ValuesAtScopes.clear();
5906 UnsignedRanges.clear();
5907 SignedRanges.clear();
5908 UniqueSCEVs.clear();
5909 SCEVAllocator.Reset();
5912 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5913 AU.setPreservesAll();
5914 AU.addRequiredTransitive<LoopInfo>();
5915 AU.addRequiredTransitive<DominatorTree>();
5918 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5919 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5922 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5924 // Print all inner loops first
5925 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5926 PrintLoopInfo(OS, SE, *I);
5929 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5932 SmallVector<BasicBlock *, 8> ExitBlocks;
5933 L->getExitBlocks(ExitBlocks);
5934 if (ExitBlocks.size() != 1)
5935 OS << "<multiple exits> ";
5937 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5938 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5940 OS << "Unpredictable backedge-taken count. ";
5945 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5948 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5949 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5951 OS << "Unpredictable max backedge-taken count. ";
5957 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
5958 // ScalarEvolution's implementation of the print method is to print
5959 // out SCEV values of all instructions that are interesting. Doing
5960 // this potentially causes it to create new SCEV objects though,
5961 // which technically conflicts with the const qualifier. This isn't
5962 // observable from outside the class though, so casting away the
5963 // const isn't dangerous.
5964 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
5966 OS << "Classifying expressions for: ";
5967 WriteAsOperand(OS, F, /*PrintType=*/false);
5969 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
5970 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
5973 const SCEV *SV = SE.getSCEV(&*I);
5976 const Loop *L = LI->getLoopFor((*I).getParent());
5978 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
5985 OS << "\t\t" "Exits: ";
5986 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
5987 if (!ExitValue->isLoopInvariant(L)) {
5988 OS << "<<Unknown>>";
5997 OS << "Determining loop execution counts for: ";
5998 WriteAsOperand(OS, F, /*PrintType=*/false);
6000 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6001 PrintLoopInfo(OS, &SE, *I);