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 static RegisterPass<ScalarEvolution>
107 R("scalar-evolution", "Scalar Evolution Analysis", false, true);
108 char ScalarEvolution::ID = 0;
110 //===----------------------------------------------------------------------===//
111 // SCEV class definitions
112 //===----------------------------------------------------------------------===//
114 //===----------------------------------------------------------------------===//
115 // Implementation of the SCEV class.
120 void SCEV::dump() const {
125 bool SCEV::isZero() const {
126 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
127 return SC->getValue()->isZero();
131 bool SCEV::isOne() const {
132 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
133 return SC->getValue()->isOne();
137 bool SCEV::isAllOnesValue() const {
138 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
139 return SC->getValue()->isAllOnesValue();
143 SCEVCouldNotCompute::SCEVCouldNotCompute() :
144 SCEV(FoldingSetNodeID(), scCouldNotCompute) {}
146 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
147 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
151 const Type *SCEVCouldNotCompute::getType() const {
152 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
156 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
157 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
161 bool SCEVCouldNotCompute::hasOperand(const SCEV *) const {
162 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
166 void SCEVCouldNotCompute::print(raw_ostream &OS) const {
167 OS << "***COULDNOTCOMPUTE***";
170 bool SCEVCouldNotCompute::classof(const SCEV *S) {
171 return S->getSCEVType() == scCouldNotCompute;
174 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
176 ID.AddInteger(scConstant);
179 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
180 SCEV *S = SCEVAllocator.Allocate<SCEVConstant>();
181 new (S) SCEVConstant(ID, V);
182 UniqueSCEVs.InsertNode(S, IP);
186 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
187 return getConstant(ConstantInt::get(getContext(), Val));
191 ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
193 ConstantInt::get(cast<IntegerType>(Ty), V, isSigned));
196 const Type *SCEVConstant::getType() const { return V->getType(); }
198 void SCEVConstant::print(raw_ostream &OS) const {
199 WriteAsOperand(OS, V, false);
202 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeID &ID,
203 unsigned SCEVTy, const SCEV *op, const Type *ty)
204 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
206 bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
207 return Op->dominates(BB, DT);
210 bool SCEVCastExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
211 return Op->properlyDominates(BB, DT);
214 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeID &ID,
215 const SCEV *op, const Type *ty)
216 : SCEVCastExpr(ID, scTruncate, op, ty) {
217 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
218 (Ty->isInteger() || isa<PointerType>(Ty)) &&
219 "Cannot truncate non-integer value!");
222 void SCEVTruncateExpr::print(raw_ostream &OS) const {
223 OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
226 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeID &ID,
227 const SCEV *op, const Type *ty)
228 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
229 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
230 (Ty->isInteger() || isa<PointerType>(Ty)) &&
231 "Cannot zero extend non-integer value!");
234 void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
235 OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
238 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeID &ID,
239 const SCEV *op, const Type *ty)
240 : SCEVCastExpr(ID, scSignExtend, op, ty) {
241 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
242 (Ty->isInteger() || isa<PointerType>(Ty)) &&
243 "Cannot sign extend non-integer value!");
246 void SCEVSignExtendExpr::print(raw_ostream &OS) const {
247 OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
250 void SCEVCommutativeExpr::print(raw_ostream &OS) const {
251 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
252 const char *OpStr = getOperationStr();
253 OS << "(" << *Operands[0];
254 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
255 OS << OpStr << *Operands[i];
259 bool SCEVNAryExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
260 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
261 if (!getOperand(i)->dominates(BB, DT))
267 bool SCEVNAryExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
268 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
269 if (!getOperand(i)->properlyDominates(BB, DT))
275 bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
276 return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
279 bool SCEVUDivExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
280 return LHS->properlyDominates(BB, DT) && RHS->properlyDominates(BB, DT);
283 void SCEVUDivExpr::print(raw_ostream &OS) const {
284 OS << "(" << *LHS << " /u " << *RHS << ")";
287 const Type *SCEVUDivExpr::getType() const {
288 // In most cases the types of LHS and RHS will be the same, but in some
289 // crazy cases one or the other may be a pointer. ScalarEvolution doesn't
290 // depend on the type for correctness, but handling types carefully can
291 // avoid extra casts in the SCEVExpander. The LHS is more likely to be
292 // a pointer type than the RHS, so use the RHS' type here.
293 return RHS->getType();
296 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
297 // Add recurrences are never invariant in the function-body (null loop).
301 // This recurrence is variant w.r.t. QueryLoop if QueryLoop contains L.
302 if (QueryLoop->contains(L))
305 // This recurrence is variant w.r.t. QueryLoop if any of its operands
307 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
308 if (!getOperand(i)->isLoopInvariant(QueryLoop))
311 // Otherwise it's loop-invariant.
315 void SCEVAddRecExpr::print(raw_ostream &OS) const {
316 OS << "{" << *Operands[0];
317 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
318 OS << ",+," << *Operands[i];
320 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
324 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
325 // All non-instruction values are loop invariant. All instructions are loop
326 // invariant if they are not contained in the specified loop.
327 // Instructions are never considered invariant in the function body
328 // (null loop) because they are defined within the "loop".
329 if (Instruction *I = dyn_cast<Instruction>(V))
330 return L && !L->contains(I);
334 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
335 if (Instruction *I = dyn_cast<Instruction>(getValue()))
336 return DT->dominates(I->getParent(), BB);
340 bool SCEVUnknown::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
341 if (Instruction *I = dyn_cast<Instruction>(getValue()))
342 return DT->properlyDominates(I->getParent(), BB);
346 const Type *SCEVUnknown::getType() const {
350 bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
351 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
352 if (VCE->getOpcode() == Instruction::PtrToInt)
353 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
354 if (CE->getOpcode() == Instruction::GetElementPtr)
355 if (CE->getOperand(0)->isNullValue()) {
357 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
358 if (CE->getNumOperands() == 2)
359 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
369 bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
370 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
371 if (VCE->getOpcode() == Instruction::PtrToInt)
372 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
373 if (CE->getOpcode() == Instruction::GetElementPtr)
374 if (CE->getOperand(0)->isNullValue()) {
376 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
377 if (const StructType *STy = dyn_cast<StructType>(Ty))
378 if (!STy->isPacked() &&
379 CE->getNumOperands() == 3 &&
380 CE->getOperand(1)->isNullValue()) {
381 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
383 STy->getNumElements() == 2 &&
384 STy->getElementType(0)->isInteger(1)) {
385 AllocTy = STy->getElementType(1);
394 bool SCEVUnknown::isOffsetOf(const Type *&CTy, Constant *&FieldNo) const {
395 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
396 if (VCE->getOpcode() == Instruction::PtrToInt)
397 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
398 if (CE->getOpcode() == Instruction::GetElementPtr &&
399 CE->getNumOperands() == 3 &&
400 CE->getOperand(0)->isNullValue() &&
401 CE->getOperand(1)->isNullValue()) {
403 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
404 // Ignore vector types here so that ScalarEvolutionExpander doesn't
405 // emit getelementptrs that index into vectors.
406 if (isa<StructType>(Ty) || isa<ArrayType>(Ty)) {
408 FieldNo = CE->getOperand(2);
416 void SCEVUnknown::print(raw_ostream &OS) const {
418 if (isSizeOf(AllocTy)) {
419 OS << "sizeof(" << *AllocTy << ")";
422 if (isAlignOf(AllocTy)) {
423 OS << "alignof(" << *AllocTy << ")";
429 if (isOffsetOf(CTy, FieldNo)) {
430 OS << "offsetof(" << *CTy << ", ";
431 WriteAsOperand(OS, FieldNo, false);
436 // Otherwise just print it normally.
437 WriteAsOperand(OS, V, false);
440 //===----------------------------------------------------------------------===//
442 //===----------------------------------------------------------------------===//
444 static bool CompareTypes(const Type *A, const Type *B) {
445 if (A->getTypeID() != B->getTypeID())
446 return A->getTypeID() < B->getTypeID();
447 if (const IntegerType *AI = dyn_cast<IntegerType>(A)) {
448 const IntegerType *BI = cast<IntegerType>(B);
449 return AI->getBitWidth() < BI->getBitWidth();
451 if (const PointerType *AI = dyn_cast<PointerType>(A)) {
452 const PointerType *BI = cast<PointerType>(B);
453 return CompareTypes(AI->getElementType(), BI->getElementType());
455 if (const ArrayType *AI = dyn_cast<ArrayType>(A)) {
456 const ArrayType *BI = cast<ArrayType>(B);
457 if (AI->getNumElements() != BI->getNumElements())
458 return AI->getNumElements() < BI->getNumElements();
459 return CompareTypes(AI->getElementType(), BI->getElementType());
461 if (const VectorType *AI = dyn_cast<VectorType>(A)) {
462 const VectorType *BI = cast<VectorType>(B);
463 if (AI->getNumElements() != BI->getNumElements())
464 return AI->getNumElements() < BI->getNumElements();
465 return CompareTypes(AI->getElementType(), BI->getElementType());
467 if (const StructType *AI = dyn_cast<StructType>(A)) {
468 const StructType *BI = cast<StructType>(B);
469 if (AI->getNumElements() != BI->getNumElements())
470 return AI->getNumElements() < BI->getNumElements();
471 for (unsigned i = 0, e = AI->getNumElements(); i != e; ++i)
472 if (CompareTypes(AI->getElementType(i), BI->getElementType(i)) ||
473 CompareTypes(BI->getElementType(i), AI->getElementType(i)))
474 return CompareTypes(AI->getElementType(i), BI->getElementType(i));
480 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
481 /// than the complexity of the RHS. This comparator is used to canonicalize
483 class SCEVComplexityCompare {
486 explicit SCEVComplexityCompare(LoopInfo *li) : LI(li) {}
488 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
489 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
493 // Primarily, sort the SCEVs by their getSCEVType().
494 if (LHS->getSCEVType() != RHS->getSCEVType())
495 return LHS->getSCEVType() < RHS->getSCEVType();
497 // Aside from the getSCEVType() ordering, the particular ordering
498 // isn't very important except that it's beneficial to be consistent,
499 // so that (a + b) and (b + a) don't end up as different expressions.
501 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
502 // not as complete as it could be.
503 if (const SCEVUnknown *LU = dyn_cast<SCEVUnknown>(LHS)) {
504 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
506 // Order pointer values after integer values. This helps SCEVExpander
508 if (isa<PointerType>(LU->getType()) && !isa<PointerType>(RU->getType()))
510 if (isa<PointerType>(RU->getType()) && !isa<PointerType>(LU->getType()))
513 // Compare getValueID values.
514 if (LU->getValue()->getValueID() != RU->getValue()->getValueID())
515 return LU->getValue()->getValueID() < RU->getValue()->getValueID();
517 // Sort arguments by their position.
518 if (const Argument *LA = dyn_cast<Argument>(LU->getValue())) {
519 const Argument *RA = cast<Argument>(RU->getValue());
520 return LA->getArgNo() < RA->getArgNo();
523 // For instructions, compare their loop depth, and their opcode.
524 // This is pretty loose.
525 if (Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
526 Instruction *RV = cast<Instruction>(RU->getValue());
528 // Compare loop depths.
529 if (LI->getLoopDepth(LV->getParent()) !=
530 LI->getLoopDepth(RV->getParent()))
531 return LI->getLoopDepth(LV->getParent()) <
532 LI->getLoopDepth(RV->getParent());
535 if (LV->getOpcode() != RV->getOpcode())
536 return LV->getOpcode() < RV->getOpcode();
538 // Compare the number of operands.
539 if (LV->getNumOperands() != RV->getNumOperands())
540 return LV->getNumOperands() < RV->getNumOperands();
546 // Compare constant values.
547 if (const SCEVConstant *LC = dyn_cast<SCEVConstant>(LHS)) {
548 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
549 if (LC->getValue()->getBitWidth() != RC->getValue()->getBitWidth())
550 return LC->getValue()->getBitWidth() < RC->getValue()->getBitWidth();
551 return LC->getValue()->getValue().ult(RC->getValue()->getValue());
554 // Compare addrec loop depths.
555 if (const SCEVAddRecExpr *LA = dyn_cast<SCEVAddRecExpr>(LHS)) {
556 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
557 if (LA->getLoop()->getLoopDepth() != RA->getLoop()->getLoopDepth())
558 return LA->getLoop()->getLoopDepth() < RA->getLoop()->getLoopDepth();
561 // Lexicographically compare n-ary expressions.
562 if (const SCEVNAryExpr *LC = dyn_cast<SCEVNAryExpr>(LHS)) {
563 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
564 for (unsigned i = 0, e = LC->getNumOperands(); i != e; ++i) {
565 if (i >= RC->getNumOperands())
567 if (operator()(LC->getOperand(i), RC->getOperand(i)))
569 if (operator()(RC->getOperand(i), LC->getOperand(i)))
572 return LC->getNumOperands() < RC->getNumOperands();
575 // Lexicographically compare udiv expressions.
576 if (const SCEVUDivExpr *LC = dyn_cast<SCEVUDivExpr>(LHS)) {
577 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
578 if (operator()(LC->getLHS(), RC->getLHS()))
580 if (operator()(RC->getLHS(), LC->getLHS()))
582 if (operator()(LC->getRHS(), RC->getRHS()))
584 if (operator()(RC->getRHS(), LC->getRHS()))
589 // Compare cast expressions by operand.
590 if (const SCEVCastExpr *LC = dyn_cast<SCEVCastExpr>(LHS)) {
591 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
592 return operator()(LC->getOperand(), RC->getOperand());
595 llvm_unreachable("Unknown SCEV kind!");
601 /// GroupByComplexity - Given a list of SCEV objects, order them by their
602 /// complexity, and group objects of the same complexity together by value.
603 /// When this routine is finished, we know that any duplicates in the vector are
604 /// consecutive and that complexity is monotonically increasing.
606 /// Note that we go take special precautions to ensure that we get determinstic
607 /// results from this routine. In other words, we don't want the results of
608 /// this to depend on where the addresses of various SCEV objects happened to
611 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
613 if (Ops.size() < 2) return; // Noop
614 if (Ops.size() == 2) {
615 // This is the common case, which also happens to be trivially simple.
617 if (SCEVComplexityCompare(LI)(Ops[1], Ops[0]))
618 std::swap(Ops[0], Ops[1]);
622 // Do the rough sort by complexity.
623 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
625 // Now that we are sorted by complexity, group elements of the same
626 // complexity. Note that this is, at worst, N^2, but the vector is likely to
627 // be extremely short in practice. Note that we take this approach because we
628 // do not want to depend on the addresses of the objects we are grouping.
629 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
630 const SCEV *S = Ops[i];
631 unsigned Complexity = S->getSCEVType();
633 // If there are any objects of the same complexity and same value as this
635 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
636 if (Ops[j] == S) { // Found a duplicate.
637 // Move it to immediately after i'th element.
638 std::swap(Ops[i+1], Ops[j]);
639 ++i; // no need to rescan it.
640 if (i == e-2) return; // Done!
648 //===----------------------------------------------------------------------===//
649 // Simple SCEV method implementations
650 //===----------------------------------------------------------------------===//
652 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
654 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
656 const Type* ResultTy) {
657 // Handle the simplest case efficiently.
659 return SE.getTruncateOrZeroExtend(It, ResultTy);
661 // We are using the following formula for BC(It, K):
663 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
665 // Suppose, W is the bitwidth of the return value. We must be prepared for
666 // overflow. Hence, we must assure that the result of our computation is
667 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
668 // safe in modular arithmetic.
670 // However, this code doesn't use exactly that formula; the formula it uses
671 // is something like the following, where T is the number of factors of 2 in
672 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
675 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
677 // This formula is trivially equivalent to the previous formula. However,
678 // this formula can be implemented much more efficiently. The trick is that
679 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
680 // arithmetic. To do exact division in modular arithmetic, all we have
681 // to do is multiply by the inverse. Therefore, this step can be done at
684 // The next issue is how to safely do the division by 2^T. The way this
685 // is done is by doing the multiplication step at a width of at least W + T
686 // bits. This way, the bottom W+T bits of the product are accurate. Then,
687 // when we perform the division by 2^T (which is equivalent to a right shift
688 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
689 // truncated out after the division by 2^T.
691 // In comparison to just directly using the first formula, this technique
692 // is much more efficient; using the first formula requires W * K bits,
693 // but this formula less than W + K bits. Also, the first formula requires
694 // a division step, whereas this formula only requires multiplies and shifts.
696 // It doesn't matter whether the subtraction step is done in the calculation
697 // width or the input iteration count's width; if the subtraction overflows,
698 // the result must be zero anyway. We prefer here to do it in the width of
699 // the induction variable because it helps a lot for certain cases; CodeGen
700 // isn't smart enough to ignore the overflow, which leads to much less
701 // efficient code if the width of the subtraction is wider than the native
704 // (It's possible to not widen at all by pulling out factors of 2 before
705 // the multiplication; for example, K=2 can be calculated as
706 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
707 // extra arithmetic, so it's not an obvious win, and it gets
708 // much more complicated for K > 3.)
710 // Protection from insane SCEVs; this bound is conservative,
711 // but it probably doesn't matter.
713 return SE.getCouldNotCompute();
715 unsigned W = SE.getTypeSizeInBits(ResultTy);
717 // Calculate K! / 2^T and T; we divide out the factors of two before
718 // multiplying for calculating K! / 2^T to avoid overflow.
719 // Other overflow doesn't matter because we only care about the bottom
720 // W bits of the result.
721 APInt OddFactorial(W, 1);
723 for (unsigned i = 3; i <= K; ++i) {
725 unsigned TwoFactors = Mult.countTrailingZeros();
727 Mult = Mult.lshr(TwoFactors);
728 OddFactorial *= Mult;
731 // We need at least W + T bits for the multiplication step
732 unsigned CalculationBits = W + T;
734 // Calcuate 2^T, at width T+W.
735 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
737 // Calculate the multiplicative inverse of K! / 2^T;
738 // this multiplication factor will perform the exact division by
740 APInt Mod = APInt::getSignedMinValue(W+1);
741 APInt MultiplyFactor = OddFactorial.zext(W+1);
742 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
743 MultiplyFactor = MultiplyFactor.trunc(W);
745 // Calculate the product, at width T+W
746 const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
748 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
749 for (unsigned i = 1; i != K; ++i) {
750 const SCEV *S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
751 Dividend = SE.getMulExpr(Dividend,
752 SE.getTruncateOrZeroExtend(S, CalculationTy));
756 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
758 // Truncate the result, and divide by K! / 2^T.
760 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
761 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
764 /// evaluateAtIteration - Return the value of this chain of recurrences at
765 /// the specified iteration number. We can evaluate this recurrence by
766 /// multiplying each element in the chain by the binomial coefficient
767 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
769 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
771 /// where BC(It, k) stands for binomial coefficient.
