1 //===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the implementation of the scalar evolution analysis
11 // engine, which is used primarily to analyze expressions involving induction
12 // variables in loops.
14 // There are several aspects to this library. First is the representation of
15 // scalar expressions, which are represented as subclasses of the SCEV class.
16 // These classes are used to represent certain types of subexpressions that we
17 // can handle. We only create one SCEV of a particular shape, so
18 // pointer-comparisons for equality are legal.
20 // One important aspect of the SCEV objects is that they are never cyclic, even
21 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
22 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
23 // recurrence) then we represent it directly as a recurrence node, otherwise we
24 // represent it as a SCEVUnknown node.
26 // In addition to being able to represent expressions of various types, we also
27 // have folders that are used to build the *canonical* representation for a
28 // particular expression. These folders are capable of using a variety of
29 // rewrite rules to simplify the expressions.
31 // Once the folders are defined, we can implement the more interesting
32 // higher-level code, such as the code that recognizes PHI nodes of various
33 // types, computes the execution count of a loop, etc.
35 // TODO: We should use these routines and value representations to implement
36 // dependence analysis!
38 //===----------------------------------------------------------------------===//
40 // There are several good references for the techniques used in this analysis.
42 // Chains of recurrences -- a method to expedite the evaluation
43 // of closed-form functions
44 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
46 // On computational properties of chains of recurrences
49 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
50 // Robert A. van Engelen
52 // Efficient Symbolic Analysis for Optimizing Compilers
53 // Robert A. van Engelen
55 // Using the chains of recurrences algebra for data dependence testing and
56 // induction variable substitution
57 // MS Thesis, Johnie Birch
59 //===----------------------------------------------------------------------===//
61 #define DEBUG_TYPE "scalar-evolution"
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/Constants.h"
64 #include "llvm/DerivedTypes.h"
65 #include "llvm/GlobalVariable.h"
66 #include "llvm/GlobalAlias.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/LLVMContext.h"
69 #include "llvm/Operator.h"
70 #include "llvm/Analysis/ConstantFolding.h"
71 #include "llvm/Analysis/Dominators.h"
72 #include "llvm/Analysis/InstructionSimplify.h"
73 #include "llvm/Analysis/LoopInfo.h"
74 #include "llvm/Analysis/ValueTracking.h"
75 #include "llvm/Assembly/Writer.h"
76 #include "llvm/Target/TargetData.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/ConstantRange.h"
79 #include "llvm/Support/Debug.h"
80 #include "llvm/Support/ErrorHandling.h"
81 #include "llvm/Support/GetElementPtrTypeIterator.h"
82 #include "llvm/Support/InstIterator.h"
83 #include "llvm/Support/MathExtras.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/ADT/Statistic.h"
86 #include "llvm/ADT/STLExtras.h"
87 #include "llvm/ADT/SmallPtrSet.h"
91 STATISTIC(NumArrayLenItCounts,
92 "Number of trip counts computed with array length");
93 STATISTIC(NumTripCountsComputed,
94 "Number of loops with predictable loop counts");
95 STATISTIC(NumTripCountsNotComputed,
96 "Number of loops without predictable loop counts");
97 STATISTIC(NumBruteForceTripCountsComputed,
98 "Number of loops with trip counts computed by force");
100 static cl::opt<unsigned>
101 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
102 cl::desc("Maximum number of iterations SCEV will "
103 "symbolically execute a constant "
107 INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
108 "Scalar Evolution Analysis", false, true)
109 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
110 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
111 INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
112 "Scalar Evolution Analysis", false, true)
113 char ScalarEvolution::ID = 0;
115 //===----------------------------------------------------------------------===//
116 // SCEV class definitions
117 //===----------------------------------------------------------------------===//
119 //===----------------------------------------------------------------------===//
120 // Implementation of the SCEV class.
123 void SCEV::dump() const {
128 void SCEV::print(raw_ostream &OS) const {
129 switch (getSCEVType()) {
131 WriteAsOperand(OS, cast<SCEVConstant>(this)->getValue(), false);
134 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(this);
135 const SCEV *Op = Trunc->getOperand();
136 OS << "(trunc " << *Op->getType() << " " << *Op << " to "
137 << *Trunc->getType() << ")";
141 const SCEVZeroExtendExpr *ZExt = cast<SCEVZeroExtendExpr>(this);
142 const SCEV *Op = ZExt->getOperand();
143 OS << "(zext " << *Op->getType() << " " << *Op << " to "
144 << *ZExt->getType() << ")";
148 const SCEVSignExtendExpr *SExt = cast<SCEVSignExtendExpr>(this);
149 const SCEV *Op = SExt->getOperand();
150 OS << "(sext " << *Op->getType() << " " << *Op << " to "
151 << *SExt->getType() << ")";
155 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(this);
156 OS << "{" << *AR->getOperand(0);
157 for (unsigned i = 1, e = AR->getNumOperands(); i != e; ++i)
158 OS << ",+," << *AR->getOperand(i);
160 if (AR->hasNoUnsignedWrap())
162 if (AR->hasNoSignedWrap())
164 WriteAsOperand(OS, AR->getLoop()->getHeader(), /*PrintType=*/false);
172 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(this);
173 const char *OpStr = 0;
174 switch (NAry->getSCEVType()) {
175 case scAddExpr: OpStr = " + "; break;
176 case scMulExpr: OpStr = " * "; break;
177 case scUMaxExpr: OpStr = " umax "; break;
178 case scSMaxExpr: OpStr = " smax "; break;
181 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
184 if (llvm::next(I) != E)
191 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(this);
192 OS << "(" << *UDiv->getLHS() << " /u " << *UDiv->getRHS() << ")";
196 const SCEVUnknown *U = cast<SCEVUnknown>(this);
198 if (U->isSizeOf(AllocTy)) {
199 OS << "sizeof(" << *AllocTy << ")";
202 if (U->isAlignOf(AllocTy)) {
203 OS << "alignof(" << *AllocTy << ")";
209 if (U->isOffsetOf(CTy, FieldNo)) {
210 OS << "offsetof(" << *CTy << ", ";
211 WriteAsOperand(OS, FieldNo, false);
216 // Otherwise just print it normally.
217 WriteAsOperand(OS, U->getValue(), false);
220 case scCouldNotCompute:
221 OS << "***COULDNOTCOMPUTE***";
225 llvm_unreachable("Unknown SCEV kind!");
228 const Type *SCEV::getType() const {
229 switch (getSCEVType()) {
231 return cast<SCEVConstant>(this)->getType();
235 return cast<SCEVCastExpr>(this)->getType();
240 return cast<SCEVNAryExpr>(this)->getType();
242 return cast<SCEVAddExpr>(this)->getType();
244 return cast<SCEVUDivExpr>(this)->getType();
246 return cast<SCEVUnknown>(this)->getType();
247 case scCouldNotCompute:
248 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
252 llvm_unreachable("Unknown SCEV kind!");
256 bool SCEV::isZero() const {
257 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
258 return SC->getValue()->isZero();
262 bool SCEV::isOne() const {
263 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
264 return SC->getValue()->isOne();
268 bool SCEV::isAllOnesValue() const {
269 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
270 return SC->getValue()->isAllOnesValue();
274 SCEVCouldNotCompute::SCEVCouldNotCompute() :
275 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
277 bool SCEVCouldNotCompute::classof(const SCEV *S) {
278 return S->getSCEVType() == scCouldNotCompute;
281 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
283 ID.AddInteger(scConstant);
286 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
287 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V);
288 UniqueSCEVs.InsertNode(S, IP);
292 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
293 return getConstant(ConstantInt::get(getContext(), Val));
297 ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
298 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
299 return getConstant(ConstantInt::get(ITy, V, isSigned));
302 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
303 unsigned SCEVTy, const SCEV *op, const Type *ty)
304 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
306 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
307 const SCEV *op, const Type *ty)
308 : SCEVCastExpr(ID, scTruncate, op, ty) {
309 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
310 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
311 "Cannot truncate non-integer value!");
314 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
315 const SCEV *op, const Type *ty)
316 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
317 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
318 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
319 "Cannot zero extend non-integer value!");
322 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
323 const SCEV *op, const Type *ty)
324 : SCEVCastExpr(ID, scSignExtend, op, ty) {
325 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
326 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
327 "Cannot sign extend non-integer value!");
330 void SCEVUnknown::deleted() {
331 // Clear this SCEVUnknown from various maps.
332 SE->forgetMemoizedResults(this);
334 // Remove this SCEVUnknown from the uniquing map.
335 SE->UniqueSCEVs.RemoveNode(this);
337 // Release the value.
341 void SCEVUnknown::allUsesReplacedWith(Value *New) {
342 // Clear this SCEVUnknown from various maps.
343 SE->forgetMemoizedResults(this);
345 // Remove this SCEVUnknown from the uniquing map.
346 SE->UniqueSCEVs.RemoveNode(this);
348 // Update this SCEVUnknown to point to the new value. This is needed
349 // because there may still be outstanding SCEVs which still point to
354 bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
355 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
356 if (VCE->getOpcode() == Instruction::PtrToInt)
357 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
358 if (CE->getOpcode() == Instruction::GetElementPtr &&
359 CE->getOperand(0)->isNullValue() &&
360 CE->getNumOperands() == 2)
361 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
363 AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
371 bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
372 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
373 if (VCE->getOpcode() == Instruction::PtrToInt)
374 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
375 if (CE->getOpcode() == Instruction::GetElementPtr &&
376 CE->getOperand(0)->isNullValue()) {
378 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
379 if (const StructType *STy = dyn_cast<StructType>(Ty))
380 if (!STy->isPacked() &&
381 CE->getNumOperands() == 3 &&
382 CE->getOperand(1)->isNullValue()) {
383 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
385 STy->getNumElements() == 2 &&
386 STy->getElementType(0)->isIntegerTy(1)) {
387 AllocTy = STy->getElementType(1);
396 bool SCEVUnknown::isOffsetOf(const Type *&CTy, Constant *&FieldNo) const {
397 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
398 if (VCE->getOpcode() == Instruction::PtrToInt)
399 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
400 if (CE->getOpcode() == Instruction::GetElementPtr &&
401 CE->getNumOperands() == 3 &&
402 CE->getOperand(0)->isNullValue() &&
403 CE->getOperand(1)->isNullValue()) {
405 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
406 // Ignore vector types here so that ScalarEvolutionExpander doesn't
407 // emit getelementptrs that index into vectors.
