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/LoopInfo.h"
73 #include "llvm/Analysis/ValueTracking.h"
74 #include "llvm/Assembly/Writer.h"
75 #include "llvm/Target/TargetData.h"
76 #include "llvm/Support/CommandLine.h"
77 #include "llvm/Support/ConstantRange.h"
78 #include "llvm/Support/Debug.h"
79 #include "llvm/Support/ErrorHandling.h"
80 #include "llvm/Support/GetElementPtrTypeIterator.h"
81 #include "llvm/Support/InstIterator.h"
82 #include "llvm/Support/MathExtras.h"
83 #include "llvm/Support/raw_ostream.h"
84 #include "llvm/ADT/Statistic.h"
85 #include "llvm/ADT/STLExtras.h"
86 #include "llvm/ADT/SmallPtrSet.h"
90 STATISTIC(NumArrayLenItCounts,
91 "Number of trip counts computed with array length");
92 STATISTIC(NumTripCountsComputed,
93 "Number of loops with predictable loop counts");
94 STATISTIC(NumTripCountsNotComputed,
95 "Number of loops without predictable loop counts");
96 STATISTIC(NumBruteForceTripCountsComputed,
97 "Number of loops with trip counts computed by force");
99 static cl::opt<unsigned>
100 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
101 cl::desc("Maximum number of iterations SCEV will "
102 "symbolically execute a constant "
106 INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
107 "Scalar Evolution Analysis", false, true)
108 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
109 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
110 INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
111 "Scalar Evolution Analysis", false, true)
112 char ScalarEvolution::ID = 0;
114 //===----------------------------------------------------------------------===//
115 // SCEV class definitions
116 //===----------------------------------------------------------------------===//
118 //===----------------------------------------------------------------------===//
119 // Implementation of the SCEV class.
124 void SCEV::dump() const {
129 bool SCEV::isZero() const {
130 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
131 return SC->getValue()->isZero();
135 bool SCEV::isOne() const {
136 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
137 return SC->getValue()->isOne();
141 bool SCEV::isAllOnesValue() const {
142 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
143 return SC->getValue()->isAllOnesValue();
147 SCEVCouldNotCompute::SCEVCouldNotCompute() :
148 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
150 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
151 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
155 const Type *SCEVCouldNotCompute::getType() const {
156 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
160 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
161 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
165 bool SCEVCouldNotCompute::hasOperand(const SCEV *) const {
166 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
170 void SCEVCouldNotCompute::print(raw_ostream &OS) const {
171 OS << "***COULDNOTCOMPUTE***";
174 bool SCEVCouldNotCompute::classof(const SCEV *S) {
175 return S->getSCEVType() == scCouldNotCompute;
178 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
180 ID.AddInteger(scConstant);
183 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
184 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V);
185 UniqueSCEVs.InsertNode(S, IP);
189 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
190 return getConstant(ConstantInt::get(getContext(), Val));
194 ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
195 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
196 return getConstant(ConstantInt::get(ITy, V, isSigned));
199 const Type *SCEVConstant::getType() const { return V->getType(); }
201 void SCEVConstant::print(raw_ostream &OS) const {
202 WriteAsOperand(OS, V, false);
205 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
206 unsigned SCEVTy, const SCEV *op, const Type *ty)
207 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
209 bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
210 return Op->dominates(BB, DT);
213 bool SCEVCastExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
214 return Op->properlyDominates(BB, DT);
217 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
218 const SCEV *op, const Type *ty)
219 : SCEVCastExpr(ID, scTruncate, op, ty) {
220 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
221 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
222 "Cannot truncate non-integer value!");
225 void SCEVTruncateExpr::print(raw_ostream &OS) const {
226 OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
229 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
230 const SCEV *op, const Type *ty)
231 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
232 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
233 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
234 "Cannot zero extend non-integer value!");
237 void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
238 OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
241 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
242 const SCEV *op, const Type *ty)
243 : SCEVCastExpr(ID, scSignExtend, op, ty) {
244 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
245 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
246 "Cannot sign extend non-integer value!");
249 void SCEVSignExtendExpr::print(raw_ostream &OS) const {
250 OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
253 void SCEVCommutativeExpr::print(raw_ostream &OS) const {
254 const char *OpStr = getOperationStr();
256 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I) {
258 if (llvm::next(I) != E)
264 bool SCEVNAryExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
265 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
266 if (!(*I)->dominates(BB, DT))
271 bool SCEVNAryExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
272 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
273 if (!(*I)->properlyDominates(BB, DT))
278 bool SCEVNAryExpr::isLoopInvariant(const Loop *L) const {
279 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
280 if (!(*I)->isLoopInvariant(L))
285 // hasComputableLoopEvolution - N-ary expressions have computable loop
286 // evolutions iff they have at least one operand that varies with the loop,
287 // but that all varying operands are computable.
288 bool SCEVNAryExpr::hasComputableLoopEvolution(const Loop *L) const {
289 bool HasVarying = false;
290 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I) {
292 if (!S->isLoopInvariant(L)) {
293 if (S->hasComputableLoopEvolution(L))
302 bool SCEVNAryExpr::hasOperand(const SCEV *O) const {
303 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I) {
305 if (O == S || S->hasOperand(O))
311 bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
312 return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
315 bool SCEVUDivExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
316 return LHS->properlyDominates(BB, DT) && RHS->properlyDominates(BB, DT);
319 void SCEVUDivExpr::print(raw_ostream &OS) const {
320 OS << "(" << *LHS << " /u " << *RHS << ")";
323 const Type *SCEVUDivExpr::getType() const {
324 // In most cases the types of LHS and RHS will be the same, but in some
325 // crazy cases one or the other may be a pointer. ScalarEvolution doesn't
326 // depend on the type for correctness, but handling types carefully can
327 // avoid extra casts in the SCEVExpander. The LHS is more likely to be
328 // a pointer type than the RHS, so use the RHS' type here.
329 return RHS->getType();
332 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
333 // Add recurrences are never invariant in the function-body (null loop).
337 // This recurrence is variant w.r.t. QueryLoop if QueryLoop contains L.
338 if (QueryLoop->contains(L))
341 // This recurrence is invariant w.r.t. QueryLoop if L contains QueryLoop.
342 if (L->contains(QueryLoop))
345 // This recurrence is variant w.r.t. QueryLoop if any of its operands
347 for (op_iterator I = op_begin(), E = op_end(); I != E; ++I)
348 if (!(*I)->isLoopInvariant(QueryLoop))
351 // Otherwise it's loop-invariant.
356 SCEVAddRecExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
357 return DT->dominates(L->getHeader(), BB) &&
358 SCEVNAryExpr::dominates(BB, DT);
362 SCEVAddRecExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
363 // This uses a "dominates" query instead of "properly dominates" query because
364 // the instruction which produces the addrec's value is a PHI, and a PHI
365 // effectively properly dominates its entire containing block.
366 return DT->dominates(L->getHeader(), BB) &&
367 SCEVNAryExpr::properlyDominates(BB, DT);
370 void SCEVAddRecExpr::print(raw_ostream &OS) const {
371 OS << "{" << *Operands[0];
372 for (unsigned i = 1, e = NumOperands; i != e; ++i)
373 OS << ",+," << *Operands[i];
375 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
379 void SCEVUnknown::deleted() {
380 // Clear this SCEVUnknown from various maps.
381 SE->ValuesAtScopes.erase(this);
382 SE->UnsignedRanges.erase(this);
383 SE->SignedRanges.erase(this);
385 // Remove this SCEVUnknown from the uniquing map.
386 SE->UniqueSCEVs.RemoveNode(this);
388 // Release the value.
392 void SCEVUnknown::allUsesReplacedWith(Value *New) {
393 // Clear this SCEVUnknown from various maps.
394 SE->ValuesAtScopes.erase(this);
395 SE->UnsignedRanges.erase(this);
396 SE->SignedRanges.erase(this);
398 // Remove this SCEVUnknown from the uniquing map.
399 SE->UniqueSCEVs.RemoveNode(this);
401 // Update this SCEVUnknown to point to the new value. This is needed
402 // because there may still be outstanding SCEVs which still point to
407 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
408 // All non-instruction values are loop invariant. All instructions are loop
409 // invariant if they are not contained in the specified loop.
410 // Instructions are never considered invariant in the function body
411 // (null loop) because they are defined within the "loop".
412 if (Instruction *I = dyn_cast<Instruction>(getValue()))
413 return L && !L->contains(I);
417 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
418 if (Instruction *I = dyn_cast<Instruction>(getValue()))
419 return DT->dominates(I->getParent(), BB);
423 bool SCEVUnknown::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
424 if (Instruction *I = dyn_cast<Instruction>(getValue()))
425 return DT->properlyDominates(I->getParent(), BB);
429 const Type *SCEVUnknown::getType() const {
430 return getValue()->getType();
433 bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
434 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
435 if (VCE->getOpcode() == Instruction::PtrToInt)
436 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
437 if (CE->getOpcode() == Instruction::GetElementPtr &&
438 CE->getOperand(0)->isNullValue() &&
439 CE->getNumOperands() == 2)
440 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
442 AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
450 bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
451 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
452 if (VCE->getOpcode() == Instruction::PtrToInt)
453 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
454 if (CE->getOpcode() == Instruction::GetElementPtr &&
455 CE->getOperand(0)->isNullValue()) {
457 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
458 if (const StructType *STy = dyn_cast<StructType>(Ty))
459 if (!STy->isPacked() &&
460 CE->getNumOperands() == 3 &&
461 CE->getOperand(1)->isNullValue()) {
462 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
464 STy->getNumElements() == 2 &&
465 STy->getElementType(0)->isIntegerTy(1)) {
466 AllocTy = STy->getElementType(1);
475 bool SCEVUnknown::isOffsetOf(const Type *&CTy, Constant *&FieldNo) const {
476 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
477 if (VCE->getOpcode() == Instruction::PtrToInt)
478 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
479 if (CE->getOpcode() == Instruction::GetElementPtr &&
480 CE->getNumOperands() == 3 &&
481 CE->getOperand(0)->isNullValue() &&
482 CE->getOperand(1)->isNullValue()) {
484 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
485 // Ignore vector types here so that ScalarEvolutionExpander doesn't
486 // emit getelementptrs that index into vectors.
487 if (Ty->isStructTy() || Ty->isArrayTy()) {
489 FieldNo = CE->getOperand(2);
497 void SCEVUnknown::print(raw_ostream &OS) const {
499 if (isSizeOf(AllocTy)) {
500 OS << "sizeof(" << *AllocTy << ")";
503 if (isAlignOf(AllocTy)) {
504 OS << "alignof(" << *AllocTy << ")";
510 if (isOffsetOf(CTy, FieldNo)) {
511 OS << "offsetof(" << *CTy << ", ";
512 WriteAsOperand(OS, FieldNo, false);
517 // Otherwise just print it normally.
518 WriteAsOperand(OS, getValue(), false);
521 //===----------------------------------------------------------------------===//
523 //===----------------------------------------------------------------------===//
526 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
527 /// than the complexity of the RHS. This comparator is used to canonicalize
529 class SCEVComplexityCompare {
530 const LoopInfo *const LI;
532 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
534 // Return true or false if LHS is less than, or at least RHS, respectively.
535 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
536 return compare(LHS, RHS) < 0;
539 // Return negative, zero, or positive, if LHS is less than, equal to, or
540 // greater than RHS, respectively. A three-way result allows recursive
541 // comparisons to be more efficient.
542 int compare(const SCEV *LHS, const SCEV *RHS) const {
543 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
547 // Primarily, sort the SCEVs by their getSCEVType().
548 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
550 return (int)LType - (int)RType;
552 // Aside from the getSCEVType() ordering, the particular ordering
553 // isn't very important except that it's beneficial to be consistent,
554 // so that (a + b) and (b + a) don't end up as different expressions.
557 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS);
558 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
560 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
561 // not as complete as it could be.
562 const Value *LV = LU->getValue(), *RV = RU->getValue();
564 // Order pointer values after integer values. This helps SCEVExpander
566 bool LIsPointer = LV->getType()->isPointerTy(),
567 RIsPointer = RV->getType()->isPointerTy();
568 if (LIsPointer != RIsPointer)
569 return (int)LIsPointer - (int)RIsPointer;
571 // Compare getValueID values.
572 unsigned LID = LV->getValueID(),
573 RID = RV->getValueID();
575 return (int)LID - (int)RID;
577 // Sort arguments by their position.
578 if (const Argument *LA = dyn_cast<Argument>(LV)) {
579 const Argument *RA = cast<Argument>(RV);
580 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo();
581 return (int)LArgNo - (int)RArgNo;
584 // For instructions, compare their loop depth, and their operand
585 // count. This is pretty loose.
586 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) {
587 const Instruction *RInst = cast<Instruction>(RV);
589 // Compare loop depths.
590 const BasicBlock *LParent = LInst->getParent(),
591 *RParent = RInst->getParent();
592 if (LParent != RParent) {
593 unsigned LDepth = LI->getLoopDepth(LParent),
594 RDepth = LI->getLoopDepth(RParent);
595 if (LDepth != RDepth)
596 return (int)LDepth - (int)RDepth;
599 // Compare the number of operands.
600 unsigned LNumOps = LInst->getNumOperands(),
601 RNumOps = RInst->getNumOperands();
602 return (int)LNumOps - (int)RNumOps;
609 const SCEVConstant *LC = cast<SCEVConstant>(LHS);
610 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
612 // Compare constant values.
613 const APInt &LA = LC->getValue()->getValue();
614 const APInt &RA = RC->getValue()->getValue();
615 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth();
616 if (LBitWidth != RBitWidth)
617 return (int)LBitWidth - (int)RBitWidth;
618 return LA.ult(RA) ? -1 : 1;
622 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS);
623 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
625 // Compare addrec loop depths.
626 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop();
627 if (LLoop != RLoop) {
628 unsigned LDepth = LLoop->getLoopDepth(),
629 RDepth = RLoop->getLoopDepth();
630 if (LDepth != RDepth)
631 return (int)LDepth - (int)RDepth;
634 // Addrec complexity grows with operand count.
635 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands();
636 if (LNumOps != RNumOps)
637 return (int)LNumOps - (int)RNumOps;
639 // Lexicographically compare.
640 for (unsigned i = 0; i != LNumOps; ++i) {
641 long X = compare(LA->getOperand(i), RA->getOperand(i));
653 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS);
654 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
656 // Lexicographically compare n-ary expressions.
657 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
658 for (unsigned i = 0; i != LNumOps; ++i) {
661 long X = compare(LC->getOperand(i), RC->getOperand(i));
665 return (int)LNumOps - (int)RNumOps;
669 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
670 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
672 // Lexicographically compare udiv expressions.
673 long X = compare(LC->getLHS(), RC->getLHS());
676 return compare(LC->getRHS(), RC->getRHS());
682 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
683 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
685 // Compare cast expressions by operand.
686 return compare(LC->getOperand(), RC->getOperand());
693 llvm_unreachable("Unknown SCEV kind!");
699 /// GroupByComplexity - Given a list of SCEV objects, order them by their
700 /// complexity, and group objects of the same complexity together by value.
701 /// When this routine is finished, we know that any duplicates in the vector are
702 /// consecutive and that complexity is monotonically increasing.
704 /// Note that we go take special precautions to ensure that we get deterministic
705 /// results from this routine. In other words, we don't want the results of
706 /// this to depend on where the addresses of various SCEV objects happened to
709 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
711 if (Ops.size() < 2) return; // Noop
712 if (Ops.size() == 2) {
713 // This is the common case, which also happens to be trivially simple.
715 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
716 if (SCEVComplexityCompare(LI)(RHS, LHS))
721 // Do the rough sort by complexity.
722 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
724 // Now that we are sorted by complexity, group elements of the same
725 // complexity. Note that this is, at worst, N^2, but the vector is likely to
726 // be extremely short in practice. Note that we take this approach because we
727 // do not want to depend on the addresses of the objects we are grouping.
728 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
729 const SCEV *S = Ops[i];
730 unsigned Complexity = S->getSCEVType();
732 // If there are any objects of the same complexity and same value as this
734 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
735 if (Ops[j] == S) { // Found a duplicate.
736 // Move it to immediately after i'th element.
737 std::swap(Ops[i+1], Ops[j]);
738 ++i; // no need to rescan it.
739 if (i == e-2) return; // Done!
747 //===----------------------------------------------------------------------===//
748 // Simple SCEV method implementations
749 //===----------------------------------------------------------------------===//
751 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
753 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
755 const Type* ResultTy) {
756 // Handle the simplest case efficiently.
758 return SE.getTruncateOrZeroExtend(It, ResultTy);
760 // We are using the following formula for BC(It, K):
762 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
764 // Suppose, W is the bitwidth of the return value. We must be prepared for
765 // overflow. Hence, we must assure that the result of our computation is
766 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
767 // safe in modular arithmetic.
769 // However, this code doesn't use exactly that formula; the formula it uses
770 // is something like the following, where T is the number of factors of 2 in
771 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
774 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
776 // This formula is trivially equivalent to the previous formula. However,
777 // this formula can be implemented much more efficiently. The trick is that
778 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
779 // arithmetic. To do exact division in modular arithmetic, all we have
780 // to do is multiply by the inverse. Therefore, this step can be done at
783 // The next issue is how to safely do the division by 2^T. The way this
784 // is done is by doing the multiplication step at a width of at least W + T
785 // bits. This way, the bottom W+T bits of the product are accurate. Then,
786 // when we perform the division by 2^T (which is equivalent to a right shift
787 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
788 // truncated out after the division by 2^T.
790 // In comparison to just directly using the first formula, this technique
791 // is much more efficient; using the first formula requires W * K bits,
792 // but this formula less than W + K bits. Also, the first formula requires
793 // a division step, whereas this formula only requires multiplies and shifts.
795 // It doesn't matter whether the subtraction step is done in the calculation
796 // width or the input iteration count's width; if the subtraction overflows,
797 // the result must be zero anyway. We prefer here to do it in the width of
798 // the induction variable because it helps a lot for certain cases; CodeGen
799 // isn't smart enough to ignore the overflow, which leads to much less
800 // efficient code if the width of the subtraction is wider than the native
803 // (It's possible to not widen at all by pulling out factors of 2 before
804 // the multiplication; for example, K=2 can be calculated as
805 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
806 // extra arithmetic, so it's not an obvious win, and it gets
807 // much more complicated for K > 3.)
809 // Protection from insane SCEVs; this bound is conservative,
810 // but it probably doesn't matter.
812 return SE.getCouldNotCompute();
814 unsigned W = SE.getTypeSizeInBits(ResultTy);
816 // Calculate K! / 2^T and T; we divide out the factors of two before
817 // multiplying for calculating K! / 2^T to avoid overflow.
818 // Other overflow doesn't matter because we only care about the bottom
819 // W bits of the result.
820 APInt OddFactorial(W, 1);
822 for (unsigned i = 3; i <= K; ++i) {
824 unsigned TwoFactors = Mult.countTrailingZeros();
826 Mult = Mult.lshr(TwoFactors);
827 OddFactorial *= Mult;
830 // We need at least W + T bits for the multiplication step
831 unsigned CalculationBits = W + T;
833 // Calculate 2^T, at width T+W.
834 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
836 // Calculate the multiplicative inverse of K! / 2^T;
837 // this multiplication factor will perform the exact division by
839 APInt Mod = APInt::getSignedMinValue(W+1);
840 APInt MultiplyFactor = OddFactorial.zext(W+1);
841 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
842 MultiplyFactor = MultiplyFactor.trunc(W);
844 // Calculate the product, at width T+W
845 const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
847 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
848 for (unsigned i = 1; i != K; ++i) {
849 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
850 Dividend = SE.getMulExpr(Dividend,
851 SE.getTruncateOrZeroExtend(S, CalculationTy));
855 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
857 // Truncate the result, and divide by K! / 2^T.
859 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
860 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
863 /// evaluateAtIteration - Return the value of this chain of recurrences at
864 /// the specified iteration number. We can evaluate this recurrence by
865 /// multiplying each element in the chain by the binomial coefficient
866 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
868 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
870 /// where BC(It, k) stands for binomial coefficient.