773 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
774 ScalarEvolution &SE) const {
775 const SCEV *Result = getStart();
776 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
777 // The computation is correct in the face of overflow provided that the
778 // multiplication is performed _after_ the evaluation of the binomial
780 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
781 if (isa<SCEVCouldNotCompute>(Coeff))
784 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
789 //===----------------------------------------------------------------------===//
790 // SCEV Expression folder implementations
791 //===----------------------------------------------------------------------===//
793 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
795 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
796 "This is not a truncating conversion!");
797 assert(isSCEVable(Ty) &&
798 "This is not a conversion to a SCEVable type!");
799 Ty = getEffectiveSCEVType(Ty);
802 ID.AddInteger(scTruncate);
806 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
808 // Fold if the operand is constant.
809 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
811 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
813 // trunc(trunc(x)) --> trunc(x)
814 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
815 return getTruncateExpr(ST->getOperand(), Ty);
817 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
818 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
819 return getTruncateOrSignExtend(SS->getOperand(), Ty);
821 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
822 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
823 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
825 // If the input value is a chrec scev, truncate the chrec's operands.
826 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
827 SmallVector<const SCEV *, 4> Operands;
828 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
829 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
830 return getAddRecExpr(Operands, AddRec->getLoop());
833 // The cast wasn't folded; create an explicit cast node.
834 // Recompute the insert position, as it may have been invalidated.
835 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
836 SCEV *S = SCEVAllocator.Allocate<SCEVTruncateExpr>();
837 new (S) SCEVTruncateExpr(ID, Op, Ty);
838 UniqueSCEVs.InsertNode(S, IP);
842 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
844 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
845 "This is not an extending conversion!");
846 assert(isSCEVable(Ty) &&
847 "This is not a conversion to a SCEVable type!");
848 Ty = getEffectiveSCEVType(Ty);
850 // Fold if the operand is constant.
851 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
852 const Type *IntTy = getEffectiveSCEVType(Ty);
853 Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
854 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
855 return getConstant(cast<ConstantInt>(C));
858 // zext(zext(x)) --> zext(x)
859 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
860 return getZeroExtendExpr(SZ->getOperand(), Ty);
862 // Before doing any expensive analysis, check to see if we've already
863 // computed a SCEV for this Op and Ty.
865 ID.AddInteger(scZeroExtend);
869 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
871 // If the input value is a chrec scev, and we can prove that the value
872 // did not overflow the old, smaller, value, we can zero extend all of the
873 // operands (often constants). This allows analysis of something like
874 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
875 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
876 if (AR->isAffine()) {
877 const SCEV *Start = AR->getStart();
878 const SCEV *Step = AR->getStepRecurrence(*this);
879 unsigned BitWidth = getTypeSizeInBits(AR->getType());
880 const Loop *L = AR->getLoop();
882 // If we have special knowledge that this addrec won't overflow,
883 // we don't need to do any further analysis.
884 if (AR->hasNoUnsignedWrap())
885 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
886 getZeroExtendExpr(Step, Ty),
889 // Check whether the backedge-taken count is SCEVCouldNotCompute.
890 // Note that this serves two purposes: It filters out loops that are
891 // simply not analyzable, and it covers the case where this code is
892 // being called from within backedge-taken count analysis, such that
893 // attempting to ask for the backedge-taken count would likely result
894 // in infinite recursion. In the later case, the analysis code will
895 // cope with a conservative value, and it will take care to purge
896 // that value once it has finished.
897 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
898 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
899 // Manually compute the final value for AR, checking for
902 // Check whether the backedge-taken count can be losslessly casted to
903 // the addrec's type. The count is always unsigned.
904 const SCEV *CastedMaxBECount =
905 getTruncateOrZeroExtend(MaxBECount, Start->getType());
906 const SCEV *RecastedMaxBECount =
907 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
908 if (MaxBECount == RecastedMaxBECount) {
909 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
910 // Check whether Start+Step*MaxBECount has no unsigned overflow.
912 getMulExpr(CastedMaxBECount,
913 getTruncateOrZeroExtend(Step, Start->getType()));
914 const SCEV *Add = getAddExpr(Start, ZMul);
915 const SCEV *OperandExtendedAdd =
916 getAddExpr(getZeroExtendExpr(Start, WideTy),
917 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
918 getZeroExtendExpr(Step, WideTy)));
919 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
920 // Return the expression with the addrec on the outside.
921 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
922 getZeroExtendExpr(Step, Ty),
925 // Similar to above, only this time treat the step value as signed.
926 // This covers loops that count down.
928 getMulExpr(CastedMaxBECount,
929 getTruncateOrSignExtend(Step, Start->getType()));
930 Add = getAddExpr(Start, SMul);
932 getAddExpr(getZeroExtendExpr(Start, WideTy),
933 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
934 getSignExtendExpr(Step, WideTy)));
935 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
936 // Return the expression with the addrec on the outside.
937 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
938 getSignExtendExpr(Step, Ty),
942 // If the backedge is guarded by a comparison with the pre-inc value
943 // the addrec is safe. Also, if the entry is guarded by a comparison
944 // with the start value and the backedge is guarded by a comparison
945 // with the post-inc value, the addrec is safe.
946 if (isKnownPositive(Step)) {
947 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
948 getUnsignedRange(Step).getUnsignedMax());
949 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
950 (isLoopGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
951 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
952 AR->getPostIncExpr(*this), N)))
953 // Return the expression with the addrec on the outside.
954 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
955 getZeroExtendExpr(Step, Ty),
957 } else if (isKnownNegative(Step)) {
958 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
959 getSignedRange(Step).getSignedMin());
960 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) &&
961 (isLoopGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) ||
962 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
963 AR->getPostIncExpr(*this), N)))
964 // Return the expression with the addrec on the outside.
965 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
966 getSignExtendExpr(Step, Ty),
972 // The cast wasn't folded; create an explicit cast node.
973 // Recompute the insert position, as it may have been invalidated.
974 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
975 SCEV *S = SCEVAllocator.Allocate<SCEVZeroExtendExpr>();
976 new (S) SCEVZeroExtendExpr(ID, Op, Ty);
977 UniqueSCEVs.InsertNode(S, IP);
981 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
983 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
984 "This is not an extending conversion!");
985 assert(isSCEVable(Ty) &&
986 "This is not a conversion to a SCEVable type!");
987 Ty = getEffectiveSCEVType(Ty);
989 // Fold if the operand is constant.
990 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
991 const Type *IntTy = getEffectiveSCEVType(Ty);
992 Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
993 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
994 return getConstant(cast<ConstantInt>(C));
997 // sext(sext(x)) --> sext(x)
998 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
999 return getSignExtendExpr(SS->getOperand(), Ty);
1001 // Before doing any expensive analysis, check to see if we've already
1002 // computed a SCEV for this Op and Ty.
1003 FoldingSetNodeID ID;
1004 ID.AddInteger(scSignExtend);
1008 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1010 // If the input value is a chrec scev, and we can prove that the value
1011 // did not overflow the old, smaller, value, we can sign extend all of the
1012 // operands (often constants). This allows analysis of something like
1013 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1014 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1015 if (AR->isAffine()) {
1016 const SCEV *Start = AR->getStart();
1017 const SCEV *Step = AR->getStepRecurrence(*this);
1018 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1019 const Loop *L = AR->getLoop();
1021 // If we have special knowledge that this addrec won't overflow,
1022 // we don't need to do any further analysis.
1023 if (AR->hasNoSignedWrap())
1024 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1025 getSignExtendExpr(Step, Ty),
1028 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1029 // Note that this serves two purposes: It filters out loops that are
1030 // simply not analyzable, and it covers the case where this code is
1031 // being called from within backedge-taken count analysis, such that
1032 // attempting to ask for the backedge-taken count would likely result
1033 // in infinite recursion. In the later case, the analysis code will
1034 // cope with a conservative value, and it will take care to purge
1035 // that value once it has finished.
1036 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1037 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1038 // Manually compute the final value for AR, checking for
1041 // Check whether the backedge-taken count can be losslessly casted to
1042 // the addrec's type. The count is always unsigned.
1043 const SCEV *CastedMaxBECount =
1044 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1045 const SCEV *RecastedMaxBECount =
1046 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1047 if (MaxBECount == RecastedMaxBECount) {
1048 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1049 // Check whether Start+Step*MaxBECount has no signed overflow.
1051 getMulExpr(CastedMaxBECount,
1052 getTruncateOrSignExtend(Step, Start->getType()));
1053 const SCEV *Add = getAddExpr(Start, SMul);
1054 const SCEV *OperandExtendedAdd =
1055 getAddExpr(getSignExtendExpr(Start, WideTy),
1056 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1057 getSignExtendExpr(Step, WideTy)));
1058 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1059 // Return the expression with the addrec on the outside.
1060 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1061 getSignExtendExpr(Step, Ty),
1064 // Similar to above, only this time treat the step value as unsigned.
1065 // This covers loops that count up with an unsigned step.
1067 getMulExpr(CastedMaxBECount,
1068 getTruncateOrZeroExtend(Step, Start->getType()));
1069 Add = getAddExpr(Start, UMul);
1070 OperandExtendedAdd =
1071 getAddExpr(getSignExtendExpr(Start, WideTy),
1072 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1073 getZeroExtendExpr(Step, WideTy)));
1074 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1075 // Return the expression with the addrec on the outside.
1076 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1077 getZeroExtendExpr(Step, Ty),
1081 // If the backedge is guarded by a comparison with the pre-inc value
1082 // the addrec is safe. Also, if the entry is guarded by a comparison
1083 // with the start value and the backedge is guarded by a comparison
1084 // with the post-inc value, the addrec is safe.
1085 if (isKnownPositive(Step)) {
1086 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1087 getSignedRange(Step).getSignedMax());
1088 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1089 (isLoopGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1090 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1091 AR->getPostIncExpr(*this), N)))
1092 // Return the expression with the addrec on the outside.
1093 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1094 getSignExtendExpr(Step, Ty),
1096 } else if (isKnownNegative(Step)) {
1097 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1098 getSignedRange(Step).getSignedMin());
1099 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1100 (isLoopGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1101 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1102 AR->getPostIncExpr(*this), N)))
1103 // Return the expression with the addrec on the outside.
1104 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1105 getSignExtendExpr(Step, Ty),
1111 // The cast wasn't folded; create an explicit cast node.
1112 // Recompute the insert position, as it may have been invalidated.
1113 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1114 SCEV *S = SCEVAllocator.Allocate<SCEVSignExtendExpr>();
1115 new (S) SCEVSignExtendExpr(ID, Op, Ty);
1116 UniqueSCEVs.InsertNode(S, IP);
1120 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1121 /// unspecified bits out to the given type.
1123 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1125 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1126 "This is not an extending conversion!");
1127 assert(isSCEVable(Ty) &&
1128 "This is not a conversion to a SCEVable type!");
1129 Ty = getEffectiveSCEVType(Ty);
1131 // Sign-extend negative constants.
1132 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1133 if (SC->getValue()->getValue().isNegative())
1134 return getSignExtendExpr(Op, Ty);
1136 // Peel off a truncate cast.
1137 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1138 const SCEV *NewOp = T->getOperand();
1139 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1140 return getAnyExtendExpr(NewOp, Ty);
1141 return getTruncateOrNoop(NewOp, Ty);
1144 // Next try a zext cast. If the cast is folded, use it.
1145 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1146 if (!isa<SCEVZeroExtendExpr>(ZExt))
1149 // Next try a sext cast. If the cast is folded, use it.
1150 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1151 if (!isa<SCEVSignExtendExpr>(SExt))
1154 // Force the cast to be folded into the operands of an addrec.
1155 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1156 SmallVector<const SCEV *, 4> Ops;
1157 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1159 Ops.push_back(getAnyExtendExpr(*I, Ty));
1160 return getAddRecExpr(Ops, AR->getLoop());
1163 // If the expression is obviously signed, use the sext cast value.
1164 if (isa<SCEVSMaxExpr>(Op))
1167 // Absent any other information, use the zext cast value.
1171 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1172 /// a list of operands to be added under the given scale, update the given
1173 /// map. This is a helper function for getAddRecExpr. As an example of
1174 /// what it does, given a sequence of operands that would form an add
1175 /// expression like this:
1177 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1179 /// where A and B are constants, update the map with these values:
1181 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1183 /// and add 13 + A*B*29 to AccumulatedConstant.
1184 /// This will allow getAddRecExpr to produce this:
1186 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1188 /// This form often exposes folding opportunities that are hidden in
1189 /// the original operand list.
1191 /// Return true iff it appears that any interesting folding opportunities
1192 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1193 /// the common case where no interesting opportunities are present, and
1194 /// is also used as a check to avoid infinite recursion.
1197 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1198 SmallVector<const SCEV *, 8> &NewOps,
1199 APInt &AccumulatedConstant,
1200 const SmallVectorImpl<const SCEV *> &Ops,
1202 ScalarEvolution &SE) {
1203 bool Interesting = false;
1205 // Iterate over the add operands.
1206 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1207 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1208 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1210 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1211 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1212 // A multiplication of a constant with another add; recurse.
1214 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1215 cast<SCEVAddExpr>(Mul->getOperand(1))
1219 // A multiplication of a constant with some other value. Update
1221 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1222 const SCEV *Key = SE.getMulExpr(MulOps);
1223 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1224 M.insert(std::make_pair(Key, NewScale));
1226 NewOps.push_back(Pair.first->first);
1228 Pair.first->second += NewScale;
1229 // The map already had an entry for this value, which may indicate
1230 // a folding opportunity.
1234 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1235 // Pull a buried constant out to the outside.
1236 if (Scale != 1 || AccumulatedConstant != 0 || C->isZero())
1238 AccumulatedConstant += Scale * C->getValue()->getValue();
1240 // An ordinary operand. Update the map.
1241 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1242 M.insert(std::make_pair(Ops[i], Scale));
1244 NewOps.push_back(Pair.first->first);
1246 Pair.first->second += Scale;
1247 // The map already had an entry for this value, which may indicate
1248 // a folding opportunity.
1258 struct APIntCompare {
1259 bool operator()(const APInt &LHS, const APInt &RHS) const {
1260 return LHS.ult(RHS);
1265 /// getAddExpr - Get a canonical add expression, or something simpler if
1267 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1268 bool HasNUW, bool HasNSW) {
1269 assert(!Ops.empty() && "Cannot get empty add!");
1270 if (Ops.size() == 1) return Ops[0];
1272 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1273 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1274 getEffectiveSCEVType(Ops[0]->getType()) &&
1275 "SCEVAddExpr operand types don't match!");
1278 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1279 if (!HasNUW && HasNSW) {
1281 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1282 if (!isKnownNonNegative(Ops[i])) {
1286 if (All) HasNUW = true;
1289 // Sort by complexity, this groups all similar expression types together.
1290 GroupByComplexity(Ops, LI);
1292 // If there are any constants, fold them together.
1294 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1296 assert(Idx < Ops.size());
1297 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1298 // We found two constants, fold them together!
1299 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1300 RHSC->getValue()->getValue());
1301 if (Ops.size() == 2) return Ops[0];
1302 Ops.erase(Ops.begin()+1); // Erase the folded element
1303 LHSC = cast<SCEVConstant>(Ops[0]);
1306 // If we are left with a constant zero being added, strip it off.
1307 if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1308 Ops.erase(Ops.begin());
1313 if (Ops.size() == 1) return Ops[0];
1315 // Okay, check to see if the same value occurs in the operand list twice. If
1316 // so, merge them together into an multiply expression. Since we sorted the
1317 // list, these values are required to be adjacent.
1318 const Type *Ty = Ops[0]->getType();
1319 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1320 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1321 // Found a match, merge the two values into a multiply, and add any
1322 // remaining values to the result.
1323 const SCEV *Two = getIntegerSCEV(2, Ty);
1324 const SCEV *Mul = getMulExpr(Ops[i], Two);
1325 if (Ops.size() == 2)
1327 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1329 return getAddExpr(Ops, HasNUW, HasNSW);
1332 // Check for truncates. If all the operands are truncated from the same
1333 // type, see if factoring out the truncate would permit the result to be
1334 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1335 // if the contents of the resulting outer trunc fold to something simple.
1336 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1337 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1338 const Type *DstType = Trunc->getType();
1339 const Type *SrcType = Trunc->getOperand()->getType();
1340 SmallVector<const SCEV *, 8> LargeOps;
1342 // Check all the operands to see if they can be represented in the
1343 // source type of the truncate.
1344 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1345 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1346 if (T->getOperand()->getType() != SrcType) {
1350 LargeOps.push_back(T->getOperand());
1351 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1352 // This could be either sign or zero extension, but sign extension
1353 // is much more likely to be foldable here.
1354 LargeOps.push_back(getSignExtendExpr(C, SrcType));
1355 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1356 SmallVector<const SCEV *, 8> LargeMulOps;
1357 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1358 if (const SCEVTruncateExpr *T =
1359 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1360 if (T->getOperand()->getType() != SrcType) {
1364 LargeMulOps.push_back(T->getOperand());
1365 } else if (const SCEVConstant *C =
1366 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1367 // This could be either sign or zero extension, but sign extension
1368 // is much more likely to be foldable here.
1369 LargeMulOps.push_back(getSignExtendExpr(C, SrcType));
1376 LargeOps.push_back(getMulExpr(LargeMulOps));
1383 // Evaluate the expression in the larger type.
1384 const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
1385 // If it folds to something simple, use it. Otherwise, don't.
1386 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1387 return getTruncateExpr(Fold, DstType);
1391 // Skip past any other cast SCEVs.
1392 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1395 // If there are add operands they would be next.
1396 if (Idx < Ops.size()) {
1397 bool DeletedAdd = false;
1398 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1399 // If we have an add, expand the add operands onto the end of the operands
1401 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
1402 Ops.erase(Ops.begin()+Idx);
1406 // If we deleted at least one add, we added operands to the end of the list,
1407 // and they are not necessarily sorted. Recurse to resort and resimplify
1408 // any operands we just aquired.
1410 return getAddExpr(Ops);
1413 // Skip over the add expression until we get to a multiply.
1414 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1417 // Check to see if there are any folding opportunities present with
1418 // operands multiplied by constant values.
1419 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1420 uint64_t BitWidth = getTypeSizeInBits(Ty);
1421 DenseMap<const SCEV *, APInt> M;
1422 SmallVector<const SCEV *, 8> NewOps;
1423 APInt AccumulatedConstant(BitWidth, 0);
1424 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1425 Ops, APInt(BitWidth, 1), *this)) {
1426 // Some interesting folding opportunity is present, so its worthwhile to
1427 // re-generate the operands list. Group the operands by constant scale,
1428 // to avoid multiplying by the same constant scale multiple times.
1429 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1430 for (SmallVector<const SCEV *, 8>::iterator I = NewOps.begin(),
1431 E = NewOps.end(); I != E; ++I)
1432 MulOpLists[M.find(*I)->second].push_back(*I);
1433 // Re-generate the operands list.