408 if (Ty->isStructTy() || Ty->isArrayTy()) {
410 FieldNo = CE->getOperand(2);
418 //===----------------------------------------------------------------------===//
420 //===----------------------------------------------------------------------===//
423 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
424 /// than the complexity of the RHS. This comparator is used to canonicalize
426 class SCEVComplexityCompare {
427 const LoopInfo *const LI;
429 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
431 // Return true or false if LHS is less than, or at least RHS, respectively.
432 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
433 return compare(LHS, RHS) < 0;
436 // Return negative, zero, or positive, if LHS is less than, equal to, or
437 // greater than RHS, respectively. A three-way result allows recursive
438 // comparisons to be more efficient.
439 int compare(const SCEV *LHS, const SCEV *RHS) const {
440 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
444 // Primarily, sort the SCEVs by their getSCEVType().
445 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
447 return (int)LType - (int)RType;
449 // Aside from the getSCEVType() ordering, the particular ordering
450 // isn't very important except that it's beneficial to be consistent,
451 // so that (a + b) and (b + a) don't end up as different expressions.
454 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS);
455 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
457 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
458 // not as complete as it could be.
459 const Value *LV = LU->getValue(), *RV = RU->getValue();
461 // Order pointer values after integer values. This helps SCEVExpander
463 bool LIsPointer = LV->getType()->isPointerTy(),
464 RIsPointer = RV->getType()->isPointerTy();
465 if (LIsPointer != RIsPointer)
466 return (int)LIsPointer - (int)RIsPointer;
468 // Compare getValueID values.
469 unsigned LID = LV->getValueID(),
470 RID = RV->getValueID();
472 return (int)LID - (int)RID;
474 // Sort arguments by their position.
475 if (const Argument *LA = dyn_cast<Argument>(LV)) {
476 const Argument *RA = cast<Argument>(RV);
477 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo();
478 return (int)LArgNo - (int)RArgNo;
481 // For instructions, compare their loop depth, and their operand
482 // count. This is pretty loose.
483 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) {
484 const Instruction *RInst = cast<Instruction>(RV);
486 // Compare loop depths.
487 const BasicBlock *LParent = LInst->getParent(),
488 *RParent = RInst->getParent();
489 if (LParent != RParent) {
490 unsigned LDepth = LI->getLoopDepth(LParent),
491 RDepth = LI->getLoopDepth(RParent);
492 if (LDepth != RDepth)
493 return (int)LDepth - (int)RDepth;
496 // Compare the number of operands.
497 unsigned LNumOps = LInst->getNumOperands(),
498 RNumOps = RInst->getNumOperands();
499 return (int)LNumOps - (int)RNumOps;
506 const SCEVConstant *LC = cast<SCEVConstant>(LHS);
507 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
509 // Compare constant values.
510 const APInt &LA = LC->getValue()->getValue();
511 const APInt &RA = RC->getValue()->getValue();
512 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth();
513 if (LBitWidth != RBitWidth)
514 return (int)LBitWidth - (int)RBitWidth;
515 return LA.ult(RA) ? -1 : 1;
519 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS);
520 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
522 // Compare addrec loop depths.
523 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop();
524 if (LLoop != RLoop) {
525 unsigned LDepth = LLoop->getLoopDepth(),
526 RDepth = RLoop->getLoopDepth();
527 if (LDepth != RDepth)
528 return (int)LDepth - (int)RDepth;
531 // Addrec complexity grows with operand count.
532 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands();
533 if (LNumOps != RNumOps)
534 return (int)LNumOps - (int)RNumOps;
536 // Lexicographically compare.
537 for (unsigned i = 0; i != LNumOps; ++i) {
538 long X = compare(LA->getOperand(i), RA->getOperand(i));
550 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS);
551 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
553 // Lexicographically compare n-ary expressions.
554 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
555 for (unsigned i = 0; i != LNumOps; ++i) {
558 long X = compare(LC->getOperand(i), RC->getOperand(i));
562 return (int)LNumOps - (int)RNumOps;
566 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
567 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
569 // Lexicographically compare udiv expressions.
570 long X = compare(LC->getLHS(), RC->getLHS());
573 return compare(LC->getRHS(), RC->getRHS());
579 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
580 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
582 // Compare cast expressions by operand.
583 return compare(LC->getOperand(), RC->getOperand());
590 llvm_unreachable("Unknown SCEV kind!");
596 /// GroupByComplexity - Given a list of SCEV objects, order them by their
597 /// complexity, and group objects of the same complexity together by value.
598 /// When this routine is finished, we know that any duplicates in the vector are
599 /// consecutive and that complexity is monotonically increasing.
601 /// Note that we go take special precautions to ensure that we get deterministic
602 /// results from this routine. In other words, we don't want the results of
603 /// this to depend on where the addresses of various SCEV objects happened to
606 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
608 if (Ops.size() < 2) return; // Noop
609 if (Ops.size() == 2) {
610 // This is the common case, which also happens to be trivially simple.
612 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
613 if (SCEVComplexityCompare(LI)(RHS, LHS))
618 // Do the rough sort by complexity.
619 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
621 // Now that we are sorted by complexity, group elements of the same
622 // complexity. Note that this is, at worst, N^2, but the vector is likely to
623 // be extremely short in practice. Note that we take this approach because we
624 // do not want to depend on the addresses of the objects we are grouping.
625 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
626 const SCEV *S = Ops[i];
627 unsigned Complexity = S->getSCEVType();
629 // If there are any objects of the same complexity and same value as this
631 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
632 if (Ops[j] == S) { // Found a duplicate.
633 // Move it to immediately after i'th element.
634 std::swap(Ops[i+1], Ops[j]);
635 ++i; // no need to rescan it.
636 if (i == e-2) return; // Done!
644 //===----------------------------------------------------------------------===//
645 // Simple SCEV method implementations
646 //===----------------------------------------------------------------------===//
648 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
650 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
652 const Type* ResultTy) {
653 // Handle the simplest case efficiently.
655 return SE.getTruncateOrZeroExtend(It, ResultTy);
657 // We are using the following formula for BC(It, K):
659 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
661 // Suppose, W is the bitwidth of the return value. We must be prepared for
662 // overflow. Hence, we must assure that the result of our computation is
663 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
664 // safe in modular arithmetic.
666 // However, this code doesn't use exactly that formula; the formula it uses
667 // is something like the following, where T is the number of factors of 2 in
668 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
671 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
673 // This formula is trivially equivalent to the previous formula. However,
674 // this formula can be implemented much more efficiently. The trick is that
675 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
676 // arithmetic. To do exact division in modular arithmetic, all we have
677 // to do is multiply by the inverse. Therefore, this step can be done at
680 // The next issue is how to safely do the division by 2^T. The way this
681 // is done is by doing the multiplication step at a width of at least W + T
682 // bits. This way, the bottom W+T bits of the product are accurate. Then,
683 // when we perform the division by 2^T (which is equivalent to a right shift
684 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
685 // truncated out after the division by 2^T.
687 // In comparison to just directly using the first formula, this technique
688 // is much more efficient; using the first formula requires W * K bits,
689 // but this formula less than W + K bits. Also, the first formula requires
690 // a division step, whereas this formula only requires multiplies and shifts.
692 // It doesn't matter whether the subtraction step is done in the calculation
693 // width or the input iteration count's width; if the subtraction overflows,
694 // the result must be zero anyway. We prefer here to do it in the width of
695 // the induction variable because it helps a lot for certain cases; CodeGen
696 // isn't smart enough to ignore the overflow, which leads to much less
697 // efficient code if the width of the subtraction is wider than the native
700 // (It's possible to not widen at all by pulling out factors of 2 before
701 // the multiplication; for example, K=2 can be calculated as
702 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
703 // extra arithmetic, so it's not an obvious win, and it gets
704 // much more complicated for K > 3.)
706 // Protection from insane SCEVs; this bound is conservative,
707 // but it probably doesn't matter.
709 return SE.getCouldNotCompute();
711 unsigned W = SE.getTypeSizeInBits(ResultTy);
713 // Calculate K! / 2^T and T; we divide out the factors of two before
714 // multiplying for calculating K! / 2^T to avoid overflow.
715 // Other overflow doesn't matter because we only care about the bottom
716 // W bits of the result.
717 APInt OddFactorial(W, 1);
719 for (unsigned i = 3; i <= K; ++i) {
721 unsigned TwoFactors = Mult.countTrailingZeros();
723 Mult = Mult.lshr(TwoFactors);
724 OddFactorial *= Mult;
727 // We need at least W + T bits for the multiplication step
728 unsigned CalculationBits = W + T;
730 // Calculate 2^T, at width T+W.
731 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
733 // Calculate the multiplicative inverse of K! / 2^T;
734 // this multiplication factor will perform the exact division by
736 APInt Mod = APInt::getSignedMinValue(W+1);
737 APInt MultiplyFactor = OddFactorial.zext(W+1);
738 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
739 MultiplyFactor = MultiplyFactor.trunc(W);
741 // Calculate the product, at width T+W
742 const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
744 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
745 for (unsigned i = 1; i != K; ++i) {
746 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
747 Dividend = SE.getMulExpr(Dividend,
748 SE.getTruncateOrZeroExtend(S, CalculationTy));
752 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
754 // Truncate the result, and divide by K! / 2^T.
756 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
757 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
760 /// evaluateAtIteration - Return the value of this chain of recurrences at
761 /// the specified iteration number. We can evaluate this recurrence by
762 /// multiplying each element in the chain by the binomial coefficient
763 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
765 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
767 /// where BC(It, k) stands for binomial coefficient.
769 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
770 ScalarEvolution &SE) const {
771 const SCEV *Result = getStart();
772 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
773 // The computation is correct in the face of overflow provided that the
774 // multiplication is performed _after_ the evaluation of the binomial
776 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
777 if (isa<SCEVCouldNotCompute>(Coeff))
780 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
785 //===----------------------------------------------------------------------===//
786 // SCEV Expression folder implementations
787 //===----------------------------------------------------------------------===//
789 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
791 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
792 "This is not a truncating conversion!");
793 assert(isSCEVable(Ty) &&
794 "This is not a conversion to a SCEVable type!");
795 Ty = getEffectiveSCEVType(Ty);
798 ID.AddInteger(scTruncate);
802 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
804 // Fold if the operand is constant.