872 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
873 ScalarEvolution &SE) const {
874 const SCEV *Result = getStart();
875 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
876 // The computation is correct in the face of overflow provided that the
877 // multiplication is performed _after_ the evaluation of the binomial
879 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
880 if (isa<SCEVCouldNotCompute>(Coeff))
883 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
888 //===----------------------------------------------------------------------===//
889 // SCEV Expression folder implementations
890 //===----------------------------------------------------------------------===//
892 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
894 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
895 "This is not a truncating conversion!");
896 assert(isSCEVable(Ty) &&
897 "This is not a conversion to a SCEVable type!");
898 Ty = getEffectiveSCEVType(Ty);
901 ID.AddInteger(scTruncate);
905 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
907 // Fold if the operand is constant.
908 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
910 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(),
911 getEffectiveSCEVType(Ty))));
913 // trunc(trunc(x)) --> trunc(x)
914 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
915 return getTruncateExpr(ST->getOperand(), Ty);
917 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
918 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
919 return getTruncateOrSignExtend(SS->getOperand(), Ty);
921 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
922 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
923 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
925 // If the input value is a chrec scev, truncate the chrec's operands.
926 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
927 SmallVector<const SCEV *, 4> Operands;
928 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
929 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
930 return getAddRecExpr(Operands, AddRec->getLoop());
933 // As a special case, fold trunc(undef) to undef. We don't want to
934 // know too much about SCEVUnknowns, but this special case is handy
936 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
937 if (isa<UndefValue>(U->getValue()))
938 return getSCEV(UndefValue::get(Ty));
940 // The cast wasn't folded; create an explicit cast node. We can reuse
941 // the existing insert position since if we get here, we won't have
942 // made any changes which would invalidate it.
943 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
945 UniqueSCEVs.InsertNode(S, IP);
949 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
951 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
952 "This is not an extending conversion!");
953 assert(isSCEVable(Ty) &&
954 "This is not a conversion to a SCEVable type!");
955 Ty = getEffectiveSCEVType(Ty);
957 // Fold if the operand is constant.
958 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
960 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(),
961 getEffectiveSCEVType(Ty))));
963 // zext(zext(x)) --> zext(x)
964 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
965 return getZeroExtendExpr(SZ->getOperand(), Ty);
967 // Before doing any expensive analysis, check to see if we've already
968 // computed a SCEV for this Op and Ty.
970 ID.AddInteger(scZeroExtend);
974 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
976 // If the input value is a chrec scev, and we can prove that the value
977 // did not overflow the old, smaller, value, we can zero extend all of the
978 // operands (often constants). This allows analysis of something like
979 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
980 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
981 if (AR->isAffine()) {
982 const SCEV *Start = AR->getStart();
983 const SCEV *Step = AR->getStepRecurrence(*this);
984 unsigned BitWidth = getTypeSizeInBits(AR->getType());
985 const Loop *L = AR->getLoop();
987 // If we have special knowledge that this addrec won't overflow,
988 // we don't need to do any further analysis.
989 if (AR->hasNoUnsignedWrap())
990 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
991 getZeroExtendExpr(Step, Ty),
994 // Check whether the backedge-taken count is SCEVCouldNotCompute.
995 // Note that this serves two purposes: It filters out loops that are
996 // simply not analyzable, and it covers the case where this code is
997 // being called from within backedge-taken count analysis, such that
998 // attempting to ask for the backedge-taken count would likely result
999 // in infinite recursion. In the later case, the analysis code will
1000 // cope with a conservative value, and it will take care to purge
1001 // that value once it has finished.
1002 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1003 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1004 // Manually compute the final value for AR, checking for
1007 // Check whether the backedge-taken count can be losslessly casted to
1008 // the addrec's type. The count is always unsigned.
1009 const SCEV *CastedMaxBECount =
1010 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1011 const SCEV *RecastedMaxBECount =
1012 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1013 if (MaxBECount == RecastedMaxBECount) {
1014 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1015 // Check whether Start+Step*MaxBECount has no unsigned overflow.
1016 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
1017 const SCEV *Add = getAddExpr(Start, ZMul);
1018 const SCEV *OperandExtendedAdd =
1019 getAddExpr(getZeroExtendExpr(Start, WideTy),
1020 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1021 getZeroExtendExpr(Step, WideTy)));
1022 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
1023 // Return the expression with the addrec on the outside.
1024 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1025 getZeroExtendExpr(Step, Ty),
1028 // Similar to above, only this time treat the step value as signed.
1029 // This covers loops that count down.
1030 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1031 Add = getAddExpr(Start, SMul);
1032 OperandExtendedAdd =
1033 getAddExpr(getZeroExtendExpr(Start, WideTy),
1034 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1035 getSignExtendExpr(Step, WideTy)));
1036 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
1037 // Return the expression with the addrec on the outside.
1038 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1039 getSignExtendExpr(Step, Ty),
1043 // If the backedge is guarded by a comparison with the pre-inc value
1044 // the addrec is safe. Also, if the entry is guarded by a comparison
1045 // with the start value and the backedge is guarded by a comparison
1046 // with the post-inc value, the addrec is safe.
1047 if (isKnownPositive(Step)) {
1048 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
1049 getUnsignedRange(Step).getUnsignedMax());
1050 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
1051 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
1052 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
1053 AR->getPostIncExpr(*this), N)))
1054 // Return the expression with the addrec on the outside.
1055 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1056 getZeroExtendExpr(Step, Ty),
1058 } else if (isKnownNegative(Step)) {
1059 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
1060 getSignedRange(Step).getSignedMin());
1061 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
1062 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
1063 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
1064 AR->getPostIncExpr(*this), N)))
1065 // Return the expression with the addrec on the outside.
1066 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1067 getSignExtendExpr(Step, Ty),
1073 // The cast wasn't folded; create an explicit cast node.
1074 // Recompute the insert position, as it may have been invalidated.
1075 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1076 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1078 UniqueSCEVs.InsertNode(S, IP);
1082 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1084 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1085 "This is not an extending conversion!");
1086 assert(isSCEVable(Ty) &&
1087 "This is not a conversion to a SCEVable type!");
1088 Ty = getEffectiveSCEVType(Ty);
1090 // Fold if the operand is constant.
1091 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1093 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(),
1094 getEffectiveSCEVType(Ty))));
1096 // sext(sext(x)) --> sext(x)
1097 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1098 return getSignExtendExpr(SS->getOperand(), Ty);
1100 // Before doing any expensive analysis, check to see if we've already
1101 // computed a SCEV for this Op and Ty.
1102 FoldingSetNodeID ID;
1103 ID.AddInteger(scSignExtend);
1107 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1109 // If the input value is a chrec scev, and we can prove that the value
1110 // did not overflow the old, smaller, value, we can sign extend all of the
1111 // operands (often constants). This allows analysis of something like
1112 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1113 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1114 if (AR->isAffine()) {
1115 const SCEV *Start = AR->getStart();
1116 const SCEV *Step = AR->getStepRecurrence(*this);
1117 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1118 const Loop *L = AR->getLoop();
1120 // If we have special knowledge that this addrec won't overflow,
1121 // we don't need to do any further analysis.
1122 if (AR->hasNoSignedWrap())
1123 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1124 getSignExtendExpr(Step, Ty),
1127 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1128 // Note that this serves two purposes: It filters out loops that are
1129 // simply not analyzable, and it covers the case where this code is
1130 // being called from within backedge-taken count analysis, such that
1131 // attempting to ask for the backedge-taken count would likely result
1132 // in infinite recursion. In the later case, the analysis code will
1133 // cope with a conservative value, and it will take care to purge
1134 // that value once it has finished.
1135 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1136 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1137 // Manually compute the final value for AR, checking for
1140 // Check whether the backedge-taken count can be losslessly casted to
1141 // the addrec's type. The count is always unsigned.
1142 const SCEV *CastedMaxBECount =
1143 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1144 const SCEV *RecastedMaxBECount =
1145 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1146 if (MaxBECount == RecastedMaxBECount) {
1147 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1148 // Check whether Start+Step*MaxBECount has no signed overflow.
1149 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1150 const SCEV *Add = getAddExpr(Start, SMul);
1151 const SCEV *OperandExtendedAdd =
1152 getAddExpr(getSignExtendExpr(Start, WideTy),
1153 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1154 getSignExtendExpr(Step, WideTy)));
1155 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1156 // Return the expression with the addrec on the outside.
1157 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1158 getSignExtendExpr(Step, Ty),
1161 // Similar to above, only this time treat the step value as unsigned.
1162 // This covers loops that count up with an unsigned step.
1163 const SCEV *UMul = getMulExpr(CastedMaxBECount, Step);
1164 Add = getAddExpr(Start, UMul);
1165 OperandExtendedAdd =
1166 getAddExpr(getSignExtendExpr(Start, WideTy),
1167 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1168 getZeroExtendExpr(Step, WideTy)));
1169 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1170 // Return the expression with the addrec on the outside.
1171 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1172 getZeroExtendExpr(Step, Ty),
1176 // If the backedge is guarded by a comparison with the pre-inc value
1177 // the addrec is safe. Also, if the entry is guarded by a comparison
1178 // with the start value and the backedge is guarded by a comparison
1179 // with the post-inc value, the addrec is safe.
1180 if (isKnownPositive(Step)) {
1181 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1182 getSignedRange(Step).getSignedMax());
1183 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1184 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1185 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1186 AR->getPostIncExpr(*this), N)))
1187 // Return the expression with the addrec on the outside.
1188 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1189 getSignExtendExpr(Step, Ty),
1191 } else if (isKnownNegative(Step)) {
1192 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1193 getSignedRange(Step).getSignedMin());
1194 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1195 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1196 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1197 AR->getPostIncExpr(*this), N)))
1198 // Return the expression with the addrec on the outside.
1199 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1200 getSignExtendExpr(Step, Ty),
1206 // The cast wasn't folded; create an explicit cast node.
1207 // Recompute the insert position, as it may have been invalidated.
1208 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1209 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1211 UniqueSCEVs.InsertNode(S, IP);
1215 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1216 /// unspecified bits out to the given type.
1218 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1220 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1221 "This is not an extending conversion!");
1222 assert(isSCEVable(Ty) &&
1223 "This is not a conversion to a SCEVable type!");
1224 Ty = getEffectiveSCEVType(Ty);
1226 // Sign-extend negative constants.
1227 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1228 if (SC->getValue()->getValue().isNegative())
1229 return getSignExtendExpr(Op, Ty);
1231 // Peel off a truncate cast.
1232 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1233 const SCEV *NewOp = T->getOperand();
1234 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1235 return getAnyExtendExpr(NewOp, Ty);
1236 return getTruncateOrNoop(NewOp, Ty);
1239 // Next try a zext cast. If the cast is folded, use it.
1240 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1241 if (!isa<SCEVZeroExtendExpr>(ZExt))
1244 // Next try a sext cast. If the cast is folded, use it.
1245 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1246 if (!isa<SCEVSignExtendExpr>(SExt))
1249 // Force the cast to be folded into the operands of an addrec.
1250 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1251 SmallVector<const SCEV *, 4> Ops;
1252 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1254 Ops.push_back(getAnyExtendExpr(*I, Ty));
1255 return getAddRecExpr(Ops, AR->getLoop());
1258 // As a special case, fold anyext(undef) to undef. We don't want to
1259 // know too much about SCEVUnknowns, but this special case is handy
1261 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
1262 if (isa<UndefValue>(U->getValue()))
1263 return getSCEV(UndefValue::get(Ty));
1265 // If the expression is obviously signed, use the sext cast value.
1266 if (isa<SCEVSMaxExpr>(Op))
1269 // Absent any other information, use the zext cast value.
1273 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1274 /// a list of operands to be added under the given scale, update the given
1275 /// map. This is a helper function for getAddRecExpr. As an example of
1276 /// what it does, given a sequence of operands that would form an add
1277 /// expression like this:
1279 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1281 /// where A and B are constants, update the map with these values:
1283 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1285 /// and add 13 + A*B*29 to AccumulatedConstant.
1286 /// This will allow getAddRecExpr to produce this:
1288 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1290 /// This form often exposes folding opportunities that are hidden in
1291 /// the original operand list.
1293 /// Return true iff it appears that any interesting folding opportunities
1294 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1295 /// the common case where no interesting opportunities are present, and
1296 /// is also used as a check to avoid infinite recursion.
1299 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1300 SmallVector<const SCEV *, 8> &NewOps,
1301 APInt &AccumulatedConstant,
1302 const SCEV *const *Ops, size_t NumOperands,
1304 ScalarEvolution &SE) {
1305 bool Interesting = false;
1307 // Iterate over the add operands. They are sorted, with constants first.
1309 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1311 // Pull a buried constant out to the outside.
1312 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1314 AccumulatedConstant += Scale * C->getValue()->getValue();
1317 // Next comes everything else. We're especially interested in multiplies
1318 // here, but they're in the middle, so just visit the rest with one loop.
1319 for (; i != NumOperands; ++i) {
1320 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1321 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1323 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1324 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1325 // A multiplication of a constant with another add; recurse.
1326 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1328 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1329 Add->op_begin(), Add->getNumOperands(),
1332 // A multiplication of a constant with some other value. Update
1334 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1335 const SCEV *Key = SE.getMulExpr(MulOps);
1336 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1337 M.insert(std::make_pair(Key, NewScale));
1339 NewOps.push_back(Pair.first->first);
1341 Pair.first->second += NewScale;
1342 // The map already had an entry for this value, which may indicate
1343 // a folding opportunity.
1348 // An ordinary operand. Update the map.
1349 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1350 M.insert(std::make_pair(Ops[i], Scale));
1352 NewOps.push_back(Pair.first->first);
1354 Pair.first->second += Scale;
1355 // The map already had an entry for this value, which may indicate
1356 // a folding opportunity.
1366 struct APIntCompare {
1367 bool operator()(const APInt &LHS, const APInt &RHS) const {
1368 return LHS.ult(RHS);
1373 /// getAddExpr - Get a canonical add expression, or something simpler if
1375 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1376 bool HasNUW, bool HasNSW) {
1377 assert(!Ops.empty() && "Cannot get empty add!");
1378 if (Ops.size() == 1) return Ops[0];
1380 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1381 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1382 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1383 "SCEVAddExpr operand types don't match!");
1386 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1387 if (!HasNUW && HasNSW) {
1389 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1390 E = Ops.end(); I != E; ++I)
1391 if (!isKnownNonNegative(*I)) {
1395 if (All) HasNUW = true;
1398 // Sort by complexity, this groups all similar expression types together.
1399 GroupByComplexity(Ops, LI);
1401 // If there are any constants, fold them together.
1403 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1405 assert(Idx < Ops.size());
1406 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1407 // We found two constants, fold them together!
1408 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1409 RHSC->getValue()->getValue());
1410 if (Ops.size() == 2) return Ops[0];
1411 Ops.erase(Ops.begin()+1); // Erase the folded element
1412 LHSC = cast<SCEVConstant>(Ops[0]);
1415 // If we are left with a constant zero being added, strip it off.
1416 if (LHSC->getValue()->isZero()) {
1417 Ops.erase(Ops.begin());
1421 if (Ops.size() == 1) return Ops[0];
1424 // Okay, check to see if the same value occurs in the operand list more than
1425 // once. If so, merge them together into an multiply expression. Since we
1426 // sorted the list, these values are required to be adjacent.
1427 const Type *Ty = Ops[0]->getType();
1428 bool FoundMatch = false;
1429 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1430 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1431 // Scan ahead to count how many equal operands there are.
1433 while (i+Count != e && Ops[i+Count] == Ops[i])
1435 // Merge the values into a multiply.
1436 const SCEV *Scale = getConstant(Ty, Count);
1437 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1438 if (Ops.size() == Count)
1441 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1442 --i; e -= Count - 1;
1446 return getAddExpr(Ops, HasNUW, HasNSW);
1448 // Check for truncates. If all the operands are truncated from the same
1449 // type, see if factoring out the truncate would permit the result to be
1450 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1451 // if the contents of the resulting outer trunc fold to something simple.
1452 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1453 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1454 const Type *DstType = Trunc->getType();
1455 const Type *SrcType = Trunc->getOperand()->getType();
1456 SmallVector<const SCEV *, 8> LargeOps;
1458 // Check all the operands to see if they can be represented in the
1459 // source type of the truncate.
1460 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1461 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1462 if (T->getOperand()->getType() != SrcType) {
1466 LargeOps.push_back(T->getOperand());
1467 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1468 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1469 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1470 SmallVector<const SCEV *, 8> LargeMulOps;
1471 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1472 if (const SCEVTruncateExpr *T =
1473 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1474 if (T->getOperand()->getType() != SrcType) {
1478 LargeMulOps.push_back(T->getOperand());
1479 } else if (const SCEVConstant *C =
1480 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1481 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1488 LargeOps.push_back(getMulExpr(LargeMulOps));
1495 // Evaluate the expression in the larger type.
1496 const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
1497 // If it folds to something simple, use it. Otherwise, don't.
1498 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1499 return getTruncateExpr(Fold, DstType);
1503 // Skip past any other cast SCEVs.
1504 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1507 // If there are add operands they would be next.
1508 if (Idx < Ops.size()) {
1509 bool DeletedAdd = false;
1510 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1511 // If we have an add, expand the add operands onto the end of the operands
1513 Ops.erase(Ops.begin()+Idx);
1514 Ops.append(Add->op_begin(), Add->op_end());
1518 // If we deleted at least one add, we added operands to the end of the list,
1519 // and they are not necessarily sorted. Recurse to resort and resimplify
1520 // any operands we just acquired.
1522 return getAddExpr(Ops);
1525 // Skip over the add expression until we get to a multiply.
1526 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1529 // Check to see if there are any folding opportunities present with
1530 // operands multiplied by constant values.
1531 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1532 uint64_t BitWidth = getTypeSizeInBits(Ty);
1533 DenseMap<const SCEV *, APInt> M;
1534 SmallVector<const SCEV *, 8> NewOps;
1535 APInt AccumulatedConstant(BitWidth, 0);
1536 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1537 Ops.data(), Ops.size(),
1538 APInt(BitWidth, 1), *this)) {
1539 // Some interesting folding opportunity is present, so its worthwhile to
1540 // re-generate the operands list. Group the operands by constant scale,
1541 // to avoid multiplying by the same constant scale multiple times.
1542 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1543 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1544 E = NewOps.end(); I != E; ++I)
1545 MulOpLists[M.find(*I)->second].push_back(*I);
1546 // Re-generate the operands list.
1548 if (AccumulatedConstant != 0)
1549 Ops.push_back(getConstant(AccumulatedConstant));
1550 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1551 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1553 Ops.push_back(getMulExpr(getConstant(I->first),
1554 getAddExpr(I->second)));
1556 return getConstant(Ty, 0);
1557 if (Ops.size() == 1)
1559 return getAddExpr(Ops);
1563 // If we are adding something to a multiply expression, make sure the
1564 // something is not already an operand of the multiply. If so, merge it into
1566 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1567 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1568 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1569 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1570 if (isa<SCEVConstant>(MulOpSCEV))
1572 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1573 if (MulOpSCEV == Ops[AddOp]) {
1574 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1575 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1576 if (Mul->getNumOperands() != 2) {
1577 // If the multiply has more than two operands, we must get the
1579 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1580 Mul->op_begin()+MulOp);
1581 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1582 InnerMul = getMulExpr(MulOps);
1584 const SCEV *One = getConstant(Ty, 1);
1585 const SCEV *AddOne = getAddExpr(One, InnerMul);
1586 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1587 if (Ops.size() == 2) return OuterMul;
1589 Ops.erase(Ops.begin()+AddOp);
1590 Ops.erase(Ops.begin()+Idx-1);
1592 Ops.erase(Ops.begin()+Idx);
1593 Ops.erase(Ops.begin()+AddOp-1);
1595 Ops.push_back(OuterMul);
1596 return getAddExpr(Ops);
1599 // Check this multiply against other multiplies being added together.
1600 for (unsigned OtherMulIdx = Idx+1;
1601 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1603 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1604 // If MulOp occurs in OtherMul, we can fold the two multiplies
1606 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1607 OMulOp != e; ++OMulOp)
1608 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1609 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1610 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1611 if (Mul->getNumOperands() != 2) {
1612 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1613 Mul->op_begin()+MulOp);
1614 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1615 InnerMul1 = getMulExpr(MulOps);
1617 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1618 if (OtherMul->getNumOperands() != 2) {
1619 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1620 OtherMul->op_begin()+OMulOp);
1621 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1622 InnerMul2 = getMulExpr(MulOps);
1624 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1625 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1626 if (Ops.size() == 2) return OuterMul;
1627 Ops.erase(Ops.begin()+Idx);
1628 Ops.erase(Ops.begin()+OtherMulIdx-1);
1629 Ops.push_back(OuterMul);
1630 return getAddExpr(Ops);
1636 // If there are any add recurrences in the operands list, see if any other
1637 // added values are loop invariant. If so, we can fold them into the
1639 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1642 // Scan over all recurrences, trying to fold loop invariants into them.