1435 if (AccumulatedConstant != 0)
1436 Ops.push_back(getConstant(AccumulatedConstant));
1437 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1438 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1440 Ops.push_back(getMulExpr(getConstant(I->first),
1441 getAddExpr(I->second)));
1443 return getIntegerSCEV(0, Ty);
1444 if (Ops.size() == 1)
1446 return getAddExpr(Ops);
1450 // If we are adding something to a multiply expression, make sure the
1451 // something is not already an operand of the multiply. If so, merge it into
1453 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1454 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1455 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1456 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1457 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1458 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(Ops[AddOp])) {
1459 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1460 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1461 if (Mul->getNumOperands() != 2) {
1462 // If the multiply has more than two operands, we must get the
1464 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(), Mul->op_end());
1465 MulOps.erase(MulOps.begin()+MulOp);
1466 InnerMul = getMulExpr(MulOps);
1468 const SCEV *One = getIntegerSCEV(1, Ty);
1469 const SCEV *AddOne = getAddExpr(InnerMul, One);
1470 const SCEV *OuterMul = getMulExpr(AddOne, Ops[AddOp]);
1471 if (Ops.size() == 2) return OuterMul;
1473 Ops.erase(Ops.begin()+AddOp);
1474 Ops.erase(Ops.begin()+Idx-1);
1476 Ops.erase(Ops.begin()+Idx);
1477 Ops.erase(Ops.begin()+AddOp-1);
1479 Ops.push_back(OuterMul);
1480 return getAddExpr(Ops);
1483 // Check this multiply against other multiplies being added together.
1484 for (unsigned OtherMulIdx = Idx+1;
1485 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1487 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1488 // If MulOp occurs in OtherMul, we can fold the two multiplies
1490 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1491 OMulOp != e; ++OMulOp)
1492 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1493 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1494 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1495 if (Mul->getNumOperands() != 2) {
1496 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1498 MulOps.erase(MulOps.begin()+MulOp);
1499 InnerMul1 = getMulExpr(MulOps);
1501 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1502 if (OtherMul->getNumOperands() != 2) {
1503 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1504 OtherMul->op_end());
1505 MulOps.erase(MulOps.begin()+OMulOp);
1506 InnerMul2 = getMulExpr(MulOps);
1508 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1509 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1510 if (Ops.size() == 2) return OuterMul;
1511 Ops.erase(Ops.begin()+Idx);
1512 Ops.erase(Ops.begin()+OtherMulIdx-1);
1513 Ops.push_back(OuterMul);
1514 return getAddExpr(Ops);
1520 // If there are any add recurrences in the operands list, see if any other
1521 // added values are loop invariant. If so, we can fold them into the
1523 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1526 // Scan over all recurrences, trying to fold loop invariants into them.
1527 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1528 // Scan all of the other operands to this add and add them to the vector if
1529 // they are loop invariant w.r.t. the recurrence.
1530 SmallVector<const SCEV *, 8> LIOps;
1531 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1532 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1533 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1534 LIOps.push_back(Ops[i]);
1535 Ops.erase(Ops.begin()+i);
1539 // If we found some loop invariants, fold them into the recurrence.
1540 if (!LIOps.empty()) {
1541 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1542 LIOps.push_back(AddRec->getStart());
1544 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1546 AddRecOps[0] = getAddExpr(LIOps);
1548 // It's tempting to propagate NUW/NSW flags here, but nuw/nsw addition
1549 // is not associative so this isn't necessarily safe.
1550 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
1552 // If all of the other operands were loop invariant, we are done.
1553 if (Ops.size() == 1) return NewRec;
1555 // Otherwise, add the folded AddRec by the non-liv parts.
1556 for (unsigned i = 0;; ++i)
1557 if (Ops[i] == AddRec) {
1561 return getAddExpr(Ops);
1564 // Okay, if there weren't any loop invariants to be folded, check to see if
1565 // there are multiple AddRec's with the same loop induction variable being
1566 // added together. If so, we can fold them.
1567 for (unsigned OtherIdx = Idx+1;
1568 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1569 if (OtherIdx != Idx) {
1570 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1571 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1572 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
1573 SmallVector<const SCEV *, 4> NewOps(AddRec->op_begin(),
1575 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
1576 if (i >= NewOps.size()) {
1577 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
1578 OtherAddRec->op_end());
1581 NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
1583 const SCEV *NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
1585 if (Ops.size() == 2) return NewAddRec;
1587 Ops.erase(Ops.begin()+Idx);
1588 Ops.erase(Ops.begin()+OtherIdx-1);
1589 Ops.push_back(NewAddRec);
1590 return getAddExpr(Ops);
1594 // Otherwise couldn't fold anything into this recurrence. Move onto the
1598 // Okay, it looks like we really DO need an add expr. Check to see if we
1599 // already have one, otherwise create a new one.
1600 FoldingSetNodeID ID;
1601 ID.AddInteger(scAddExpr);
1602 ID.AddInteger(Ops.size());
1603 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1604 ID.AddPointer(Ops[i]);
1607 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1609 S = SCEVAllocator.Allocate<SCEVAddExpr>();
1610 new (S) SCEVAddExpr(ID, Ops);
1611 UniqueSCEVs.InsertNode(S, IP);
1613 if (HasNUW) S->setHasNoUnsignedWrap(true);
1614 if (HasNSW) S->setHasNoSignedWrap(true);
1618 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1620 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1621 bool HasNUW, bool HasNSW) {
1622 assert(!Ops.empty() && "Cannot get empty mul!");
1623 if (Ops.size() == 1) return Ops[0];
1625 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1626 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1627 getEffectiveSCEVType(Ops[0]->getType()) &&
1628 "SCEVMulExpr operand types don't match!");
1631 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1632 if (!HasNUW && HasNSW) {
1634 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1635 if (!isKnownNonNegative(Ops[i])) {
1639 if (All) HasNUW = true;
1642 // Sort by complexity, this groups all similar expression types together.
1643 GroupByComplexity(Ops, LI);
1645 // If there are any constants, fold them together.
1647 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1649 // C1*(C2+V) -> C1*C2 + C1*V
1650 if (Ops.size() == 2)
1651 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1652 if (Add->getNumOperands() == 2 &&
1653 isa<SCEVConstant>(Add->getOperand(0)))
1654 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1655 getMulExpr(LHSC, Add->getOperand(1)));
1658 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1659 // We found two constants, fold them together!
1660 ConstantInt *Fold = ConstantInt::get(getContext(),
1661 LHSC->getValue()->getValue() *
1662 RHSC->getValue()->getValue());
1663 Ops[0] = getConstant(Fold);
1664 Ops.erase(Ops.begin()+1); // Erase the folded element
1665 if (Ops.size() == 1) return Ops[0];
1666 LHSC = cast<SCEVConstant>(Ops[0]);
1669 // If we are left with a constant one being multiplied, strip it off.
1670 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1671 Ops.erase(Ops.begin());
1673 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1674 // If we have a multiply of zero, it will always be zero.
1676 } else if (Ops[0]->isAllOnesValue()) {
1677 // If we have a mul by -1 of an add, try distributing the -1 among the
1679 if (Ops.size() == 2)
1680 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1681 SmallVector<const SCEV *, 4> NewOps;
1682 bool AnyFolded = false;
1683 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1685 const SCEV *Mul = getMulExpr(Ops[0], *I);
1686 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1687 NewOps.push_back(Mul);
1690 return getAddExpr(NewOps);
1695 // Skip over the add expression until we get to a multiply.
1696 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1699 if (Ops.size() == 1)
1702 // If there are mul operands inline them all into this expression.
1703 if (Idx < Ops.size()) {
1704 bool DeletedMul = false;
1705 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1706 // If we have an mul, expand the mul operands onto the end of the operands
1708 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1709 Ops.erase(Ops.begin()+Idx);
1713 // If we deleted at least one mul, we added operands to the end of the list,
1714 // and they are not necessarily sorted. Recurse to resort and resimplify
1715 // any operands we just aquired.
1717 return getMulExpr(Ops);
1720 // If there are any add recurrences in the operands list, see if any other
1721 // added values are loop invariant. If so, we can fold them into the
1723 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1726 // Scan over all recurrences, trying to fold loop invariants into them.
1727 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1728 // Scan all of the other operands to this mul and add them to the vector if
1729 // they are loop invariant w.r.t. the recurrence.
1730 SmallVector<const SCEV *, 8> LIOps;
1731 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1732 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1733 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1734 LIOps.push_back(Ops[i]);
1735 Ops.erase(Ops.begin()+i);
1739 // If we found some loop invariants, fold them into the recurrence.
1740 if (!LIOps.empty()) {
1741 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1742 SmallVector<const SCEV *, 4> NewOps;
1743 NewOps.reserve(AddRec->getNumOperands());
1744 if (LIOps.size() == 1) {
1745 const SCEV *Scale = LIOps[0];
1746 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1747 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1749 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1750 SmallVector<const SCEV *, 4> MulOps(LIOps.begin(), LIOps.end());
1751 MulOps.push_back(AddRec->getOperand(i));
1752 NewOps.push_back(getMulExpr(MulOps));
1756 // It's tempting to propagate the NSW flag here, but nsw multiplication
1757 // is not associative so this isn't necessarily safe.
1758 const SCEV *NewRec = getAddRecExpr(NewOps, AddRec->getLoop(),
1759 HasNUW && AddRec->hasNoUnsignedWrap(),
1762 // If all of the other operands were loop invariant, we are done.
1763 if (Ops.size() == 1) return NewRec;
1765 // Otherwise, multiply the folded AddRec by the non-liv parts.
1766 for (unsigned i = 0;; ++i)
1767 if (Ops[i] == AddRec) {
1771 return getMulExpr(Ops);
1774 // Okay, if there weren't any loop invariants to be folded, check to see if
1775 // there are multiple AddRec's with the same loop induction variable being
1776 // multiplied together. If so, we can fold them.
1777 for (unsigned OtherIdx = Idx+1;
1778 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1779 if (OtherIdx != Idx) {
1780 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1781 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1782 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1783 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1784 const SCEV *NewStart = getMulExpr(F->getStart(),
1786 const SCEV *B = F->getStepRecurrence(*this);
1787 const SCEV *D = G->getStepRecurrence(*this);
1788 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1791 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1793 if (Ops.size() == 2) return NewAddRec;
1795 Ops.erase(Ops.begin()+Idx);
1796 Ops.erase(Ops.begin()+OtherIdx-1);
1797 Ops.push_back(NewAddRec);
1798 return getMulExpr(Ops);
1802 // Otherwise couldn't fold anything into this recurrence. Move onto the
1806 // Okay, it looks like we really DO need an mul expr. Check to see if we
1807 // already have one, otherwise create a new one.
1808 FoldingSetNodeID ID;
1809 ID.AddInteger(scMulExpr);
1810 ID.AddInteger(Ops.size());
1811 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1812 ID.AddPointer(Ops[i]);
1815 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1817 S = SCEVAllocator.Allocate<SCEVMulExpr>();
1818 new (S) SCEVMulExpr(ID, Ops);
1819 UniqueSCEVs.InsertNode(S, IP);
1821 if (HasNUW) S->setHasNoUnsignedWrap(true);
1822 if (HasNSW) S->setHasNoSignedWrap(true);
1826 /// getUDivExpr - Get a canonical unsigned division expression, or something
1827 /// simpler if possible.
1828 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1830 assert(getEffectiveSCEVType(LHS->getType()) ==
1831 getEffectiveSCEVType(RHS->getType()) &&
1832 "SCEVUDivExpr operand types don't match!");
1834 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1835 if (RHSC->getValue()->equalsInt(1))
1836 return LHS; // X udiv 1 --> x
1838 return getIntegerSCEV(0, LHS->getType()); // value is undefined
1840 // Determine if the division can be folded into the operands of
1842 // TODO: Generalize this to non-constants by using known-bits information.
1843 const Type *Ty = LHS->getType();
1844 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1845 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ;
1846 // For non-power-of-two values, effectively round the value up to the
1847 // nearest power of two.
1848 if (!RHSC->getValue()->getValue().isPowerOf2())
1850 const IntegerType *ExtTy =
1851 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1852 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1853 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1854 if (const SCEVConstant *Step =
1855 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1856 if (!Step->getValue()->getValue()
1857 .urem(RHSC->getValue()->getValue()) &&
1858 getZeroExtendExpr(AR, ExtTy) ==
1859 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1860 getZeroExtendExpr(Step, ExtTy),
1862 SmallVector<const SCEV *, 4> Operands;
1863 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1864 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1865 return getAddRecExpr(Operands, AR->getLoop());
1867 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1868 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1869 SmallVector<const SCEV *, 4> Operands;
1870 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1871 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1872 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1873 // Find an operand that's safely divisible.
1874 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1875 const SCEV *Op = M->getOperand(i);
1876 const SCEV *Div = getUDivExpr(Op, RHSC);
1877 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1878 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
1879 Operands = SmallVector<const SCEV *, 4>(MOperands.begin(),
1882 return getMulExpr(Operands);
1886 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1887 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1888 SmallVector<const SCEV *, 4> Operands;
1889 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1890 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1891 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1893 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1894 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
1895 if (isa<SCEVUDivExpr>(Op) || getMulExpr(Op, RHS) != A->getOperand(i))
1897 Operands.push_back(Op);
1899 if (Operands.size() == A->getNumOperands())
1900 return getAddExpr(Operands);
1904 // Fold if both operands are constant.
1905 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1906 Constant *LHSCV = LHSC->getValue();
1907 Constant *RHSCV = RHSC->getValue();
1908 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
1913 FoldingSetNodeID ID;
1914 ID.AddInteger(scUDivExpr);
1918 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1919 SCEV *S = SCEVAllocator.Allocate<SCEVUDivExpr>();
1920 new (S) SCEVUDivExpr(ID, LHS, RHS);
1921 UniqueSCEVs.InsertNode(S, IP);
1926 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1927 /// Simplify the expression as much as possible.
1928 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
1929 const SCEV *Step, const Loop *L,
1930 bool HasNUW, bool HasNSW) {
1931 SmallVector<const SCEV *, 4> Operands;
1932 Operands.push_back(Start);
1933 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1934 if (StepChrec->getLoop() == L) {
1935 Operands.insert(Operands.end(), StepChrec->op_begin(),
1936 StepChrec->op_end());
1937 return getAddRecExpr(Operands, L);
1940 Operands.push_back(Step);
1941 return getAddRecExpr(Operands, L, HasNUW, HasNSW);
1944 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1945 /// Simplify the expression as much as possible.
1947 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
1949 bool HasNUW, bool HasNSW) {
1950 if (Operands.size() == 1) return Operands[0];
1952 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
1953 assert(getEffectiveSCEVType(Operands[i]->getType()) ==
1954 getEffectiveSCEVType(Operands[0]->getType()) &&
1955 "SCEVAddRecExpr operand types don't match!");
1958 if (Operands.back()->isZero()) {
1959 Operands.pop_back();
1960 return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
1963 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1964 if (!HasNUW && HasNSW) {
1966 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1967 if (!isKnownNonNegative(Operands[i])) {
1971 if (All) HasNUW = true;
1974 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
1975 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
1976 const Loop *NestedLoop = NestedAR->getLoop();
1977 if (L->contains(NestedLoop->getHeader()) ?
1978 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
1979 (!NestedLoop->contains(L->getHeader()) &&
1980 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
1981 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
1982 NestedAR->op_end());
1983 Operands[0] = NestedAR->getStart();
1984 // AddRecs require their operands be loop-invariant with respect to their
1985 // loops. Don't perform this transformation if it would break this
1987 bool AllInvariant = true;
1988 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1989 if (!Operands[i]->isLoopInvariant(L)) {
1990 AllInvariant = false;
1994 NestedOperands[0] = getAddRecExpr(Operands, L);
1995 AllInvariant = true;
1996 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
1997 if (!NestedOperands[i]->isLoopInvariant(NestedLoop)) {
1998 AllInvariant = false;
2002 // Ok, both add recurrences are valid after the transformation.
2003 return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
2005 // Reset Operands to its original state.
2006 Operands[0] = NestedAR;
2010 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2011 // already have one, otherwise create a new one.
2012 FoldingSetNodeID ID;
2013 ID.AddInteger(scAddRecExpr);
2014 ID.AddInteger(Operands.size());
2015 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2016 ID.AddPointer(Operands[i]);
2020 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2022 S = SCEVAllocator.Allocate<SCEVAddRecExpr>();
2023 new (S) SCEVAddRecExpr(ID, Operands, L);
2024 UniqueSCEVs.InsertNode(S, IP);
2026 if (HasNUW) S->setHasNoUnsignedWrap(true);
2027 if (HasNSW) S->setHasNoSignedWrap(true);
2031 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2033 SmallVector<const SCEV *, 2> Ops;
2036 return getSMaxExpr(Ops);
2040 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2041 assert(!Ops.empty() && "Cannot get empty smax!");
2042 if (Ops.size() == 1) return Ops[0];
2044 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2045 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
2046 getEffectiveSCEVType(Ops[0]->getType()) &&
2047 "SCEVSMaxExpr operand types don't match!");
2050 // Sort by complexity, this groups all similar expression types together.
2051 GroupByComplexity(Ops, LI);
2053 // If there are any constants, fold them together.
2055 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2057 assert(Idx < Ops.size());
2058 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2059 // We found two constants, fold them together!
2060 ConstantInt *Fold = ConstantInt::get(getContext(),
2061 APIntOps::smax(LHSC->getValue()->getValue(),
2062 RHSC->getValue()->getValue()));
2063 Ops[0] = getConstant(Fold);
2064 Ops.erase(Ops.begin()+1); // Erase the folded element
2065 if (Ops.size() == 1) return Ops[0];
2066 LHSC = cast<SCEVConstant>(Ops[0]);
2069 // If we are left with a constant minimum-int, strip it off.
2070 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2071 Ops.erase(Ops.begin());
2073 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2074 // If we have an smax with a constant maximum-int, it will always be
2080 if (Ops.size() == 1) return Ops[0];
2082 // Find the first SMax
2083 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2086 // Check to see if one of the operands is an SMax. If so, expand its operands
2087 // onto our operand list, and recurse to simplify.
2088 if (Idx < Ops.size()) {
2089 bool DeletedSMax = false;
2090 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2091 Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
2092 Ops.erase(Ops.begin()+Idx);
2097 return getSMaxExpr(Ops);
2100 // Okay, check to see if the same value occurs in the operand list twice. If
2101 // so, delete one. Since we sorted the list, these values are required to
2103 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2104 if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
2105 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2109 if (Ops.size() == 1) return Ops[0];
2111 assert(!Ops.empty() && "Reduced smax down to nothing!");
2113 // Okay, it looks like we really DO need an smax expr. Check to see if we
2114 // already have one, otherwise create a new one.