805 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
807 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(),
808 getEffectiveSCEVType(Ty))));
810 // trunc(trunc(x)) --> trunc(x)
811 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
812 return getTruncateExpr(ST->getOperand(), Ty);
814 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
815 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
816 return getTruncateOrSignExtend(SS->getOperand(), Ty);
818 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
819 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
820 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
822 // trunc(x1+x2+...+xN) --> trunc(x1)+trunc(x2)+...+trunc(xN) if we can
823 // eliminate all the truncates.
824 if (const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Op)) {
825 SmallVector<const SCEV *, 4> Operands;
826 bool hasTrunc = false;
827 for (unsigned i = 0, e = SA->getNumOperands(); i != e && !hasTrunc; ++i) {
828 const SCEV *S = getTruncateExpr(SA->getOperand(i), Ty);
829 hasTrunc = isa<SCEVTruncateExpr>(S);
830 Operands.push_back(S);
833 return getAddExpr(Operands, false, false);
836 // If the input value is a chrec scev, truncate the chrec's operands.
837 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
838 SmallVector<const SCEV *, 4> Operands;
839 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
840 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
841 return getAddRecExpr(Operands, AddRec->getLoop());
844 // As a special case, fold trunc(undef) to undef. We don't want to
845 // know too much about SCEVUnknowns, but this special case is handy
847 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
848 if (isa<UndefValue>(U->getValue()))
849 return getSCEV(UndefValue::get(Ty));
851 // The cast wasn't folded; create an explicit cast node. We can reuse
852 // the existing insert position since if we get here, we won't have
853 // made any changes which would invalidate it.
854 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
856 UniqueSCEVs.InsertNode(S, IP);
860 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
862 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
863 "This is not an extending conversion!");
864 assert(isSCEVable(Ty) &&
865 "This is not a conversion to a SCEVable type!");
866 Ty = getEffectiveSCEVType(Ty);
868 // Fold if the operand is constant.
869 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
871 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(),
872 getEffectiveSCEVType(Ty))));
874 // zext(zext(x)) --> zext(x)
875 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
876 return getZeroExtendExpr(SZ->getOperand(), Ty);
878 // Before doing any expensive analysis, check to see if we've already
879 // computed a SCEV for this Op and Ty.
881 ID.AddInteger(scZeroExtend);
885 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
887 // If the input value is a chrec scev, and we can prove that the value
888 // did not overflow the old, smaller, value, we can zero extend all of the
889 // operands (often constants). This allows analysis of something like
890 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
891 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
892 if (AR->isAffine()) {
893 const SCEV *Start = AR->getStart();
894 const SCEV *Step = AR->getStepRecurrence(*this);
895 unsigned BitWidth = getTypeSizeInBits(AR->getType());
896 const Loop *L = AR->getLoop();
898 // If we have special knowledge that this addrec won't overflow,
899 // we don't need to do any further analysis.
900 if (AR->hasNoUnsignedWrap())
901 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
902 getZeroExtendExpr(Step, Ty),
905 // Check whether the backedge-taken count is SCEVCouldNotCompute.
906 // Note that this serves two purposes: It filters out loops that are
907 // simply not analyzable, and it covers the case where this code is
908 // being called from within backedge-taken count analysis, such that
909 // attempting to ask for the backedge-taken count would likely result
910 // in infinite recursion. In the later case, the analysis code will
911 // cope with a conservative value, and it will take care to purge
912 // that value once it has finished.
913 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
914 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
915 // Manually compute the final value for AR, checking for
918 // Check whether the backedge-taken count can be losslessly casted to
919 // the addrec's type. The count is always unsigned.
920 const SCEV *CastedMaxBECount =
921 getTruncateOrZeroExtend(MaxBECount, Start->getType());
922 const SCEV *RecastedMaxBECount =
923 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
924 if (MaxBECount == RecastedMaxBECount) {
925 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
926 // Check whether Start+Step*MaxBECount has no unsigned overflow.
927 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
928 const SCEV *Add = getAddExpr(Start, ZMul);
929 const SCEV *OperandExtendedAdd =
930 getAddExpr(getZeroExtendExpr(Start, WideTy),
931 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
932 getZeroExtendExpr(Step, WideTy)));
933 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
934 // Return the expression with the addrec on the outside.
935 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
936 getZeroExtendExpr(Step, Ty),
939 // Similar to above, only this time treat the step value as signed.
940 // This covers loops that count down.
941 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
942 Add = getAddExpr(Start, SMul);
944 getAddExpr(getZeroExtendExpr(Start, WideTy),
945 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
946 getSignExtendExpr(Step, WideTy)));
947 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
948 // Return the expression with the addrec on the outside.
949 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
950 getSignExtendExpr(Step, Ty),
954 // If the backedge is guarded by a comparison with the pre-inc value
955 // the addrec is safe. Also, if the entry is guarded by a comparison
956 // with the start value and the backedge is guarded by a comparison
957 // with the post-inc value, the addrec is safe.
958 if (isKnownPositive(Step)) {
959 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
960 getUnsignedRange(Step).getUnsignedMax());
961 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
962 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
963 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
964 AR->getPostIncExpr(*this), N)))
965 // Return the expression with the addrec on the outside.
966 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
967 getZeroExtendExpr(Step, Ty),
969 } else if (isKnownNegative(Step)) {
970 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
971 getSignedRange(Step).getSignedMin());
972 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
973 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
974 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
975 AR->getPostIncExpr(*this), N)))
976 // Return the expression with the addrec on the outside.
977 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
978 getSignExtendExpr(Step, Ty),
984 // The cast wasn't folded; create an explicit cast node.
985 // Recompute the insert position, as it may have been invalidated.
986 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
987 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
989 UniqueSCEVs.InsertNode(S, IP);
993 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
995 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
996 "This is not an extending conversion!");
997 assert(isSCEVable(Ty) &&
998 "This is not a conversion to a SCEVable type!");
999 Ty = getEffectiveSCEVType(Ty);
1001 // Fold if the operand is constant.
1002 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1004 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(),
1005 getEffectiveSCEVType(Ty))));
1007 // sext(sext(x)) --> sext(x)
1008 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1009 return getSignExtendExpr(SS->getOperand(), Ty);
1011 // sext(zext(x)) --> zext(x)
1012 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
1013 return getZeroExtendExpr(SZ->getOperand(), Ty);
1015 // Before doing any expensive analysis, check to see if we've already
1016 // computed a SCEV for this Op and Ty.
1017 FoldingSetNodeID ID;
1018 ID.AddInteger(scSignExtend);
1022 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1024 // If the input value is a chrec scev, and we can prove that the value
1025 // did not overflow the old, smaller, value, we can sign extend all of the
1026 // operands (often constants). This allows analysis of something like
1027 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1028 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1029 if (AR->isAffine()) {
1030 const SCEV *Start = AR->getStart();
1031 const SCEV *Step = AR->getStepRecurrence(*this);
1032 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1033 const Loop *L = AR->getLoop();
1035 // If we have special knowledge that this addrec won't overflow,
1036 // we don't need to do any further analysis.
1037 if (AR->hasNoSignedWrap())
1038 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1039 getSignExtendExpr(Step, Ty),
1042 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1043 // Note that this serves two purposes: It filters out loops that are
1044 // simply not analyzable, and it covers the case where this code is
1045 // being called from within backedge-taken count analysis, such that
1046 // attempting to ask for the backedge-taken count would likely result
1047 // in infinite recursion. In the later case, the analysis code will
1048 // cope with a conservative value, and it will take care to purge
1049 // that value once it has finished.
1050 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1051 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1052 // Manually compute the final value for AR, checking for
1055 // Check whether the backedge-taken count can be losslessly casted to
1056 // the addrec's type. The count is always unsigned.
1057 const SCEV *CastedMaxBECount =
1058 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1059 const SCEV *RecastedMaxBECount =
1060 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1061 if (MaxBECount == RecastedMaxBECount) {
1062 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1063 // Check whether Start+Step*MaxBECount has no signed overflow.
1064 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1065 const SCEV *Add = getAddExpr(Start, SMul);
1066 const SCEV *OperandExtendedAdd =
1067 getAddExpr(getSignExtendExpr(Start, WideTy),
1068 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1069 getSignExtendExpr(Step, WideTy)));
1070 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1071 // Return the expression with the addrec on the outside.
1072 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1073 getSignExtendExpr(Step, Ty),
1076 // Similar to above, only this time treat the step value as unsigned.
1077 // This covers loops that count up with an unsigned step.
1078 const SCEV *UMul = getMulExpr(CastedMaxBECount, Step);
1079 Add = getAddExpr(Start, UMul);
1080 OperandExtendedAdd =
1081 getAddExpr(getSignExtendExpr(Start, WideTy),
1082 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1083 getZeroExtendExpr(Step, WideTy)));
1084 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1085 // Return the expression with the addrec on the outside.
1086 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1087 getZeroExtendExpr(Step, Ty),
1091 // If the backedge is guarded by a comparison with the pre-inc value
1092 // the addrec is safe. Also, if the entry is guarded by a comparison
1093 // with the start value and the backedge is guarded by a comparison
1094 // with the post-inc value, the addrec is safe.
1095 if (isKnownPositive(Step)) {
1096 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1097 getSignedRange(Step).getSignedMax());
1098 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1099 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1100 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1101 AR->getPostIncExpr(*this), N)))
1102 // Return the expression with the addrec on the outside.
1103 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1104 getSignExtendExpr(Step, Ty),
1106 } else if (isKnownNegative(Step)) {
1107 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1108 getSignedRange(Step).getSignedMin());
1109 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1110 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1111 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1112 AR->getPostIncExpr(*this), N)))
1113 // Return the expression with the addrec on the outside.
1114 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1115 getSignExtendExpr(Step, Ty),
1121 // The cast wasn't folded; create an explicit cast node.
1122 // Recompute the insert position, as it may have been invalidated.
1123 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1124 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1126 UniqueSCEVs.InsertNode(S, IP);
1130 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1131 /// unspecified bits out to the given type.
1133 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1135 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1136 "This is not an extending conversion!");
1137 assert(isSCEVable(Ty) &&
1138 "This is not a conversion to a SCEVable type!");
1139 Ty = getEffectiveSCEVType(Ty);
1141 // Sign-extend negative constants.
1142 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1143 if (SC->getValue()->getValue().isNegative())
1144 return getSignExtendExpr(Op, Ty);
1146 // Peel off a truncate cast.
1147 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1148 const SCEV *NewOp = T->getOperand();
1149 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1150 return getAnyExtendExpr(NewOp, Ty);
1151 return getTruncateOrNoop(NewOp, Ty);
1154 // Next try a zext cast. If the cast is folded, use it.