1643 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1644 // Scan all of the other operands to this add and add them to the vector if
1645 // they are loop invariant w.r.t. the recurrence.
1646 SmallVector<const SCEV *, 8> LIOps;
1647 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1648 const Loop *AddRecLoop = AddRec->getLoop();
1649 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1650 if (Ops[i]->isLoopInvariant(AddRecLoop)) {
1651 LIOps.push_back(Ops[i]);
1652 Ops.erase(Ops.begin()+i);
1656 // If we found some loop invariants, fold them into the recurrence.
1657 if (!LIOps.empty()) {
1658 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1659 LIOps.push_back(AddRec->getStart());
1661 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1663 AddRecOps[0] = getAddExpr(LIOps);
1665 // Build the new addrec. Propagate the NUW and NSW flags if both the
1666 // outer add and the inner addrec are guaranteed to have no overflow.
1667 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop,
1668 HasNUW && AddRec->hasNoUnsignedWrap(),
1669 HasNSW && AddRec->hasNoSignedWrap());
1671 // If all of the other operands were loop invariant, we are done.
1672 if (Ops.size() == 1) return NewRec;
1674 // Otherwise, add the folded AddRec by the non-liv parts.
1675 for (unsigned i = 0;; ++i)
1676 if (Ops[i] == AddRec) {
1680 return getAddExpr(Ops);
1683 // Okay, if there weren't any loop invariants to be folded, check to see if
1684 // there are multiple AddRec's with the same loop induction variable being
1685 // added together. If so, we can fold them.
1686 for (unsigned OtherIdx = Idx+1;
1687 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1689 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1690 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1691 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1693 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1695 if (const SCEVAddRecExpr *OtherAddRec =
1696 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1697 if (OtherAddRec->getLoop() == AddRecLoop) {
1698 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1700 if (i >= AddRecOps.size()) {
1701 AddRecOps.append(OtherAddRec->op_begin()+i,
1702 OtherAddRec->op_end());
1705 AddRecOps[i] = getAddExpr(AddRecOps[i],
1706 OtherAddRec->getOperand(i));
1708 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1710 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop);
1711 return getAddExpr(Ops);
1714 // Otherwise couldn't fold anything into this recurrence. Move onto the
1718 // Okay, it looks like we really DO need an add expr. Check to see if we
1719 // already have one, otherwise create a new one.
1720 FoldingSetNodeID ID;
1721 ID.AddInteger(scAddExpr);
1722 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1723 ID.AddPointer(Ops[i]);
1726 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1728 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1729 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1730 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1732 UniqueSCEVs.InsertNode(S, IP);
1734 if (HasNUW) S->setHasNoUnsignedWrap(true);
1735 if (HasNSW) S->setHasNoSignedWrap(true);
1739 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1741 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1742 bool HasNUW, bool HasNSW) {
1743 assert(!Ops.empty() && "Cannot get empty mul!");
1744 if (Ops.size() == 1) return Ops[0];
1746 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1747 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1748 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1749 "SCEVMulExpr operand types don't match!");
1752 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1753 if (!HasNUW && HasNSW) {
1755 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1756 E = Ops.end(); I != E; ++I)
1757 if (!isKnownNonNegative(*I)) {
1761 if (All) HasNUW = true;
1764 // Sort by complexity, this groups all similar expression types together.
1765 GroupByComplexity(Ops, LI);
1767 // If there are any constants, fold them together.
1769 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1771 // C1*(C2+V) -> C1*C2 + C1*V
1772 if (Ops.size() == 2)
1773 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1774 if (Add->getNumOperands() == 2 &&
1775 isa<SCEVConstant>(Add->getOperand(0)))
1776 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1777 getMulExpr(LHSC, Add->getOperand(1)));
1780 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1781 // We found two constants, fold them together!
1782 ConstantInt *Fold = ConstantInt::get(getContext(),
1783 LHSC->getValue()->getValue() *
1784 RHSC->getValue()->getValue());
1785 Ops[0] = getConstant(Fold);
1786 Ops.erase(Ops.begin()+1); // Erase the folded element
1787 if (Ops.size() == 1) return Ops[0];
1788 LHSC = cast<SCEVConstant>(Ops[0]);
1791 // If we are left with a constant one being multiplied, strip it off.
1792 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1793 Ops.erase(Ops.begin());
1795 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1796 // If we have a multiply of zero, it will always be zero.
1798 } else if (Ops[0]->isAllOnesValue()) {
1799 // If we have a mul by -1 of an add, try distributing the -1 among the
1801 if (Ops.size() == 2)
1802 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1803 SmallVector<const SCEV *, 4> NewOps;
1804 bool AnyFolded = false;
1805 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1807 const SCEV *Mul = getMulExpr(Ops[0], *I);
1808 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1809 NewOps.push_back(Mul);
1812 return getAddExpr(NewOps);
1816 if (Ops.size() == 1)
1820 // Skip over the add expression until we get to a multiply.
1821 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1824 // If there are mul operands inline them all into this expression.
1825 if (Idx < Ops.size()) {
1826 bool DeletedMul = false;
1827 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1828 // If we have an mul, expand the mul operands onto the end of the operands
1830 Ops.erase(Ops.begin()+Idx);
1831 Ops.append(Mul->op_begin(), Mul->op_end());
1835 // If we deleted at least one mul, we added operands to the end of the list,
1836 // and they are not necessarily sorted. Recurse to resort and resimplify
1837 // any operands we just acquired.
1839 return getMulExpr(Ops);
1842 // If there are any add recurrences in the operands list, see if any other
1843 // added values are loop invariant. If so, we can fold them into the
1845 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1848 // Scan over all recurrences, trying to fold loop invariants into them.
1849 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1850 // Scan all of the other operands to this mul and add them to the vector if
1851 // they are loop invariant w.r.t. the recurrence.
1852 SmallVector<const SCEV *, 8> LIOps;
1853 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1854 const Loop *AddRecLoop = AddRec->getLoop();
1855 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1856 if (Ops[i]->isLoopInvariant(AddRecLoop)) {
1857 LIOps.push_back(Ops[i]);
1858 Ops.erase(Ops.begin()+i);
1862 // If we found some loop invariants, fold them into the recurrence.
1863 if (!LIOps.empty()) {
1864 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1865 SmallVector<const SCEV *, 4> NewOps;
1866 NewOps.reserve(AddRec->getNumOperands());
1867 const SCEV *Scale = getMulExpr(LIOps);
1868 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1869 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1871 // Build the new addrec. Propagate the NUW and NSW flags if both the
1872 // outer mul and the inner addrec are guaranteed to have no overflow.
1873 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop,
1874 HasNUW && AddRec->hasNoUnsignedWrap(),
1875 HasNSW && AddRec->hasNoSignedWrap());
1877 // If all of the other operands were loop invariant, we are done.
1878 if (Ops.size() == 1) return NewRec;
1880 // Otherwise, multiply the folded AddRec by the non-liv parts.
1881 for (unsigned i = 0;; ++i)
1882 if (Ops[i] == AddRec) {
1886 return getMulExpr(Ops);
1889 // Okay, if there weren't any loop invariants to be folded, check to see if
1890 // there are multiple AddRec's with the same loop induction variable being
1891 // multiplied together. If so, we can fold them.
1892 for (unsigned OtherIdx = Idx+1;
1893 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1895 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1896 // F * G, where F = {A,+,B}<L> and G = {C,+,D}<L> -->
1897 // {A*C,+,F*D + G*B + B*D}<L>
1898 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1900 if (const SCEVAddRecExpr *OtherAddRec =
1901 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1902 if (OtherAddRec->getLoop() == AddRecLoop) {
1903 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1904 const SCEV *NewStart = getMulExpr(F->getStart(), G->getStart());
1905 const SCEV *B = F->getStepRecurrence(*this);
1906 const SCEV *D = G->getStepRecurrence(*this);
1907 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1910 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1912 if (Ops.size() == 2) return NewAddRec;
1913 Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
1914 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1916 return getMulExpr(Ops);
1919 // Otherwise couldn't fold anything into this recurrence. Move onto the
1923 // Okay, it looks like we really DO need an mul expr. Check to see if we
1924 // already have one, otherwise create a new one.
1925 FoldingSetNodeID ID;
1926 ID.AddInteger(scMulExpr);
1927 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1928 ID.AddPointer(Ops[i]);
1931 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1933 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1934 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1935 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
1937 UniqueSCEVs.InsertNode(S, IP);
1939 if (HasNUW) S->setHasNoUnsignedWrap(true);
1940 if (HasNSW) S->setHasNoSignedWrap(true);
1944 /// getUDivExpr - Get a canonical unsigned division expression, or something
1945 /// simpler if possible.
1946 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1948 assert(getEffectiveSCEVType(LHS->getType()) ==
1949 getEffectiveSCEVType(RHS->getType()) &&
1950 "SCEVUDivExpr operand types don't match!");
1952 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1953 if (RHSC->getValue()->equalsInt(1))
1954 return LHS; // X udiv 1 --> x
1955 // If the denominator is zero, the result of the udiv is undefined. Don't
1956 // try to analyze it, because the resolution chosen here may differ from
1957 // the resolution chosen in other parts of the compiler.
1958 if (!RHSC->getValue()->isZero()) {
1959 // Determine if the division can be folded into the operands of
1961 // TODO: Generalize this to non-constants by using known-bits information.
1962 const Type *Ty = LHS->getType();
1963 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1964 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
1965 // For non-power-of-two values, effectively round the value up to the
1966 // nearest power of two.
1967 if (!RHSC->getValue()->getValue().isPowerOf2())
1969 const IntegerType *ExtTy =
1970 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1971 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1972 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1973 if (const SCEVConstant *Step =
1974 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1975 if (!Step->getValue()->getValue()
1976 .urem(RHSC->getValue()->getValue()) &&
1977 getZeroExtendExpr(AR, ExtTy) ==
1978 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1979 getZeroExtendExpr(Step, ExtTy),
1981 SmallVector<const SCEV *, 4> Operands;
1982 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1983 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1984 return getAddRecExpr(Operands, AR->getLoop());
1986 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1987 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1988 SmallVector<const SCEV *, 4> Operands;
1989 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1990 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1991 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1992 // Find an operand that's safely divisible.
1993 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1994 const SCEV *Op = M->getOperand(i);
1995 const SCEV *Div = getUDivExpr(Op, RHSC);
1996 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1997 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
2000 return getMulExpr(Operands);
2004 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
2005 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
2006 SmallVector<const SCEV *, 4> Operands;
2007 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
2008 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
2009 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
2011 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2012 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
2013 if (isa<SCEVUDivExpr>(Op) ||
2014 getMulExpr(Op, RHS) != A->getOperand(i))
2016 Operands.push_back(Op);
2018 if (Operands.size() == A->getNumOperands())
2019 return getAddExpr(Operands);
2023 // Fold if both operands are constant.
2024 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
2025 Constant *LHSCV = LHSC->getValue();
2026 Constant *RHSCV = RHSC->getValue();
2027 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
2033 FoldingSetNodeID ID;
2034 ID.AddInteger(scUDivExpr);
2038 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2039 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
2041 UniqueSCEVs.InsertNode(S, IP);
2046 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2047 /// Simplify the expression as much as possible.
2048 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
2049 const SCEV *Step, const Loop *L,
2050 bool HasNUW, bool HasNSW) {
2051 SmallVector<const SCEV *, 4> Operands;
2052 Operands.push_back(Start);
2053 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
2054 if (StepChrec->getLoop() == L) {
2055 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
2056 return getAddRecExpr(Operands, L);
2059 Operands.push_back(Step);
2060 return getAddRecExpr(Operands, L, HasNUW, HasNSW);
2063 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2064 /// Simplify the expression as much as possible.
2066 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2068 bool HasNUW, bool HasNSW) {
2069 if (Operands.size() == 1) return Operands[0];
2071 const Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2072 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2073 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2074 "SCEVAddRecExpr operand types don't match!");
2077 if (Operands.back()->isZero()) {
2078 Operands.pop_back();
2079 return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
2082 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2083 // use that information to infer NUW and NSW flags. However, computing a
2084 // BE count requires calling getAddRecExpr, so we may not yet have a
2085 // meaningful BE count at this point (and if we don't, we'd be stuck
2086 // with a SCEVCouldNotCompute as the cached BE count).
2088 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
2089 if (!HasNUW && HasNSW) {
2091 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2092 E = Operands.end(); I != E; ++I)
2093 if (!isKnownNonNegative(*I)) {
2097 if (All) HasNUW = true;
2100 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2101 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2102 const Loop *NestedLoop = NestedAR->getLoop();
2103 if (L->contains(NestedLoop) ?
2104 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2105 (!NestedLoop->contains(L) &&
2106 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2107 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2108 NestedAR->op_end());
2109 Operands[0] = NestedAR->getStart();
2110 // AddRecs require their operands be loop-invariant with respect to their
2111 // loops. Don't perform this transformation if it would break this
2113 bool AllInvariant = true;
2114 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2115 if (!Operands[i]->isLoopInvariant(L)) {
2116 AllInvariant = false;
2120 NestedOperands[0] = getAddRecExpr(Operands, L);
2121 AllInvariant = true;
2122 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2123 if (!NestedOperands[i]->isLoopInvariant(NestedLoop)) {
2124 AllInvariant = false;
2128 // Ok, both add recurrences are valid after the transformation.
2129 return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
2131 // Reset Operands to its original state.
2132 Operands[0] = NestedAR;
2136 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2137 // already have one, otherwise create a new one.
2138 FoldingSetNodeID ID;
2139 ID.AddInteger(scAddRecExpr);
2140 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2141 ID.AddPointer(Operands[i]);
2145 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2147 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2148 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2149 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2150 O, Operands.size(), L);
2151 UniqueSCEVs.InsertNode(S, IP);
2153 if (HasNUW) S->setHasNoUnsignedWrap(true);
2154 if (HasNSW) S->setHasNoSignedWrap(true);
2158 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2160 SmallVector<const SCEV *, 2> Ops;
2163 return getSMaxExpr(Ops);
2167 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2168 assert(!Ops.empty() && "Cannot get empty smax!");
2169 if (Ops.size() == 1) return Ops[0];
2171 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2172 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2173 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2174 "SCEVSMaxExpr operand types don't match!");
2177 // Sort by complexity, this groups all similar expression types together.
2178 GroupByComplexity(Ops, LI);
2180 // If there are any constants, fold them together.
2182 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2184 assert(Idx < Ops.size());
2185 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2186 // We found two constants, fold them together!
2187 ConstantInt *Fold = ConstantInt::get(getContext(),
2188 APIntOps::smax(LHSC->getValue()->getValue(),
2189 RHSC->getValue()->getValue()));
2190 Ops[0] = getConstant(Fold);
2191 Ops.erase(Ops.begin()+1); // Erase the folded element
2192 if (Ops.size() == 1) return Ops[0];
2193 LHSC = cast<SCEVConstant>(Ops[0]);
2196 // If we are left with a constant minimum-int, strip it off.
2197 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2198 Ops.erase(Ops.begin());
2200 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2201 // If we have an smax with a constant maximum-int, it will always be
2206 if (Ops.size() == 1) return Ops[0];
2209 // Find the first SMax
2210 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2213 // Check to see if one of the operands is an SMax. If so, expand its operands
2214 // onto our operand list, and recurse to simplify.
2215 if (Idx < Ops.size()) {
2216 bool DeletedSMax = false;
2217 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2218 Ops.erase(Ops.begin()+Idx);
2219 Ops.append(SMax->op_begin(), SMax->op_end());
2224 return getSMaxExpr(Ops);
2227 // Okay, check to see if the same value occurs in the operand list twice. If
2228 // so, delete one. Since we sorted the list, these values are required to
2230 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2231 // X smax Y smax Y --> X smax Y
2232 // X smax Y --> X, if X is always greater than Y
2233 if (Ops[i] == Ops[i+1] ||
2234 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2235 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2237 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2238 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2242 if (Ops.size() == 1) return Ops[0];
2244 assert(!Ops.empty() && "Reduced smax down to nothing!");
2246 // Okay, it looks like we really DO need an smax expr. Check to see if we
2247 // already have one, otherwise create a new one.
2248 FoldingSetNodeID ID;
2249 ID.AddInteger(scSMaxExpr);
2250 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2251 ID.AddPointer(Ops[i]);
2253 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2254 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2255 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2256 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2258 UniqueSCEVs.InsertNode(S, IP);
2262 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2264 SmallVector<const SCEV *, 2> Ops;
2267 return getUMaxExpr(Ops);
2271 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2272 assert(!Ops.empty() && "Cannot get empty umax!");
2273 if (Ops.size() == 1) return Ops[0];
2275 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2276 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2277 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2278 "SCEVUMaxExpr operand types don't match!");
2281 // Sort by complexity, this groups all similar expression types together.
2282 GroupByComplexity(Ops, LI);
2284 // If there are any constants, fold them together.
2286 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2288 assert(Idx < Ops.size());
2289 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2290 // We found two constants, fold them together!
2291 ConstantInt *Fold = ConstantInt::get(getContext(),
2292 APIntOps::umax(LHSC->getValue()->getValue(),
2293 RHSC->getValue()->getValue()));
2294 Ops[0] = getConstant(Fold);
2295 Ops.erase(Ops.begin()+1); // Erase the folded element
2296 if (Ops.size() == 1) return Ops[0];
2297 LHSC = cast<SCEVConstant>(Ops[0]);
2300 // If we are left with a constant minimum-int, strip it off.
2301 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2302 Ops.erase(Ops.begin());
2304 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2305 // If we have an umax with a constant maximum-int, it will always be
2310 if (Ops.size() == 1) return Ops[0];
2313 // Find the first UMax
2314 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2317 // Check to see if one of the operands is a UMax. If so, expand its operands
2318 // onto our operand list, and recurse to simplify.
2319 if (Idx < Ops.size()) {
2320 bool DeletedUMax = false;
2321 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2322 Ops.erase(Ops.begin()+Idx);
2323 Ops.append(UMax->op_begin(), UMax->op_end());
2328 return getUMaxExpr(Ops);
2331 // Okay, check to see if the same value occurs in the operand list twice. If
2332 // so, delete one. Since we sorted the list, these values are required to
2334 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2335 // X umax Y umax Y --> X umax Y
2336 // X umax Y --> X, if X is always greater than Y
2337 if (Ops[i] == Ops[i+1] ||
2338 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2339 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2341 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2342 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2346 if (Ops.size() == 1) return Ops[0];
2348 assert(!Ops.empty() && "Reduced umax down to nothing!");
2350 // Okay, it looks like we really DO need a umax expr. Check to see if we
2351 // already have one, otherwise create a new one.
2352 FoldingSetNodeID ID;
2353 ID.AddInteger(scUMaxExpr);
2354 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2355 ID.AddPointer(Ops[i]);
2357 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2358 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2359 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2360 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2362 UniqueSCEVs.InsertNode(S, IP);
2366 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2368 // ~smax(~x, ~y) == smin(x, y).
2369 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2372 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2374 // ~umax(~x, ~y) == umin(x, y)
2375 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2378 const SCEV *ScalarEvolution::getSizeOfExpr(const Type *AllocTy) {
2379 // If we have TargetData, we can bypass creating a target-independent
2380 // constant expression and then folding it back into a ConstantInt.
2381 // This is just a compile-time optimization.
2383 return getConstant(TD->getIntPtrType(getContext()),
2384 TD->getTypeAllocSize(AllocTy));
2386 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2387 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2388 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2390 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2391 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2394 const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
2395 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2396 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2397 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2399 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2400 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2403 const SCEV *ScalarEvolution::getOffsetOfExpr(const StructType *STy,
2405 // If we have TargetData, we can bypass creating a target-independent
2406 // constant expression and then folding it back into a ConstantInt.
2407 // This is just a compile-time optimization.