2115 FoldingSetNodeID ID;
2116 ID.AddInteger(scSMaxExpr);
2117 ID.AddInteger(Ops.size());
2118 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2119 ID.AddPointer(Ops[i]);
2121 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2122 SCEV *S = SCEVAllocator.Allocate<SCEVSMaxExpr>();
2123 new (S) SCEVSMaxExpr(ID, Ops);
2124 UniqueSCEVs.InsertNode(S, IP);
2128 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2130 SmallVector<const SCEV *, 2> Ops;
2133 return getUMaxExpr(Ops);
2137 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2138 assert(!Ops.empty() && "Cannot get empty umax!");
2139 if (Ops.size() == 1) return Ops[0];
2141 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2142 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
2143 getEffectiveSCEVType(Ops[0]->getType()) &&
2144 "SCEVUMaxExpr operand types don't match!");
2147 // Sort by complexity, this groups all similar expression types together.
2148 GroupByComplexity(Ops, LI);
2150 // If there are any constants, fold them together.
2152 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2154 assert(Idx < Ops.size());
2155 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2156 // We found two constants, fold them together!
2157 ConstantInt *Fold = ConstantInt::get(getContext(),
2158 APIntOps::umax(LHSC->getValue()->getValue(),
2159 RHSC->getValue()->getValue()));
2160 Ops[0] = getConstant(Fold);
2161 Ops.erase(Ops.begin()+1); // Erase the folded element
2162 if (Ops.size() == 1) return Ops[0];
2163 LHSC = cast<SCEVConstant>(Ops[0]);
2166 // If we are left with a constant minimum-int, strip it off.
2167 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2168 Ops.erase(Ops.begin());
2170 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2171 // If we have an umax with a constant maximum-int, it will always be
2177 if (Ops.size() == 1) return Ops[0];
2179 // Find the first UMax
2180 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2183 // Check to see if one of the operands is a UMax. If so, expand its operands
2184 // onto our operand list, and recurse to simplify.
2185 if (Idx < Ops.size()) {
2186 bool DeletedUMax = false;
2187 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2188 Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
2189 Ops.erase(Ops.begin()+Idx);
2194 return getUMaxExpr(Ops);
2197 // Okay, check to see if the same value occurs in the operand list twice. If
2198 // so, delete one. Since we sorted the list, these values are required to
2200 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2201 if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
2202 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2206 if (Ops.size() == 1) return Ops[0];
2208 assert(!Ops.empty() && "Reduced umax down to nothing!");
2210 // Okay, it looks like we really DO need a umax expr. Check to see if we
2211 // already have one, otherwise create a new one.
2212 FoldingSetNodeID ID;
2213 ID.AddInteger(scUMaxExpr);
2214 ID.AddInteger(Ops.size());
2215 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2216 ID.AddPointer(Ops[i]);
2218 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2219 SCEV *S = SCEVAllocator.Allocate<SCEVUMaxExpr>();
2220 new (S) SCEVUMaxExpr(ID, Ops);
2221 UniqueSCEVs.InsertNode(S, IP);
2225 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2227 // ~smax(~x, ~y) == smin(x, y).
2228 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2231 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2233 // ~umax(~x, ~y) == umin(x, y)
2234 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2237 const SCEV *ScalarEvolution::getSizeOfExpr(const Type *AllocTy) {
2238 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2239 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2240 C = ConstantFoldConstantExpression(CE, TD);
2241 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2242 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2245 const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
2246 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2247 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2248 C = ConstantFoldConstantExpression(CE, TD);
2249 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2250 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2253 const SCEV *ScalarEvolution::getOffsetOfExpr(const StructType *STy,
2255 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2256 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2257 C = ConstantFoldConstantExpression(CE, TD);
2258 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2259 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2262 const SCEV *ScalarEvolution::getOffsetOfExpr(const Type *CTy,
2263 Constant *FieldNo) {
2264 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2265 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2266 C = ConstantFoldConstantExpression(CE, TD);
2267 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2268 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2271 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2272 // Don't attempt to do anything other than create a SCEVUnknown object
2273 // here. createSCEV only calls getUnknown after checking for all other
2274 // interesting possibilities, and any other code that calls getUnknown
2275 // is doing so in order to hide a value from SCEV canonicalization.
2277 FoldingSetNodeID ID;
2278 ID.AddInteger(scUnknown);
2281 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2282 SCEV *S = SCEVAllocator.Allocate<SCEVUnknown>();
2283 new (S) SCEVUnknown(ID, V);
2284 UniqueSCEVs.InsertNode(S, IP);
2288 //===----------------------------------------------------------------------===//
2289 // Basic SCEV Analysis and PHI Idiom Recognition Code
2292 /// isSCEVable - Test if values of the given type are analyzable within
2293 /// the SCEV framework. This primarily includes integer types, and it
2294 /// can optionally include pointer types if the ScalarEvolution class
2295 /// has access to target-specific information.
2296 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2297 // Integers and pointers are always SCEVable.
2298 return Ty->isInteger() || isa<PointerType>(Ty);
2301 /// getTypeSizeInBits - Return the size in bits of the specified type,
2302 /// for which isSCEVable must return true.
2303 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2304 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2306 // If we have a TargetData, use it!
2308 return TD->getTypeSizeInBits(Ty);
2310 // Integer types have fixed sizes.
2311 if (Ty->isInteger())
2312 return Ty->getPrimitiveSizeInBits();
2314 // The only other support type is pointer. Without TargetData, conservatively
2315 // assume pointers are 64-bit.
2316 assert(isa<PointerType>(Ty) && "isSCEVable permitted a non-SCEVable type!");
2320 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2321 /// the given type and which represents how SCEV will treat the given
2322 /// type, for which isSCEVable must return true. For pointer types,
2323 /// this is the pointer-sized integer type.
2324 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2325 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2327 if (Ty->isInteger())
2330 // The only other support type is pointer.
2331 assert(isa<PointerType>(Ty) && "Unexpected non-pointer non-integer type!");
2332 if (TD) return TD->getIntPtrType(getContext());
2334 // Without TargetData, conservatively assume pointers are 64-bit.
2335 return Type::getInt64Ty(getContext());
2338 const SCEV *ScalarEvolution::getCouldNotCompute() {
2339 return &CouldNotCompute;
2342 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2343 /// expression and create a new one.
2344 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2345 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2347 std::map<SCEVCallbackVH, const SCEV *>::iterator I = Scalars.find(V);
2348 if (I != Scalars.end()) return I->second;
2349 const SCEV *S = createSCEV(V);
2350 Scalars.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2354 /// getIntegerSCEV - Given a SCEVable type, create a constant for the
2355 /// specified signed integer value and return a SCEV for the constant.
2356 const SCEV *ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
2357 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
2358 return getConstant(ConstantInt::get(ITy, Val));
2361 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2363 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2364 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2366 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2368 const Type *Ty = V->getType();
2369 Ty = getEffectiveSCEVType(Ty);
2370 return getMulExpr(V,
2371 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2374 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2375 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2376 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2378 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2380 const Type *Ty = V->getType();
2381 Ty = getEffectiveSCEVType(Ty);
2382 const SCEV *AllOnes =
2383 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2384 return getMinusSCEV(AllOnes, V);
2387 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
2389 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS,
2392 return getAddExpr(LHS, getNegativeSCEV(RHS));
2395 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2396 /// input value to the specified type. If the type must be extended, it is zero
2399 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V,
2401 const Type *SrcTy = V->getType();
2402 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2403 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2404 "Cannot truncate or zero extend with non-integer arguments!");
2405 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2406 return V; // No conversion
2407 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2408 return getTruncateExpr(V, Ty);
2409 return getZeroExtendExpr(V, Ty);
2412 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2413 /// input value to the specified type. If the type must be extended, it is sign
2416 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2418 const Type *SrcTy = V->getType();
2419 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2420 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2421 "Cannot truncate or zero extend with non-integer arguments!");
2422 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2423 return V; // No conversion
2424 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2425 return getTruncateExpr(V, Ty);
2426 return getSignExtendExpr(V, Ty);
2429 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2430 /// input value to the specified type. If the type must be extended, it is zero
2431 /// extended. The conversion must not be narrowing.
2433 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2434 const Type *SrcTy = V->getType();
2435 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2436 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2437 "Cannot noop or zero extend with non-integer arguments!");
2438 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2439 "getNoopOrZeroExtend cannot truncate!");
2440 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2441 return V; // No conversion
2442 return getZeroExtendExpr(V, Ty);
2445 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2446 /// input value to the specified type. If the type must be extended, it is sign
2447 /// extended. The conversion must not be narrowing.
2449 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
2450 const Type *SrcTy = V->getType();
2451 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2452 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2453 "Cannot noop or sign extend with non-integer arguments!");
2454 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2455 "getNoopOrSignExtend cannot truncate!");
2456 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2457 return V; // No conversion
2458 return getSignExtendExpr(V, Ty);
2461 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2462 /// the input value to the specified type. If the type must be extended,
2463 /// it is extended with unspecified bits. The conversion must not be
2466 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2467 const Type *SrcTy = V->getType();
2468 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2469 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2470 "Cannot noop or any extend with non-integer arguments!");
2471 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2472 "getNoopOrAnyExtend cannot truncate!");
2473 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2474 return V; // No conversion
2475 return getAnyExtendExpr(V, Ty);
2478 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2479 /// input value to the specified type. The conversion must not be widening.
2481 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2482 const Type *SrcTy = V->getType();
2483 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2484 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2485 "Cannot truncate or noop with non-integer arguments!");
2486 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2487 "getTruncateOrNoop cannot extend!");
2488 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2489 return V; // No conversion
2490 return getTruncateExpr(V, Ty);
2493 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2494 /// the types using zero-extension, and then perform a umax operation
2496 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2498 const SCEV *PromotedLHS = LHS;
2499 const SCEV *PromotedRHS = RHS;
2501 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2502 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2504 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2506 return getUMaxExpr(PromotedLHS, PromotedRHS);
2509 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2510 /// the types using zero-extension, and then perform a umin operation
2512 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2514 const SCEV *PromotedLHS = LHS;
2515 const SCEV *PromotedRHS = RHS;
2517 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2518 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2520 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2522 return getUMinExpr(PromotedLHS, PromotedRHS);
2525 /// PushDefUseChildren - Push users of the given Instruction
2526 /// onto the given Worklist.
2528 PushDefUseChildren(Instruction *I,
2529 SmallVectorImpl<Instruction *> &Worklist) {
2530 // Push the def-use children onto the Worklist stack.
2531 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2533 Worklist.push_back(cast<Instruction>(UI));
2536 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2537 /// instructions that depend on the given instruction and removes them from
2538 /// the Scalars map if they reference SymName. This is used during PHI
2541 ScalarEvolution::ForgetSymbolicName(Instruction *I, const SCEV *SymName) {
2542 SmallVector<Instruction *, 16> Worklist;
2543 PushDefUseChildren(I, Worklist);
2545 SmallPtrSet<Instruction *, 8> Visited;
2547 while (!Worklist.empty()) {
2548 Instruction *I = Worklist.pop_back_val();
2549 if (!Visited.insert(I)) continue;
2551 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
2552 Scalars.find(static_cast<Value *>(I));
2553 if (It != Scalars.end()) {
2554 // Short-circuit the def-use traversal if the symbolic name
2555 // ceases to appear in expressions.
2556 if (!It->second->hasOperand(SymName))
2559 // SCEVUnknown for a PHI either means that it has an unrecognized
2560 // structure, or it's a PHI that's in the progress of being computed
2561 // by createNodeForPHI. In the former case, additional loop trip
2562 // count information isn't going to change anything. In the later
2563 // case, createNodeForPHI will perform the necessary updates on its
2564 // own when it gets to that point.
2565 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
2566 ValuesAtScopes.erase(It->second);
2571 PushDefUseChildren(I, Worklist);
2575 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2576 /// a loop header, making it a potential recurrence, or it doesn't.
2578 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2579 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
2580 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2581 if (L->getHeader() == PN->getParent()) {
2582 // If it lives in the loop header, it has two incoming values, one
2583 // from outside the loop, and one from inside.
2584 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
2585 unsigned BackEdge = IncomingEdge^1;
2587 // While we are analyzing this PHI node, handle its value symbolically.
2588 const SCEV *SymbolicName = getUnknown(PN);
2589 assert(Scalars.find(PN) == Scalars.end() &&
2590 "PHI node already processed?");
2591 Scalars.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2593 // Using this symbolic name for the PHI, analyze the value coming around
2595 Value *BEValueV = PN->getIncomingValue(BackEdge);
2596 const SCEV *BEValue = getSCEV(BEValueV);
2598 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2599 // has a special value for the first iteration of the loop.
2601 // If the value coming around the backedge is an add with the symbolic
2602 // value we just inserted, then we found a simple induction variable!
2603 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2604 // If there is a single occurrence of the symbolic value, replace it
2605 // with a recurrence.
2606 unsigned FoundIndex = Add->getNumOperands();
2607 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2608 if (Add->getOperand(i) == SymbolicName)
2609 if (FoundIndex == e) {
2614 if (FoundIndex != Add->getNumOperands()) {
2615 // Create an add with everything but the specified operand.
2616 SmallVector<const SCEV *, 8> Ops;
2617 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2618 if (i != FoundIndex)
2619 Ops.push_back(Add->getOperand(i));
2620 const SCEV *Accum = getAddExpr(Ops);
2622 // This is not a valid addrec if the step amount is varying each
2623 // loop iteration, but is not itself an addrec in this loop.
2624 if (Accum->isLoopInvariant(L) ||
2625 (isa<SCEVAddRecExpr>(Accum) &&
2626 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2627 bool HasNUW = false;
2628 bool HasNSW = false;
2630 // If the increment doesn't overflow, then neither the addrec nor
2631 // the post-increment will overflow.
2632 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
2633 if (OBO->hasNoUnsignedWrap())
2635 if (OBO->hasNoSignedWrap())
2639 const SCEV *StartVal =
2640 getSCEV(PN->getIncomingValue(IncomingEdge));
2641 const SCEV *PHISCEV =
2642 getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
2644 // Since the no-wrap flags are on the increment, they apply to the
2645 // post-incremented value as well.
2646 if (Accum->isLoopInvariant(L))
2647 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
2648 Accum, L, HasNUW, HasNSW);
2650 // Okay, for the entire analysis of this edge we assumed the PHI
2651 // to be symbolic. We now need to go back and purge all of the
2652 // entries for the scalars that use the symbolic expression.
2653 ForgetSymbolicName(PN, SymbolicName);
2654 Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
2658 } else if (const SCEVAddRecExpr *AddRec =
2659 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2660 // Otherwise, this could be a loop like this:
2661 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2662 // In this case, j = {1,+,1} and BEValue is j.
2663 // Because the other in-value of i (0) fits the evolution of BEValue
2664 // i really is an addrec evolution.
2665 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2666 const SCEV *StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
2668 // If StartVal = j.start - j.stride, we can use StartVal as the
2669 // initial step of the addrec evolution.
2670 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2671 AddRec->getOperand(1))) {
2672 const SCEV *PHISCEV =
2673 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2675 // Okay, for the entire analysis of this edge we assumed the PHI
2676 // to be symbolic. We now need to go back and purge all of the
2677 // entries for the scalars that use the symbolic expression.
2678 ForgetSymbolicName(PN, SymbolicName);
2679 Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
2685 return SymbolicName;
2688 // It's tempting to recognize PHIs with a unique incoming value, however
2689 // this leads passes like indvars to break LCSSA form. Fortunately, such
2690 // PHIs are rare, as instcombine zaps them.
2692 // If it's not a loop phi, we can't handle it yet.
2693 return getUnknown(PN);
2696 /// createNodeForGEP - Expand GEP instructions into add and multiply
2697 /// operations. This allows them to be analyzed by regular SCEV code.
2699 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
2701 bool InBounds = GEP->isInBounds();
2702 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2703 Value *Base = GEP->getOperand(0);
2704 // Don't attempt to analyze GEPs over unsized objects.
2705 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2706 return getUnknown(GEP);
2707 const SCEV *TotalOffset = getIntegerSCEV(0, IntPtrTy);
2708 gep_type_iterator GTI = gep_type_begin(GEP);
2709 for (GetElementPtrInst::op_iterator I = next(GEP->op_begin()),
2713 // Compute the (potentially symbolic) offset in bytes for this index.
2714 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2715 // For a struct, add the member offset.
2716 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2717 TotalOffset = getAddExpr(TotalOffset,
2718 getOffsetOfExpr(STy, FieldNo),
2719 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2721 // For an array, add the element offset, explicitly scaled.
2722 const SCEV *LocalOffset = getSCEV(Index);
2723 if (!isa<PointerType>(LocalOffset->getType()))
2724 // Getelementptr indicies are signed.
2725 LocalOffset = getTruncateOrSignExtend(LocalOffset, IntPtrTy);
2726 // Lower "inbounds" GEPs to NSW arithmetic.
2727 LocalOffset = getMulExpr(LocalOffset, getSizeOfExpr(*GTI),
2728 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2729 TotalOffset = getAddExpr(TotalOffset, LocalOffset,
2730 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2733 return getAddExpr(getSCEV(Base), TotalOffset,
2734 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2737 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2738 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2739 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2740 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2742 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2743 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2744 return C->getValue()->getValue().countTrailingZeros();
2746 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2747 return std::min(GetMinTrailingZeros(T->getOperand()),
2748 (uint32_t)getTypeSizeInBits(T->getType()));
2750 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2751 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2752 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2753 getTypeSizeInBits(E->getType()) : OpRes;
2756 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2757 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2758 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2759 getTypeSizeInBits(E->getType()) : OpRes;
2762 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2763 // The result is the min of all operands results.
2764 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2765 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2766 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2770 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2771 // The result is the sum of all operands results.
2772 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2773 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2774 for (unsigned i = 1, e = M->getNumOperands();
2775 SumOpRes != BitWidth && i != e; ++i)
2776 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2781 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2782 // The result is the min of all operands results.
2783 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2784 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2785 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2789 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2790 // The result is the min of all operands results.
2791 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2792 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2793 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2797 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2798 // The result is the min of all operands results.
2799 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2800 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2801 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2805 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2806 // For a SCEVUnknown, ask ValueTracking.
2807 unsigned BitWidth = getTypeSizeInBits(U->getType());
2808 APInt Mask = APInt::getAllOnesValue(BitWidth);
2809 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2810 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
2811 return Zeros.countTrailingOnes();
2818 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
2821 ScalarEvolution::getUnsignedRange(const SCEV *S) {
2823 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2824 return ConstantRange(C->getValue()->getValue());
2826 unsigned BitWidth = getTypeSizeInBits(S->getType());
2827 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
2829 // If the value has known zeros, the maximum unsigned value will have those
2830 // known zeros as well.