1155 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1156 if (!isa<SCEVZeroExtendExpr>(ZExt))
1159 // Next try a sext cast. If the cast is folded, use it.
1160 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1161 if (!isa<SCEVSignExtendExpr>(SExt))
1164 // Force the cast to be folded into the operands of an addrec.
1165 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1166 SmallVector<const SCEV *, 4> Ops;
1167 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1169 Ops.push_back(getAnyExtendExpr(*I, Ty));
1170 return getAddRecExpr(Ops, AR->getLoop());
1173 // As a special case, fold anyext(undef) to undef. We don't want to
1174 // know too much about SCEVUnknowns, but this special case is handy
1176 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
1177 if (isa<UndefValue>(U->getValue()))
1178 return getSCEV(UndefValue::get(Ty));
1180 // If the expression is obviously signed, use the sext cast value.
1181 if (isa<SCEVSMaxExpr>(Op))
1184 // Absent any other information, use the zext cast value.
1188 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1189 /// a list of operands to be added under the given scale, update the given
1190 /// map. This is a helper function for getAddRecExpr. As an example of
1191 /// what it does, given a sequence of operands that would form an add
1192 /// expression like this:
1194 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1196 /// where A and B are constants, update the map with these values:
1198 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1200 /// and add 13 + A*B*29 to AccumulatedConstant.
1201 /// This will allow getAddRecExpr to produce this:
1203 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1205 /// This form often exposes folding opportunities that are hidden in
1206 /// the original operand list.
1208 /// Return true iff it appears that any interesting folding opportunities
1209 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1210 /// the common case where no interesting opportunities are present, and
1211 /// is also used as a check to avoid infinite recursion.
1214 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1215 SmallVector<const SCEV *, 8> &NewOps,
1216 APInt &AccumulatedConstant,
1217 const SCEV *const *Ops, size_t NumOperands,
1219 ScalarEvolution &SE) {
1220 bool Interesting = false;
1222 // Iterate over the add operands. They are sorted, with constants first.
1224 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1226 // Pull a buried constant out to the outside.
1227 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1229 AccumulatedConstant += Scale * C->getValue()->getValue();
1232 // Next comes everything else. We're especially interested in multiplies
1233 // here, but they're in the middle, so just visit the rest with one loop.
1234 for (; i != NumOperands; ++i) {
1235 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1236 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1238 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1239 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1240 // A multiplication of a constant with another add; recurse.
1241 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1243 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1244 Add->op_begin(), Add->getNumOperands(),
1247 // A multiplication of a constant with some other value. Update
1249 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1250 const SCEV *Key = SE.getMulExpr(MulOps);
1251 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1252 M.insert(std::make_pair(Key, NewScale));
1254 NewOps.push_back(Pair.first->first);
1256 Pair.first->second += NewScale;
1257 // The map already had an entry for this value, which may indicate
1258 // a folding opportunity.
1263 // An ordinary operand. Update the map.
1264 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1265 M.insert(std::make_pair(Ops[i], Scale));
1267 NewOps.push_back(Pair.first->first);
1269 Pair.first->second += Scale;
1270 // The map already had an entry for this value, which may indicate
1271 // a folding opportunity.
1281 struct APIntCompare {
1282 bool operator()(const APInt &LHS, const APInt &RHS) const {
1283 return LHS.ult(RHS);
1288 /// getAddExpr - Get a canonical add expression, or something simpler if
1290 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1291 bool HasNUW, bool HasNSW) {
1292 assert(!Ops.empty() && "Cannot get empty add!");
1293 if (Ops.size() == 1) return Ops[0];
1295 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1296 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1297 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1298 "SCEVAddExpr operand types don't match!");
1301 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1302 if (!HasNUW && HasNSW) {
1304 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1305 E = Ops.end(); I != E; ++I)
1306 if (!isKnownNonNegative(*I)) {
1310 if (All) HasNUW = true;
1313 // Sort by complexity, this groups all similar expression types together.
1314 GroupByComplexity(Ops, LI);
1316 // If there are any constants, fold them together.
1318 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1320 assert(Idx < Ops.size());
1321 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1322 // We found two constants, fold them together!
1323 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1324 RHSC->getValue()->getValue());
1325 if (Ops.size() == 2) return Ops[0];
1326 Ops.erase(Ops.begin()+1); // Erase the folded element
1327 LHSC = cast<SCEVConstant>(Ops[0]);
1330 // If we are left with a constant zero being added, strip it off.
1331 if (LHSC->getValue()->isZero()) {
1332 Ops.erase(Ops.begin());
1336 if (Ops.size() == 1) return Ops[0];
1339 // Okay, check to see if the same value occurs in the operand list more than
1340 // once. If so, merge them together into an multiply expression. Since we
1341 // sorted the list, these values are required to be adjacent.
1342 const Type *Ty = Ops[0]->getType();
1343 bool FoundMatch = false;
1344 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1345 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1346 // Scan ahead to count how many equal operands there are.
1348 while (i+Count != e && Ops[i+Count] == Ops[i])
1350 // Merge the values into a multiply.
1351 const SCEV *Scale = getConstant(Ty, Count);
1352 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1353 if (Ops.size() == Count)
1356 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1357 --i; e -= Count - 1;
1361 return getAddExpr(Ops, HasNUW, HasNSW);
1363 // Check for truncates. If all the operands are truncated from the same
1364 // type, see if factoring out the truncate would permit the result to be
1365 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1366 // if the contents of the resulting outer trunc fold to something simple.
1367 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1368 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1369 const Type *DstType = Trunc->getType();
1370 const Type *SrcType = Trunc->getOperand()->getType();
1371 SmallVector<const SCEV *, 8> LargeOps;
1373 // Check all the operands to see if they can be represented in the
1374 // source type of the truncate.
1375 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1376 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1377 if (T->getOperand()->getType() != SrcType) {
1381 LargeOps.push_back(T->getOperand());
1382 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1383 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1384 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1385 SmallVector<const SCEV *, 8> LargeMulOps;
1386 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1387 if (const SCEVTruncateExpr *T =
1388 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1389 if (T->getOperand()->getType() != SrcType) {
1393 LargeMulOps.push_back(T->getOperand());
1394 } else if (const SCEVConstant *C =
1395 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1396 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1403 LargeOps.push_back(getMulExpr(LargeMulOps));
1410 // Evaluate the expression in the larger type.
1411 const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
1412 // If it folds to something simple, use it. Otherwise, don't.
1413 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1414 return getTruncateExpr(Fold, DstType);
1418 // Skip past any other cast SCEVs.
1419 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1422 // If there are add operands they would be next.
1423 if (Idx < Ops.size()) {
1424 bool DeletedAdd = false;
1425 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1426 // If we have an add, expand the add operands onto the end of the operands
1428 Ops.erase(Ops.begin()+Idx);
1429 Ops.append(Add->op_begin(), Add->op_end());
1433 // If we deleted at least one add, we added operands to the end of the list,
1434 // and they are not necessarily sorted. Recurse to resort and resimplify
1435 // any operands we just acquired.
1437 return getAddExpr(Ops);
1440 // Skip over the add expression until we get to a multiply.
1441 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1444 // Check to see if there are any folding opportunities present with
1445 // operands multiplied by constant values.
1446 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1447 uint64_t BitWidth = getTypeSizeInBits(Ty);
1448 DenseMap<const SCEV *, APInt> M;
1449 SmallVector<const SCEV *, 8> NewOps;
1450 APInt AccumulatedConstant(BitWidth, 0);
1451 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1452 Ops.data(), Ops.size(),
1453 APInt(BitWidth, 1), *this)) {
1454 // Some interesting folding opportunity is present, so its worthwhile to
1455 // re-generate the operands list. Group the operands by constant scale,
1456 // to avoid multiplying by the same constant scale multiple times.
1457 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1458 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1459 E = NewOps.end(); I != E; ++I)
1460 MulOpLists[M.find(*I)->second].push_back(*I);
1461 // Re-generate the operands list.
1463 if (AccumulatedConstant != 0)
1464 Ops.push_back(getConstant(AccumulatedConstant));
1465 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1466 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1468 Ops.push_back(getMulExpr(getConstant(I->first),
1469 getAddExpr(I->second)));
1471 return getConstant(Ty, 0);
1472 if (Ops.size() == 1)
1474 return getAddExpr(Ops);
1478 // If we are adding something to a multiply expression, make sure the
1479 // something is not already an operand of the multiply. If so, merge it into
1481 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1482 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1483 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1484 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1485 if (isa<SCEVConstant>(MulOpSCEV))
1487 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1488 if (MulOpSCEV == Ops[AddOp]) {
1489 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1490 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1491 if (Mul->getNumOperands() != 2) {
1492 // If the multiply has more than two operands, we must get the
1494 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1495 Mul->op_begin()+MulOp);
1496 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1497 InnerMul = getMulExpr(MulOps);
1499 const SCEV *One = getConstant(Ty, 1);
1500 const SCEV *AddOne = getAddExpr(One, InnerMul);
1501 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1502 if (Ops.size() == 2) return OuterMul;
1504 Ops.erase(Ops.begin()+AddOp);
1505 Ops.erase(Ops.begin()+Idx-1);
1507 Ops.erase(Ops.begin()+Idx);
1508 Ops.erase(Ops.begin()+AddOp-1);
1510 Ops.push_back(OuterMul);
1511 return getAddExpr(Ops);
1514 // Check this multiply against other multiplies being added together.
1515 for (unsigned OtherMulIdx = Idx+1;
1516 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1518 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1519 // If MulOp occurs in OtherMul, we can fold the two multiplies
1521 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1522 OMulOp != e; ++OMulOp)
1523 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1524 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1525 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1526 if (Mul->getNumOperands() != 2) {
1527 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1528 Mul->op_begin()+MulOp);
1529 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1530 InnerMul1 = getMulExpr(MulOps);
1532 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1533 if (OtherMul->getNumOperands() != 2) {
1534 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1535 OtherMul->op_begin()+OMulOp);
1536 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1537 InnerMul2 = getMulExpr(MulOps);
1539 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1540 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1541 if (Ops.size() == 2) return OuterMul;
1542 Ops.erase(Ops.begin()+Idx);
1543 Ops.erase(Ops.begin()+OtherMulIdx-1);
1544 Ops.push_back(OuterMul);
1545 return getAddExpr(Ops);
1551 // If there are any add recurrences in the operands list, see if any other
1552 // added values are loop invariant. If so, we can fold them into the
1554 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1557 // Scan over all recurrences, trying to fold loop invariants into them.