2409 return getConstant(TD->getIntPtrType(getContext()),
2410 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2412 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2413 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2414 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2416 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2417 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2420 const SCEV *ScalarEvolution::getOffsetOfExpr(const Type *CTy,
2421 Constant *FieldNo) {
2422 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2423 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2424 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2426 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2427 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2430 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2431 // Don't attempt to do anything other than create a SCEVUnknown object
2432 // here. createSCEV only calls getUnknown after checking for all other
2433 // interesting possibilities, and any other code that calls getUnknown
2434 // is doing so in order to hide a value from SCEV canonicalization.
2436 FoldingSetNodeID ID;
2437 ID.AddInteger(scUnknown);
2440 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2441 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2442 "Stale SCEVUnknown in uniquing map!");
2445 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2447 FirstUnknown = cast<SCEVUnknown>(S);
2448 UniqueSCEVs.InsertNode(S, IP);
2452 //===----------------------------------------------------------------------===//
2453 // Basic SCEV Analysis and PHI Idiom Recognition Code
2456 /// isSCEVable - Test if values of the given type are analyzable within
2457 /// the SCEV framework. This primarily includes integer types, and it
2458 /// can optionally include pointer types if the ScalarEvolution class
2459 /// has access to target-specific information.
2460 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2461 // Integers and pointers are always SCEVable.
2462 return Ty->isIntegerTy() || Ty->isPointerTy();
2465 /// getTypeSizeInBits - Return the size in bits of the specified type,
2466 /// for which isSCEVable must return true.
2467 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2468 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2470 // If we have a TargetData, use it!
2472 return TD->getTypeSizeInBits(Ty);
2474 // Integer types have fixed sizes.
2475 if (Ty->isIntegerTy())
2476 return Ty->getPrimitiveSizeInBits();
2478 // The only other support type is pointer. Without TargetData, conservatively
2479 // assume pointers are 64-bit.
2480 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2484 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2485 /// the given type and which represents how SCEV will treat the given
2486 /// type, for which isSCEVable must return true. For pointer types,
2487 /// this is the pointer-sized integer type.
2488 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2489 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2491 if (Ty->isIntegerTy())
2494 // The only other support type is pointer.
2495 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2496 if (TD) return TD->getIntPtrType(getContext());
2498 // Without TargetData, conservatively assume pointers are 64-bit.
2499 return Type::getInt64Ty(getContext());
2502 const SCEV *ScalarEvolution::getCouldNotCompute() {
2503 return &CouldNotCompute;
2506 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2507 /// expression and create a new one.
2508 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2509 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2511 ValueExprMapType::const_iterator I = ValueExprMap.find(V);
2512 if (I != ValueExprMap.end()) return I->second;
2513 const SCEV *S = createSCEV(V);
2515 // The process of creating a SCEV for V may have caused other SCEVs
2516 // to have been created, so it's necessary to insert the new entry
2517 // from scratch, rather than trying to remember the insert position
2519 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2523 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2525 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2526 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2528 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2530 const Type *Ty = V->getType();
2531 Ty = getEffectiveSCEVType(Ty);
2532 return getMulExpr(V,
2533 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2536 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2537 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2538 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2540 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2542 const Type *Ty = V->getType();
2543 Ty = getEffectiveSCEVType(Ty);
2544 const SCEV *AllOnes =
2545 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2546 return getMinusSCEV(AllOnes, V);
2549 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
2551 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS,
2553 // Fast path: X - X --> 0.
2555 return getConstant(LHS->getType(), 0);
2558 return getAddExpr(LHS, getNegativeSCEV(RHS));
2561 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2562 /// input value to the specified type. If the type must be extended, it is zero
2565 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V,
2567 const Type *SrcTy = V->getType();
2568 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2569 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2570 "Cannot truncate or zero extend with non-integer arguments!");
2571 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2572 return V; // No conversion
2573 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2574 return getTruncateExpr(V, Ty);
2575 return getZeroExtendExpr(V, Ty);
2578 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2579 /// input value to the specified type. If the type must be extended, it is sign
2582 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2584 const Type *SrcTy = V->getType();
2585 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2586 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2587 "Cannot truncate or zero extend with non-integer arguments!");
2588 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2589 return V; // No conversion
2590 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2591 return getTruncateExpr(V, Ty);
2592 return getSignExtendExpr(V, Ty);
2595 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2596 /// input value to the specified type. If the type must be extended, it is zero
2597 /// extended. The conversion must not be narrowing.
2599 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2600 const Type *SrcTy = V->getType();
2601 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2602 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2603 "Cannot noop or zero extend with non-integer arguments!");
2604 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2605 "getNoopOrZeroExtend cannot truncate!");
2606 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2607 return V; // No conversion
2608 return getZeroExtendExpr(V, Ty);
2611 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2612 /// input value to the specified type. If the type must be extended, it is sign
2613 /// extended. The conversion must not be narrowing.
2615 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
2616 const Type *SrcTy = V->getType();
2617 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2618 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2619 "Cannot noop or sign extend with non-integer arguments!");
2620 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2621 "getNoopOrSignExtend cannot truncate!");
2622 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2623 return V; // No conversion
2624 return getSignExtendExpr(V, Ty);
2627 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2628 /// the input value to the specified type. If the type must be extended,
2629 /// it is extended with unspecified bits. The conversion must not be
2632 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2633 const Type *SrcTy = V->getType();
2634 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2635 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2636 "Cannot noop or any extend with non-integer arguments!");
2637 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2638 "getNoopOrAnyExtend cannot truncate!");
2639 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2640 return V; // No conversion
2641 return getAnyExtendExpr(V, Ty);
2644 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2645 /// input value to the specified type. The conversion must not be widening.
2647 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2648 const Type *SrcTy = V->getType();
2649 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2650 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2651 "Cannot truncate or noop with non-integer arguments!");
2652 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2653 "getTruncateOrNoop cannot extend!");
2654 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2655 return V; // No conversion
2656 return getTruncateExpr(V, Ty);
2659 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2660 /// the types using zero-extension, and then perform a umax operation
2662 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2664 const SCEV *PromotedLHS = LHS;
2665 const SCEV *PromotedRHS = RHS;
2667 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2668 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2670 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2672 return getUMaxExpr(PromotedLHS, PromotedRHS);
2675 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2676 /// the types using zero-extension, and then perform a umin operation
2678 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2680 const SCEV *PromotedLHS = LHS;
2681 const SCEV *PromotedRHS = RHS;
2683 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2684 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2686 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2688 return getUMinExpr(PromotedLHS, PromotedRHS);
2691 /// PushDefUseChildren - Push users of the given Instruction
2692 /// onto the given Worklist.
2694 PushDefUseChildren(Instruction *I,
2695 SmallVectorImpl<Instruction *> &Worklist) {
2696 // Push the def-use children onto the Worklist stack.
2697 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2699 Worklist.push_back(cast<Instruction>(*UI));
2702 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2703 /// instructions that depend on the given instruction and removes them from
2704 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2707 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2708 SmallVector<Instruction *, 16> Worklist;
2709 PushDefUseChildren(PN, Worklist);
2711 SmallPtrSet<Instruction *, 8> Visited;
2713 while (!Worklist.empty()) {
2714 Instruction *I = Worklist.pop_back_val();
2715 if (!Visited.insert(I)) continue;
2717 ValueExprMapType::iterator It =
2718 ValueExprMap.find(static_cast<Value *>(I));
2719 if (It != ValueExprMap.end()) {
2720 const SCEV *Old = It->second;
2722 // Short-circuit the def-use traversal if the symbolic name
2723 // ceases to appear in expressions.
2724 if (Old != SymName && !Old->hasOperand(SymName))
2727 // SCEVUnknown for a PHI either means that it has an unrecognized
2728 // structure, it's a PHI that's in the progress of being computed
2729 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2730 // additional loop trip count information isn't going to change anything.
2731 // In the second case, createNodeForPHI will perform the necessary
2732 // updates on its own when it gets to that point. In the third, we do
2733 // want to forget the SCEVUnknown.
2734 if (!isa<PHINode>(I) ||
2735 !isa<SCEVUnknown>(Old) ||
2736 (I != PN && Old == SymName)) {
2737 ValuesAtScopes.erase(Old);
2738 UnsignedRanges.erase(Old);
2739 SignedRanges.erase(Old);
2740 ValueExprMap.erase(It);
2744 PushDefUseChildren(I, Worklist);
2748 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2749 /// a loop header, making it a potential recurrence, or it doesn't.
2751 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2752 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2753 if (L->getHeader() == PN->getParent()) {
2754 // The loop may have multiple entrances or multiple exits; we can analyze
2755 // this phi as an addrec if it has a unique entry value and a unique
2757 Value *BEValueV = 0, *StartValueV = 0;
2758 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2759 Value *V = PN->getIncomingValue(i);
2760 if (L->contains(PN->getIncomingBlock(i))) {
2763 } else if (BEValueV != V) {
2767 } else if (!StartValueV) {
2769 } else if (StartValueV != V) {
2774 if (BEValueV && StartValueV) {
2775 // While we are analyzing this PHI node, handle its value symbolically.
2776 const SCEV *SymbolicName = getUnknown(PN);
2777 assert(ValueExprMap.find(PN) == ValueExprMap.end() &&
2778 "PHI node already processed?");
2779 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2781 // Using this symbolic name for the PHI, analyze the value coming around
2783 const SCEV *BEValue = getSCEV(BEValueV);
2785 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2786 // has a special value for the first iteration of the loop.
2788 // If the value coming around the backedge is an add with the symbolic
2789 // value we just inserted, then we found a simple induction variable!
2790 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2791 // If there is a single occurrence of the symbolic value, replace it
2792 // with a recurrence.
2793 unsigned FoundIndex = Add->getNumOperands();
2794 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2795 if (Add->getOperand(i) == SymbolicName)
2796 if (FoundIndex == e) {
2801 if (FoundIndex != Add->getNumOperands()) {
2802 // Create an add with everything but the specified operand.
2803 SmallVector<const SCEV *, 8> Ops;
2804 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2805 if (i != FoundIndex)
2806 Ops.push_back(Add->getOperand(i));
2807 const SCEV *Accum = getAddExpr(Ops);
2809 // This is not a valid addrec if the step amount is varying each
2810 // loop iteration, but is not itself an addrec in this loop.
2811 if (Accum->isLoopInvariant(L) ||
2812 (isa<SCEVAddRecExpr>(Accum) &&
2813 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2814 bool HasNUW = false;
2815 bool HasNSW = false;
2817 // If the increment doesn't overflow, then neither the addrec nor
2818 // the post-increment will overflow.
2819 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
2820 if (OBO->hasNoUnsignedWrap())
2822 if (OBO->hasNoSignedWrap())
2826 const SCEV *StartVal = getSCEV(StartValueV);
2827 const SCEV *PHISCEV =
2828 getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
2830 // Since the no-wrap flags are on the increment, they apply to the
2831 // post-incremented value as well.
2832 if (Accum->isLoopInvariant(L))
2833 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
2834 Accum, L, HasNUW, HasNSW);
2836 // Okay, for the entire analysis of this edge we assumed the PHI
2837 // to be symbolic. We now need to go back and purge all of the
2838 // entries for the scalars that use the symbolic expression.
2839 ForgetSymbolicName(PN, SymbolicName);
2840 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2844 } else if (const SCEVAddRecExpr *AddRec =
2845 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2846 // Otherwise, this could be a loop like this:
2847 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2848 // In this case, j = {1,+,1} and BEValue is j.
2849 // Because the other in-value of i (0) fits the evolution of BEValue
2850 // i really is an addrec evolution.
2851 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2852 const SCEV *StartVal = getSCEV(StartValueV);
2854 // If StartVal = j.start - j.stride, we can use StartVal as the
2855 // initial step of the addrec evolution.
2856 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2857 AddRec->getOperand(1))) {
2858 const SCEV *PHISCEV =
2859 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2861 // Okay, for the entire analysis of this edge we assumed the PHI
2862 // to be symbolic. We now need to go back and purge all of the
2863 // entries for the scalars that use the symbolic expression.
2864 ForgetSymbolicName(PN, SymbolicName);
2865 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2873 // If the PHI has a single incoming value, follow that value, unless the
2874 // PHI's incoming blocks are in a different loop, in which case doing so
2875 // risks breaking LCSSA form. Instcombine would normally zap these, but
2876 // it doesn't have DominatorTree information, so it may miss cases.
2877 if (Value *V = PN->hasConstantValue(DT)) {
2878 bool AllSameLoop = true;
2879 Loop *PNLoop = LI->getLoopFor(PN->getParent());
2880 for (size_t i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
2881 if (LI->getLoopFor(PN->getIncomingBlock(i)) != PNLoop) {
2882 AllSameLoop = false;
2889 // If it's not a loop phi, we can't handle it yet.
2890 return getUnknown(PN);
2893 /// createNodeForGEP - Expand GEP instructions into add and multiply
2894 /// operations. This allows them to be analyzed by regular SCEV code.
2896 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
2898 // Don't blindly transfer the inbounds flag from the GEP instruction to the
2899 // Add expression, because the Instruction may be guarded by control flow
2900 // and the no-overflow bits may not be valid for the expression in any
2903 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2904 Value *Base = GEP->getOperand(0);
2905 // Don't attempt to analyze GEPs over unsized objects.
2906 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2907 return getUnknown(GEP);
2908 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
2909 gep_type_iterator GTI = gep_type_begin(GEP);
2910 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
2914 // Compute the (potentially symbolic) offset in bytes for this index.
2915 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2916 // For a struct, add the member offset.
2917 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2918 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
2920 // Add the field offset to the running total offset.
2921 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
2923 // For an array, add the element offset, explicitly scaled.
2924 const SCEV *ElementSize = getSizeOfExpr(*GTI);
2925 const SCEV *IndexS = getSCEV(Index);
2926 // Getelementptr indices are signed.
2927 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
2929 // Multiply the index by the element size to compute the element offset.
2930 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize);
2932 // Add the element offset to the running total offset.
2933 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2937 // Get the SCEV for the GEP base.
2938 const SCEV *BaseS = getSCEV(Base);
2940 // Add the total offset from all the GEP indices to the base.
2941 return getAddExpr(BaseS, TotalOffset);
2944 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2945 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2946 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2947 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2949 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2950 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2951 return C->getValue()->getValue().countTrailingZeros();
2953 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2954 return std::min(GetMinTrailingZeros(T->getOperand()),
2955 (uint32_t)getTypeSizeInBits(T->getType()));
2957 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2958 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2959 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2960 getTypeSizeInBits(E->getType()) : OpRes;
2963 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2964 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2965 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2966 getTypeSizeInBits(E->getType()) : OpRes;
2969 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2970 // The result is the min of all operands results.
2971 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2972 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2973 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2977 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2978 // The result is the sum of all operands results.
2979 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2980 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2981 for (unsigned i = 1, e = M->getNumOperands();
2982 SumOpRes != BitWidth && i != e; ++i)
2983 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2988 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2989 // The result is the min of all operands results.
2990 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2991 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2992 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2996 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2997 // The result is the min of all operands results.
2998 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2999 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3000 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3004 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
3005 // The result is the min of all operands results.
3006 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3007 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3008 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3012 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3013 // For a SCEVUnknown, ask ValueTracking.
3014 unsigned BitWidth = getTypeSizeInBits(U->getType());
3015 APInt Mask = APInt::getAllOnesValue(BitWidth);
3016 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3017 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
3018 return Zeros.countTrailingOnes();
3025 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
3028 ScalarEvolution::getUnsignedRange(const SCEV *S) {
3029 // See if we've computed this range already.
3030 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
3031 if (I != UnsignedRanges.end())
3034 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3035 return UnsignedRanges[C] = ConstantRange(C->getValue()->getValue());
3037 unsigned BitWidth = getTypeSizeInBits(S->getType());
3038 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3040 // If the value has known zeros, the maximum unsigned value will have those
3041 // known zeros as well.
3042 uint32_t TZ = GetMinTrailingZeros(S);
3044 ConservativeResult =
3045 ConstantRange(APInt::getMinValue(BitWidth),
3046 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
3048 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3049 ConstantRange X = getUnsignedRange(Add->getOperand(0));
3050 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3051 X = X.add(getUnsignedRange(Add->getOperand(i)));
3052 return UnsignedRanges[Add] = ConservativeResult.intersectWith(X);
3055 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3056 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
3057 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3058 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
3059 return UnsignedRanges[Mul] = ConservativeResult.intersectWith(X);
3062 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3063 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
3064 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3065 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
3066 return UnsignedRanges[SMax] = ConservativeResult.intersectWith(X);
3069 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3070 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3071 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3072 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3073 return UnsignedRanges[UMax] = ConservativeResult.intersectWith(X);
3076 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3077 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3078 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3079 return UnsignedRanges[UDiv] = ConservativeResult.intersectWith(X.udiv(Y));
3082 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3083 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3084 return UnsignedRanges[ZExt] =
3085 ConservativeResult.intersectWith(X.zeroExtend(BitWidth));
3088 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3089 ConstantRange X = getUnsignedRange(SExt->getOperand());
3090 return UnsignedRanges[SExt] =
3091 ConservativeResult.intersectWith(X.signExtend(BitWidth));
3094 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3095 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3096 return UnsignedRanges[Trunc] =
3097 ConservativeResult.intersectWith(X.truncate(BitWidth));
3100 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3101 // If there's no unsigned wrap, the value will never be less than its
3103 if (AddRec->hasNoUnsignedWrap())
3104 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3105 if (!C->getValue()->isZero())
3106 ConservativeResult =
3107 ConservativeResult.intersectWith(
3108 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3110 // TODO: non-affine addrec
3111 if (AddRec->isAffine()) {
3112 const Type *Ty = AddRec->getType();
3113 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3114 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3115 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3116 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3118 const SCEV *Start = AddRec->getStart();
3119 const SCEV *Step = AddRec->getStepRecurrence(*this);
3121 ConstantRange StartRange = getUnsignedRange(Start);
3122 ConstantRange StepRange = getSignedRange(Step);
3123 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3124 ConstantRange EndRange =
3125 StartRange.add(MaxBECountRange.multiply(StepRange));
3127 // Check for overflow. This must be done with ConstantRange arithmetic
3128 // because we could be called from within the ScalarEvolution overflow
3130 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3131 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3132 ConstantRange ExtMaxBECountRange =
3133 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3134 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3135 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3137 return UnsignedRanges[AddRec] = ConservativeResult;
3139 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3140 EndRange.getUnsignedMin());
3141 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3142 EndRange.getUnsignedMax());
3143 if (Min.isMinValue() && Max.isMaxValue())
3144 return UnsignedRanges[AddRec] = ConservativeResult;
3145 return UnsignedRanges[AddRec] =
3146 ConservativeResult.intersectWith(ConstantRange(Min, Max+1));
3150 return UnsignedRanges[AddRec] = ConservativeResult;
3153 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3154 // For a SCEVUnknown, ask ValueTracking.
3155 APInt Mask = APInt::getAllOnesValue(BitWidth);
3156 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3157 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
3158 if (Ones == ~Zeros + 1)
3159 return UnsignedRanges[U] = ConservativeResult;
3160 return UnsignedRanges[U] =
3161 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1));
3164 return UnsignedRanges[S] = ConservativeResult;
3167 /// getSignedRange - Determine the signed range for a particular SCEV.
3170 ScalarEvolution::getSignedRange(const SCEV *S) {
3171 // See if we've computed this range already.
3172 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3173 if (I != SignedRanges.end())
3176 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3177 return SignedRanges[C] = ConstantRange(C->getValue()->getValue());
3179 unsigned BitWidth = getTypeSizeInBits(S->getType());
3180 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3182 // If the value has known zeros, the maximum signed value will have those
3183 // known zeros as well.