2831 uint32_t TZ = GetMinTrailingZeros(S);
2833 ConservativeResult =
2834 ConstantRange(APInt::getMinValue(BitWidth),
2835 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
2837 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2838 ConstantRange X = getUnsignedRange(Add->getOperand(0));
2839 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2840 X = X.add(getUnsignedRange(Add->getOperand(i)));
2841 return ConservativeResult.intersectWith(X);
2844 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2845 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
2846 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2847 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
2848 return ConservativeResult.intersectWith(X);
2851 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2852 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
2853 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2854 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
2855 return ConservativeResult.intersectWith(X);
2858 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2859 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
2860 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2861 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
2862 return ConservativeResult.intersectWith(X);
2865 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2866 ConstantRange X = getUnsignedRange(UDiv->getLHS());
2867 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
2868 return ConservativeResult.intersectWith(X.udiv(Y));
2871 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2872 ConstantRange X = getUnsignedRange(ZExt->getOperand());
2873 return ConservativeResult.intersectWith(X.zeroExtend(BitWidth));
2876 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2877 ConstantRange X = getUnsignedRange(SExt->getOperand());
2878 return ConservativeResult.intersectWith(X.signExtend(BitWidth));
2881 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
2882 ConstantRange X = getUnsignedRange(Trunc->getOperand());
2883 return ConservativeResult.intersectWith(X.truncate(BitWidth));
2886 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
2887 // If there's no unsigned wrap, the value will never be less than its
2889 if (AddRec->hasNoUnsignedWrap())
2890 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
2891 ConservativeResult =
2892 ConstantRange(C->getValue()->getValue(),
2893 APInt(getTypeSizeInBits(C->getType()), 0));
2895 // TODO: non-affine addrec
2896 if (AddRec->isAffine()) {
2897 const Type *Ty = AddRec->getType();
2898 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
2899 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
2900 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
2901 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
2903 const SCEV *Start = AddRec->getStart();
2904 const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
2906 // Check for overflow.
2907 if (!AddRec->hasNoUnsignedWrap())
2908 return ConservativeResult;
2910 ConstantRange StartRange = getUnsignedRange(Start);
2911 ConstantRange EndRange = getUnsignedRange(End);
2912 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
2913 EndRange.getUnsignedMin());
2914 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
2915 EndRange.getUnsignedMax());
2916 if (Min.isMinValue() && Max.isMaxValue())
2917 return ConservativeResult;
2918 return ConservativeResult.intersectWith(ConstantRange(Min, Max+1));
2922 return ConservativeResult;
2925 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2926 // For a SCEVUnknown, ask ValueTracking.
2927 unsigned BitWidth = getTypeSizeInBits(U->getType());
2928 APInt Mask = APInt::getAllOnesValue(BitWidth);
2929 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2930 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
2931 if (Ones == ~Zeros + 1)
2932 return ConservativeResult;
2933 return ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1));
2936 return ConservativeResult;
2939 /// getSignedRange - Determine the signed range for a particular SCEV.
2942 ScalarEvolution::getSignedRange(const SCEV *S) {
2944 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2945 return ConstantRange(C->getValue()->getValue());
2947 unsigned BitWidth = getTypeSizeInBits(S->getType());
2948 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
2950 // If the value has known zeros, the maximum signed value will have those
2951 // known zeros as well.
2952 uint32_t TZ = GetMinTrailingZeros(S);
2954 ConservativeResult =
2955 ConstantRange(APInt::getSignedMinValue(BitWidth),
2956 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
2958 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2959 ConstantRange X = getSignedRange(Add->getOperand(0));
2960 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2961 X = X.add(getSignedRange(Add->getOperand(i)));
2962 return ConservativeResult.intersectWith(X);
2965 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2966 ConstantRange X = getSignedRange(Mul->getOperand(0));
2967 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2968 X = X.multiply(getSignedRange(Mul->getOperand(i)));
2969 return ConservativeResult.intersectWith(X);
2972 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2973 ConstantRange X = getSignedRange(SMax->getOperand(0));
2974 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2975 X = X.smax(getSignedRange(SMax->getOperand(i)));
2976 return ConservativeResult.intersectWith(X);
2979 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2980 ConstantRange X = getSignedRange(UMax->getOperand(0));
2981 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2982 X = X.umax(getSignedRange(UMax->getOperand(i)));
2983 return ConservativeResult.intersectWith(X);
2986 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2987 ConstantRange X = getSignedRange(UDiv->getLHS());
2988 ConstantRange Y = getSignedRange(UDiv->getRHS());
2989 return ConservativeResult.intersectWith(X.udiv(Y));
2992 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2993 ConstantRange X = getSignedRange(ZExt->getOperand());
2994 return ConservativeResult.intersectWith(X.zeroExtend(BitWidth));
2997 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2998 ConstantRange X = getSignedRange(SExt->getOperand());
2999 return ConservativeResult.intersectWith(X.signExtend(BitWidth));
3002 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3003 ConstantRange X = getSignedRange(Trunc->getOperand());
3004 return ConservativeResult.intersectWith(X.truncate(BitWidth));
3007 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3008 // If there's no signed wrap, and all the operands have the same sign or
3009 // zero, the value won't ever change sign.
3010 if (AddRec->hasNoSignedWrap()) {
3011 bool AllNonNeg = true;
3012 bool AllNonPos = true;
3013 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3014 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3015 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3018 ConservativeResult = ConservativeResult.intersectWith(
3019 ConstantRange(APInt(BitWidth, 0),
3020 APInt::getSignedMinValue(BitWidth)));
3022 ConservativeResult = ConservativeResult.intersectWith(
3023 ConstantRange(APInt::getSignedMinValue(BitWidth),
3024 APInt(BitWidth, 1)));
3027 // TODO: non-affine addrec
3028 if (AddRec->isAffine()) {
3029 const Type *Ty = AddRec->getType();
3030 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3031 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3032 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3033 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3035 const SCEV *Start = AddRec->getStart();
3036 const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
3038 // Check for overflow.
3039 if (!AddRec->hasNoSignedWrap())
3040 return ConservativeResult;
3042 ConstantRange StartRange = getSignedRange(Start);
3043 ConstantRange EndRange = getSignedRange(End);
3044 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3045 EndRange.getSignedMin());
3046 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3047 EndRange.getSignedMax());
3048 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3049 return ConservativeResult;
3050 return ConservativeResult.intersectWith(ConstantRange(Min, Max+1));
3054 return ConservativeResult;
3057 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3058 // For a SCEVUnknown, ask ValueTracking.
3059 if (!U->getValue()->getType()->isInteger() && !TD)
3060 return ConservativeResult;
3061 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3063 return ConservativeResult;
3064 return ConservativeResult.intersectWith(
3065 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3066 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1));
3069 return ConservativeResult;
3072 /// createSCEV - We know that there is no SCEV for the specified value.
3073 /// Analyze the expression.
3075 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3076 if (!isSCEVable(V->getType()))
3077 return getUnknown(V);
3079 unsigned Opcode = Instruction::UserOp1;
3080 if (Instruction *I = dyn_cast<Instruction>(V))
3081 Opcode = I->getOpcode();
3082 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3083 Opcode = CE->getOpcode();
3084 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3085 return getConstant(CI);
3086 else if (isa<ConstantPointerNull>(V))
3087 return getIntegerSCEV(0, V->getType());
3088 else if (isa<UndefValue>(V))
3089 return getIntegerSCEV(0, V->getType());
3090 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3091 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3093 return getUnknown(V);
3095 Operator *U = cast<Operator>(V);
3097 case Instruction::Add:
3098 // Don't transfer the NSW and NUW bits from the Add instruction to the
3099 // Add expression, because the Instruction may be guarded by control
3100 // flow and the no-overflow bits may not be valid for the expression in
3102 return getAddExpr(getSCEV(U->getOperand(0)),
3103 getSCEV(U->getOperand(1)));
3104 case Instruction::Mul:
3105 // Don't transfer the NSW and NUW bits from the Mul instruction to the
3106 // Mul expression, as with Add.
3107 return getMulExpr(getSCEV(U->getOperand(0)),
3108 getSCEV(U->getOperand(1)));
3109 case Instruction::UDiv:
3110 return getUDivExpr(getSCEV(U->getOperand(0)),
3111 getSCEV(U->getOperand(1)));
3112 case Instruction::Sub:
3113 return getMinusSCEV(getSCEV(U->getOperand(0)),
3114 getSCEV(U->getOperand(1)));
3115 case Instruction::And:
3116 // For an expression like x&255 that merely masks off the high bits,
3117 // use zext(trunc(x)) as the SCEV expression.
3118 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3119 if (CI->isNullValue())
3120 return getSCEV(U->getOperand(1));
3121 if (CI->isAllOnesValue())
3122 return getSCEV(U->getOperand(0));
3123 const APInt &A = CI->getValue();
3125 // Instcombine's ShrinkDemandedConstant may strip bits out of
3126 // constants, obscuring what would otherwise be a low-bits mask.
3127 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3128 // knew about to reconstruct a low-bits mask value.
3129 unsigned LZ = A.countLeadingZeros();
3130 unsigned BitWidth = A.getBitWidth();
3131 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
3132 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3133 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
3135 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3137 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3139 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3140 IntegerType::get(getContext(), BitWidth - LZ)),
3145 case Instruction::Or:
3146 // If the RHS of the Or is a constant, we may have something like:
3147 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3148 // optimizations will transparently handle this case.
3150 // In order for this transformation to be safe, the LHS must be of the
3151 // form X*(2^n) and the Or constant must be less than 2^n.
3152 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3153 const SCEV *LHS = getSCEV(U->getOperand(0));
3154 const APInt &CIVal = CI->getValue();
3155 if (GetMinTrailingZeros(LHS) >=
3156 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3157 // Build a plain add SCEV.
3158 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3159 // If the LHS of the add was an addrec and it has no-wrap flags,
3160 // transfer the no-wrap flags, since an or won't introduce a wrap.
3161 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3162 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3163 if (OldAR->hasNoUnsignedWrap())
3164 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3165 if (OldAR->hasNoSignedWrap())
3166 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3172 case Instruction::Xor:
3173 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3174 // If the RHS of the xor is a signbit, then this is just an add.
3175 // Instcombine turns add of signbit into xor as a strength reduction step.
3176 if (CI->getValue().isSignBit())
3177 return getAddExpr(getSCEV(U->getOperand(0)),
3178 getSCEV(U->getOperand(1)));
3180 // If the RHS of xor is -1, then this is a not operation.
3181 if (CI->isAllOnesValue())
3182 return getNotSCEV(getSCEV(U->getOperand(0)));
3184 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3185 // This is a variant of the check for xor with -1, and it handles
3186 // the case where instcombine has trimmed non-demanded bits out
3187 // of an xor with -1.
3188 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3189 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3190 if (BO->getOpcode() == Instruction::And &&
3191 LCI->getValue() == CI->getValue())
3192 if (const SCEVZeroExtendExpr *Z =
3193 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3194 const Type *UTy = U->getType();
3195 const SCEV *Z0 = Z->getOperand();
3196 const Type *Z0Ty = Z0->getType();
3197 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3199 // If C is a low-bits mask, the zero extend is zerving to
3200 // mask off the high bits. Complement the operand and
3201 // re-apply the zext.
3202 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3203 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3205 // If C is a single bit, it may be in the sign-bit position
3206 // before the zero-extend. In this case, represent the xor
3207 // using an add, which is equivalent, and re-apply the zext.
3208 APInt Trunc = APInt(CI->getValue()).trunc(Z0TySize);
3209 if (APInt(Trunc).zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3211 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3217 case Instruction::Shl:
3218 // Turn shift left of a constant amount into a multiply.
3219 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3220 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3221 Constant *X = ConstantInt::get(getContext(),
3222 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
3223 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3227 case Instruction::LShr:
3228 // Turn logical shift right of a constant into a unsigned divide.
3229 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3230 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3231 Constant *X = ConstantInt::get(getContext(),
3232 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
3233 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3237 case Instruction::AShr:
3238 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3239 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3240 if (Instruction *L = dyn_cast<Instruction>(U->getOperand(0)))
3241 if (L->getOpcode() == Instruction::Shl &&
3242 L->getOperand(1) == U->getOperand(1)) {
3243 unsigned BitWidth = getTypeSizeInBits(U->getType());
3244 uint64_t Amt = BitWidth - CI->getZExtValue();
3245 if (Amt == BitWidth)
3246 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3248 return getIntegerSCEV(0, U->getType()); // value is undefined
3250 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3251 IntegerType::get(getContext(), Amt)),
3256 case Instruction::Trunc:
3257 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3259 case Instruction::ZExt:
3260 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3262 case Instruction::SExt:
3263 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3265 case Instruction::BitCast:
3266 // BitCasts are no-op casts so we just eliminate the cast.
3267 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3268 return getSCEV(U->getOperand(0));
3271 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3272 // lead to pointer expressions which cannot safely be expanded to GEPs,
3273 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3274 // simplifying integer expressions.
3276 case Instruction::GetElementPtr:
3277 return createNodeForGEP(cast<GEPOperator>(U));
3279 case Instruction::PHI:
3280 return createNodeForPHI(cast<PHINode>(U));
3282 case Instruction::Select:
3283 // This could be a smax or umax that was lowered earlier.
3284 // Try to recover it.
3285 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3286 Value *LHS = ICI->getOperand(0);
3287 Value *RHS = ICI->getOperand(1);
3288 switch (ICI->getPredicate()) {
3289 case ICmpInst::ICMP_SLT:
3290 case ICmpInst::ICMP_SLE:
3291 std::swap(LHS, RHS);
3293 case ICmpInst::ICMP_SGT:
3294 case ICmpInst::ICMP_SGE:
3295 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
3296 return getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
3297 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
3298 return getSMinExpr(getSCEV(LHS), getSCEV(RHS));
3300 case ICmpInst::ICMP_ULT:
3301 case ICmpInst::ICMP_ULE:
3302 std::swap(LHS, RHS);
3304 case ICmpInst::ICMP_UGT:
3305 case ICmpInst::ICMP_UGE:
3306 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
3307 return getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
3308 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
3309 return getUMinExpr(getSCEV(LHS), getSCEV(RHS));
3311 case ICmpInst::ICMP_NE:
3312 // n != 0 ? n : 1 -> umax(n, 1)
3313 if (LHS == U->getOperand(1) &&
3314 isa<ConstantInt>(U->getOperand(2)) &&
3315 cast<ConstantInt>(U->getOperand(2))->isOne() &&
3316 isa<ConstantInt>(RHS) &&
3317 cast<ConstantInt>(RHS)->isZero())
3318 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(2)));
3320 case ICmpInst::ICMP_EQ:
3321 // n == 0 ? 1 : n -> umax(n, 1)
3322 if (LHS == U->getOperand(2) &&
3323 isa<ConstantInt>(U->getOperand(1)) &&
3324 cast<ConstantInt>(U->getOperand(1))->isOne() &&
3325 isa<ConstantInt>(RHS) &&
3326 cast<ConstantInt>(RHS)->isZero())
3327 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(1)));
3334 default: // We cannot analyze this expression.
3338 return getUnknown(V);
3343 //===----------------------------------------------------------------------===//
3344 // Iteration Count Computation Code
3347 /// getBackedgeTakenCount - If the specified loop has a predictable
3348 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3349 /// object. The backedge-taken count is the number of times the loop header
3350 /// will be branched to from within the loop. This is one less than the
3351 /// trip count of the loop, since it doesn't count the first iteration,
3352 /// when the header is branched to from outside the loop.
3354 /// Note that it is not valid to call this method on a loop without a
3355 /// loop-invariant backedge-taken count (see
3356 /// hasLoopInvariantBackedgeTakenCount).
3358 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3359 return getBackedgeTakenInfo(L).Exact;
3362 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3363 /// return the least SCEV value that is known never to be less than the
3364 /// actual backedge taken count.
3365 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3366 return getBackedgeTakenInfo(L).Max;
3369 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3370 /// onto the given Worklist.
3372 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3373 BasicBlock *Header = L->getHeader();
3375 // Push all Loop-header PHIs onto the Worklist stack.
3376 for (BasicBlock::iterator I = Header->begin();
3377 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3378 Worklist.push_back(PN);
3381 const ScalarEvolution::BackedgeTakenInfo &
3382 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3383 // Initially insert a CouldNotCompute for this loop. If the insertion
3384 // succeeds, procede to actually compute a backedge-taken count and
3385 // update the value. The temporary CouldNotCompute value tells SCEV
3386 // code elsewhere that it shouldn't attempt to request a new
3387 // backedge-taken count, which could result in infinite recursion.
3388 std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3389 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3391 BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
3392 if (BECount.Exact != getCouldNotCompute()) {
3393 assert(BECount.Exact->isLoopInvariant(L) &&
3394 BECount.Max->isLoopInvariant(L) &&
3395 "Computed backedge-taken count isn't loop invariant for loop!");
3396 ++NumTripCountsComputed;
3398 // Update the value in the map.
3399 Pair.first->second = BECount;
3401 if (BECount.Max != getCouldNotCompute())
3402 // Update the value in the map.
3403 Pair.first->second = BECount;
3404 if (isa<PHINode>(L->getHeader()->begin()))
3405 // Only count loops that have phi nodes as not being computable.
3406 ++NumTripCountsNotComputed;
3409 // Now that we know more about the trip count for this loop, forget any
3410 // existing SCEV values for PHI nodes in this loop since they are only
3411 // conservative estimates made without the benefit of trip count
3412 // information. This is similar to the code in forgetLoop, except that
3413 // it handles SCEVUnknown PHI nodes specially.
3414 if (BECount.hasAnyInfo()) {
3415 SmallVector<Instruction *, 16> Worklist;
3416 PushLoopPHIs(L, Worklist);
3418 SmallPtrSet<Instruction *, 8> Visited;
3419 while (!Worklist.empty()) {
3420 Instruction *I = Worklist.pop_back_val();
3421 if (!Visited.insert(I)) continue;
3423 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
3424 Scalars.find(static_cast<Value *>(I));
3425 if (It != Scalars.end()) {
3426 // SCEVUnknown for a PHI either means that it has an unrecognized
3427 // structure, or it's a PHI that's in the progress of being computed
3428 // by createNodeForPHI. In the former case, additional loop trip
3429 // count information isn't going to change anything. In the later
3430 // case, createNodeForPHI will perform the necessary updates on its
3431 // own when it gets to that point.
3432 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
3433 ValuesAtScopes.erase(It->second);
3436 if (PHINode *PN = dyn_cast<PHINode>(I))
3437 ConstantEvolutionLoopExitValue.erase(PN);
3440 PushDefUseChildren(I, Worklist);
3444 return Pair.first->second;
3447 /// forgetLoop - This method should be called by the client when it has
3448 /// changed a loop in a way that may effect ScalarEvolution's ability to
3449 /// compute a trip count, or if the loop is deleted.
3450 void ScalarEvolution::forgetLoop(const Loop *L) {
3451 // Drop any stored trip count value.
3452 BackedgeTakenCounts.erase(L);
3454 // Drop information about expressions based on loop-header PHIs.
3455 SmallVector<Instruction *, 16> Worklist;
3456 PushLoopPHIs(L, Worklist);
3458 SmallPtrSet<Instruction *, 8> Visited;
3459 while (!Worklist.empty()) {
3460 Instruction *I = Worklist.pop_back_val();
3461 if (!Visited.insert(I)) continue;
3463 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
3464 Scalars.find(static_cast<Value *>(I));
3465 if (It != Scalars.end()) {
3466 ValuesAtScopes.erase(It->second);
3468 if (PHINode *PN = dyn_cast<PHINode>(I))
3469 ConstantEvolutionLoopExitValue.erase(PN);
3472 PushDefUseChildren(I, Worklist);
3476 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3477 /// of the specified loop will execute.