1558 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1559 // Scan all of the other operands to this add and add them to the vector if
1560 // they are loop invariant w.r.t. the recurrence.
1561 SmallVector<const SCEV *, 8> LIOps;
1562 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1563 const Loop *AddRecLoop = AddRec->getLoop();
1564 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1565 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1566 LIOps.push_back(Ops[i]);
1567 Ops.erase(Ops.begin()+i);
1571 // If we found some loop invariants, fold them into the recurrence.
1572 if (!LIOps.empty()) {
1573 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1574 LIOps.push_back(AddRec->getStart());
1576 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1578 AddRecOps[0] = getAddExpr(LIOps);
1580 // Build the new addrec. Propagate the NUW and NSW flags if both the
1581 // outer add and the inner addrec are guaranteed to have no overflow.
1582 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop,
1583 HasNUW && AddRec->hasNoUnsignedWrap(),
1584 HasNSW && AddRec->hasNoSignedWrap());
1586 // If all of the other operands were loop invariant, we are done.
1587 if (Ops.size() == 1) return NewRec;
1589 // Otherwise, add the folded AddRec by the non-liv parts.
1590 for (unsigned i = 0;; ++i)
1591 if (Ops[i] == AddRec) {
1595 return getAddExpr(Ops);
1598 // Okay, if there weren't any loop invariants to be folded, check to see if
1599 // there are multiple AddRec's with the same loop induction variable being
1600 // added together. If so, we can fold them.
1601 for (unsigned OtherIdx = Idx+1;
1602 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1604 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1605 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1606 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1608 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1610 if (const SCEVAddRecExpr *OtherAddRec =
1611 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1612 if (OtherAddRec->getLoop() == AddRecLoop) {
1613 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1615 if (i >= AddRecOps.size()) {
1616 AddRecOps.append(OtherAddRec->op_begin()+i,
1617 OtherAddRec->op_end());
1620 AddRecOps[i] = getAddExpr(AddRecOps[i],
1621 OtherAddRec->getOperand(i));
1623 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1625 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop);
1626 return getAddExpr(Ops);
1629 // Otherwise couldn't fold anything into this recurrence. Move onto the
1633 // Okay, it looks like we really DO need an add expr. Check to see if we
1634 // already have one, otherwise create a new one.
1635 FoldingSetNodeID ID;
1636 ID.AddInteger(scAddExpr);
1637 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1638 ID.AddPointer(Ops[i]);
1641 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1643 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1644 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1645 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1647 UniqueSCEVs.InsertNode(S, IP);
1649 if (HasNUW) S->setHasNoUnsignedWrap(true);
1650 if (HasNSW) S->setHasNoSignedWrap(true);
1654 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1656 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1657 bool HasNUW, bool HasNSW) {
1658 assert(!Ops.empty() && "Cannot get empty mul!");
1659 if (Ops.size() == 1) return Ops[0];
1661 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1662 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1663 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1664 "SCEVMulExpr operand types don't match!");
1667 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1668 if (!HasNUW && HasNSW) {
1670 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1671 E = Ops.end(); I != E; ++I)
1672 if (!isKnownNonNegative(*I)) {
1676 if (All) HasNUW = true;
1679 // Sort by complexity, this groups all similar expression types together.
1680 GroupByComplexity(Ops, LI);
1682 // If there are any constants, fold them together.
1684 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1686 // C1*(C2+V) -> C1*C2 + C1*V
1687 if (Ops.size() == 2)
1688 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1689 if (Add->getNumOperands() == 2 &&
1690 isa<SCEVConstant>(Add->getOperand(0)))
1691 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1692 getMulExpr(LHSC, Add->getOperand(1)));
1695 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1696 // We found two constants, fold them together!
1697 ConstantInt *Fold = ConstantInt::get(getContext(),
1698 LHSC->getValue()->getValue() *
1699 RHSC->getValue()->getValue());
1700 Ops[0] = getConstant(Fold);
1701 Ops.erase(Ops.begin()+1); // Erase the folded element
1702 if (Ops.size() == 1) return Ops[0];
1703 LHSC = cast<SCEVConstant>(Ops[0]);
1706 // If we are left with a constant one being multiplied, strip it off.
1707 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1708 Ops.erase(Ops.begin());
1710 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1711 // If we have a multiply of zero, it will always be zero.
1713 } else if (Ops[0]->isAllOnesValue()) {
1714 // If we have a mul by -1 of an add, try distributing the -1 among the
1716 if (Ops.size() == 2)
1717 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1718 SmallVector<const SCEV *, 4> NewOps;
1719 bool AnyFolded = false;
1720 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1722 const SCEV *Mul = getMulExpr(Ops[0], *I);
1723 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1724 NewOps.push_back(Mul);
1727 return getAddExpr(NewOps);
1731 if (Ops.size() == 1)
1735 // Skip over the add expression until we get to a multiply.
1736 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1739 // If there are mul operands inline them all into this expression.
1740 if (Idx < Ops.size()) {
1741 bool DeletedMul = false;
1742 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1743 // If we have an mul, expand the mul operands onto the end of the operands
1745 Ops.erase(Ops.begin()+Idx);
1746 Ops.append(Mul->op_begin(), Mul->op_end());
1750 // If we deleted at least one mul, we added operands to the end of the list,
1751 // and they are not necessarily sorted. Recurse to resort and resimplify
1752 // any operands we just acquired.
1754 return getMulExpr(Ops);
1757 // If there are any add recurrences in the operands list, see if any other
1758 // added values are loop invariant. If so, we can fold them into the
1760 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1763 // Scan over all recurrences, trying to fold loop invariants into them.
1764 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1765 // Scan all of the other operands to this mul and add them to the vector if
1766 // they are loop invariant w.r.t. the recurrence.
1767 SmallVector<const SCEV *, 8> LIOps;
1768 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1769 const Loop *AddRecLoop = AddRec->getLoop();
1770 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1771 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1772 LIOps.push_back(Ops[i]);
1773 Ops.erase(Ops.begin()+i);
1777 // If we found some loop invariants, fold them into the recurrence.
1778 if (!LIOps.empty()) {
1779 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1780 SmallVector<const SCEV *, 4> NewOps;
1781 NewOps.reserve(AddRec->getNumOperands());
1782 const SCEV *Scale = getMulExpr(LIOps);
1783 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1784 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1786 // Build the new addrec. Propagate the NUW and NSW flags if both the
1787 // outer mul and the inner addrec are guaranteed to have no overflow.
1788 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop,
1789 HasNUW && AddRec->hasNoUnsignedWrap(),
1790 HasNSW && AddRec->hasNoSignedWrap());
1792 // If all of the other operands were loop invariant, we are done.
1793 if (Ops.size() == 1) return NewRec;
1795 // Otherwise, multiply the folded AddRec by the non-liv parts.
1796 for (unsigned i = 0;; ++i)
1797 if (Ops[i] == AddRec) {
1801 return getMulExpr(Ops);
1804 // Okay, if there weren't any loop invariants to be folded, check to see if
1805 // there are multiple AddRec's with the same loop induction variable being
1806 // multiplied together. If so, we can fold them.
1807 for (unsigned OtherIdx = Idx+1;
1808 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1810 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1811 // F * G, where F = {A,+,B}<L> and G = {C,+,D}<L> -->
1812 // {A*C,+,F*D + G*B + B*D}<L>
1813 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1815 if (const SCEVAddRecExpr *OtherAddRec =
1816 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1817 if (OtherAddRec->getLoop() == AddRecLoop) {
1818 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1819 const SCEV *NewStart = getMulExpr(F->getStart(), G->getStart());
1820 const SCEV *B = F->getStepRecurrence(*this);
1821 const SCEV *D = G->getStepRecurrence(*this);
1822 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1825 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1827 if (Ops.size() == 2) return NewAddRec;
1828 Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
1829 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1831 return getMulExpr(Ops);
1834 // Otherwise couldn't fold anything into this recurrence. Move onto the
1838 // Okay, it looks like we really DO need an mul expr. Check to see if we
1839 // already have one, otherwise create a new one.
1840 FoldingSetNodeID ID;
1841 ID.AddInteger(scMulExpr);
1842 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1843 ID.AddPointer(Ops[i]);
1846 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1848 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1849 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1850 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
1852 UniqueSCEVs.InsertNode(S, IP);
1854 if (HasNUW) S->setHasNoUnsignedWrap(true);
1855 if (HasNSW) S->setHasNoSignedWrap(true);
1859 /// getUDivExpr - Get a canonical unsigned division expression, or something
1860 /// simpler if possible.
1861 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1863 assert(getEffectiveSCEVType(LHS->getType()) ==
1864 getEffectiveSCEVType(RHS->getType()) &&
1865 "SCEVUDivExpr operand types don't match!");
1867 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1868 if (RHSC->getValue()->equalsInt(1))
1869 return LHS; // X udiv 1 --> x
1870 // If the denominator is zero, the result of the udiv is undefined. Don't
1871 // try to analyze it, because the resolution chosen here may differ from
1872 // the resolution chosen in other parts of the compiler.
1873 if (!RHSC->getValue()->isZero()) {
1874 // Determine if the division can be folded into the operands of
1876 // TODO: Generalize this to non-constants by using known-bits information.
1877 const Type *Ty = LHS->getType();
1878 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1879 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
1880 // For non-power-of-two values, effectively round the value up to the
1881 // nearest power of two.
1882 if (!RHSC->getValue()->getValue().isPowerOf2())
1884 const IntegerType *ExtTy =
1885 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1886 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1887 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1888 if (const SCEVConstant *Step =
1889 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1890 if (!Step->getValue()->getValue()
1891 .urem(RHSC->getValue()->getValue()) &&
1892 getZeroExtendExpr(AR, ExtTy) ==
1893 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1894 getZeroExtendExpr(Step, ExtTy),
1896 SmallVector<const SCEV *, 4> Operands;
1897 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1898 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1899 return getAddRecExpr(Operands, AR->getLoop());
1901 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1902 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1903 SmallVector<const SCEV *, 4> Operands;
1904 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1905 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1906 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1907 // Find an operand that's safely divisible.
1908 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1909 const SCEV *Op = M->getOperand(i);
1910 const SCEV *Div = getUDivExpr(Op, RHSC);
1911 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1912 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
1915 return getMulExpr(Operands);
1919 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1920 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1921 SmallVector<const SCEV *, 4> Operands;
1922 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1923 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1924 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1926 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1927 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
1928 if (isa<SCEVUDivExpr>(Op) ||
1929 getMulExpr(Op, RHS) != A->getOperand(i))
1931 Operands.push_back(Op);
1933 if (Operands.size() == A->getNumOperands())
1934 return getAddExpr(Operands);
1938 // Fold if both operands are constant.