3184 uint32_t TZ = GetMinTrailingZeros(S);
3186 ConservativeResult =
3187 ConstantRange(APInt::getSignedMinValue(BitWidth),
3188 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3190 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3191 ConstantRange X = getSignedRange(Add->getOperand(0));
3192 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3193 X = X.add(getSignedRange(Add->getOperand(i)));
3194 return SignedRanges[Add] = ConservativeResult.intersectWith(X);
3197 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3198 ConstantRange X = getSignedRange(Mul->getOperand(0));
3199 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3200 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3201 return SignedRanges[Mul] = ConservativeResult.intersectWith(X);
3204 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3205 ConstantRange X = getSignedRange(SMax->getOperand(0));
3206 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3207 X = X.smax(getSignedRange(SMax->getOperand(i)));
3208 return SignedRanges[SMax] = ConservativeResult.intersectWith(X);
3211 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3212 ConstantRange X = getSignedRange(UMax->getOperand(0));
3213 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3214 X = X.umax(getSignedRange(UMax->getOperand(i)));
3215 return SignedRanges[UMax] = ConservativeResult.intersectWith(X);
3218 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3219 ConstantRange X = getSignedRange(UDiv->getLHS());
3220 ConstantRange Y = getSignedRange(UDiv->getRHS());
3221 return SignedRanges[UDiv] = ConservativeResult.intersectWith(X.udiv(Y));
3224 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3225 ConstantRange X = getSignedRange(ZExt->getOperand());
3226 return SignedRanges[ZExt] =
3227 ConservativeResult.intersectWith(X.zeroExtend(BitWidth));
3230 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3231 ConstantRange X = getSignedRange(SExt->getOperand());
3232 return SignedRanges[SExt] =
3233 ConservativeResult.intersectWith(X.signExtend(BitWidth));
3236 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3237 ConstantRange X = getSignedRange(Trunc->getOperand());
3238 return SignedRanges[Trunc] =
3239 ConservativeResult.intersectWith(X.truncate(BitWidth));
3242 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3243 // If there's no signed wrap, and all the operands have the same sign or
3244 // zero, the value won't ever change sign.
3245 if (AddRec->hasNoSignedWrap()) {
3246 bool AllNonNeg = true;
3247 bool AllNonPos = true;
3248 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3249 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3250 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3253 ConservativeResult = ConservativeResult.intersectWith(
3254 ConstantRange(APInt(BitWidth, 0),
3255 APInt::getSignedMinValue(BitWidth)));
3257 ConservativeResult = ConservativeResult.intersectWith(
3258 ConstantRange(APInt::getSignedMinValue(BitWidth),
3259 APInt(BitWidth, 1)));
3262 // TODO: non-affine addrec
3263 if (AddRec->isAffine()) {
3264 const Type *Ty = AddRec->getType();
3265 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3266 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3267 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3268 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3270 const SCEV *Start = AddRec->getStart();
3271 const SCEV *Step = AddRec->getStepRecurrence(*this);
3273 ConstantRange StartRange = getSignedRange(Start);
3274 ConstantRange StepRange = getSignedRange(Step);
3275 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3276 ConstantRange EndRange =
3277 StartRange.add(MaxBECountRange.multiply(StepRange));
3279 // Check for overflow. This must be done with ConstantRange arithmetic
3280 // because we could be called from within the ScalarEvolution overflow
3282 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3283 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3284 ConstantRange ExtMaxBECountRange =
3285 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3286 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3287 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3289 return SignedRanges[AddRec] = ConservativeResult;
3291 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3292 EndRange.getSignedMin());
3293 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3294 EndRange.getSignedMax());
3295 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3296 return SignedRanges[AddRec] = ConservativeResult;
3297 return SignedRanges[AddRec] =
3298 ConservativeResult.intersectWith(ConstantRange(Min, Max+1));
3302 return SignedRanges[AddRec] = ConservativeResult;
3305 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3306 // For a SCEVUnknown, ask ValueTracking.
3307 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3308 return SignedRanges[U] = ConservativeResult;
3309 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3311 return SignedRanges[U] = ConservativeResult;
3312 return SignedRanges[U] = ConservativeResult.intersectWith(
3313 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3314 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1));
3317 return SignedRanges[S] = ConservativeResult;
3320 /// createSCEV - We know that there is no SCEV for the specified value.
3321 /// Analyze the expression.
3323 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3324 if (!isSCEVable(V->getType()))
3325 return getUnknown(V);
3327 unsigned Opcode = Instruction::UserOp1;
3328 if (Instruction *I = dyn_cast<Instruction>(V)) {
3329 Opcode = I->getOpcode();
3331 // Don't attempt to analyze instructions in blocks that aren't
3332 // reachable. Such instructions don't matter, and they aren't required
3333 // to obey basic rules for definitions dominating uses which this
3334 // analysis depends on.
3335 if (!DT->isReachableFromEntry(I->getParent()))
3336 return getUnknown(V);
3337 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3338 Opcode = CE->getOpcode();
3339 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3340 return getConstant(CI);
3341 else if (isa<ConstantPointerNull>(V))
3342 return getConstant(V->getType(), 0);
3343 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3344 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3346 return getUnknown(V);
3348 Operator *U = cast<Operator>(V);
3350 case Instruction::Add: {
3351 // The simple thing to do would be to just call getSCEV on both operands
3352 // and call getAddExpr with the result. However if we're looking at a
3353 // bunch of things all added together, this can be quite inefficient,
3354 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3355 // Instead, gather up all the operands and make a single getAddExpr call.
3356 // LLVM IR canonical form means we need only traverse the left operands.
3357 SmallVector<const SCEV *, 4> AddOps;
3358 AddOps.push_back(getSCEV(U->getOperand(1)));
3359 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3360 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3361 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3363 U = cast<Operator>(Op);
3364 const SCEV *Op1 = getSCEV(U->getOperand(1));
3365 if (Opcode == Instruction::Sub)
3366 AddOps.push_back(getNegativeSCEV(Op1));
3368 AddOps.push_back(Op1);
3370 AddOps.push_back(getSCEV(U->getOperand(0)));
3371 return getAddExpr(AddOps);
3373 case Instruction::Mul: {
3374 // See the Add code above.
3375 SmallVector<const SCEV *, 4> MulOps;
3376 MulOps.push_back(getSCEV(U->getOperand(1)));
3377 for (Value *Op = U->getOperand(0);
3378 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3379 Op = U->getOperand(0)) {
3380 U = cast<Operator>(Op);
3381 MulOps.push_back(getSCEV(U->getOperand(1)));
3383 MulOps.push_back(getSCEV(U->getOperand(0)));
3384 return getMulExpr(MulOps);
3386 case Instruction::UDiv:
3387 return getUDivExpr(getSCEV(U->getOperand(0)),
3388 getSCEV(U->getOperand(1)));
3389 case Instruction::Sub:
3390 return getMinusSCEV(getSCEV(U->getOperand(0)),
3391 getSCEV(U->getOperand(1)));
3392 case Instruction::And:
3393 // For an expression like x&255 that merely masks off the high bits,
3394 // use zext(trunc(x)) as the SCEV expression.
3395 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3396 if (CI->isNullValue())
3397 return getSCEV(U->getOperand(1));
3398 if (CI->isAllOnesValue())
3399 return getSCEV(U->getOperand(0));
3400 const APInt &A = CI->getValue();
3402 // Instcombine's ShrinkDemandedConstant may strip bits out of
3403 // constants, obscuring what would otherwise be a low-bits mask.
3404 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3405 // knew about to reconstruct a low-bits mask value.
3406 unsigned LZ = A.countLeadingZeros();
3407 unsigned BitWidth = A.getBitWidth();
3408 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
3409 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3410 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
3412 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3414 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3416 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3417 IntegerType::get(getContext(), BitWidth - LZ)),
3422 case Instruction::Or:
3423 // If the RHS of the Or is a constant, we may have something like:
3424 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3425 // optimizations will transparently handle this case.
3427 // In order for this transformation to be safe, the LHS must be of the
3428 // form X*(2^n) and the Or constant must be less than 2^n.
3429 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3430 const SCEV *LHS = getSCEV(U->getOperand(0));
3431 const APInt &CIVal = CI->getValue();
3432 if (GetMinTrailingZeros(LHS) >=
3433 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3434 // Build a plain add SCEV.
3435 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3436 // If the LHS of the add was an addrec and it has no-wrap flags,
3437 // transfer the no-wrap flags, since an or won't introduce a wrap.
3438 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3439 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3440 if (OldAR->hasNoUnsignedWrap())
3441 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3442 if (OldAR->hasNoSignedWrap())
3443 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3449 case Instruction::Xor:
3450 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3451 // If the RHS of the xor is a signbit, then this is just an add.
3452 // Instcombine turns add of signbit into xor as a strength reduction step.
3453 if (CI->getValue().isSignBit())
3454 return getAddExpr(getSCEV(U->getOperand(0)),
3455 getSCEV(U->getOperand(1)));
3457 // If the RHS of xor is -1, then this is a not operation.
3458 if (CI->isAllOnesValue())
3459 return getNotSCEV(getSCEV(U->getOperand(0)));
3461 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3462 // This is a variant of the check for xor with -1, and it handles
3463 // the case where instcombine has trimmed non-demanded bits out
3464 // of an xor with -1.
3465 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3466 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3467 if (BO->getOpcode() == Instruction::And &&
3468 LCI->getValue() == CI->getValue())
3469 if (const SCEVZeroExtendExpr *Z =
3470 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3471 const Type *UTy = U->getType();
3472 const SCEV *Z0 = Z->getOperand();
3473 const Type *Z0Ty = Z0->getType();
3474 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3476 // If C is a low-bits mask, the zero extend is serving to
3477 // mask off the high bits. Complement the operand and
3478 // re-apply the zext.
3479 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3480 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3482 // If C is a single bit, it may be in the sign-bit position
3483 // before the zero-extend. In this case, represent the xor
3484 // using an add, which is equivalent, and re-apply the zext.
3485 APInt Trunc = APInt(CI->getValue()).trunc(Z0TySize);
3486 if (APInt(Trunc).zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3488 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3494 case Instruction::Shl:
3495 // Turn shift left of a constant amount into a multiply.
3496 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3497 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3499 // If the shift count is not less than the bitwidth, the result of
3500 // the shift is undefined. Don't try to analyze it, because the
3501 // resolution chosen here may differ from the resolution chosen in
3502 // other parts of the compiler.
3503 if (SA->getValue().uge(BitWidth))
3506 Constant *X = ConstantInt::get(getContext(),
3507 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3508 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3512 case Instruction::LShr:
3513 // Turn logical shift right of a constant into a unsigned divide.
3514 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3515 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3517 // If the shift count is not less than the bitwidth, the result of
3518 // the shift is undefined. Don't try to analyze it, because the
3519 // resolution chosen here may differ from the resolution chosen in
3520 // other parts of the compiler.
3521 if (SA->getValue().uge(BitWidth))
3524 Constant *X = ConstantInt::get(getContext(),
3525 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3526 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3530 case Instruction::AShr:
3531 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3532 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3533 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3534 if (L->getOpcode() == Instruction::Shl &&
3535 L->getOperand(1) == U->getOperand(1)) {
3536 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3538 // If the shift count is not less than the bitwidth, the result of
3539 // the shift is undefined. Don't try to analyze it, because the
3540 // resolution chosen here may differ from the resolution chosen in
3541 // other parts of the compiler.
3542 if (CI->getValue().uge(BitWidth))
3545 uint64_t Amt = BitWidth - CI->getZExtValue();
3546 if (Amt == BitWidth)
3547 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3549 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3550 IntegerType::get(getContext(),
3556 case Instruction::Trunc:
3557 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3559 case Instruction::ZExt:
3560 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3562 case Instruction::SExt:
3563 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3565 case Instruction::BitCast:
3566 // BitCasts are no-op casts so we just eliminate the cast.
3567 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3568 return getSCEV(U->getOperand(0));
3571 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3572 // lead to pointer expressions which cannot safely be expanded to GEPs,
3573 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3574 // simplifying integer expressions.
3576 case Instruction::GetElementPtr:
3577 return createNodeForGEP(cast<GEPOperator>(U));
3579 case Instruction::PHI:
3580 return createNodeForPHI(cast<PHINode>(U));
3582 case Instruction::Select:
3583 // This could be a smax or umax that was lowered earlier.
3584 // Try to recover it.
3585 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3586 Value *LHS = ICI->getOperand(0);
3587 Value *RHS = ICI->getOperand(1);
3588 switch (ICI->getPredicate()) {
3589 case ICmpInst::ICMP_SLT:
3590 case ICmpInst::ICMP_SLE:
3591 std::swap(LHS, RHS);
3593 case ICmpInst::ICMP_SGT:
3594 case ICmpInst::ICMP_SGE:
3595 // a >s b ? a+x : b+x -> smax(a, b)+x
3596 // a >s b ? b+x : a+x -> smin(a, b)+x
3597 if (LHS->getType() == U->getType()) {
3598 const SCEV *LS = getSCEV(LHS);
3599 const SCEV *RS = getSCEV(RHS);
3600 const SCEV *LA = getSCEV(U->getOperand(1));
3601 const SCEV *RA = getSCEV(U->getOperand(2));
3602 const SCEV *LDiff = getMinusSCEV(LA, LS);
3603 const SCEV *RDiff = getMinusSCEV(RA, RS);
3605 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3606 LDiff = getMinusSCEV(LA, RS);
3607 RDiff = getMinusSCEV(RA, LS);
3609 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3612 case ICmpInst::ICMP_ULT:
3613 case ICmpInst::ICMP_ULE:
3614 std::swap(LHS, RHS);
3616 case ICmpInst::ICMP_UGT:
3617 case ICmpInst::ICMP_UGE:
3618 // a >u b ? a+x : b+x -> umax(a, b)+x
3619 // a >u b ? b+x : a+x -> umin(a, b)+x
3620 if (LHS->getType() == U->getType()) {
3621 const SCEV *LS = getSCEV(LHS);
3622 const SCEV *RS = getSCEV(RHS);
3623 const SCEV *LA = getSCEV(U->getOperand(1));
3624 const SCEV *RA = getSCEV(U->getOperand(2));
3625 const SCEV *LDiff = getMinusSCEV(LA, LS);
3626 const SCEV *RDiff = getMinusSCEV(RA, RS);
3628 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3629 LDiff = getMinusSCEV(LA, RS);
3630 RDiff = getMinusSCEV(RA, LS);
3632 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3635 case ICmpInst::ICMP_NE:
3636 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3637 if (LHS->getType() == U->getType() &&
3638 isa<ConstantInt>(RHS) &&
3639 cast<ConstantInt>(RHS)->isZero()) {
3640 const SCEV *One = getConstant(LHS->getType(), 1);
3641 const SCEV *LS = getSCEV(LHS);
3642 const SCEV *LA = getSCEV(U->getOperand(1));
3643 const SCEV *RA = getSCEV(U->getOperand(2));
3644 const SCEV *LDiff = getMinusSCEV(LA, LS);
3645 const SCEV *RDiff = getMinusSCEV(RA, One);
3647 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3650 case ICmpInst::ICMP_EQ:
3651 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3652 if (LHS->getType() == U->getType() &&
3653 isa<ConstantInt>(RHS) &&
3654 cast<ConstantInt>(RHS)->isZero()) {
3655 const SCEV *One = getConstant(LHS->getType(), 1);
3656 const SCEV *LS = getSCEV(LHS);
3657 const SCEV *LA = getSCEV(U->getOperand(1));
3658 const SCEV *RA = getSCEV(U->getOperand(2));
3659 const SCEV *LDiff = getMinusSCEV(LA, One);
3660 const SCEV *RDiff = getMinusSCEV(RA, LS);
3662 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3670 default: // We cannot analyze this expression.
3674 return getUnknown(V);
3679 //===----------------------------------------------------------------------===//
3680 // Iteration Count Computation Code
3683 /// getBackedgeTakenCount - If the specified loop has a predictable
3684 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3685 /// object. The backedge-taken count is the number of times the loop header
3686 /// will be branched to from within the loop. This is one less than the
3687 /// trip count of the loop, since it doesn't count the first iteration,
3688 /// when the header is branched to from outside the loop.
3690 /// Note that it is not valid to call this method on a loop without a
3691 /// loop-invariant backedge-taken count (see
3692 /// hasLoopInvariantBackedgeTakenCount).
3694 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3695 return getBackedgeTakenInfo(L).Exact;
3698 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3699 /// return the least SCEV value that is known never to be less than the
3700 /// actual backedge taken count.
3701 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3702 return getBackedgeTakenInfo(L).Max;
3705 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3706 /// onto the given Worklist.
3708 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3709 BasicBlock *Header = L->getHeader();
3711 // Push all Loop-header PHIs onto the Worklist stack.
3712 for (BasicBlock::iterator I = Header->begin();
3713 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3714 Worklist.push_back(PN);
3717 const ScalarEvolution::BackedgeTakenInfo &
3718 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3719 // Initially insert a CouldNotCompute for this loop. If the insertion
3720 // succeeds, proceed to actually compute a backedge-taken count and
3721 // update the value. The temporary CouldNotCompute value tells SCEV
3722 // code elsewhere that it shouldn't attempt to request a new
3723 // backedge-taken count, which could result in infinite recursion.
3724 std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3725 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3727 BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
3728 if (BECount.Exact != getCouldNotCompute()) {
3729 assert(BECount.Exact->isLoopInvariant(L) &&
3730 BECount.Max->isLoopInvariant(L) &&
3731 "Computed backedge-taken count isn't loop invariant for loop!");
3732 ++NumTripCountsComputed;
3734 // Update the value in the map.
3735 Pair.first->second = BECount;
3737 if (BECount.Max != getCouldNotCompute())
3738 // Update the value in the map.
3739 Pair.first->second = BECount;
3740 if (isa<PHINode>(L->getHeader()->begin()))
3741 // Only count loops that have phi nodes as not being computable.
3742 ++NumTripCountsNotComputed;
3745 // Now that we know more about the trip count for this loop, forget any
3746 // existing SCEV values for PHI nodes in this loop since they are only
3747 // conservative estimates made without the benefit of trip count
3748 // information. This is similar to the code in forgetLoop, except that
3749 // it handles SCEVUnknown PHI nodes specially.
3750 if (BECount.hasAnyInfo()) {
3751 SmallVector<Instruction *, 16> Worklist;
3752 PushLoopPHIs(L, Worklist);
3754 SmallPtrSet<Instruction *, 8> Visited;
3755 while (!Worklist.empty()) {
3756 Instruction *I = Worklist.pop_back_val();
3757 if (!Visited.insert(I)) continue;
3759 ValueExprMapType::iterator It =
3760 ValueExprMap.find(static_cast<Value *>(I));
3761 if (It != ValueExprMap.end()) {
3762 const SCEV *Old = It->second;
3764 // SCEVUnknown for a PHI either means that it has an unrecognized
3765 // structure, or it's a PHI that's in the progress of being computed
3766 // by createNodeForPHI. In the former case, additional loop trip
3767 // count information isn't going to change anything. In the later
3768 // case, createNodeForPHI will perform the necessary updates on its
3769 // own when it gets to that point.
3770 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
3771 ValuesAtScopes.erase(Old);
3772 UnsignedRanges.erase(Old);
3773 SignedRanges.erase(Old);
3774 ValueExprMap.erase(It);
3776 if (PHINode *PN = dyn_cast<PHINode>(I))
3777 ConstantEvolutionLoopExitValue.erase(PN);
3780 PushDefUseChildren(I, Worklist);
3784 return Pair.first->second;
3787 /// forgetLoop - This method should be called by the client when it has
3788 /// changed a loop in a way that may effect ScalarEvolution's ability to
3789 /// compute a trip count, or if the loop is deleted.
3790 void ScalarEvolution::forgetLoop(const Loop *L) {
3791 // Drop any stored trip count value.
3792 BackedgeTakenCounts.erase(L);
3794 // Drop information about expressions based on loop-header PHIs.
3795 SmallVector<Instruction *, 16> Worklist;
3796 PushLoopPHIs(L, Worklist);
3798 SmallPtrSet<Instruction *, 8> Visited;
3799 while (!Worklist.empty()) {
3800 Instruction *I = Worklist.pop_back_val();
3801 if (!Visited.insert(I)) continue;
3803 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3804 if (It != ValueExprMap.end()) {
3805 const SCEV *Old = It->second;
3806 ValuesAtScopes.erase(Old);
3807 UnsignedRanges.erase(Old);
3808 SignedRanges.erase(Old);
3809 ValueExprMap.erase(It);
3810 if (PHINode *PN = dyn_cast<PHINode>(I))
3811 ConstantEvolutionLoopExitValue.erase(PN);
3814 PushDefUseChildren(I, Worklist);
3817 // Forget all contained loops too, to avoid dangling entries in the
3818 // ValuesAtScopes map.
3819 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
3823 /// forgetValue - This method should be called by the client when it has
3824 /// changed a value in a way that may effect its value, or which may
3825 /// disconnect it from a def-use chain linking it to a loop.
3826 void ScalarEvolution::forgetValue(Value *V) {
3827 Instruction *I = dyn_cast<Instruction>(V);
3830 // Drop information about expressions based on loop-header PHIs.