3478 ScalarEvolution::BackedgeTakenInfo
3479 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3480 SmallVector<BasicBlock *, 8> ExitingBlocks;
3481 L->getExitingBlocks(ExitingBlocks);
3483 // Examine all exits and pick the most conservative values.
3484 const SCEV *BECount = getCouldNotCompute();
3485 const SCEV *MaxBECount = getCouldNotCompute();
3486 bool CouldNotComputeBECount = false;
3487 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3488 BackedgeTakenInfo NewBTI =
3489 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3491 if (NewBTI.Exact == getCouldNotCompute()) {
3492 // We couldn't compute an exact value for this exit, so
3493 // we won't be able to compute an exact value for the loop.
3494 CouldNotComputeBECount = true;
3495 BECount = getCouldNotCompute();
3496 } else if (!CouldNotComputeBECount) {
3497 if (BECount == getCouldNotCompute())
3498 BECount = NewBTI.Exact;
3500 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3502 if (MaxBECount == getCouldNotCompute())
3503 MaxBECount = NewBTI.Max;
3504 else if (NewBTI.Max != getCouldNotCompute())
3505 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3508 return BackedgeTakenInfo(BECount, MaxBECount);
3511 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3512 /// of the specified loop will execute if it exits via the specified block.
3513 ScalarEvolution::BackedgeTakenInfo
3514 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3515 BasicBlock *ExitingBlock) {
3517 // Okay, we've chosen an exiting block. See what condition causes us to
3518 // exit at this block.
3520 // FIXME: we should be able to handle switch instructions (with a single exit)
3521 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3522 if (ExitBr == 0) return getCouldNotCompute();
3523 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3525 // At this point, we know we have a conditional branch that determines whether
3526 // the loop is exited. However, we don't know if the branch is executed each
3527 // time through the loop. If not, then the execution count of the branch will
3528 // not be equal to the trip count of the loop.
3530 // Currently we check for this by checking to see if the Exit branch goes to
3531 // the loop header. If so, we know it will always execute the same number of
3532 // times as the loop. We also handle the case where the exit block *is* the
3533 // loop header. This is common for un-rotated loops.
3535 // If both of those tests fail, walk up the unique predecessor chain to the
3536 // header, stopping if there is an edge that doesn't exit the loop. If the
3537 // header is reached, the execution count of the branch will be equal to the
3538 // trip count of the loop.
3540 // More extensive analysis could be done to handle more cases here.
3542 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3543 ExitBr->getSuccessor(1) != L->getHeader() &&
3544 ExitBr->getParent() != L->getHeader()) {
3545 // The simple checks failed, try climbing the unique predecessor chain
3546 // up to the header.
3548 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3549 BasicBlock *Pred = BB->getUniquePredecessor();
3551 return getCouldNotCompute();
3552 TerminatorInst *PredTerm = Pred->getTerminator();
3553 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3554 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3557 // If the predecessor has a successor that isn't BB and isn't
3558 // outside the loop, assume the worst.
3559 if (L->contains(PredSucc))
3560 return getCouldNotCompute();
3562 if (Pred == L->getHeader()) {
3569 return getCouldNotCompute();
3572 // Procede to the next level to examine the exit condition expression.
3573 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3574 ExitBr->getSuccessor(0),
3575 ExitBr->getSuccessor(1));
3578 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3579 /// backedge of the specified loop will execute if its exit condition
3580 /// were a conditional branch of ExitCond, TBB, and FBB.
3581 ScalarEvolution::BackedgeTakenInfo
3582 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3586 // Check if the controlling expression for this loop is an And or Or.
3587 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3588 if (BO->getOpcode() == Instruction::And) {
3589 // Recurse on the operands of the and.
3590 BackedgeTakenInfo BTI0 =
3591 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3592 BackedgeTakenInfo BTI1 =
3593 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3594 const SCEV *BECount = getCouldNotCompute();
3595 const SCEV *MaxBECount = getCouldNotCompute();
3596 if (L->contains(TBB)) {
3597 // Both conditions must be true for the loop to continue executing.
3598 // Choose the less conservative count.
3599 if (BTI0.Exact == getCouldNotCompute() ||
3600 BTI1.Exact == getCouldNotCompute())
3601 BECount = getCouldNotCompute();
3603 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3604 if (BTI0.Max == getCouldNotCompute())
3605 MaxBECount = BTI1.Max;
3606 else if (BTI1.Max == getCouldNotCompute())
3607 MaxBECount = BTI0.Max;
3609 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3611 // Both conditions must be true for the loop to exit.
3612 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3613 if (BTI0.Exact != getCouldNotCompute() &&
3614 BTI1.Exact != getCouldNotCompute())
3615 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3616 if (BTI0.Max != getCouldNotCompute() &&
3617 BTI1.Max != getCouldNotCompute())
3618 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
3621 return BackedgeTakenInfo(BECount, MaxBECount);
3623 if (BO->getOpcode() == Instruction::Or) {
3624 // Recurse on the operands of the or.
3625 BackedgeTakenInfo BTI0 =
3626 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3627 BackedgeTakenInfo BTI1 =
3628 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3629 const SCEV *BECount = getCouldNotCompute();
3630 const SCEV *MaxBECount = getCouldNotCompute();
3631 if (L->contains(FBB)) {
3632 // Both conditions must be false for the loop to continue executing.
3633 // Choose the less conservative count.
3634 if (BTI0.Exact == getCouldNotCompute() ||
3635 BTI1.Exact == getCouldNotCompute())
3636 BECount = getCouldNotCompute();
3638 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3639 if (BTI0.Max == getCouldNotCompute())
3640 MaxBECount = BTI1.Max;
3641 else if (BTI1.Max == getCouldNotCompute())
3642 MaxBECount = BTI0.Max;
3644 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3646 // Both conditions must be false for the loop to exit.
3647 assert(L->contains(TBB) && "Loop block has no successor in loop!");
3648 if (BTI0.Exact != getCouldNotCompute() &&
3649 BTI1.Exact != getCouldNotCompute())
3650 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3651 if (BTI0.Max != getCouldNotCompute() &&
3652 BTI1.Max != getCouldNotCompute())
3653 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
3656 return BackedgeTakenInfo(BECount, MaxBECount);
3660 // With an icmp, it may be feasible to compute an exact backedge-taken count.
3661 // Procede to the next level to examine the icmp.
3662 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
3663 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
3665 // If it's not an integer or pointer comparison then compute it the hard way.
3666 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3669 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
3670 /// backedge of the specified loop will execute if its exit condition
3671 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
3672 ScalarEvolution::BackedgeTakenInfo
3673 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
3678 // If the condition was exit on true, convert the condition to exit on false
3679 ICmpInst::Predicate Cond;
3680 if (!L->contains(FBB))
3681 Cond = ExitCond->getPredicate();
3683 Cond = ExitCond->getInversePredicate();
3685 // Handle common loops like: for (X = "string"; *X; ++X)
3686 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
3687 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
3689 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
3690 if (!isa<SCEVCouldNotCompute>(ItCnt)) {
3691 unsigned BitWidth = getTypeSizeInBits(ItCnt->getType());
3692 return BackedgeTakenInfo(ItCnt,
3693 isa<SCEVConstant>(ItCnt) ? ItCnt :
3694 getConstant(APInt::getMaxValue(BitWidth)-1));
3698 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
3699 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
3701 // Try to evaluate any dependencies out of the loop.
3702 LHS = getSCEVAtScope(LHS, L);
3703 RHS = getSCEVAtScope(RHS, L);
3705 // At this point, we would like to compute how many iterations of the
3706 // loop the predicate will return true for these inputs.
3707 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
3708 // If there is a loop-invariant, force it into the RHS.
3709 std::swap(LHS, RHS);
3710 Cond = ICmpInst::getSwappedPredicate(Cond);
3713 // If we have a comparison of a chrec against a constant, try to use value
3714 // ranges to answer this query.
3715 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
3716 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
3717 if (AddRec->getLoop() == L) {
3718 // Form the constant range.
3719 ConstantRange CompRange(
3720 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
3722 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
3723 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
3727 case ICmpInst::ICMP_NE: { // while (X != Y)
3728 // Convert to: while (X-Y != 0)
3729 const SCEV *TC = HowFarToZero(getMinusSCEV(LHS, RHS), L);
3730 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3733 case ICmpInst::ICMP_EQ: { // while (X == Y)
3734 // Convert to: while (X-Y == 0)
3735 const SCEV *TC = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
3736 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3739 case ICmpInst::ICMP_SLT: {
3740 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
3741 if (BTI.hasAnyInfo()) return BTI;
3744 case ICmpInst::ICMP_SGT: {
3745 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3746 getNotSCEV(RHS), L, true);
3747 if (BTI.hasAnyInfo()) return BTI;
3750 case ICmpInst::ICMP_ULT: {
3751 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
3752 if (BTI.hasAnyInfo()) return BTI;
3755 case ICmpInst::ICMP_UGT: {
3756 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3757 getNotSCEV(RHS), L, false);
3758 if (BTI.hasAnyInfo()) return BTI;
3763 dbgs() << "ComputeBackedgeTakenCount ";
3764 if (ExitCond->getOperand(0)->getType()->isUnsigned())
3765 dbgs() << "[unsigned] ";
3766 dbgs() << *LHS << " "
3767 << Instruction::getOpcodeName(Instruction::ICmp)
3768 << " " << *RHS << "\n";
3773 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3776 static ConstantInt *
3777 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
3778 ScalarEvolution &SE) {
3779 const SCEV *InVal = SE.getConstant(C);
3780 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
3781 assert(isa<SCEVConstant>(Val) &&
3782 "Evaluation of SCEV at constant didn't fold correctly?");
3783 return cast<SCEVConstant>(Val)->getValue();
3786 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
3787 /// and a GEP expression (missing the pointer index) indexing into it, return
3788 /// the addressed element of the initializer or null if the index expression is
3791 GetAddressedElementFromGlobal(GlobalVariable *GV,
3792 const std::vector<ConstantInt*> &Indices) {
3793 Constant *Init = GV->getInitializer();
3794 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
3795 uint64_t Idx = Indices[i]->getZExtValue();
3796 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
3797 assert(Idx < CS->getNumOperands() && "Bad struct index!");
3798 Init = cast<Constant>(CS->getOperand(Idx));
3799 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
3800 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
3801 Init = cast<Constant>(CA->getOperand(Idx));
3802 } else if (isa<ConstantAggregateZero>(Init)) {
3803 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
3804 assert(Idx < STy->getNumElements() && "Bad struct index!");
3805 Init = Constant::getNullValue(STy->getElementType(Idx));
3806 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
3807 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
3808 Init = Constant::getNullValue(ATy->getElementType());
3810 llvm_unreachable("Unknown constant aggregate type!");
3814 return 0; // Unknown initializer type
3820 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
3821 /// 'icmp op load X, cst', try to see if we can compute the backedge
3822 /// execution count.
3824 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
3828 ICmpInst::Predicate predicate) {
3829 if (LI->isVolatile()) return getCouldNotCompute();
3831 // Check to see if the loaded pointer is a getelementptr of a global.
3832 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
3833 if (!GEP) return getCouldNotCompute();
3835 // Make sure that it is really a constant global we are gepping, with an
3836 // initializer, and make sure the first IDX is really 0.
3837 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
3838 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
3839 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
3840 !cast<Constant>(GEP->getOperand(1))->isNullValue())
3841 return getCouldNotCompute();
3843 // Okay, we allow one non-constant index into the GEP instruction.
3845 std::vector<ConstantInt*> Indexes;
3846 unsigned VarIdxNum = 0;
3847 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
3848 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
3849 Indexes.push_back(CI);
3850 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
3851 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
3852 VarIdx = GEP->getOperand(i);
3854 Indexes.push_back(0);
3857 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
3858 // Check to see if X is a loop variant variable value now.
3859 const SCEV *Idx = getSCEV(VarIdx);
3860 Idx = getSCEVAtScope(Idx, L);
3862 // We can only recognize very limited forms of loop index expressions, in
3863 // particular, only affine AddRec's like {C1,+,C2}.
3864 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
3865 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
3866 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
3867 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
3868 return getCouldNotCompute();
3870 unsigned MaxSteps = MaxBruteForceIterations;
3871 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
3872 ConstantInt *ItCst = ConstantInt::get(
3873 cast<IntegerType>(IdxExpr->getType()), IterationNum);
3874 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
3876 // Form the GEP offset.
3877 Indexes[VarIdxNum] = Val;
3879 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
3880 if (Result == 0) break; // Cannot compute!
3882 // Evaluate the condition for this iteration.
3883 Result = ConstantExpr::getICmp(predicate, Result, RHS);
3884 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
3885 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
3887 dbgs() << "\n***\n*** Computed loop count " << *ItCst
3888 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
3891 ++NumArrayLenItCounts;
3892 return getConstant(ItCst); // Found terminating iteration!
3895 return getCouldNotCompute();
3899 /// CanConstantFold - Return true if we can constant fold an instruction of the
3900 /// specified type, assuming that all operands were constants.
3901 static bool CanConstantFold(const Instruction *I) {
3902 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
3903 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
3906 if (const CallInst *CI = dyn_cast<CallInst>(I))
3907 if (const Function *F = CI->getCalledFunction())
3908 return canConstantFoldCallTo(F);
3912 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
3913 /// in the loop that V is derived from. We allow arbitrary operations along the
3914 /// way, but the operands of an operation must either be constants or a value
3915 /// derived from a constant PHI. If this expression does not fit with these
3916 /// constraints, return null.
3917 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
3918 // If this is not an instruction, or if this is an instruction outside of the
3919 // loop, it can't be derived from a loop PHI.
3920 Instruction *I = dyn_cast<Instruction>(V);
3921 if (I == 0 || !L->contains(I)) return 0;
3923 if (PHINode *PN = dyn_cast<PHINode>(I)) {
3924 if (L->getHeader() == I->getParent())
3927 // We don't currently keep track of the control flow needed to evaluate
3928 // PHIs, so we cannot handle PHIs inside of loops.
3932 // If we won't be able to constant fold this expression even if the operands
3933 // are constants, return early.
3934 if (!CanConstantFold(I)) return 0;
3936 // Otherwise, we can evaluate this instruction if all of its operands are
3937 // constant or derived from a PHI node themselves.
3939 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
3940 if (!(isa<Constant>(I->getOperand(Op)) ||
3941 isa<GlobalValue>(I->getOperand(Op)))) {
3942 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
3943 if (P == 0) return 0; // Not evolving from PHI
3947 return 0; // Evolving from multiple different PHIs.
3950 // This is a expression evolving from a constant PHI!
3954 /// EvaluateExpression - Given an expression that passes the
3955 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
3956 /// in the loop has the value PHIVal. If we can't fold this expression for some
3957 /// reason, return null.
3958 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
3959 const TargetData *TD) {
3960 if (isa<PHINode>(V)) return PHIVal;
3961 if (Constant *C = dyn_cast<Constant>(V)) return C;
3962 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
3963 Instruction *I = cast<Instruction>(V);
3965 std::vector<Constant*> Operands;
3966 Operands.resize(I->getNumOperands());
3968 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3969 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
3970 if (Operands[i] == 0) return 0;
3973 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
3974 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
3976 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
3977 &Operands[0], Operands.size(), TD);
3980 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
3981 /// in the header of its containing loop, we know the loop executes a
3982 /// constant number of times, and the PHI node is just a recurrence
3983 /// involving constants, fold it.
3985 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
3988 std::map<PHINode*, Constant*>::iterator I =
3989 ConstantEvolutionLoopExitValue.find(PN);
3990 if (I != ConstantEvolutionLoopExitValue.end())
3993 if (BEs.ugt(APInt(BEs.getBitWidth(),MaxBruteForceIterations)))
3994 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
3996 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
3998 // Since the loop is canonicalized, the PHI node must have two entries. One
3999 // entry must be a constant (coming in from outside of the loop), and the
4000 // second must be derived from the same PHI.
4001 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4002 Constant *StartCST =
4003 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4005 return RetVal = 0; // Must be a constant.
4007 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4008 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
4010 return RetVal = 0; // Not derived from same PHI.
4012 // Execute the loop symbolically to determine the exit value.
4013 if (BEs.getActiveBits() >= 32)
4014 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4016 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4017 unsigned IterationNum = 0;
4018 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
4019 if (IterationNum == NumIterations)
4020 return RetVal = PHIVal; // Got exit value!
4022 // Compute the value of the PHI node for the next iteration.
4023 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4024 if (NextPHI == PHIVal)
4025 return RetVal = NextPHI; // Stopped evolving!
4027 return 0; // Couldn't evaluate!
4032 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
4033 /// constant number of times (the condition evolves only from constants),
4034 /// try to evaluate a few iterations of the loop until we get the exit
4035 /// condition gets a value of ExitWhen (true or false). If we cannot
4036 /// evaluate the trip count of the loop, return getCouldNotCompute().
4038 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
4041 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4042 if (PN == 0) return getCouldNotCompute();
4044 // Since the loop is canonicalized, the PHI node must have two entries. One
4045 // entry must be a constant (coming in from outside of the loop), and the
4046 // second must be derived from the same PHI.
4047 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4048 Constant *StartCST =
4049 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4050 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
4052 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4053 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
4054 if (PN2 != PN) return getCouldNotCompute(); // Not derived from same PHI.
4056 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4057 // the loop symbolically to determine when the condition gets a value of
4059 unsigned IterationNum = 0;
4060 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4061 for (Constant *PHIVal = StartCST;
4062 IterationNum != MaxIterations; ++IterationNum) {
4063 ConstantInt *CondVal =
4064 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
4066 // Couldn't symbolically evaluate.
4067 if (!CondVal) return getCouldNotCompute();
4069 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4070 ++NumBruteForceTripCountsComputed;
4071 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4074 // Compute the value of the PHI node for the next iteration.
4075 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4076 if (NextPHI == 0 || NextPHI == PHIVal)
4077 return getCouldNotCompute();// Couldn't evaluate or not making progress...
4081 // Too many iterations were needed to evaluate.
4082 return getCouldNotCompute();
4085 /// getSCEVAtScope - Return a SCEV expression for the specified value
4086 /// at the specified scope in the program. The L value specifies a loop
4087 /// nest to evaluate the expression at, where null is the top-level or a
4088 /// specified loop is immediately inside of the loop.
4090 /// This method can be used to compute the exit value for a variable defined
4091 /// in a loop by querying what the value will hold in the parent loop.
4093 /// In the case that a relevant loop exit value cannot be computed, the
4094 /// original value V is returned.
4095 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4096 // Check to see if we've folded this expression at this loop before.
4097 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4098 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4099 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4101 return Pair.first->second ? Pair.first->second : V;
4103 // Otherwise compute it.