1939 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1940 Constant *LHSCV = LHSC->getValue();
1941 Constant *RHSCV = RHSC->getValue();
1942 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
1948 FoldingSetNodeID ID;
1949 ID.AddInteger(scUDivExpr);
1953 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1954 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
1956 UniqueSCEVs.InsertNode(S, IP);
1961 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1962 /// Simplify the expression as much as possible.
1963 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
1964 const SCEV *Step, const Loop *L,
1965 bool HasNUW, bool HasNSW) {
1966 SmallVector<const SCEV *, 4> Operands;
1967 Operands.push_back(Start);
1968 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1969 if (StepChrec->getLoop() == L) {
1970 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
1971 return getAddRecExpr(Operands, L);
1974 Operands.push_back(Step);
1975 return getAddRecExpr(Operands, L, HasNUW, HasNSW);
1978 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1979 /// Simplify the expression as much as possible.
1981 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
1983 bool HasNUW, bool HasNSW) {
1984 if (Operands.size() == 1) return Operands[0];
1986 const Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
1987 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
1988 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
1989 "SCEVAddRecExpr operand types don't match!");
1990 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1991 assert(isLoopInvariant(Operands[i], L) &&
1992 "SCEVAddRecExpr operand is not loop-invariant!");
1995 if (Operands.back()->isZero()) {
1996 Operands.pop_back();
1997 return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
2000 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2001 // use that information to infer NUW and NSW flags. However, computing a
2002 // BE count requires calling getAddRecExpr, so we may not yet have a
2003 // meaningful BE count at this point (and if we don't, we'd be stuck
2004 // with a SCEVCouldNotCompute as the cached BE count).
2006 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
2007 if (!HasNUW && HasNSW) {
2009 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2010 E = Operands.end(); I != E; ++I)
2011 if (!isKnownNonNegative(*I)) {
2015 if (All) HasNUW = true;
2018 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2019 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2020 const Loop *NestedLoop = NestedAR->getLoop();
2021 if (L->contains(NestedLoop) ?
2022 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2023 (!NestedLoop->contains(L) &&
2024 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2025 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2026 NestedAR->op_end());
2027 Operands[0] = NestedAR->getStart();
2028 // AddRecs require their operands be loop-invariant with respect to their
2029 // loops. Don't perform this transformation if it would break this
2031 bool AllInvariant = true;
2032 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2033 if (!isLoopInvariant(Operands[i], L)) {
2034 AllInvariant = false;
2038 NestedOperands[0] = getAddRecExpr(Operands, L);
2039 AllInvariant = true;
2040 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2041 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2042 AllInvariant = false;
2046 // Ok, both add recurrences are valid after the transformation.
2047 return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
2049 // Reset Operands to its original state.
2050 Operands[0] = NestedAR;
2054 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2055 // already have one, otherwise create a new one.
2056 FoldingSetNodeID ID;
2057 ID.AddInteger(scAddRecExpr);
2058 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2059 ID.AddPointer(Operands[i]);
2063 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2065 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2066 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2067 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2068 O, Operands.size(), L);
2069 UniqueSCEVs.InsertNode(S, IP);
2071 if (HasNUW) S->setHasNoUnsignedWrap(true);
2072 if (HasNSW) S->setHasNoSignedWrap(true);
2076 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2078 SmallVector<const SCEV *, 2> Ops;
2081 return getSMaxExpr(Ops);
2085 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2086 assert(!Ops.empty() && "Cannot get empty smax!");
2087 if (Ops.size() == 1) return Ops[0];
2089 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2090 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2091 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2092 "SCEVSMaxExpr operand types don't match!");
2095 // Sort by complexity, this groups all similar expression types together.
2096 GroupByComplexity(Ops, LI);
2098 // If there are any constants, fold them together.
2100 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2102 assert(Idx < Ops.size());
2103 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2104 // We found two constants, fold them together!
2105 ConstantInt *Fold = ConstantInt::get(getContext(),
2106 APIntOps::smax(LHSC->getValue()->getValue(),
2107 RHSC->getValue()->getValue()));
2108 Ops[0] = getConstant(Fold);
2109 Ops.erase(Ops.begin()+1); // Erase the folded element
2110 if (Ops.size() == 1) return Ops[0];
2111 LHSC = cast<SCEVConstant>(Ops[0]);
2114 // If we are left with a constant minimum-int, strip it off.
2115 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2116 Ops.erase(Ops.begin());
2118 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2119 // If we have an smax with a constant maximum-int, it will always be
2124 if (Ops.size() == 1) return Ops[0];
2127 // Find the first SMax
2128 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2131 // Check to see if one of the operands is an SMax. If so, expand its operands
2132 // onto our operand list, and recurse to simplify.
2133 if (Idx < Ops.size()) {
2134 bool DeletedSMax = false;
2135 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2136 Ops.erase(Ops.begin()+Idx);
2137 Ops.append(SMax->op_begin(), SMax->op_end());
2142 return getSMaxExpr(Ops);
2145 // Okay, check to see if the same value occurs in the operand list twice. If
2146 // so, delete one. Since we sorted the list, these values are required to
2148 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2149 // X smax Y smax Y --> X smax Y
2150 // X smax Y --> X, if X is always greater than Y
2151 if (Ops[i] == Ops[i+1] ||
2152 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2153 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2155 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2156 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2160 if (Ops.size() == 1) return Ops[0];
2162 assert(!Ops.empty() && "Reduced smax down to nothing!");
2164 // Okay, it looks like we really DO need an smax expr. Check to see if we
2165 // already have one, otherwise create a new one.
2166 FoldingSetNodeID ID;
2167 ID.AddInteger(scSMaxExpr);
2168 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2169 ID.AddPointer(Ops[i]);
2171 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2172 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2173 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2174 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2176 UniqueSCEVs.InsertNode(S, IP);
2180 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2182 SmallVector<const SCEV *, 2> Ops;
2185 return getUMaxExpr(Ops);
2189 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2190 assert(!Ops.empty() && "Cannot get empty umax!");
2191 if (Ops.size() == 1) return Ops[0];
2193 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2194 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2195 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2196 "SCEVUMaxExpr operand types don't match!");
2199 // Sort by complexity, this groups all similar expression types together.
2200 GroupByComplexity(Ops, LI);
2202 // If there are any constants, fold them together.
2204 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2206 assert(Idx < Ops.size());
2207 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2208 // We found two constants, fold them together!
2209 ConstantInt *Fold = ConstantInt::get(getContext(),
2210 APIntOps::umax(LHSC->getValue()->getValue(),
2211 RHSC->getValue()->getValue()));
2212 Ops[0] = getConstant(Fold);
2213 Ops.erase(Ops.begin()+1); // Erase the folded element
2214 if (Ops.size() == 1) return Ops[0];
2215 LHSC = cast<SCEVConstant>(Ops[0]);
2218 // If we are left with a constant minimum-int, strip it off.
2219 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2220 Ops.erase(Ops.begin());
2222 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2223 // If we have an umax with a constant maximum-int, it will always be
2228 if (Ops.size() == 1) return Ops[0];
2231 // Find the first UMax
2232 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2235 // Check to see if one of the operands is a UMax. If so, expand its operands
2236 // onto our operand list, and recurse to simplify.
2237 if (Idx < Ops.size()) {
2238 bool DeletedUMax = false;
2239 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2240 Ops.erase(Ops.begin()+Idx);
2241 Ops.append(UMax->op_begin(), UMax->op_end());
2246 return getUMaxExpr(Ops);
2249 // Okay, check to see if the same value occurs in the operand list twice. If
2250 // so, delete one. Since we sorted the list, these values are required to
2252 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2253 // X umax Y umax Y --> X umax Y
2254 // X umax Y --> X, if X is always greater than Y
2255 if (Ops[i] == Ops[i+1] ||
2256 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2257 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2259 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2260 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2264 if (Ops.size() == 1) return Ops[0];
2266 assert(!Ops.empty() && "Reduced umax down to nothing!");
2268 // Okay, it looks like we really DO need a umax expr. Check to see if we
2269 // already have one, otherwise create a new one.
2270 FoldingSetNodeID ID;
2271 ID.AddInteger(scUMaxExpr);
2272 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2273 ID.AddPointer(Ops[i]);
2275 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2276 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2277 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2278 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2280 UniqueSCEVs.InsertNode(S, IP);
2284 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2286 // ~smax(~x, ~y) == smin(x, y).
2287 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2290 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2292 // ~umax(~x, ~y) == umin(x, y)
2293 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2296 const SCEV *ScalarEvolution::getSizeOfExpr(const Type *AllocTy) {
2297 // If we have TargetData, we can bypass creating a target-independent
2298 // constant expression and then folding it back into a ConstantInt.
2299 // This is just a compile-time optimization.
2301 return getConstant(TD->getIntPtrType(getContext()),
2302 TD->getTypeAllocSize(AllocTy));
2304 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2305 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2306 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2308 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2309 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2312 const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
2313 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2314 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2315 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2317 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2318 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2321 const SCEV *ScalarEvolution::getOffsetOfExpr(const StructType *STy,
2323 // If we have TargetData, we can bypass creating a target-independent
2324 // constant expression and then folding it back into a ConstantInt.
2325 // This is just a compile-time optimization.
2327 return getConstant(TD->getIntPtrType(getContext()),
2328 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2330 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2331 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2332 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2334 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2335 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2338 const SCEV *ScalarEvolution::getOffsetOfExpr(const Type *CTy,
2339 Constant *FieldNo) {
2340 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2341 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2342 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2344 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2345 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2348 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2349 // Don't attempt to do anything other than create a SCEVUnknown object
2350 // here. createSCEV only calls getUnknown after checking for all other
2351 // interesting possibilities, and any other code that calls getUnknown
2352 // is doing so in order to hide a value from SCEV canonicalization.
2354 FoldingSetNodeID ID;
2355 ID.AddInteger(scUnknown);
2358 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2359 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2360 "Stale SCEVUnknown in uniquing map!");
2363 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2365 FirstUnknown = cast<SCEVUnknown>(S);
2366 UniqueSCEVs.InsertNode(S, IP);
2370 //===----------------------------------------------------------------------===//
2371 // Basic SCEV Analysis and PHI Idiom Recognition Code
2374 /// isSCEVable - Test if values of the given type are analyzable within
2375 /// the SCEV framework. This primarily includes integer types, and it
2376 /// can optionally include pointer types if the ScalarEvolution class
2377 /// has access to target-specific information.