3831 SmallVector<Instruction *, 16> Worklist;
3832 Worklist.push_back(I);
3834 SmallPtrSet<Instruction *, 8> Visited;
3835 while (!Worklist.empty()) {
3836 I = Worklist.pop_back_val();
3837 if (!Visited.insert(I)) continue;
3839 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3840 if (It != ValueExprMap.end()) {
3841 const SCEV *Old = It->second;
3842 ValuesAtScopes.erase(Old);
3843 UnsignedRanges.erase(Old);
3844 SignedRanges.erase(Old);
3845 ValueExprMap.erase(It);
3846 if (PHINode *PN = dyn_cast<PHINode>(I))
3847 ConstantEvolutionLoopExitValue.erase(PN);
3850 PushDefUseChildren(I, Worklist);
3854 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3855 /// of the specified loop will execute.
3856 ScalarEvolution::BackedgeTakenInfo
3857 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3858 SmallVector<BasicBlock *, 8> ExitingBlocks;
3859 L->getExitingBlocks(ExitingBlocks);
3861 // Examine all exits and pick the most conservative values.
3862 const SCEV *BECount = getCouldNotCompute();
3863 const SCEV *MaxBECount = getCouldNotCompute();
3864 bool CouldNotComputeBECount = false;
3865 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3866 BackedgeTakenInfo NewBTI =
3867 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3869 if (NewBTI.Exact == getCouldNotCompute()) {
3870 // We couldn't compute an exact value for this exit, so
3871 // we won't be able to compute an exact value for the loop.
3872 CouldNotComputeBECount = true;
3873 BECount = getCouldNotCompute();
3874 } else if (!CouldNotComputeBECount) {
3875 if (BECount == getCouldNotCompute())
3876 BECount = NewBTI.Exact;
3878 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3880 if (MaxBECount == getCouldNotCompute())
3881 MaxBECount = NewBTI.Max;
3882 else if (NewBTI.Max != getCouldNotCompute())
3883 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3886 return BackedgeTakenInfo(BECount, MaxBECount);
3889 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3890 /// of the specified loop will execute if it exits via the specified block.
3891 ScalarEvolution::BackedgeTakenInfo
3892 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3893 BasicBlock *ExitingBlock) {
3895 // Okay, we've chosen an exiting block. See what condition causes us to
3896 // exit at this block.
3898 // FIXME: we should be able to handle switch instructions (with a single exit)
3899 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3900 if (ExitBr == 0) return getCouldNotCompute();
3901 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3903 // At this point, we know we have a conditional branch that determines whether
3904 // the loop is exited. However, we don't know if the branch is executed each
3905 // time through the loop. If not, then the execution count of the branch will
3906 // not be equal to the trip count of the loop.
3908 // Currently we check for this by checking to see if the Exit branch goes to
3909 // the loop header. If so, we know it will always execute the same number of
3910 // times as the loop. We also handle the case where the exit block *is* the
3911 // loop header. This is common for un-rotated loops.
3913 // If both of those tests fail, walk up the unique predecessor chain to the
3914 // header, stopping if there is an edge that doesn't exit the loop. If the
3915 // header is reached, the execution count of the branch will be equal to the
3916 // trip count of the loop.
3918 // More extensive analysis could be done to handle more cases here.
3920 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3921 ExitBr->getSuccessor(1) != L->getHeader() &&
3922 ExitBr->getParent() != L->getHeader()) {
3923 // The simple checks failed, try climbing the unique predecessor chain
3924 // up to the header.
3926 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3927 BasicBlock *Pred = BB->getUniquePredecessor();
3929 return getCouldNotCompute();
3930 TerminatorInst *PredTerm = Pred->getTerminator();
3931 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3932 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3935 // If the predecessor has a successor that isn't BB and isn't
3936 // outside the loop, assume the worst.
3937 if (L->contains(PredSucc))
3938 return getCouldNotCompute();
3940 if (Pred == L->getHeader()) {
3947 return getCouldNotCompute();
3950 // Proceed to the next level to examine the exit condition expression.
3951 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3952 ExitBr->getSuccessor(0),
3953 ExitBr->getSuccessor(1));
3956 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3957 /// backedge of the specified loop will execute if its exit condition
3958 /// were a conditional branch of ExitCond, TBB, and FBB.
3959 ScalarEvolution::BackedgeTakenInfo
3960 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3964 // Check if the controlling expression for this loop is an And or Or.
3965 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3966 if (BO->getOpcode() == Instruction::And) {
3967 // Recurse on the operands of the and.
3968 BackedgeTakenInfo BTI0 =
3969 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3970 BackedgeTakenInfo BTI1 =
3971 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3972 const SCEV *BECount = getCouldNotCompute();
3973 const SCEV *MaxBECount = getCouldNotCompute();
3974 if (L->contains(TBB)) {
3975 // Both conditions must be true for the loop to continue executing.
3976 // Choose the less conservative count.
3977 if (BTI0.Exact == getCouldNotCompute() ||
3978 BTI1.Exact == getCouldNotCompute())
3979 BECount = getCouldNotCompute();
3981 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3982 if (BTI0.Max == getCouldNotCompute())
3983 MaxBECount = BTI1.Max;
3984 else if (BTI1.Max == getCouldNotCompute())
3985 MaxBECount = BTI0.Max;
3987 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3989 // Both conditions must be true at the same time for the loop to exit.
3990 // For now, be conservative.
3991 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3992 if (BTI0.Max == BTI1.Max)
3993 MaxBECount = BTI0.Max;
3994 if (BTI0.Exact == BTI1.Exact)
3995 BECount = BTI0.Exact;
3998 return BackedgeTakenInfo(BECount, MaxBECount);
4000 if (BO->getOpcode() == Instruction::Or) {
4001 // Recurse on the operands of the or.
4002 BackedgeTakenInfo BTI0 =
4003 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
4004 BackedgeTakenInfo BTI1 =
4005 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
4006 const SCEV *BECount = getCouldNotCompute();
4007 const SCEV *MaxBECount = getCouldNotCompute();
4008 if (L->contains(FBB)) {
4009 // Both conditions must be false for the loop to continue executing.
4010 // Choose the less conservative count.
4011 if (BTI0.Exact == getCouldNotCompute() ||
4012 BTI1.Exact == getCouldNotCompute())
4013 BECount = getCouldNotCompute();
4015 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
4016 if (BTI0.Max == getCouldNotCompute())
4017 MaxBECount = BTI1.Max;
4018 else if (BTI1.Max == getCouldNotCompute())
4019 MaxBECount = BTI0.Max;
4021 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
4023 // Both conditions must be false at the same time for the loop to exit.
4024 // For now, be conservative.
4025 assert(L->contains(TBB) && "Loop block has no successor in loop!");
4026 if (BTI0.Max == BTI1.Max)
4027 MaxBECount = BTI0.Max;
4028 if (BTI0.Exact == BTI1.Exact)
4029 BECount = BTI0.Exact;
4032 return BackedgeTakenInfo(BECount, MaxBECount);
4036 // With an icmp, it may be feasible to compute an exact backedge-taken count.
4037 // Proceed to the next level to examine the icmp.
4038 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
4039 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
4041 // Check for a constant condition. These are normally stripped out by
4042 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
4043 // preserve the CFG and is temporarily leaving constant conditions
4045 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
4046 if (L->contains(FBB) == !CI->getZExtValue())
4047 // The backedge is always taken.
4048 return getCouldNotCompute();
4050 // The backedge is never taken.
4051 return getConstant(CI->getType(), 0);
4054 // If it's not an integer or pointer comparison then compute it the hard way.
4055 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
4058 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
4059 /// backedge of the specified loop will execute if its exit condition
4060 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4061 ScalarEvolution::BackedgeTakenInfo
4062 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
4067 // If the condition was exit on true, convert the condition to exit on false
4068 ICmpInst::Predicate Cond;
4069 if (!L->contains(FBB))
4070 Cond = ExitCond->getPredicate();
4072 Cond = ExitCond->getInversePredicate();
4074 // Handle common loops like: for (X = "string"; *X; ++X)
4075 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4076 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4077 BackedgeTakenInfo ItCnt =
4078 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
4079 if (ItCnt.hasAnyInfo())
4083 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4084 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4086 // Try to evaluate any dependencies out of the loop.
4087 LHS = getSCEVAtScope(LHS, L);
4088 RHS = getSCEVAtScope(RHS, L);
4090 // At this point, we would like to compute how many iterations of the
4091 // loop the predicate will return true for these inputs.
4092 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
4093 // If there is a loop-invariant, force it into the RHS.
4094 std::swap(LHS, RHS);
4095 Cond = ICmpInst::getSwappedPredicate(Cond);
4098 // Simplify the operands before analyzing them.
4099 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4101 // If we have a comparison of a chrec against a constant, try to use value
4102 // ranges to answer this query.
4103 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4104 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4105 if (AddRec->getLoop() == L) {
4106 // Form the constant range.
4107 ConstantRange CompRange(
4108 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4110 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4111 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4115 case ICmpInst::ICMP_NE: { // while (X != Y)
4116 // Convert to: while (X-Y != 0)
4117 BackedgeTakenInfo BTI = HowFarToZero(getMinusSCEV(LHS, RHS), L);
4118 if (BTI.hasAnyInfo()) return BTI;
4121 case ICmpInst::ICMP_EQ: { // while (X == Y)
4122 // Convert to: while (X-Y == 0)
4123 BackedgeTakenInfo BTI = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4124 if (BTI.hasAnyInfo()) return BTI;
4127 case ICmpInst::ICMP_SLT: {
4128 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
4129 if (BTI.hasAnyInfo()) return BTI;
4132 case ICmpInst::ICMP_SGT: {
4133 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4134 getNotSCEV(RHS), L, true);
4135 if (BTI.hasAnyInfo()) return BTI;
4138 case ICmpInst::ICMP_ULT: {
4139 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
4140 if (BTI.hasAnyInfo()) return BTI;
4143 case ICmpInst::ICMP_UGT: {
4144 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4145 getNotSCEV(RHS), L, false);
4146 if (BTI.hasAnyInfo()) return BTI;
4151 dbgs() << "ComputeBackedgeTakenCount ";
4152 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4153 dbgs() << "[unsigned] ";
4154 dbgs() << *LHS << " "
4155 << Instruction::getOpcodeName(Instruction::ICmp)
4156 << " " << *RHS << "\n";
4161 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
4164 static ConstantInt *
4165 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4166 ScalarEvolution &SE) {
4167 const SCEV *InVal = SE.getConstant(C);
4168 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4169 assert(isa<SCEVConstant>(Val) &&
4170 "Evaluation of SCEV at constant didn't fold correctly?");
4171 return cast<SCEVConstant>(Val)->getValue();
4174 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
4175 /// and a GEP expression (missing the pointer index) indexing into it, return
4176 /// the addressed element of the initializer or null if the index expression is
4179 GetAddressedElementFromGlobal(GlobalVariable *GV,
4180 const std::vector<ConstantInt*> &Indices) {
4181 Constant *Init = GV->getInitializer();
4182 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
4183 uint64_t Idx = Indices[i]->getZExtValue();
4184 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
4185 assert(Idx < CS->getNumOperands() && "Bad struct index!");
4186 Init = cast<Constant>(CS->getOperand(Idx));
4187 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
4188 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
4189 Init = cast<Constant>(CA->getOperand(Idx));
4190 } else if (isa<ConstantAggregateZero>(Init)) {
4191 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
4192 assert(Idx < STy->getNumElements() && "Bad struct index!");
4193 Init = Constant::getNullValue(STy->getElementType(Idx));
4194 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
4195 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
4196 Init = Constant::getNullValue(ATy->getElementType());
4198 llvm_unreachable("Unknown constant aggregate type!");
4202 return 0; // Unknown initializer type
4208 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
4209 /// 'icmp op load X, cst', try to see if we can compute the backedge
4210 /// execution count.
4211 ScalarEvolution::BackedgeTakenInfo
4212 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
4216 ICmpInst::Predicate predicate) {
4217 if (LI->isVolatile()) return getCouldNotCompute();
4219 // Check to see if the loaded pointer is a getelementptr of a global.
4220 // TODO: Use SCEV instead of manually grubbing with GEPs.
4221 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4222 if (!GEP) return getCouldNotCompute();
4224 // Make sure that it is really a constant global we are gepping, with an
4225 // initializer, and make sure the first IDX is really 0.
4226 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4227 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4228 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4229 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4230 return getCouldNotCompute();
4232 // Okay, we allow one non-constant index into the GEP instruction.
4234 std::vector<ConstantInt*> Indexes;
4235 unsigned VarIdxNum = 0;
4236 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4237 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4238 Indexes.push_back(CI);
4239 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4240 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4241 VarIdx = GEP->getOperand(i);
4243 Indexes.push_back(0);
4246 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4247 // Check to see if X is a loop variant variable value now.
4248 const SCEV *Idx = getSCEV(VarIdx);
4249 Idx = getSCEVAtScope(Idx, L);
4251 // We can only recognize very limited forms of loop index expressions, in
4252 // particular, only affine AddRec's like {C1,+,C2}.
4253 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4254 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
4255 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4256 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4257 return getCouldNotCompute();
4259 unsigned MaxSteps = MaxBruteForceIterations;
4260 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4261 ConstantInt *ItCst = ConstantInt::get(
4262 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4263 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4265 // Form the GEP offset.
4266 Indexes[VarIdxNum] = Val;
4268 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
4269 if (Result == 0) break; // Cannot compute!
4271 // Evaluate the condition for this iteration.
4272 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4273 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4274 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4276 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4277 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4280 ++NumArrayLenItCounts;
4281 return getConstant(ItCst); // Found terminating iteration!
4284 return getCouldNotCompute();
4288 /// CanConstantFold - Return true if we can constant fold an instruction of the
4289 /// specified type, assuming that all operands were constants.
4290 static bool CanConstantFold(const Instruction *I) {
4291 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4292 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
4295 if (const CallInst *CI = dyn_cast<CallInst>(I))
4296 if (const Function *F = CI->getCalledFunction())
4297 return canConstantFoldCallTo(F);
4301 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4302 /// in the loop that V is derived from. We allow arbitrary operations along the
4303 /// way, but the operands of an operation must either be constants or a value
4304 /// derived from a constant PHI. If this expression does not fit with these
4305 /// constraints, return null.
4306 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4307 // If this is not an instruction, or if this is an instruction outside of the
4308 // loop, it can't be derived from a loop PHI.
4309 Instruction *I = dyn_cast<Instruction>(V);
4310 if (I == 0 || !L->contains(I)) return 0;
4312 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4313 if (L->getHeader() == I->getParent())
4316 // We don't currently keep track of the control flow needed to evaluate
4317 // PHIs, so we cannot handle PHIs inside of loops.
4321 // If we won't be able to constant fold this expression even if the operands
4322 // are constants, return early.
4323 if (!CanConstantFold(I)) return 0;
4325 // Otherwise, we can evaluate this instruction if all of its operands are
4326 // constant or derived from a PHI node themselves.
4328 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
4329 if (!isa<Constant>(I->getOperand(Op))) {
4330 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
4331 if (P == 0) return 0; // Not evolving from PHI
4335 return 0; // Evolving from multiple different PHIs.
4338 // This is a expression evolving from a constant PHI!
4342 /// EvaluateExpression - Given an expression that passes the
4343 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4344 /// in the loop has the value PHIVal. If we can't fold this expression for some
4345 /// reason, return null.
4346 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
4347 const TargetData *TD) {
4348 if (isa<PHINode>(V)) return PHIVal;
4349 if (Constant *C = dyn_cast<Constant>(V)) return C;
4350 Instruction *I = cast<Instruction>(V);
4352 std::vector<Constant*> Operands(I->getNumOperands());
4354 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4355 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
4356 if (Operands[i] == 0) return 0;
4359 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4360 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4362 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4363 &Operands[0], Operands.size(), TD);
4366 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4367 /// in the header of its containing loop, we know the loop executes a
4368 /// constant number of times, and the PHI node is just a recurrence
4369 /// involving constants, fold it.
4371 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4374 std::map<PHINode*, Constant*>::const_iterator I =
4375 ConstantEvolutionLoopExitValue.find(PN);
4376 if (I != ConstantEvolutionLoopExitValue.end())
4379 if (BEs.ugt(MaxBruteForceIterations))
4380 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4382 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4384 // Since the loop is canonicalized, the PHI node must have two entries. One
4385 // entry must be a constant (coming in from outside of the loop), and the
4386 // second must be derived from the same PHI.
4387 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4388 Constant *StartCST =
4389 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4391 return RetVal = 0; // Must be a constant.
4393 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4394 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4395 !isa<Constant>(BEValue))
4396 return RetVal = 0; // Not derived from same PHI.
4398 // Execute the loop symbolically to determine the exit value.
4399 if (BEs.getActiveBits() >= 32)
4400 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4402 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4403 unsigned IterationNum = 0;
4404 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
4405 if (IterationNum == NumIterations)
4406 return RetVal = PHIVal; // Got exit value!
4408 // Compute the value of the PHI node for the next iteration.
4409 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4410 if (NextPHI == PHIVal)
4411 return RetVal = NextPHI; // Stopped evolving!
4413 return 0; // Couldn't evaluate!
4418 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
4419 /// constant number of times (the condition evolves only from constants),
4420 /// try to evaluate a few iterations of the loop until we get the exit
4421 /// condition gets a value of ExitWhen (true or false). If we cannot
4422 /// evaluate the trip count of the loop, return getCouldNotCompute().
4424 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
4427 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4428 if (PN == 0) return getCouldNotCompute();
4430 // If the loop is canonicalized, the PHI will have exactly two entries.
4431 // That's the only form we support here.
4432 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4434 // One entry must be a constant (coming in from outside of the loop), and the
4435 // second must be derived from the same PHI.
4436 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4437 Constant *StartCST =
4438 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4439 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
4441 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4442 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4443 !isa<Constant>(BEValue))
4444 return getCouldNotCompute(); // Not derived from same PHI.
4446 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4447 // the loop symbolically to determine when the condition gets a value of
4449 unsigned IterationNum = 0;
4450 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4451 for (Constant *PHIVal = StartCST;
4452 IterationNum != MaxIterations; ++IterationNum) {
4453 ConstantInt *CondVal =
4454 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
4456 // Couldn't symbolically evaluate.
4457 if (!CondVal) return getCouldNotCompute();
4459 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4460 ++NumBruteForceTripCountsComputed;
4461 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4464 // Compute the value of the PHI node for the next iteration.
4465 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4466 if (NextPHI == 0 || NextPHI == PHIVal)
4467 return getCouldNotCompute();// Couldn't evaluate or not making progress...
4471 // Too many iterations were needed to evaluate.
4472 return getCouldNotCompute();
4475 /// getSCEVAtScope - Return a SCEV expression for the specified value
4476 /// at the specified scope in the program. The L value specifies a loop
4477 /// nest to evaluate the expression at, where null is the top-level or a
4478 /// specified loop is immediately inside of the loop.
4480 /// This method can be used to compute the exit value for a variable defined
4481 /// in a loop by querying what the value will hold in the parent loop.
4483 /// In the case that a relevant loop exit value cannot be computed, the
4484 /// original value V is returned.
4485 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4486 // Check to see if we've folded this expression at this loop before.
4487 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4488 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4489 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4491 return Pair.first->second ? Pair.first->second : V;
4493 // Otherwise compute it.
4494 const SCEV *C = computeSCEVAtScope(V, L);
4495 ValuesAtScopes[V][L] = C;
4499 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
4500 if (isa<SCEVConstant>(V)) return V;
4502 // If this instruction is evolved from a constant-evolving PHI, compute the
4503 // exit value from the loop without using SCEVs.
4504 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
4505 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
4506 const Loop *LI = (*this->LI)[I->getParent()];
4507 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
4508 if (PHINode *PN = dyn_cast<PHINode>(I))
4509 if (PN->getParent() == LI->getHeader()) {
4510 // Okay, there is no closed form solution for the PHI node. Check
4511 // to see if the loop that contains it has a known backedge-taken
4512 // count. If so, we may be able to force computation of the exit
4514 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
4515 if (const SCEVConstant *BTCC =
4516 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
4517 // Okay, we know how many times the containing loop executes. If
4518 // this is a constant evolving PHI node, get the final value at
4519 // the specified iteration number.