4104 const SCEV *C = computeSCEVAtScope(V, L);
4105 ValuesAtScopes[V][L] = C;
4109 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
4110 if (isa<SCEVConstant>(V)) return V;
4112 // If this instruction is evolved from a constant-evolving PHI, compute the
4113 // exit value from the loop without using SCEVs.
4114 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
4115 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
4116 const Loop *LI = (*this->LI)[I->getParent()];
4117 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
4118 if (PHINode *PN = dyn_cast<PHINode>(I))
4119 if (PN->getParent() == LI->getHeader()) {
4120 // Okay, there is no closed form solution for the PHI node. Check
4121 // to see if the loop that contains it has a known backedge-taken
4122 // count. If so, we may be able to force computation of the exit
4124 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
4125 if (const SCEVConstant *BTCC =
4126 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
4127 // Okay, we know how many times the containing loop executes. If
4128 // this is a constant evolving PHI node, get the final value at
4129 // the specified iteration number.
4130 Constant *RV = getConstantEvolutionLoopExitValue(PN,
4131 BTCC->getValue()->getValue(),
4133 if (RV) return getSCEV(RV);
4137 // Okay, this is an expression that we cannot symbolically evaluate
4138 // into a SCEV. Check to see if it's possible to symbolically evaluate
4139 // the arguments into constants, and if so, try to constant propagate the
4140 // result. This is particularly useful for computing loop exit values.
4141 if (CanConstantFold(I)) {
4142 std::vector<Constant*> Operands;
4143 Operands.reserve(I->getNumOperands());
4144 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4145 Value *Op = I->getOperand(i);
4146 if (Constant *C = dyn_cast<Constant>(Op)) {
4147 Operands.push_back(C);
4149 // If any of the operands is non-constant and if they are
4150 // non-integer and non-pointer, don't even try to analyze them
4151 // with scev techniques.
4152 if (!isSCEVable(Op->getType()))
4155 const SCEV *OpV = getSCEVAtScope(Op, L);
4156 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
4157 Constant *C = SC->getValue();
4158 if (C->getType() != Op->getType())
4159 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4163 Operands.push_back(C);
4164 } else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
4165 if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
4166 if (C->getType() != Op->getType())
4168 ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4172 Operands.push_back(C);
4182 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4183 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4184 Operands[0], Operands[1], TD);
4186 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4187 &Operands[0], Operands.size(), TD);
4192 // This is some other type of SCEVUnknown, just return it.
4196 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4197 // Avoid performing the look-up in the common case where the specified
4198 // expression has no loop-variant portions.
4199 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4200 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4201 if (OpAtScope != Comm->getOperand(i)) {
4202 // Okay, at least one of these operands is loop variant but might be
4203 // foldable. Build a new instance of the folded commutative expression.
4204 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4205 Comm->op_begin()+i);
4206 NewOps.push_back(OpAtScope);
4208 for (++i; i != e; ++i) {
4209 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4210 NewOps.push_back(OpAtScope);
4212 if (isa<SCEVAddExpr>(Comm))
4213 return getAddExpr(NewOps);
4214 if (isa<SCEVMulExpr>(Comm))
4215 return getMulExpr(NewOps);
4216 if (isa<SCEVSMaxExpr>(Comm))
4217 return getSMaxExpr(NewOps);
4218 if (isa<SCEVUMaxExpr>(Comm))
4219 return getUMaxExpr(NewOps);
4220 llvm_unreachable("Unknown commutative SCEV type!");
4223 // If we got here, all operands are loop invariant.
4227 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4228 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4229 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4230 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4231 return Div; // must be loop invariant
4232 return getUDivExpr(LHS, RHS);
4235 // If this is a loop recurrence for a loop that does not contain L, then we
4236 // are dealing with the final value computed by the loop.
4237 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4238 if (!L || !AddRec->getLoop()->contains(L)) {
4239 // To evaluate this recurrence, we need to know how many times the AddRec
4240 // loop iterates. Compute this now.
4241 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4242 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4244 // Then, evaluate the AddRec.
4245 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4250 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4251 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4252 if (Op == Cast->getOperand())
4253 return Cast; // must be loop invariant
4254 return getZeroExtendExpr(Op, Cast->getType());
4257 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4258 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4259 if (Op == Cast->getOperand())
4260 return Cast; // must be loop invariant
4261 return getSignExtendExpr(Op, Cast->getType());
4264 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4265 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4266 if (Op == Cast->getOperand())
4267 return Cast; // must be loop invariant
4268 return getTruncateExpr(Op, Cast->getType());
4271 llvm_unreachable("Unknown SCEV type!");
4275 /// getSCEVAtScope - This is a convenience function which does
4276 /// getSCEVAtScope(getSCEV(V), L).
4277 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4278 return getSCEVAtScope(getSCEV(V), L);
4281 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4282 /// following equation:
4284 /// A * X = B (mod N)
4286 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4287 /// A and B isn't important.
4289 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4290 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4291 ScalarEvolution &SE) {
4292 uint32_t BW = A.getBitWidth();
4293 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4294 assert(A != 0 && "A must be non-zero.");
4298 // The gcd of A and N may have only one prime factor: 2. The number of
4299 // trailing zeros in A is its multiplicity
4300 uint32_t Mult2 = A.countTrailingZeros();
4303 // 2. Check if B is divisible by D.
4305 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4306 // is not less than multiplicity of this prime factor for D.
4307 if (B.countTrailingZeros() < Mult2)
4308 return SE.getCouldNotCompute();
4310 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4313 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4314 // bit width during computations.
4315 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4316 APInt Mod(BW + 1, 0);
4317 Mod.set(BW - Mult2); // Mod = N / D
4318 APInt I = AD.multiplicativeInverse(Mod);
4320 // 4. Compute the minimum unsigned root of the equation:
4321 // I * (B / D) mod (N / D)
4322 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4324 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4326 return SE.getConstant(Result.trunc(BW));
4329 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4330 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4331 /// might be the same) or two SCEVCouldNotCompute objects.
4333 static std::pair<const SCEV *,const SCEV *>
4334 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4335 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4336 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4337 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4338 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4340 // We currently can only solve this if the coefficients are constants.
4341 if (!LC || !MC || !NC) {
4342 const SCEV *CNC = SE.getCouldNotCompute();
4343 return std::make_pair(CNC, CNC);
4346 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4347 const APInt &L = LC->getValue()->getValue();
4348 const APInt &M = MC->getValue()->getValue();
4349 const APInt &N = NC->getValue()->getValue();
4350 APInt Two(BitWidth, 2);
4351 APInt Four(BitWidth, 4);
4354 using namespace APIntOps;
4356 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4357 // The B coefficient is M-N/2
4361 // The A coefficient is N/2
4362 APInt A(N.sdiv(Two));
4364 // Compute the B^2-4ac term.
4367 SqrtTerm -= Four * (A * C);
4369 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4370 // integer value or else APInt::sqrt() will assert.
4371 APInt SqrtVal(SqrtTerm.sqrt());
4373 // Compute the two solutions for the quadratic formula.
4374 // The divisions must be performed as signed divisions.
4376 APInt TwoA( A << 1 );
4377 if (TwoA.isMinValue()) {
4378 const SCEV *CNC = SE.getCouldNotCompute();
4379 return std::make_pair(CNC, CNC);
4382 LLVMContext &Context = SE.getContext();
4384 ConstantInt *Solution1 =
4385 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4386 ConstantInt *Solution2 =
4387 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4389 return std::make_pair(SE.getConstant(Solution1),
4390 SE.getConstant(Solution2));
4391 } // end APIntOps namespace
4394 /// HowFarToZero - Return the number of times a backedge comparing the specified
4395 /// value to zero will execute. If not computable, return CouldNotCompute.
4396 const SCEV *ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4397 // If the value is a constant
4398 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4399 // If the value is already zero, the branch will execute zero times.
4400 if (C->getValue()->isZero()) return C;
4401 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4404 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4405 if (!AddRec || AddRec->getLoop() != L)
4406 return getCouldNotCompute();
4408 if (AddRec->isAffine()) {
4409 // If this is an affine expression, the execution count of this branch is
4410 // the minimum unsigned root of the following equation:
4412 // Start + Step*N = 0 (mod 2^BW)
4416 // Step*N = -Start (mod 2^BW)
4418 // where BW is the common bit width of Start and Step.
4420 // Get the initial value for the loop.
4421 const SCEV *Start = getSCEVAtScope(AddRec->getStart(),
4422 L->getParentLoop());
4423 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1),
4424 L->getParentLoop());
4426 if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
4427 // For now we handle only constant steps.
4429 // First, handle unitary steps.
4430 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4431 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4432 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4433 return Start; // N = Start (as unsigned)
4435 // Then, try to solve the above equation provided that Start is constant.
4436 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4437 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4438 -StartC->getValue()->getValue(),
4441 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
4442 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4443 // the quadratic equation to solve it.
4444 std::pair<const SCEV *,const SCEV *> Roots = SolveQuadraticEquation(AddRec,
4446 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4447 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4450 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
4451 << " sol#2: " << *R2 << "\n";
4453 // Pick the smallest positive root value.
4454 if (ConstantInt *CB =
4455 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
4456 R1->getValue(), R2->getValue()))) {
4457 if (CB->getZExtValue() == false)
4458 std::swap(R1, R2); // R1 is the minimum root now.
4460 // We can only use this value if the chrec ends up with an exact zero
4461 // value at this index. When solving for "X*X != 5", for example, we
4462 // should not accept a root of 2.
4463 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4465 return R1; // We found a quadratic root!
4470 return getCouldNotCompute();
4473 /// HowFarToNonZero - Return the number of times a backedge checking the
4474 /// specified value for nonzero will execute. If not computable, return
4476 const SCEV *ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4477 // Loops that look like: while (X == 0) are very strange indeed. We don't
4478 // handle them yet except for the trivial case. This could be expanded in the
4479 // future as needed.
4481 // If the value is a constant, check to see if it is known to be non-zero
4482 // already. If so, the backedge will execute zero times.
4483 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4484 if (!C->getValue()->isNullValue())
4485 return getIntegerSCEV(0, C->getType());
4486 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4489 // We could implement others, but I really doubt anyone writes loops like
4490 // this, and if they did, they would already be constant folded.
4491 return getCouldNotCompute();
4494 /// getLoopPredecessor - If the given loop's header has exactly one unique
4495 /// predecessor outside the loop, return it. Otherwise return null.
4497 BasicBlock *ScalarEvolution::getLoopPredecessor(const Loop *L) {
4498 BasicBlock *Header = L->getHeader();
4499 BasicBlock *Pred = 0;
4500 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header);
4502 if (!L->contains(*PI)) {
4503 if (Pred && Pred != *PI) return 0; // Multiple predecessors.
4509 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4510 /// (which may not be an immediate predecessor) which has exactly one
4511 /// successor from which BB is reachable, or null if no such block is
4515 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4516 // If the block has a unique predecessor, then there is no path from the
4517 // predecessor to the block that does not go through the direct edge
4518 // from the predecessor to the block.
4519 if (BasicBlock *Pred = BB->getSinglePredecessor())
4522 // A loop's header is defined to be a block that dominates the loop.
4523 // If the header has a unique predecessor outside the loop, it must be
4524 // a block that has exactly one successor that can reach the loop.
4525 if (Loop *L = LI->getLoopFor(BB))
4526 return getLoopPredecessor(L);
4531 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4532 /// testing whether two expressions are equal, however for the purposes of
4533 /// looking for a condition guarding a loop, it can be useful to be a little
4534 /// more general, since a front-end may have replicated the controlling
4537 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4538 // Quick check to see if they are the same SCEV.
4539 if (A == B) return true;
4541 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4542 // two different instructions with the same value. Check for this case.
4543 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4544 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4545 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4546 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4547 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4550 // Otherwise assume they may have a different value.
4554 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
4555 return getSignedRange(S).getSignedMax().isNegative();
4558 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
4559 return getSignedRange(S).getSignedMin().isStrictlyPositive();
4562 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
4563 return !getSignedRange(S).getSignedMin().isNegative();
4566 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
4567 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
4570 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
4571 return isKnownNegative(S) || isKnownPositive(S);
4574 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
4575 const SCEV *LHS, const SCEV *RHS) {
4577 if (HasSameValue(LHS, RHS))
4578 return ICmpInst::isTrueWhenEqual(Pred);
4582 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4584 case ICmpInst::ICMP_SGT:
4585 Pred = ICmpInst::ICMP_SLT;
4586 std::swap(LHS, RHS);
4587 case ICmpInst::ICMP_SLT: {
4588 ConstantRange LHSRange = getSignedRange(LHS);
4589 ConstantRange RHSRange = getSignedRange(RHS);
4590 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
4592 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
4596 case ICmpInst::ICMP_SGE:
4597 Pred = ICmpInst::ICMP_SLE;
4598 std::swap(LHS, RHS);
4599 case ICmpInst::ICMP_SLE: {
4600 ConstantRange LHSRange = getSignedRange(LHS);
4601 ConstantRange RHSRange = getSignedRange(RHS);
4602 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
4604 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
4608 case ICmpInst::ICMP_UGT:
4609 Pred = ICmpInst::ICMP_ULT;
4610 std::swap(LHS, RHS);
4611 case ICmpInst::ICMP_ULT: {
4612 ConstantRange LHSRange = getUnsignedRange(LHS);
4613 ConstantRange RHSRange = getUnsignedRange(RHS);
4614 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
4616 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
4620 case ICmpInst::ICMP_UGE:
4621 Pred = ICmpInst::ICMP_ULE;
4622 std::swap(LHS, RHS);
4623 case ICmpInst::ICMP_ULE: {
4624 ConstantRange LHSRange = getUnsignedRange(LHS);
4625 ConstantRange RHSRange = getUnsignedRange(RHS);
4626 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
4628 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
4632 case ICmpInst::ICMP_NE: {
4633 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
4635 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
4638 const SCEV *Diff = getMinusSCEV(LHS, RHS);
4639 if (isKnownNonZero(Diff))
4643 case ICmpInst::ICMP_EQ:
4644 // The check at the top of the function catches the case where
4645 // the values are known to be equal.
4651 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
4652 /// protected by a conditional between LHS and RHS. This is used to
4653 /// to eliminate casts.
4655 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
4656 ICmpInst::Predicate Pred,
4657 const SCEV *LHS, const SCEV *RHS) {
4658 // Interpret a null as meaning no loop, where there is obviously no guard
4659 // (interprocedural conditions notwithstanding).
4660 if (!L) return true;
4662 BasicBlock *Latch = L->getLoopLatch();
4666 BranchInst *LoopContinuePredicate =
4667 dyn_cast<BranchInst>(Latch->getTerminator());
4668 if (!LoopContinuePredicate ||
4669 LoopContinuePredicate->isUnconditional())
4672 return isImpliedCond(LoopContinuePredicate->getCondition(), Pred, LHS, RHS,
4673 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
4676 /// isLoopGuardedByCond - Test whether entry to the loop is protected
4677 /// by a conditional between LHS and RHS. This is used to help avoid max
4678 /// expressions in loop trip counts, and to eliminate casts.
4680 ScalarEvolution::isLoopGuardedByCond(const Loop *L,
4681 ICmpInst::Predicate Pred,
4682 const SCEV *LHS, const SCEV *RHS) {
4683 // Interpret a null as meaning no loop, where there is obviously no guard
4684 // (interprocedural conditions notwithstanding).
4685 if (!L) return false;
4687 BasicBlock *Predecessor = getLoopPredecessor(L);
4688 BasicBlock *PredecessorDest = L->getHeader();
4690 // Starting at the loop predecessor, climb up the predecessor chain, as long
4691 // as there are predecessors that can be found that have unique successors
4692 // leading to the original header.
4694 PredecessorDest = Predecessor,
4695 Predecessor = getPredecessorWithUniqueSuccessorForBB(Predecessor)) {
4697 BranchInst *LoopEntryPredicate =
4698 dyn_cast<BranchInst>(Predecessor->getTerminator());
4699 if (!LoopEntryPredicate ||
4700 LoopEntryPredicate->isUnconditional())
4703 if (isImpliedCond(LoopEntryPredicate->getCondition(), Pred, LHS, RHS,
4704 LoopEntryPredicate->getSuccessor(0) != PredecessorDest))
4711 /// isImpliedCond - Test whether the condition described by Pred, LHS,
4712 /// and RHS is true whenever the given Cond value evaluates to true.
4713 bool ScalarEvolution::isImpliedCond(Value *CondValue,
4714 ICmpInst::Predicate Pred,
4715 const SCEV *LHS, const SCEV *RHS,
4717 // Recursivly handle And and Or conditions.
4718 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CondValue)) {
4719 if (BO->getOpcode() == Instruction::And) {
4721 return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
4722 isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
4723 } else if (BO->getOpcode() == Instruction::Or) {
4725 return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
4726 isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
4730 ICmpInst *ICI = dyn_cast<ICmpInst>(CondValue);
4731 if (!ICI) return false;
4733 // Bail if the ICmp's operands' types are wider than the needed type
4734 // before attempting to call getSCEV on them. This avoids infinite
4735 // recursion, since the analysis of widening casts can require loop
4736 // exit condition information for overflow checking, which would
4738 if (getTypeSizeInBits(LHS->getType()) <
4739 getTypeSizeInBits(ICI->getOperand(0)->getType()))
4742 // Now that we found a conditional branch that dominates the loop, check to
4743 // see if it is the comparison we are looking for.
4744 ICmpInst::Predicate FoundPred;
4746 FoundPred = ICI->getInversePredicate();
4748 FoundPred = ICI->getPredicate();
4750 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
4751 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
4753 // Balance the types. The case where FoundLHS' type is wider than
4754 // LHS' type is checked for above.
4755 if (getTypeSizeInBits(LHS->getType()) >
4756 getTypeSizeInBits(FoundLHS->getType())) {
4757 if (CmpInst::isSigned(Pred)) {
4758 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
4759 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
4761 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
4762 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
4766 // Canonicalize the query to match the way instcombine will have
4767 // canonicalized the comparison.
4768 // First, put a constant operand on the right.
4769 if (isa<SCEVConstant>(LHS)) {
4770 std::swap(LHS, RHS);
4771 Pred = ICmpInst::getSwappedPredicate(Pred);
4773 // Then, canonicalize comparisons with boundary cases.