2378 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2379 // Integers and pointers are always SCEVable.
2380 return Ty->isIntegerTy() || Ty->isPointerTy();
2383 /// getTypeSizeInBits - Return the size in bits of the specified type,
2384 /// for which isSCEVable must return true.
2385 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2386 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2388 // If we have a TargetData, use it!
2390 return TD->getTypeSizeInBits(Ty);
2392 // Integer types have fixed sizes.
2393 if (Ty->isIntegerTy())
2394 return Ty->getPrimitiveSizeInBits();
2396 // The only other support type is pointer. Without TargetData, conservatively
2397 // assume pointers are 64-bit.
2398 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2402 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2403 /// the given type and which represents how SCEV will treat the given
2404 /// type, for which isSCEVable must return true. For pointer types,
2405 /// this is the pointer-sized integer type.
2406 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2407 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2409 if (Ty->isIntegerTy())
2412 // The only other support type is pointer.
2413 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2414 if (TD) return TD->getIntPtrType(getContext());
2416 // Without TargetData, conservatively assume pointers are 64-bit.
2417 return Type::getInt64Ty(getContext());
2420 const SCEV *ScalarEvolution::getCouldNotCompute() {
2421 return &CouldNotCompute;
2424 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2425 /// expression and create a new one.
2426 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2427 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2429 ValueExprMapType::const_iterator I = ValueExprMap.find(V);
2430 if (I != ValueExprMap.end()) return I->second;
2431 const SCEV *S = createSCEV(V);
2433 // The process of creating a SCEV for V may have caused other SCEVs
2434 // to have been created, so it's necessary to insert the new entry
2435 // from scratch, rather than trying to remember the insert position
2437 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2441 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2443 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2444 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2446 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2448 const Type *Ty = V->getType();
2449 Ty = getEffectiveSCEVType(Ty);
2450 return getMulExpr(V,
2451 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2454 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2455 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2456 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2458 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2460 const Type *Ty = V->getType();
2461 Ty = getEffectiveSCEVType(Ty);
2462 const SCEV *AllOnes =
2463 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2464 return getMinusSCEV(AllOnes, V);
2467 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1,
2468 /// and thus the HasNUW and HasNSW bits apply to the resultant add, not
2469 /// whether the sub would have overflowed.
2470 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2471 bool HasNUW, bool HasNSW) {
2472 // Fast path: X - X --> 0.
2474 return getConstant(LHS->getType(), 0);
2477 return getAddExpr(LHS, getNegativeSCEV(RHS), HasNUW, HasNSW);
2480 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2481 /// input value to the specified type. If the type must be extended, it is zero
2484 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, const Type *Ty) {
2485 const Type *SrcTy = V->getType();
2486 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2487 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2488 "Cannot truncate or zero extend with non-integer arguments!");
2489 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2490 return V; // No conversion
2491 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2492 return getTruncateExpr(V, Ty);
2493 return getZeroExtendExpr(V, Ty);
2496 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2497 /// input value to the specified type. If the type must be extended, it is sign
2500 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2502 const Type *SrcTy = V->getType();
2503 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2504 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2505 "Cannot truncate or zero extend with non-integer arguments!");
2506 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2507 return V; // No conversion
2508 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2509 return getTruncateExpr(V, Ty);
2510 return getSignExtendExpr(V, Ty);
2513 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2514 /// input value to the specified type. If the type must be extended, it is zero
2515 /// extended. The conversion must not be narrowing.
2517 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2518 const Type *SrcTy = V->getType();
2519 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2520 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2521 "Cannot noop or zero extend with non-integer arguments!");
2522 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2523 "getNoopOrZeroExtend cannot truncate!");
2524 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2525 return V; // No conversion
2526 return getZeroExtendExpr(V, Ty);
2529 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2530 /// input value to the specified type. If the type must be extended, it is sign
2531 /// extended. The conversion must not be narrowing.
2533 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
2534 const Type *SrcTy = V->getType();
2535 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2536 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2537 "Cannot noop or sign extend with non-integer arguments!");
2538 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2539 "getNoopOrSignExtend cannot truncate!");
2540 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2541 return V; // No conversion
2542 return getSignExtendExpr(V, Ty);
2545 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2546 /// the input value to the specified type. If the type must be extended,
2547 /// it is extended with unspecified bits. The conversion must not be
2550 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2551 const Type *SrcTy = V->getType();
2552 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2553 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2554 "Cannot noop or any extend with non-integer arguments!");
2555 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2556 "getNoopOrAnyExtend cannot truncate!");
2557 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2558 return V; // No conversion
2559 return getAnyExtendExpr(V, Ty);
2562 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2563 /// input value to the specified type. The conversion must not be widening.
2565 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2566 const Type *SrcTy = V->getType();
2567 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2568 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2569 "Cannot truncate or noop with non-integer arguments!");
2570 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2571 "getTruncateOrNoop cannot extend!");
2572 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2573 return V; // No conversion
2574 return getTruncateExpr(V, Ty);
2577 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2578 /// the types using zero-extension, and then perform a umax operation
2580 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2582 const SCEV *PromotedLHS = LHS;
2583 const SCEV *PromotedRHS = RHS;
2585 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2586 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2588 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2590 return getUMaxExpr(PromotedLHS, PromotedRHS);
2593 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2594 /// the types using zero-extension, and then perform a umin operation
2596 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2598 const SCEV *PromotedLHS = LHS;
2599 const SCEV *PromotedRHS = RHS;
2601 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2602 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2604 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2606 return getUMinExpr(PromotedLHS, PromotedRHS);
2609 /// PushDefUseChildren - Push users of the given Instruction
2610 /// onto the given Worklist.
2612 PushDefUseChildren(Instruction *I,
2613 SmallVectorImpl<Instruction *> &Worklist) {
2614 // Push the def-use children onto the Worklist stack.
2615 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2617 Worklist.push_back(cast<Instruction>(*UI));
2620 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2621 /// instructions that depend on the given instruction and removes them from
2622 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2625 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2626 SmallVector<Instruction *, 16> Worklist;
2627 PushDefUseChildren(PN, Worklist);
2629 SmallPtrSet<Instruction *, 8> Visited;
2631 while (!Worklist.empty()) {
2632 Instruction *I = Worklist.pop_back_val();
2633 if (!Visited.insert(I)) continue;
2635 ValueExprMapType::iterator It =
2636 ValueExprMap.find(static_cast<Value *>(I));
2637 if (It != ValueExprMap.end()) {
2638 const SCEV *Old = It->second;
2640 // Short-circuit the def-use traversal if the symbolic name
2641 // ceases to appear in expressions.
2642 if (Old != SymName && !hasOperand(Old, SymName))
2645 // SCEVUnknown for a PHI either means that it has an unrecognized
2646 // structure, it's a PHI that's in the progress of being computed
2647 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2648 // additional loop trip count information isn't going to change anything.
2649 // In the second case, createNodeForPHI will perform the necessary
2650 // updates on its own when it gets to that point. In the third, we do
2651 // want to forget the SCEVUnknown.
2652 if (!isa<PHINode>(I) ||
2653 !isa<SCEVUnknown>(Old) ||
2654 (I != PN && Old == SymName)) {
2655 forgetMemoizedResults(Old);
2656 ValueExprMap.erase(It);
2660 PushDefUseChildren(I, Worklist);
2664 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2665 /// a loop header, making it a potential recurrence, or it doesn't.
2667 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2668 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2669 if (L->getHeader() == PN->getParent()) {
2670 // The loop may have multiple entrances or multiple exits; we can analyze
2671 // this phi as an addrec if it has a unique entry value and a unique
2673 Value *BEValueV = 0, *StartValueV = 0;
2674 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2675 Value *V = PN->getIncomingValue(i);
2676 if (L->contains(PN->getIncomingBlock(i))) {
2679 } else if (BEValueV != V) {
2683 } else if (!StartValueV) {
2685 } else if (StartValueV != V) {
2690 if (BEValueV && StartValueV) {
2691 // While we are analyzing this PHI node, handle its value symbolically.
2692 const SCEV *SymbolicName = getUnknown(PN);
2693 assert(ValueExprMap.find(PN) == ValueExprMap.end() &&
2694 "PHI node already processed?");
2695 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2697 // Using this symbolic name for the PHI, analyze the value coming around
2699 const SCEV *BEValue = getSCEV(BEValueV);
2701 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2702 // has a special value for the first iteration of the loop.
2704 // If the value coming around the backedge is an add with the symbolic
2705 // value we just inserted, then we found a simple induction variable!
2706 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2707 // If there is a single occurrence of the symbolic value, replace it
2708 // with a recurrence.
2709 unsigned FoundIndex = Add->getNumOperands();
2710 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2711 if (Add->getOperand(i) == SymbolicName)
2712 if (FoundIndex == e) {
2717 if (FoundIndex != Add->getNumOperands()) {
2718 // Create an add with everything but the specified operand.
2719 SmallVector<const SCEV *, 8> Ops;
2720 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2721 if (i != FoundIndex)
2722 Ops.push_back(Add->getOperand(i));
2723 const SCEV *Accum = getAddExpr(Ops);
2725 // This is not a valid addrec if the step amount is varying each
2726 // loop iteration, but is not itself an addrec in this loop.
2727 if (isLoopInvariant(Accum, L) ||
2728 (isa<SCEVAddRecExpr>(Accum) &&
2729 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2730 bool HasNUW = false;
2731 bool HasNSW = false;
2733 // If the increment doesn't overflow, then neither the addrec nor
2734 // the post-increment will overflow.
2735 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
2736 if (OBO->hasNoUnsignedWrap())
2738 if (OBO->hasNoSignedWrap())
2740 } else if (const GEPOperator *GEP =
2741 dyn_cast<GEPOperator>(BEValueV)) {
2742 // If the increment is a GEP, then we know it won't perform an
2743 // unsigned overflow, because the address space cannot be
2745 HasNUW |= GEP->isInBounds();
2748 const SCEV *StartVal = getSCEV(StartValueV);
2749 const SCEV *PHISCEV =
2750 getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
2752 // Since the no-wrap flags are on the increment, they apply to the
2753 // post-incremented value as well.