4520 Constant *RV = getConstantEvolutionLoopExitValue(PN,
4521 BTCC->getValue()->getValue(),
4523 if (RV) return getSCEV(RV);
4527 // Okay, this is an expression that we cannot symbolically evaluate
4528 // into a SCEV. Check to see if it's possible to symbolically evaluate
4529 // the arguments into constants, and if so, try to constant propagate the
4530 // result. This is particularly useful for computing loop exit values.
4531 if (CanConstantFold(I)) {
4532 SmallVector<Constant *, 4> Operands;
4533 bool MadeImprovement = false;
4534 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4535 Value *Op = I->getOperand(i);
4536 if (Constant *C = dyn_cast<Constant>(Op)) {
4537 Operands.push_back(C);
4541 // If any of the operands is non-constant and if they are
4542 // non-integer and non-pointer, don't even try to analyze them
4543 // with scev techniques.
4544 if (!isSCEVable(Op->getType()))
4547 const SCEV *OrigV = getSCEV(Op);
4548 const SCEV *OpV = getSCEVAtScope(OrigV, L);
4549 MadeImprovement |= OrigV != OpV;
4552 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
4554 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV))
4555 C = dyn_cast<Constant>(SU->getValue());
4557 if (C->getType() != Op->getType())
4558 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4562 Operands.push_back(C);
4565 // Check to see if getSCEVAtScope actually made an improvement.
4566 if (MadeImprovement) {
4568 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4569 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4570 Operands[0], Operands[1], TD);
4572 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4573 &Operands[0], Operands.size(), TD);
4580 // This is some other type of SCEVUnknown, just return it.
4584 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4585 // Avoid performing the look-up in the common case where the specified
4586 // expression has no loop-variant portions.
4587 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4588 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4589 if (OpAtScope != Comm->getOperand(i)) {
4590 // Okay, at least one of these operands is loop variant but might be
4591 // foldable. Build a new instance of the folded commutative expression.
4592 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4593 Comm->op_begin()+i);
4594 NewOps.push_back(OpAtScope);
4596 for (++i; i != e; ++i) {
4597 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4598 NewOps.push_back(OpAtScope);
4600 if (isa<SCEVAddExpr>(Comm))
4601 return getAddExpr(NewOps);
4602 if (isa<SCEVMulExpr>(Comm))
4603 return getMulExpr(NewOps);
4604 if (isa<SCEVSMaxExpr>(Comm))
4605 return getSMaxExpr(NewOps);
4606 if (isa<SCEVUMaxExpr>(Comm))
4607 return getUMaxExpr(NewOps);
4608 llvm_unreachable("Unknown commutative SCEV type!");
4611 // If we got here, all operands are loop invariant.
4615 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4616 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4617 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4618 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4619 return Div; // must be loop invariant
4620 return getUDivExpr(LHS, RHS);
4623 // If this is a loop recurrence for a loop that does not contain L, then we
4624 // are dealing with the final value computed by the loop.
4625 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4626 // First, attempt to evaluate each operand.
4627 // Avoid performing the look-up in the common case where the specified
4628 // expression has no loop-variant portions.
4629 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
4630 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
4631 if (OpAtScope == AddRec->getOperand(i))
4634 // Okay, at least one of these operands is loop variant but might be
4635 // foldable. Build a new instance of the folded commutative expression.
4636 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
4637 AddRec->op_begin()+i);
4638 NewOps.push_back(OpAtScope);
4639 for (++i; i != e; ++i)
4640 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
4642 AddRec = cast<SCEVAddRecExpr>(getAddRecExpr(NewOps, AddRec->getLoop()));
4646 // If the scope is outside the addrec's loop, evaluate it by using the
4647 // loop exit value of the addrec.
4648 if (!AddRec->getLoop()->contains(L)) {
4649 // To evaluate this recurrence, we need to know how many times the AddRec
4650 // loop iterates. Compute this now.
4651 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4652 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4654 // Then, evaluate the AddRec.
4655 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4661 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4662 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4663 if (Op == Cast->getOperand())
4664 return Cast; // must be loop invariant
4665 return getZeroExtendExpr(Op, Cast->getType());
4668 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4669 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4670 if (Op == Cast->getOperand())
4671 return Cast; // must be loop invariant
4672 return getSignExtendExpr(Op, Cast->getType());
4675 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4676 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4677 if (Op == Cast->getOperand())
4678 return Cast; // must be loop invariant
4679 return getTruncateExpr(Op, Cast->getType());
4682 llvm_unreachable("Unknown SCEV type!");
4686 /// getSCEVAtScope - This is a convenience function which does
4687 /// getSCEVAtScope(getSCEV(V), L).
4688 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4689 return getSCEVAtScope(getSCEV(V), L);
4692 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4693 /// following equation:
4695 /// A * X = B (mod N)
4697 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4698 /// A and B isn't important.
4700 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4701 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4702 ScalarEvolution &SE) {
4703 uint32_t BW = A.getBitWidth();
4704 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4705 assert(A != 0 && "A must be non-zero.");
4709 // The gcd of A and N may have only one prime factor: 2. The number of
4710 // trailing zeros in A is its multiplicity
4711 uint32_t Mult2 = A.countTrailingZeros();
4714 // 2. Check if B is divisible by D.
4716 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4717 // is not less than multiplicity of this prime factor for D.
4718 if (B.countTrailingZeros() < Mult2)
4719 return SE.getCouldNotCompute();
4721 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4724 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4725 // bit width during computations.
4726 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4727 APInt Mod(BW + 1, 0);
4728 Mod.set(BW - Mult2); // Mod = N / D
4729 APInt I = AD.multiplicativeInverse(Mod);
4731 // 4. Compute the minimum unsigned root of the equation:
4732 // I * (B / D) mod (N / D)
4733 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4735 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4737 return SE.getConstant(Result.trunc(BW));
4740 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4741 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4742 /// might be the same) or two SCEVCouldNotCompute objects.
4744 static std::pair<const SCEV *,const SCEV *>
4745 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4746 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4747 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4748 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4749 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4751 // We currently can only solve this if the coefficients are constants.
4752 if (!LC || !MC || !NC) {
4753 const SCEV *CNC = SE.getCouldNotCompute();
4754 return std::make_pair(CNC, CNC);
4757 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4758 const APInt &L = LC->getValue()->getValue();
4759 const APInt &M = MC->getValue()->getValue();
4760 const APInt &N = NC->getValue()->getValue();
4761 APInt Two(BitWidth, 2);
4762 APInt Four(BitWidth, 4);
4765 using namespace APIntOps;
4767 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4768 // The B coefficient is M-N/2
4772 // The A coefficient is N/2
4773 APInt A(N.sdiv(Two));
4775 // Compute the B^2-4ac term.
4778 SqrtTerm -= Four * (A * C);
4780 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4781 // integer value or else APInt::sqrt() will assert.
4782 APInt SqrtVal(SqrtTerm.sqrt());
4784 // Compute the two solutions for the quadratic formula.
4785 // The divisions must be performed as signed divisions.
4787 APInt TwoA( A << 1 );
4788 if (TwoA.isMinValue()) {
4789 const SCEV *CNC = SE.getCouldNotCompute();
4790 return std::make_pair(CNC, CNC);
4793 LLVMContext &Context = SE.getContext();
4795 ConstantInt *Solution1 =
4796 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4797 ConstantInt *Solution2 =
4798 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4800 return std::make_pair(SE.getConstant(Solution1),
4801 SE.getConstant(Solution2));
4802 } // end APIntOps namespace
4805 /// HowFarToZero - Return the number of times a backedge comparing the specified
4806 /// value to zero will execute. If not computable, return CouldNotCompute.
4807 ScalarEvolution::BackedgeTakenInfo
4808 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4809 // If the value is a constant
4810 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4811 // If the value is already zero, the branch will execute zero times.
4812 if (C->getValue()->isZero()) return C;
4813 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4816 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4817 if (!AddRec || AddRec->getLoop() != L)
4818 return getCouldNotCompute();
4820 if (AddRec->isAffine()) {
4821 // If this is an affine expression, the execution count of this branch is
4822 // the minimum unsigned root of the following equation:
4824 // Start + Step*N = 0 (mod 2^BW)
4828 // Step*N = -Start (mod 2^BW)
4830 // where BW is the common bit width of Start and Step.
4832 // Get the initial value for the loop.
4833 const SCEV *Start = getSCEVAtScope(AddRec->getStart(),
4834 L->getParentLoop());
4835 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1),
4836 L->getParentLoop());
4838 if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
4839 // For now we handle only constant steps.
4841 // First, handle unitary steps.
4842 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4843 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4844 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4845 return Start; // N = Start (as unsigned)
4847 // Then, try to solve the above equation provided that Start is constant.
4848 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4849 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4850 -StartC->getValue()->getValue(),
4853 } else if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
4854 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4855 // the quadratic equation to solve it.
4856 std::pair<const SCEV *,const SCEV *> Roots = SolveQuadraticEquation(AddRec,
4858 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4859 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4862 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
4863 << " sol#2: " << *R2 << "\n";
4865 // Pick the smallest positive root value.
4866 if (ConstantInt *CB =
4867 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
4868 R1->getValue(), R2->getValue()))) {
4869 if (CB->getZExtValue() == false)
4870 std::swap(R1, R2); // R1 is the minimum root now.
4872 // We can only use this value if the chrec ends up with an exact zero
4873 // value at this index. When solving for "X*X != 5", for example, we
4874 // should not accept a root of 2.
4875 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4877 return R1; // We found a quadratic root!
4882 return getCouldNotCompute();
4885 /// HowFarToNonZero - Return the number of times a backedge checking the
4886 /// specified value for nonzero will execute. If not computable, return
4888 ScalarEvolution::BackedgeTakenInfo
4889 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4890 // Loops that look like: while (X == 0) are very strange indeed. We don't
4891 // handle them yet except for the trivial case. This could be expanded in the
4892 // future as needed.
4894 // If the value is a constant, check to see if it is known to be non-zero
4895 // already. If so, the backedge will execute zero times.
4896 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4897 if (!C->getValue()->isNullValue())
4898 return getConstant(C->getType(), 0);
4899 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4902 // We could implement others, but I really doubt anyone writes loops like
4903 // this, and if they did, they would already be constant folded.
4904 return getCouldNotCompute();
4907 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4908 /// (which may not be an immediate predecessor) which has exactly one
4909 /// successor from which BB is reachable, or null if no such block is
4912 std::pair<BasicBlock *, BasicBlock *>
4913 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4914 // If the block has a unique predecessor, then there is no path from the
4915 // predecessor to the block that does not go through the direct edge
4916 // from the predecessor to the block.
4917 if (BasicBlock *Pred = BB->getSinglePredecessor())
4918 return std::make_pair(Pred, BB);
4920 // A loop's header is defined to be a block that dominates the loop.
4921 // If the header has a unique predecessor outside the loop, it must be
4922 // a block that has exactly one successor that can reach the loop.
4923 if (Loop *L = LI->getLoopFor(BB))
4924 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
4926 return std::pair<BasicBlock *, BasicBlock *>();
4929 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4930 /// testing whether two expressions are equal, however for the purposes of
4931 /// looking for a condition guarding a loop, it can be useful to be a little
4932 /// more general, since a front-end may have replicated the controlling
4935 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4936 // Quick check to see if they are the same SCEV.
4937 if (A == B) return true;
4939 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4940 // two different instructions with the same value. Check for this case.
4941 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4942 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4943 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4944 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4945 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4948 // Otherwise assume they may have a different value.
4952 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
4953 /// predicate Pred. Return true iff any changes were made.
4955 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
4956 const SCEV *&LHS, const SCEV *&RHS) {
4957 bool Changed = false;
4959 // Canonicalize a constant to the right side.
4960 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
4961 // Check for both operands constant.
4962 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
4963 if (ConstantExpr::getICmp(Pred,
4965 RHSC->getValue())->isNullValue())
4966 goto trivially_false;
4968 goto trivially_true;
4970 // Otherwise swap the operands to put the constant on the right.
4971 std::swap(LHS, RHS);
4972 Pred = ICmpInst::getSwappedPredicate(Pred);
4976 // If we're comparing an addrec with a value which is loop-invariant in the
4977 // addrec's loop, put the addrec on the left. Also make a dominance check,
4978 // as both operands could be addrecs loop-invariant in each other's loop.
4979 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
4980 const Loop *L = AR->getLoop();
4981 if (LHS->isLoopInvariant(L) && LHS->properlyDominates(L->getHeader(), DT)) {
4982 std::swap(LHS, RHS);
4983 Pred = ICmpInst::getSwappedPredicate(Pred);
4988 // If there's a constant operand, canonicalize comparisons with boundary
4989 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
4990 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
4991 const APInt &RA = RC->getValue()->getValue();
4993 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4994 case ICmpInst::ICMP_EQ:
4995 case ICmpInst::ICMP_NE:
4997 case ICmpInst::ICMP_UGE:
4998 if ((RA - 1).isMinValue()) {
4999 Pred = ICmpInst::ICMP_NE;
5000 RHS = getConstant(RA - 1);
5004 if (RA.isMaxValue()) {
5005 Pred = ICmpInst::ICMP_EQ;
5009 if (RA.isMinValue()) goto trivially_true;
5011 Pred = ICmpInst::ICMP_UGT;
5012 RHS = getConstant(RA - 1);
5015 case ICmpInst::ICMP_ULE:
5016 if ((RA + 1).isMaxValue()) {
5017 Pred = ICmpInst::ICMP_NE;
5018 RHS = getConstant(RA + 1);
5022 if (RA.isMinValue()) {
5023 Pred = ICmpInst::ICMP_EQ;
5027 if (RA.isMaxValue()) goto trivially_true;
5029 Pred = ICmpInst::ICMP_ULT;
5030 RHS = getConstant(RA + 1);
5033 case ICmpInst::ICMP_SGE:
5034 if ((RA - 1).isMinSignedValue()) {
5035 Pred = ICmpInst::ICMP_NE;
5036 RHS = getConstant(RA - 1);
5040 if (RA.isMaxSignedValue()) {
5041 Pred = ICmpInst::ICMP_EQ;
5045 if (RA.isMinSignedValue()) goto trivially_true;
5047 Pred = ICmpInst::ICMP_SGT;
5048 RHS = getConstant(RA - 1);
5051 case ICmpInst::ICMP_SLE:
5052 if ((RA + 1).isMaxSignedValue()) {
5053 Pred = ICmpInst::ICMP_NE;
5054 RHS = getConstant(RA + 1);
5058 if (RA.isMinSignedValue()) {
5059 Pred = ICmpInst::ICMP_EQ;
5063 if (RA.isMaxSignedValue()) goto trivially_true;
5065 Pred = ICmpInst::ICMP_SLT;
5066 RHS = getConstant(RA + 1);
5069 case ICmpInst::ICMP_UGT:
5070 if (RA.isMinValue()) {
5071 Pred = ICmpInst::ICMP_NE;
5075 if ((RA + 1).isMaxValue()) {
5076 Pred = ICmpInst::ICMP_EQ;
5077 RHS = getConstant(RA + 1);
5081 if (RA.isMaxValue()) goto trivially_false;
5083 case ICmpInst::ICMP_ULT:
5084 if (RA.isMaxValue()) {
5085 Pred = ICmpInst::ICMP_NE;
5089 if ((RA - 1).isMinValue()) {
5090 Pred = ICmpInst::ICMP_EQ;
5091 RHS = getConstant(RA - 1);
5095 if (RA.isMinValue()) goto trivially_false;
5097 case ICmpInst::ICMP_SGT:
5098 if (RA.isMinSignedValue()) {
5099 Pred = ICmpInst::ICMP_NE;
5103 if ((RA + 1).isMaxSignedValue()) {
5104 Pred = ICmpInst::ICMP_EQ;
5105 RHS = getConstant(RA + 1);
5109 if (RA.isMaxSignedValue()) goto trivially_false;
5111 case ICmpInst::ICMP_SLT:
5112 if (RA.isMaxSignedValue()) {
5113 Pred = ICmpInst::ICMP_NE;
5117 if ((RA - 1).isMinSignedValue()) {
5118 Pred = ICmpInst::ICMP_EQ;
5119 RHS = getConstant(RA - 1);
5123 if (RA.isMinSignedValue()) goto trivially_false;
5128 // Check for obvious equality.
5129 if (HasSameValue(LHS, RHS)) {
5130 if (ICmpInst::isTrueWhenEqual(Pred))
5131 goto trivially_true;
5132 if (ICmpInst::isFalseWhenEqual(Pred))
5133 goto trivially_false;
5136 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5137 // adding or subtracting 1 from one of the operands.
5139 case ICmpInst::ICMP_SLE:
5140 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5141 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5142 /*HasNUW=*/false, /*HasNSW=*/true);
5143 Pred = ICmpInst::ICMP_SLT;
5145 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5146 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5147 /*HasNUW=*/false, /*HasNSW=*/true);
5148 Pred = ICmpInst::ICMP_SLT;
5152 case ICmpInst::ICMP_SGE:
5153 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5154 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5155 /*HasNUW=*/false, /*HasNSW=*/true);
5156 Pred = ICmpInst::ICMP_SGT;
5158 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5159 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5160 /*HasNUW=*/false, /*HasNSW=*/true);
5161 Pred = ICmpInst::ICMP_SGT;
5165 case ICmpInst::ICMP_ULE:
5166 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5167 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5168 /*HasNUW=*/true, /*HasNSW=*/false);
5169 Pred = ICmpInst::ICMP_ULT;
5171 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5172 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5173 /*HasNUW=*/true, /*HasNSW=*/false);
5174 Pred = ICmpInst::ICMP_ULT;
5178 case ICmpInst::ICMP_UGE:
5179 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5180 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5181 /*HasNUW=*/true, /*HasNSW=*/false);
5182 Pred = ICmpInst::ICMP_UGT;
5184 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5185 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5186 /*HasNUW=*/true, /*HasNSW=*/false);
5187 Pred = ICmpInst::ICMP_UGT;
5195 // TODO: More simplifications are possible here.
5201 LHS = RHS = getConstant(Type::getInt1Ty(getContext()), 0);
5202 Pred = ICmpInst::ICMP_EQ;
5207 LHS = RHS = getConstant(Type::getInt1Ty(getContext()), 0);
5208 Pred = ICmpInst::ICMP_NE;
5212 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5213 return getSignedRange(S).getSignedMax().isNegative();
5216 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5217 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5220 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5221 return !getSignedRange(S).getSignedMin().isNegative();
5224 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5225 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5228 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5229 return isKnownNegative(S) || isKnownPositive(S);
5232 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5233 const SCEV *LHS, const SCEV *RHS) {
5234 // Canonicalize the inputs first.
5235 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5237 // If LHS or RHS is an addrec, check to see if the condition is true in
5238 // every iteration of the loop.
5239 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5240 if (isLoopEntryGuardedByCond(
5241 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5242 isLoopBackedgeGuardedByCond(
5243 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5245 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5246 if (isLoopEntryGuardedByCond(
5247 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5248 isLoopBackedgeGuardedByCond(
5249 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5252 // Otherwise see what can be done with known constant ranges.
5253 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5257 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5258 const SCEV *LHS, const SCEV *RHS) {
5259 if (HasSameValue(LHS, RHS))
5260 return ICmpInst::isTrueWhenEqual(Pred);
5262 // This code is split out from isKnownPredicate because it is called from
5263 // within isLoopEntryGuardedByCond.
5266 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5268 case ICmpInst::ICMP_SGT:
5269 Pred = ICmpInst::ICMP_SLT;
5270 std::swap(LHS, RHS);
5271 case ICmpInst::ICMP_SLT: {
5272 ConstantRange LHSRange = getSignedRange(LHS);
5273 ConstantRange RHSRange = getSignedRange(RHS);
5274 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5276 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5280 case ICmpInst::ICMP_SGE:
5281 Pred = ICmpInst::ICMP_SLE;
5282 std::swap(LHS, RHS);
5283 case ICmpInst::ICMP_SLE: {
5284 ConstantRange LHSRange = getSignedRange(LHS);
5285 ConstantRange RHSRange = getSignedRange(RHS);
5286 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5288 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
5292 case ICmpInst::ICMP_UGT:
5293 Pred = ICmpInst::ICMP_ULT;
5294 std::swap(LHS, RHS);
5295 case ICmpInst::ICMP_ULT: {
5296 ConstantRange LHSRange = getUnsignedRange(LHS);
5297 ConstantRange RHSRange = getUnsignedRange(RHS);
5298 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
5300 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
5304 case ICmpInst::ICMP_UGE:
5305 Pred = ICmpInst::ICMP_ULE;
5306 std::swap(LHS, RHS);
5307 case ICmpInst::ICMP_ULE: {
5308 ConstantRange LHSRange = getUnsignedRange(LHS);
5309 ConstantRange RHSRange = getUnsignedRange(RHS);
5310 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
5312 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
5316 case ICmpInst::ICMP_NE: {
5317 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
5319 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
5322 const SCEV *Diff = getMinusSCEV(LHS, RHS);
5323 if (isKnownNonZero(Diff))
5327 case ICmpInst::ICMP_EQ:
5328 // The check at the top of the function catches the case where
5329 // the values are known to be equal.