4774 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
4775 const APInt &RA = RC->getValue()->getValue();
4777 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4778 case ICmpInst::ICMP_EQ:
4779 case ICmpInst::ICMP_NE:
4781 case ICmpInst::ICMP_UGE:
4782 if ((RA - 1).isMinValue()) {
4783 Pred = ICmpInst::ICMP_NE;
4784 RHS = getConstant(RA - 1);
4787 if (RA.isMaxValue()) {
4788 Pred = ICmpInst::ICMP_EQ;
4791 if (RA.isMinValue()) return true;
4793 case ICmpInst::ICMP_ULE:
4794 if ((RA + 1).isMaxValue()) {
4795 Pred = ICmpInst::ICMP_NE;
4796 RHS = getConstant(RA + 1);
4799 if (RA.isMinValue()) {
4800 Pred = ICmpInst::ICMP_EQ;
4803 if (RA.isMaxValue()) return true;
4805 case ICmpInst::ICMP_SGE:
4806 if ((RA - 1).isMinSignedValue()) {
4807 Pred = ICmpInst::ICMP_NE;
4808 RHS = getConstant(RA - 1);
4811 if (RA.isMaxSignedValue()) {
4812 Pred = ICmpInst::ICMP_EQ;
4815 if (RA.isMinSignedValue()) return true;
4817 case ICmpInst::ICMP_SLE:
4818 if ((RA + 1).isMaxSignedValue()) {
4819 Pred = ICmpInst::ICMP_NE;
4820 RHS = getConstant(RA + 1);
4823 if (RA.isMinSignedValue()) {
4824 Pred = ICmpInst::ICMP_EQ;
4827 if (RA.isMaxSignedValue()) return true;
4829 case ICmpInst::ICMP_UGT:
4830 if (RA.isMinValue()) {
4831 Pred = ICmpInst::ICMP_NE;
4834 if ((RA + 1).isMaxValue()) {
4835 Pred = ICmpInst::ICMP_EQ;
4836 RHS = getConstant(RA + 1);
4839 if (RA.isMaxValue()) return false;
4841 case ICmpInst::ICMP_ULT:
4842 if (RA.isMaxValue()) {
4843 Pred = ICmpInst::ICMP_NE;
4846 if ((RA - 1).isMinValue()) {
4847 Pred = ICmpInst::ICMP_EQ;
4848 RHS = getConstant(RA - 1);
4851 if (RA.isMinValue()) return false;
4853 case ICmpInst::ICMP_SGT:
4854 if (RA.isMinSignedValue()) {
4855 Pred = ICmpInst::ICMP_NE;
4858 if ((RA + 1).isMaxSignedValue()) {
4859 Pred = ICmpInst::ICMP_EQ;
4860 RHS = getConstant(RA + 1);
4863 if (RA.isMaxSignedValue()) return false;
4865 case ICmpInst::ICMP_SLT:
4866 if (RA.isMaxSignedValue()) {
4867 Pred = ICmpInst::ICMP_NE;
4870 if ((RA - 1).isMinSignedValue()) {
4871 Pred = ICmpInst::ICMP_EQ;
4872 RHS = getConstant(RA - 1);
4875 if (RA.isMinSignedValue()) return false;
4880 // Check to see if we can make the LHS or RHS match.
4881 if (LHS == FoundRHS || RHS == FoundLHS) {
4882 if (isa<SCEVConstant>(RHS)) {
4883 std::swap(FoundLHS, FoundRHS);
4884 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
4886 std::swap(LHS, RHS);
4887 Pred = ICmpInst::getSwappedPredicate(Pred);
4891 // Check whether the found predicate is the same as the desired predicate.
4892 if (FoundPred == Pred)
4893 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
4895 // Check whether swapping the found predicate makes it the same as the
4896 // desired predicate.
4897 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
4898 if (isa<SCEVConstant>(RHS))
4899 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
4901 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
4902 RHS, LHS, FoundLHS, FoundRHS);
4905 // Check whether the actual condition is beyond sufficient.
4906 if (FoundPred == ICmpInst::ICMP_EQ)
4907 if (ICmpInst::isTrueWhenEqual(Pred))
4908 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
4910 if (Pred == ICmpInst::ICMP_NE)
4911 if (!ICmpInst::isTrueWhenEqual(FoundPred))
4912 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
4915 // Otherwise assume the worst.
4919 /// isImpliedCondOperands - Test whether the condition described by Pred,
4920 /// LHS, and RHS is true whenever the condition desribed by Pred, FoundLHS,
4921 /// and FoundRHS is true.
4922 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
4923 const SCEV *LHS, const SCEV *RHS,
4924 const SCEV *FoundLHS,
4925 const SCEV *FoundRHS) {
4926 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
4927 FoundLHS, FoundRHS) ||
4928 // ~x < ~y --> x > y
4929 isImpliedCondOperandsHelper(Pred, LHS, RHS,
4930 getNotSCEV(FoundRHS),
4931 getNotSCEV(FoundLHS));
4934 /// isImpliedCondOperandsHelper - Test whether the condition described by
4935 /// Pred, LHS, and RHS is true whenever the condition desribed by Pred,
4936 /// FoundLHS, and FoundRHS is true.
4938 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
4939 const SCEV *LHS, const SCEV *RHS,
4940 const SCEV *FoundLHS,
4941 const SCEV *FoundRHS) {
4943 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4944 case ICmpInst::ICMP_EQ:
4945 case ICmpInst::ICMP_NE:
4946 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
4949 case ICmpInst::ICMP_SLT:
4950 case ICmpInst::ICMP_SLE:
4951 if (isKnownPredicate(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
4952 isKnownPredicate(ICmpInst::ICMP_SGE, RHS, FoundRHS))
4955 case ICmpInst::ICMP_SGT:
4956 case ICmpInst::ICMP_SGE:
4957 if (isKnownPredicate(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
4958 isKnownPredicate(ICmpInst::ICMP_SLE, RHS, FoundRHS))
4961 case ICmpInst::ICMP_ULT:
4962 case ICmpInst::ICMP_ULE:
4963 if (isKnownPredicate(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
4964 isKnownPredicate(ICmpInst::ICMP_UGE, RHS, FoundRHS))
4967 case ICmpInst::ICMP_UGT:
4968 case ICmpInst::ICMP_UGE:
4969 if (isKnownPredicate(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
4970 isKnownPredicate(ICmpInst::ICMP_ULE, RHS, FoundRHS))
4978 /// getBECount - Subtract the end and start values and divide by the step,
4979 /// rounding up, to get the number of times the backedge is executed. Return
4980 /// CouldNotCompute if an intermediate computation overflows.
4981 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
4985 assert(!isKnownNegative(Step) &&
4986 "This code doesn't handle negative strides yet!");
4988 const Type *Ty = Start->getType();
4989 const SCEV *NegOne = getIntegerSCEV(-1, Ty);
4990 const SCEV *Diff = getMinusSCEV(End, Start);
4991 const SCEV *RoundUp = getAddExpr(Step, NegOne);
4993 // Add an adjustment to the difference between End and Start so that
4994 // the division will effectively round up.
4995 const SCEV *Add = getAddExpr(Diff, RoundUp);
4998 // Check Add for unsigned overflow.
4999 // TODO: More sophisticated things could be done here.
5000 const Type *WideTy = IntegerType::get(getContext(),
5001 getTypeSizeInBits(Ty) + 1);
5002 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
5003 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
5004 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
5005 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
5006 return getCouldNotCompute();
5009 return getUDivExpr(Add, Step);
5012 /// HowManyLessThans - Return the number of times a backedge containing the
5013 /// specified less-than comparison will execute. If not computable, return
5014 /// CouldNotCompute.
5015 ScalarEvolution::BackedgeTakenInfo
5016 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
5017 const Loop *L, bool isSigned) {
5018 // Only handle: "ADDREC < LoopInvariant".
5019 if (!RHS->isLoopInvariant(L)) return getCouldNotCompute();
5021 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
5022 if (!AddRec || AddRec->getLoop() != L)
5023 return getCouldNotCompute();
5025 // Check to see if we have a flag which makes analysis easy.
5026 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
5027 AddRec->hasNoUnsignedWrap();
5029 if (AddRec->isAffine()) {
5030 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
5031 const SCEV *Step = AddRec->getStepRecurrence(*this);
5034 return getCouldNotCompute();
5035 if (Step->isOne()) {
5036 // With unit stride, the iteration never steps past the limit value.
5037 } else if (isKnownPositive(Step)) {
5038 // Test whether a positive iteration iteration can step past the limit
5039 // value and past the maximum value for its type in a single step.
5040 // Note that it's not sufficient to check NoWrap here, because even
5041 // though the value after a wrap is undefined, it's not undefined
5042 // behavior, so if wrap does occur, the loop could either terminate or
5043 // loop infinitely, but in either case, the loop is guaranteed to
5044 // iterate at least until the iteration where the wrapping occurs.
5045 const SCEV *One = getIntegerSCEV(1, Step->getType());
5047 APInt Max = APInt::getSignedMaxValue(BitWidth);
5048 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
5049 .slt(getSignedRange(RHS).getSignedMax()))
5050 return getCouldNotCompute();
5052 APInt Max = APInt::getMaxValue(BitWidth);
5053 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
5054 .ult(getUnsignedRange(RHS).getUnsignedMax()))
5055 return getCouldNotCompute();
5058 // TODO: Handle negative strides here and below.
5059 return getCouldNotCompute();
5061 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
5062 // m. So, we count the number of iterations in which {n,+,s} < m is true.
5063 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
5064 // treat m-n as signed nor unsigned due to overflow possibility.
5066 // First, we get the value of the LHS in the first iteration: n
5067 const SCEV *Start = AddRec->getOperand(0);
5069 // Determine the minimum constant start value.
5070 const SCEV *MinStart = getConstant(isSigned ?
5071 getSignedRange(Start).getSignedMin() :
5072 getUnsignedRange(Start).getUnsignedMin());
5074 // If we know that the condition is true in order to enter the loop,
5075 // then we know that it will run exactly (m-n)/s times. Otherwise, we
5076 // only know that it will execute (max(m,n)-n)/s times. In both cases,
5077 // the division must round up.
5078 const SCEV *End = RHS;
5079 if (!isLoopGuardedByCond(L,
5080 isSigned ? ICmpInst::ICMP_SLT :
5082 getMinusSCEV(Start, Step), RHS))
5083 End = isSigned ? getSMaxExpr(RHS, Start)
5084 : getUMaxExpr(RHS, Start);
5086 // Determine the maximum constant end value.
5087 const SCEV *MaxEnd = getConstant(isSigned ?
5088 getSignedRange(End).getSignedMax() :
5089 getUnsignedRange(End).getUnsignedMax());
5091 // If MaxEnd is within a step of the maximum integer value in its type,
5092 // adjust it down to the minimum value which would produce the same effect.
5093 // This allows the subsequent ceiling divison of (N+(step-1))/step to
5094 // compute the correct value.
5095 const SCEV *StepMinusOne = getMinusSCEV(Step,
5096 getIntegerSCEV(1, Step->getType()));
5099 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
5102 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
5105 // Finally, we subtract these two values and divide, rounding up, to get
5106 // the number of times the backedge is executed.
5107 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
5109 // The maximum backedge count is similar, except using the minimum start
5110 // value and the maximum end value.
5111 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
5113 return BackedgeTakenInfo(BECount, MaxBECount);
5116 return getCouldNotCompute();
5119 /// getNumIterationsInRange - Return the number of iterations of this loop that
5120 /// produce values in the specified constant range. Another way of looking at
5121 /// this is that it returns the first iteration number where the value is not in
5122 /// the condition, thus computing the exit count. If the iteration count can't
5123 /// be computed, an instance of SCEVCouldNotCompute is returned.
5124 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
5125 ScalarEvolution &SE) const {
5126 if (Range.isFullSet()) // Infinite loop.
5127 return SE.getCouldNotCompute();
5129 // If the start is a non-zero constant, shift the range to simplify things.
5130 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
5131 if (!SC->getValue()->isZero()) {
5132 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
5133 Operands[0] = SE.getIntegerSCEV(0, SC->getType());
5134 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
5135 if (const SCEVAddRecExpr *ShiftedAddRec =
5136 dyn_cast<SCEVAddRecExpr>(Shifted))
5137 return ShiftedAddRec->getNumIterationsInRange(
5138 Range.subtract(SC->getValue()->getValue()), SE);
5139 // This is strange and shouldn't happen.
5140 return SE.getCouldNotCompute();
5143 // The only time we can solve this is when we have all constant indices.
5144 // Otherwise, we cannot determine the overflow conditions.
5145 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5146 if (!isa<SCEVConstant>(getOperand(i)))
5147 return SE.getCouldNotCompute();
5150 // Okay at this point we know that all elements of the chrec are constants and
5151 // that the start element is zero.
5153 // First check to see if the range contains zero. If not, the first
5155 unsigned BitWidth = SE.getTypeSizeInBits(getType());
5156 if (!Range.contains(APInt(BitWidth, 0)))
5157 return SE.getIntegerSCEV(0, getType());
5160 // If this is an affine expression then we have this situation:
5161 // Solve {0,+,A} in Range === Ax in Range
5163 // We know that zero is in the range. If A is positive then we know that
5164 // the upper value of the range must be the first possible exit value.
5165 // If A is negative then the lower of the range is the last possible loop
5166 // value. Also note that we already checked for a full range.
5167 APInt One(BitWidth,1);
5168 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5169 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5171 // The exit value should be (End+A)/A.
5172 APInt ExitVal = (End + A).udiv(A);
5173 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5175 // Evaluate at the exit value. If we really did fall out of the valid
5176 // range, then we computed our trip count, otherwise wrap around or other
5177 // things must have happened.
5178 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5179 if (Range.contains(Val->getValue()))
5180 return SE.getCouldNotCompute(); // Something strange happened
5182 // Ensure that the previous value is in the range. This is a sanity check.
5183 assert(Range.contains(
5184 EvaluateConstantChrecAtConstant(this,
5185 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5186 "Linear scev computation is off in a bad way!");
5187 return SE.getConstant(ExitValue);
5188 } else if (isQuadratic()) {
5189 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5190 // quadratic equation to solve it. To do this, we must frame our problem in
5191 // terms of figuring out when zero is crossed, instead of when
5192 // Range.getUpper() is crossed.
5193 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5194 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5195 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5197 // Next, solve the constructed addrec
5198 std::pair<const SCEV *,const SCEV *> Roots =
5199 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5200 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5201 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5203 // Pick the smallest positive root value.
5204 if (ConstantInt *CB =
5205 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5206 R1->getValue(), R2->getValue()))) {
5207 if (CB->getZExtValue() == false)
5208 std::swap(R1, R2); // R1 is the minimum root now.
5210 // Make sure the root is not off by one. The returned iteration should
5211 // not be in the range, but the previous one should be. When solving
5212 // for "X*X < 5", for example, we should not return a root of 2.
5213 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5216 if (Range.contains(R1Val->getValue())) {
5217 // The next iteration must be out of the range...
5218 ConstantInt *NextVal =
5219 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5221 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5222 if (!Range.contains(R1Val->getValue()))
5223 return SE.getConstant(NextVal);
5224 return SE.getCouldNotCompute(); // Something strange happened
5227 // If R1 was not in the range, then it is a good return value. Make
5228 // sure that R1-1 WAS in the range though, just in case.
5229 ConstantInt *NextVal =
5230 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5231 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5232 if (Range.contains(R1Val->getValue()))
5234 return SE.getCouldNotCompute(); // Something strange happened
5239 return SE.getCouldNotCompute();
5244 //===----------------------------------------------------------------------===//
5245 // SCEVCallbackVH Class Implementation
5246 //===----------------------------------------------------------------------===//
5248 void ScalarEvolution::SCEVCallbackVH::deleted() {
5249 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5250 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5251 SE->ConstantEvolutionLoopExitValue.erase(PN);
5252 SE->Scalars.erase(getValPtr());
5253 // this now dangles!
5256 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *) {
5257 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5259 // Forget all the expressions associated with users of the old value,
5260 // so that future queries will recompute the expressions using the new
5262 SmallVector<User *, 16> Worklist;
5263 SmallPtrSet<User *, 8> Visited;
5264 Value *Old = getValPtr();
5265 bool DeleteOld = false;
5266 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5268 Worklist.push_back(*UI);
5269 while (!Worklist.empty()) {
5270 User *U = Worklist.pop_back_val();
5271 // Deleting the Old value will cause this to dangle. Postpone
5272 // that until everything else is done.
5277 if (!Visited.insert(U))
5279 if (PHINode *PN = dyn_cast<PHINode>(U))
5280 SE->ConstantEvolutionLoopExitValue.erase(PN);
5281 SE->Scalars.erase(U);
5282 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5284 Worklist.push_back(*UI);
5286 // Delete the Old value if it (indirectly) references itself.
5288 if (PHINode *PN = dyn_cast<PHINode>(Old))
5289 SE->ConstantEvolutionLoopExitValue.erase(PN);
5290 SE->Scalars.erase(Old);
5291 // this now dangles!
5296 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5297 : CallbackVH(V), SE(se) {}
5299 //===----------------------------------------------------------------------===//
5300 // ScalarEvolution Class Implementation
5301 //===----------------------------------------------------------------------===//
5303 ScalarEvolution::ScalarEvolution()
5304 : FunctionPass(&ID) {
5307 bool ScalarEvolution::runOnFunction(Function &F) {
5309 LI = &getAnalysis<LoopInfo>();
5310 DT = &getAnalysis<DominatorTree>();
5311 TD = getAnalysisIfAvailable<TargetData>();
5315 void ScalarEvolution::releaseMemory() {
5317 BackedgeTakenCounts.clear();
5318 ConstantEvolutionLoopExitValue.clear();
5319 ValuesAtScopes.clear();
5320 UniqueSCEVs.clear();
5321 SCEVAllocator.Reset();
5324 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5325 AU.setPreservesAll();
5326 AU.addRequiredTransitive<LoopInfo>();
5327 AU.addRequiredTransitive<DominatorTree>();
5330 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5331 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5334 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5336 // Print all inner loops first
5337 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5338 PrintLoopInfo(OS, SE, *I);
5341 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5344 SmallVector<BasicBlock *, 8> ExitBlocks;
5345 L->getExitBlocks(ExitBlocks);
5346 if (ExitBlocks.size() != 1)
5347 OS << "<multiple exits> ";
5349 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5350 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5352 OS << "Unpredictable backedge-taken count. ";
5357 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5360 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5361 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5363 OS << "Unpredictable max backedge-taken count. ";
5369 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
5370 // ScalarEvolution's implementaiton of the print method is to print
5371 // out SCEV values of all instructions that are interesting. Doing
5372 // this potentially causes it to create new SCEV objects though,
5373 // which technically conflicts with the const qualifier. This isn't
5374 // observable from outside the class though, so casting away the
5375 // const isn't dangerous.
5376 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
5378 OS << "Classifying expressions for: ";
5379 WriteAsOperand(OS, F, /*PrintType=*/false);
5381 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
5382 if (isSCEVable(I->getType())) {
5385 const SCEV *SV = SE.getSCEV(&*I);
5388 const Loop *L = LI->getLoopFor((*I).getParent());
5390 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
5397 OS << "\t\t" "Exits: ";
5398 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
5399 if (!ExitValue->isLoopInvariant(L)) {
5400 OS << "<<Unknown>>";
5409 OS << "Determining loop execution counts for: ";
5410 WriteAsOperand(OS, F, /*PrintType=*/false);
5412 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
5413 PrintLoopInfo(OS, &SE, *I);