5335 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
5336 /// protected by a conditional between LHS and RHS. This is used to
5337 /// to eliminate casts.
5339 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
5340 ICmpInst::Predicate Pred,
5341 const SCEV *LHS, const SCEV *RHS) {
5342 // Interpret a null as meaning no loop, where there is obviously no guard
5343 // (interprocedural conditions notwithstanding).
5344 if (!L) return true;
5346 BasicBlock *Latch = L->getLoopLatch();
5350 BranchInst *LoopContinuePredicate =
5351 dyn_cast<BranchInst>(Latch->getTerminator());
5352 if (!LoopContinuePredicate ||
5353 LoopContinuePredicate->isUnconditional())
5356 return isImpliedCond(Pred, LHS, RHS,
5357 LoopContinuePredicate->getCondition(),
5358 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
5361 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
5362 /// by a conditional between LHS and RHS. This is used to help avoid max
5363 /// expressions in loop trip counts, and to eliminate casts.
5365 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
5366 ICmpInst::Predicate Pred,
5367 const SCEV *LHS, const SCEV *RHS) {
5368 // Interpret a null as meaning no loop, where there is obviously no guard
5369 // (interprocedural conditions notwithstanding).
5370 if (!L) return false;
5372 // Starting at the loop predecessor, climb up the predecessor chain, as long
5373 // as there are predecessors that can be found that have unique successors
5374 // leading to the original header.
5375 for (std::pair<BasicBlock *, BasicBlock *>
5376 Pair(L->getLoopPredecessor(), L->getHeader());
5378 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
5380 BranchInst *LoopEntryPredicate =
5381 dyn_cast<BranchInst>(Pair.first->getTerminator());
5382 if (!LoopEntryPredicate ||
5383 LoopEntryPredicate->isUnconditional())
5386 if (isImpliedCond(Pred, LHS, RHS,
5387 LoopEntryPredicate->getCondition(),
5388 LoopEntryPredicate->getSuccessor(0) != Pair.second))
5395 /// isImpliedCond - Test whether the condition described by Pred, LHS,
5396 /// and RHS is true whenever the given Cond value evaluates to true.
5397 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
5398 const SCEV *LHS, const SCEV *RHS,
5399 Value *FoundCondValue,
5401 // Recursively handle And and Or conditions.
5402 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
5403 if (BO->getOpcode() == Instruction::And) {
5405 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5406 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5407 } else if (BO->getOpcode() == Instruction::Or) {
5409 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5410 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5414 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
5415 if (!ICI) return false;
5417 // Bail if the ICmp's operands' types are wider than the needed type
5418 // before attempting to call getSCEV on them. This avoids infinite
5419 // recursion, since the analysis of widening casts can require loop
5420 // exit condition information for overflow checking, which would
5422 if (getTypeSizeInBits(LHS->getType()) <
5423 getTypeSizeInBits(ICI->getOperand(0)->getType()))
5426 // Now that we found a conditional branch that dominates the loop, check to
5427 // see if it is the comparison we are looking for.
5428 ICmpInst::Predicate FoundPred;
5430 FoundPred = ICI->getInversePredicate();
5432 FoundPred = ICI->getPredicate();
5434 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
5435 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
5437 // Balance the types. The case where FoundLHS' type is wider than
5438 // LHS' type is checked for above.
5439 if (getTypeSizeInBits(LHS->getType()) >
5440 getTypeSizeInBits(FoundLHS->getType())) {
5441 if (CmpInst::isSigned(Pred)) {
5442 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
5443 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
5445 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
5446 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
5450 // Canonicalize the query to match the way instcombine will have
5451 // canonicalized the comparison.
5452 if (SimplifyICmpOperands(Pred, LHS, RHS))
5454 return CmpInst::isTrueWhenEqual(Pred);
5455 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
5456 if (FoundLHS == FoundRHS)
5457 return CmpInst::isFalseWhenEqual(Pred);
5459 // Check to see if we can make the LHS or RHS match.
5460 if (LHS == FoundRHS || RHS == FoundLHS) {
5461 if (isa<SCEVConstant>(RHS)) {
5462 std::swap(FoundLHS, FoundRHS);
5463 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
5465 std::swap(LHS, RHS);
5466 Pred = ICmpInst::getSwappedPredicate(Pred);
5470 // Check whether the found predicate is the same as the desired predicate.
5471 if (FoundPred == Pred)
5472 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
5474 // Check whether swapping the found predicate makes it the same as the
5475 // desired predicate.
5476 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
5477 if (isa<SCEVConstant>(RHS))
5478 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
5480 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
5481 RHS, LHS, FoundLHS, FoundRHS);
5484 // Check whether the actual condition is beyond sufficient.
5485 if (FoundPred == ICmpInst::ICMP_EQ)
5486 if (ICmpInst::isTrueWhenEqual(Pred))
5487 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
5489 if (Pred == ICmpInst::ICMP_NE)
5490 if (!ICmpInst::isTrueWhenEqual(FoundPred))
5491 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
5494 // Otherwise assume the worst.
5498 /// isImpliedCondOperands - Test whether the condition described by Pred,
5499 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
5500 /// and FoundRHS is true.
5501 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
5502 const SCEV *LHS, const SCEV *RHS,
5503 const SCEV *FoundLHS,
5504 const SCEV *FoundRHS) {
5505 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
5506 FoundLHS, FoundRHS) ||
5507 // ~x < ~y --> x > y
5508 isImpliedCondOperandsHelper(Pred, LHS, RHS,
5509 getNotSCEV(FoundRHS),
5510 getNotSCEV(FoundLHS));
5513 /// isImpliedCondOperandsHelper - Test whether the condition described by
5514 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
5515 /// FoundLHS, and FoundRHS is true.
5517 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
5518 const SCEV *LHS, const SCEV *RHS,
5519 const SCEV *FoundLHS,
5520 const SCEV *FoundRHS) {
5522 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5523 case ICmpInst::ICMP_EQ:
5524 case ICmpInst::ICMP_NE:
5525 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
5528 case ICmpInst::ICMP_SLT:
5529 case ICmpInst::ICMP_SLE:
5530 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
5531 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
5534 case ICmpInst::ICMP_SGT:
5535 case ICmpInst::ICMP_SGE:
5536 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
5537 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
5540 case ICmpInst::ICMP_ULT:
5541 case ICmpInst::ICMP_ULE:
5542 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
5543 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
5546 case ICmpInst::ICMP_UGT:
5547 case ICmpInst::ICMP_UGE:
5548 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
5549 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
5557 /// getBECount - Subtract the end and start values and divide by the step,
5558 /// rounding up, to get the number of times the backedge is executed. Return
5559 /// CouldNotCompute if an intermediate computation overflows.
5560 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
5564 assert(!isKnownNegative(Step) &&
5565 "This code doesn't handle negative strides yet!");
5567 const Type *Ty = Start->getType();
5568 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
5569 const SCEV *Diff = getMinusSCEV(End, Start);
5570 const SCEV *RoundUp = getAddExpr(Step, NegOne);
5572 // Add an adjustment to the difference between End and Start so that
5573 // the division will effectively round up.
5574 const SCEV *Add = getAddExpr(Diff, RoundUp);
5577 // Check Add for unsigned overflow.
5578 // TODO: More sophisticated things could be done here.
5579 const Type *WideTy = IntegerType::get(getContext(),
5580 getTypeSizeInBits(Ty) + 1);
5581 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
5582 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
5583 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
5584 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
5585 return getCouldNotCompute();
5588 return getUDivExpr(Add, Step);
5591 /// HowManyLessThans - Return the number of times a backedge containing the
5592 /// specified less-than comparison will execute. If not computable, return
5593 /// CouldNotCompute.
5594 ScalarEvolution::BackedgeTakenInfo
5595 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
5596 const Loop *L, bool isSigned) {
5597 // Only handle: "ADDREC < LoopInvariant".
5598 if (!RHS->isLoopInvariant(L)) return getCouldNotCompute();
5600 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
5601 if (!AddRec || AddRec->getLoop() != L)
5602 return getCouldNotCompute();
5604 // Check to see if we have a flag which makes analysis easy.
5605 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
5606 AddRec->hasNoUnsignedWrap();
5608 if (AddRec->isAffine()) {
5609 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
5610 const SCEV *Step = AddRec->getStepRecurrence(*this);
5613 return getCouldNotCompute();
5614 if (Step->isOne()) {
5615 // With unit stride, the iteration never steps past the limit value.
5616 } else if (isKnownPositive(Step)) {
5617 // Test whether a positive iteration can step past the limit
5618 // value and past the maximum value for its type in a single step.
5619 // Note that it's not sufficient to check NoWrap here, because even
5620 // though the value after a wrap is undefined, it's not undefined
5621 // behavior, so if wrap does occur, the loop could either terminate or
5622 // loop infinitely, but in either case, the loop is guaranteed to
5623 // iterate at least until the iteration where the wrapping occurs.
5624 const SCEV *One = getConstant(Step->getType(), 1);
5626 APInt Max = APInt::getSignedMaxValue(BitWidth);
5627 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
5628 .slt(getSignedRange(RHS).getSignedMax()))
5629 return getCouldNotCompute();
5631 APInt Max = APInt::getMaxValue(BitWidth);
5632 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
5633 .ult(getUnsignedRange(RHS).getUnsignedMax()))
5634 return getCouldNotCompute();
5637 // TODO: Handle negative strides here and below.
5638 return getCouldNotCompute();
5640 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
5641 // m. So, we count the number of iterations in which {n,+,s} < m is true.
5642 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
5643 // treat m-n as signed nor unsigned due to overflow possibility.
5645 // First, we get the value of the LHS in the first iteration: n
5646 const SCEV *Start = AddRec->getOperand(0);
5648 // Determine the minimum constant start value.
5649 const SCEV *MinStart = getConstant(isSigned ?
5650 getSignedRange(Start).getSignedMin() :
5651 getUnsignedRange(Start).getUnsignedMin());
5653 // If we know that the condition is true in order to enter the loop,
5654 // then we know that it will run exactly (m-n)/s times. Otherwise, we
5655 // only know that it will execute (max(m,n)-n)/s times. In both cases,
5656 // the division must round up.
5657 const SCEV *End = RHS;
5658 if (!isLoopEntryGuardedByCond(L,
5659 isSigned ? ICmpInst::ICMP_SLT :
5661 getMinusSCEV(Start, Step), RHS))
5662 End = isSigned ? getSMaxExpr(RHS, Start)
5663 : getUMaxExpr(RHS, Start);
5665 // Determine the maximum constant end value.
5666 const SCEV *MaxEnd = getConstant(isSigned ?
5667 getSignedRange(End).getSignedMax() :
5668 getUnsignedRange(End).getUnsignedMax());
5670 // If MaxEnd is within a step of the maximum integer value in its type,
5671 // adjust it down to the minimum value which would produce the same effect.
5672 // This allows the subsequent ceiling division of (N+(step-1))/step to
5673 // compute the correct value.
5674 const SCEV *StepMinusOne = getMinusSCEV(Step,
5675 getConstant(Step->getType(), 1));
5678 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
5681 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
5684 // Finally, we subtract these two values and divide, rounding up, to get
5685 // the number of times the backedge is executed.
5686 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
5688 // The maximum backedge count is similar, except using the minimum start
5689 // value and the maximum end value.
5690 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
5692 return BackedgeTakenInfo(BECount, MaxBECount);
5695 return getCouldNotCompute();
5698 /// getNumIterationsInRange - Return the number of iterations of this loop that
5699 /// produce values in the specified constant range. Another way of looking at
5700 /// this is that it returns the first iteration number where the value is not in
5701 /// the condition, thus computing the exit count. If the iteration count can't
5702 /// be computed, an instance of SCEVCouldNotCompute is returned.
5703 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
5704 ScalarEvolution &SE) const {
5705 if (Range.isFullSet()) // Infinite loop.
5706 return SE.getCouldNotCompute();
5708 // If the start is a non-zero constant, shift the range to simplify things.
5709 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
5710 if (!SC->getValue()->isZero()) {
5711 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
5712 Operands[0] = SE.getConstant(SC->getType(), 0);
5713 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
5714 if (const SCEVAddRecExpr *ShiftedAddRec =
5715 dyn_cast<SCEVAddRecExpr>(Shifted))
5716 return ShiftedAddRec->getNumIterationsInRange(
5717 Range.subtract(SC->getValue()->getValue()), SE);
5718 // This is strange and shouldn't happen.
5719 return SE.getCouldNotCompute();
5722 // The only time we can solve this is when we have all constant indices.
5723 // Otherwise, we cannot determine the overflow conditions.
5724 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5725 if (!isa<SCEVConstant>(getOperand(i)))
5726 return SE.getCouldNotCompute();
5729 // Okay at this point we know that all elements of the chrec are constants and
5730 // that the start element is zero.
5732 // First check to see if the range contains zero. If not, the first
5734 unsigned BitWidth = SE.getTypeSizeInBits(getType());
5735 if (!Range.contains(APInt(BitWidth, 0)))
5736 return SE.getConstant(getType(), 0);
5739 // If this is an affine expression then we have this situation:
5740 // Solve {0,+,A} in Range === Ax in Range
5742 // We know that zero is in the range. If A is positive then we know that
5743 // the upper value of the range must be the first possible exit value.
5744 // If A is negative then the lower of the range is the last possible loop
5745 // value. Also note that we already checked for a full range.
5746 APInt One(BitWidth,1);
5747 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5748 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5750 // The exit value should be (End+A)/A.
5751 APInt ExitVal = (End + A).udiv(A);
5752 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5754 // Evaluate at the exit value. If we really did fall out of the valid
5755 // range, then we computed our trip count, otherwise wrap around or other
5756 // things must have happened.
5757 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5758 if (Range.contains(Val->getValue()))
5759 return SE.getCouldNotCompute(); // Something strange happened
5761 // Ensure that the previous value is in the range. This is a sanity check.
5762 assert(Range.contains(
5763 EvaluateConstantChrecAtConstant(this,
5764 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5765 "Linear scev computation is off in a bad way!");
5766 return SE.getConstant(ExitValue);
5767 } else if (isQuadratic()) {
5768 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5769 // quadratic equation to solve it. To do this, we must frame our problem in
5770 // terms of figuring out when zero is crossed, instead of when
5771 // Range.getUpper() is crossed.
5772 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5773 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5774 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5776 // Next, solve the constructed addrec
5777 std::pair<const SCEV *,const SCEV *> Roots =
5778 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5779 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5780 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5782 // Pick the smallest positive root value.
5783 if (ConstantInt *CB =
5784 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5785 R1->getValue(), R2->getValue()))) {
5786 if (CB->getZExtValue() == false)
5787 std::swap(R1, R2); // R1 is the minimum root now.
5789 // Make sure the root is not off by one. The returned iteration should
5790 // not be in the range, but the previous one should be. When solving
5791 // for "X*X < 5", for example, we should not return a root of 2.
5792 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5795 if (Range.contains(R1Val->getValue())) {
5796 // The next iteration must be out of the range...
5797 ConstantInt *NextVal =
5798 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5800 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5801 if (!Range.contains(R1Val->getValue()))
5802 return SE.getConstant(NextVal);
5803 return SE.getCouldNotCompute(); // Something strange happened
5806 // If R1 was not in the range, then it is a good return value. Make
5807 // sure that R1-1 WAS in the range though, just in case.
5808 ConstantInt *NextVal =
5809 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5810 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5811 if (Range.contains(R1Val->getValue()))
5813 return SE.getCouldNotCompute(); // Something strange happened
5818 return SE.getCouldNotCompute();
5823 //===----------------------------------------------------------------------===//
5824 // SCEVCallbackVH Class Implementation
5825 //===----------------------------------------------------------------------===//
5827 void ScalarEvolution::SCEVCallbackVH::deleted() {
5828 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5829 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5830 SE->ConstantEvolutionLoopExitValue.erase(PN);
5831 SE->ValueExprMap.erase(getValPtr());
5832 // this now dangles!
5835 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
5836 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5838 // Forget all the expressions associated with users of the old value,
5839 // so that future queries will recompute the expressions using the new
5841 Value *Old = getValPtr();
5842 SmallVector<User *, 16> Worklist;
5843 SmallPtrSet<User *, 8> Visited;
5844 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5846 Worklist.push_back(*UI);
5847 while (!Worklist.empty()) {
5848 User *U = Worklist.pop_back_val();
5849 // Deleting the Old value will cause this to dangle. Postpone
5850 // that until everything else is done.
5853 if (!Visited.insert(U))
5855 if (PHINode *PN = dyn_cast<PHINode>(U))
5856 SE->ConstantEvolutionLoopExitValue.erase(PN);
5857 SE->ValueExprMap.erase(U);
5858 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5860 Worklist.push_back(*UI);
5862 // Delete the Old value.
5863 if (PHINode *PN = dyn_cast<PHINode>(Old))
5864 SE->ConstantEvolutionLoopExitValue.erase(PN);
5865 SE->ValueExprMap.erase(Old);
5866 // this now dangles!
5869 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5870 : CallbackVH(V), SE(se) {}
5872 //===----------------------------------------------------------------------===//
5873 // ScalarEvolution Class Implementation
5874 //===----------------------------------------------------------------------===//
5876 ScalarEvolution::ScalarEvolution()
5877 : FunctionPass(ID), FirstUnknown(0) {
5878 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
5881 bool ScalarEvolution::runOnFunction(Function &F) {
5883 LI = &getAnalysis<LoopInfo>();
5884 TD = getAnalysisIfAvailable<TargetData>();
5885 DT = &getAnalysis<DominatorTree>();
5889 void ScalarEvolution::releaseMemory() {
5890 // Iterate through all the SCEVUnknown instances and call their
5891 // destructors, so that they release their references to their values.
5892 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
5896 ValueExprMap.clear();
5897 BackedgeTakenCounts.clear();
5898 ConstantEvolutionLoopExitValue.clear();
5899 ValuesAtScopes.clear();
5900 UnsignedRanges.clear();
5901 SignedRanges.clear();
5902 UniqueSCEVs.clear();
5903 SCEVAllocator.Reset();
5906 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5907 AU.setPreservesAll();
5908 AU.addRequiredTransitive<LoopInfo>();
5909 AU.addRequiredTransitive<DominatorTree>();
5912 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5913 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5916 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5918 // Print all inner loops first
5919 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5920 PrintLoopInfo(OS, SE, *I);
5923 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5926 SmallVector<BasicBlock *, 8> ExitBlocks;
5927 L->getExitBlocks(ExitBlocks);
5928 if (ExitBlocks.size() != 1)
5929 OS << "<multiple exits> ";
5931 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5932 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5934 OS << "Unpredictable backedge-taken count. ";
5939 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5942 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5943 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5945 OS << "Unpredictable max backedge-taken count. ";
5951 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
5952 // ScalarEvolution's implementation of the print method is to print
5953 // out SCEV values of all instructions that are interesting. Doing
5954 // this potentially causes it to create new SCEV objects though,
5955 // which technically conflicts with the const qualifier. This isn't
5956 // observable from outside the class though, so casting away the
5957 // const isn't dangerous.
5958 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
5960 OS << "Classifying expressions for: ";
5961 WriteAsOperand(OS, F, /*PrintType=*/false);
5963 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
5964 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
5967 const SCEV *SV = SE.getSCEV(&*I);
5970 const Loop *L = LI->getLoopFor((*I).getParent());
5972 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
5979 OS << "\t\t" "Exits: ";
5980 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
5981 if (!ExitValue->isLoopInvariant(L)) {
5982 OS << "<<Unknown>>";
5991 OS << "Determining loop execution counts for: ";
5992 WriteAsOperand(OS, F, /*PrintType=*/false);
5994 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
5995 PrintLoopInfo(OS, &SE, *I);