2754 if (isLoopInvariant(Accum, L))
2755 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
2756 Accum, L, HasNUW, HasNSW);
2758 // Okay, for the entire analysis of this edge we assumed the PHI
2759 // to be symbolic. We now need to go back and purge all of the
2760 // entries for the scalars that use the symbolic expression.
2761 ForgetSymbolicName(PN, SymbolicName);
2762 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2766 } else if (const SCEVAddRecExpr *AddRec =
2767 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2768 // Otherwise, this could be a loop like this:
2769 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2770 // In this case, j = {1,+,1} and BEValue is j.
2771 // Because the other in-value of i (0) fits the evolution of BEValue
2772 // i really is an addrec evolution.
2773 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2774 const SCEV *StartVal = getSCEV(StartValueV);
2776 // If StartVal = j.start - j.stride, we can use StartVal as the
2777 // initial step of the addrec evolution.
2778 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2779 AddRec->getOperand(1))) {
2780 const SCEV *PHISCEV =
2781 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2783 // Okay, for the entire analysis of this edge we assumed the PHI
2784 // to be symbolic. We now need to go back and purge all of the
2785 // entries for the scalars that use the symbolic expression.
2786 ForgetSymbolicName(PN, SymbolicName);
2787 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2795 // If the PHI has a single incoming value, follow that value, unless the
2796 // PHI's incoming blocks are in a different loop, in which case doing so
2797 // risks breaking LCSSA form. Instcombine would normally zap these, but
2798 // it doesn't have DominatorTree information, so it may miss cases.
2799 if (Value *V = SimplifyInstruction(PN, TD, DT))
2800 if (LI->replacementPreservesLCSSAForm(PN, V))
2803 // If it's not a loop phi, we can't handle it yet.
2804 return getUnknown(PN);
2807 /// createNodeForGEP - Expand GEP instructions into add and multiply
2808 /// operations. This allows them to be analyzed by regular SCEV code.
2810 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
2812 // Don't blindly transfer the inbounds flag from the GEP instruction to the
2813 // Add expression, because the Instruction may be guarded by control flow
2814 // and the no-overflow bits may not be valid for the expression in any
2817 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2818 Value *Base = GEP->getOperand(0);
2819 // Don't attempt to analyze GEPs over unsized objects.
2820 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2821 return getUnknown(GEP);
2822 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
2823 gep_type_iterator GTI = gep_type_begin(GEP);
2824 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
2828 // Compute the (potentially symbolic) offset in bytes for this index.
2829 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2830 // For a struct, add the member offset.
2831 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2832 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
2834 // Add the field offset to the running total offset.
2835 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
2837 // For an array, add the element offset, explicitly scaled.
2838 const SCEV *ElementSize = getSizeOfExpr(*GTI);
2839 const SCEV *IndexS = getSCEV(Index);
2840 // Getelementptr indices are signed.
2841 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
2843 // Multiply the index by the element size to compute the element offset.
2844 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize);
2846 // Add the element offset to the running total offset.
2847 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2851 // Get the SCEV for the GEP base.
2852 const SCEV *BaseS = getSCEV(Base);
2854 // Add the total offset from all the GEP indices to the base.
2855 return getAddExpr(BaseS, TotalOffset);
2858 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2859 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2860 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2861 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2863 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2864 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2865 return C->getValue()->getValue().countTrailingZeros();
2867 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2868 return std::min(GetMinTrailingZeros(T->getOperand()),
2869 (uint32_t)getTypeSizeInBits(T->getType()));
2871 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2872 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2873 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2874 getTypeSizeInBits(E->getType()) : OpRes;
2877 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2878 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2879 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2880 getTypeSizeInBits(E->getType()) : OpRes;
2883 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2884 // The result is the min of all operands results.
2885 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2886 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2887 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2891 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2892 // The result is the sum of all operands results.
2893 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2894 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2895 for (unsigned i = 1, e = M->getNumOperands();
2896 SumOpRes != BitWidth && i != e; ++i)
2897 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2902 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2903 // The result is the min of all operands results.
2904 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2905 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2906 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2910 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2911 // The result is the min of all operands results.
2912 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2913 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2914 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2918 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2919 // The result is the min of all operands results.
2920 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2921 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2922 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2926 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2927 // For a SCEVUnknown, ask ValueTracking.
2928 unsigned BitWidth = getTypeSizeInBits(U->getType());
2929 APInt Mask = APInt::getAllOnesValue(BitWidth);
2930 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2931 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
2932 return Zeros.countTrailingOnes();
2939 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
2942 ScalarEvolution::getUnsignedRange(const SCEV *S) {
2943 // See if we've computed this range already.
2944 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
2945 if (I != UnsignedRanges.end())
2948 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2949 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
2951 unsigned BitWidth = getTypeSizeInBits(S->getType());
2952 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
2954 // If the value has known zeros, the maximum unsigned value will have those
2955 // known zeros as well.
2956 uint32_t TZ = GetMinTrailingZeros(S);
2958 ConservativeResult =
2959 ConstantRange(APInt::getMinValue(BitWidth),
2960 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
2962 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2963 ConstantRange X = getUnsignedRange(Add->getOperand(0));
2964 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2965 X = X.add(getUnsignedRange(Add->getOperand(i)));
2966 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
2969 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2970 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
2971 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2972 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
2973 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
2976 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2977 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
2978 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2979 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
2980 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
2983 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2984 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
2985 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2986 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
2987 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
2990 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2991 ConstantRange X = getUnsignedRange(UDiv->getLHS());
2992 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
2993 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
2996 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2997 ConstantRange X = getUnsignedRange(ZExt->getOperand());
2998 return setUnsignedRange(ZExt,
2999 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3002 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3003 ConstantRange X = getUnsignedRange(SExt->getOperand());
3004 return setUnsignedRange(SExt,
3005 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3008 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3009 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3010 return setUnsignedRange(Trunc,
3011 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3014 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3015 // If there's no unsigned wrap, the value will never be less than its
3017 if (AddRec->hasNoUnsignedWrap())
3018 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3019 if (!C->getValue()->isZero())
3020 ConservativeResult =
3021 ConservativeResult.intersectWith(
3022 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3024 // TODO: non-affine addrec
3025 if (AddRec->isAffine()) {
3026 const Type *Ty = AddRec->getType();
3027 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3028 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3029 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3030 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3032 const SCEV *Start = AddRec->getStart();
3033 const SCEV *Step = AddRec->getStepRecurrence(*this);
3035 ConstantRange StartRange = getUnsignedRange(Start);
3036 ConstantRange StepRange = getSignedRange(Step);
3037 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3038 ConstantRange EndRange =
3039 StartRange.add(MaxBECountRange.multiply(StepRange));
3041 // Check for overflow. This must be done with ConstantRange arithmetic
3042 // because we could be called from within the ScalarEvolution overflow
3044 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3045 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3046 ConstantRange ExtMaxBECountRange =
3047 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3048 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3049 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3051 return setUnsignedRange(AddRec, ConservativeResult);
3053 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3054 EndRange.getUnsignedMin());
3055 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3056 EndRange.getUnsignedMax());
3057 if (Min.isMinValue() && Max.isMaxValue())
3058 return setUnsignedRange(AddRec, ConservativeResult);
3059 return setUnsignedRange(AddRec,
3060 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3064 return setUnsignedRange(AddRec, ConservativeResult);
3067 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3068 // For a SCEVUnknown, ask ValueTracking.
3069 APInt Mask = APInt::getAllOnesValue(BitWidth);
3070 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3071 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
3072 if (Ones == ~Zeros + 1)
3073 return setUnsignedRange(U, ConservativeResult);
3074 return setUnsignedRange(U,
3075 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3078 return setUnsignedRange(S, ConservativeResult);
3081 /// getSignedRange - Determine the signed range for a particular SCEV.
3084 ScalarEvolution::getSignedRange(const SCEV *S) {
3085 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3086 if (I != SignedRanges.end())
3089 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3090 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3092 unsigned BitWidth = getTypeSizeInBits(S->getType());
3093 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3095 // If the value has known zeros, the maximum signed value will have those
3096 // known zeros as well.
3097 uint32_t TZ = GetMinTrailingZeros(S);
3099 ConservativeResult =
3100 ConstantRange(APInt::getSignedMinValue(BitWidth),
3101 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3103 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3104 ConstantRange X = getSignedRange(Add->getOperand(0));
3105 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3106 X = X.add(getSignedRange(Add->getOperand(i)));
3107 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3110 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3111 ConstantRange X = getSignedRange(Mul->getOperand(0));
3112 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3113 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3114 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3117 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3118 ConstantRange X = getSignedRange(SMax->getOperand(0));
3119 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3120 X = X.smax(getSignedRange(SMax->getOperand(i)));
3121 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3124 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3125 ConstantRange X = getSignedRange(UMax->getOperand(0));
3126 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3127 X = X.umax(getSignedRange(UMax->getOperand(i)));
3128 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3131 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3132 ConstantRange X = getSignedRange(UDiv->getLHS());
3133 ConstantRange Y = getSignedRange(UDiv->getRHS());
3134 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3137 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3138 ConstantRange X = getSignedRange(ZExt->getOperand());
3139 return setSignedRange(ZExt,
3140 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3143 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3144 ConstantRange X = getSignedRange(SExt->getOperand());
3145 return setSignedRange(SExt,
3146 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3149 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3150 ConstantRange X = getSignedRange(Trunc->getOperand());
3151 return setSignedRange(Trunc,
3152 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3155 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3156 // If there's no signed wrap, and all the operands have the same sign or
3157 // zero, the value won't ever change sign.
3158 if (AddRec->hasNoSignedWrap()) {
3159 bool AllNonNeg = true;
3160 bool AllNonPos = true;
3161 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3162 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3163 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3166 ConservativeResult = ConservativeResult.intersectWith(
3167 ConstantRange(APInt(BitWidth, 0),
3168 APInt::getSignedMinValue(BitWidth)));
3170 ConservativeResult = ConservativeResult.intersectWith(
3171 ConstantRange(APInt::getSignedMinValue(BitWidth),
3172 APInt(BitWidth, 1)));
3175 // TODO: non-affine addrec
3176 if (AddRec->isAffine()) {
3177 const Type *Ty = AddRec->getType();
3178 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3179 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3180 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3181 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3183 const SCEV *Start = AddRec->getStart();
3184 const SCEV *Step = AddRec->getStepRecurrence(*this);
3186 ConstantRange StartRange = getSignedRange(Start);
3187 ConstantRange StepRange = getSignedRange(Step);
3188 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3189 ConstantRange EndRange =
3190 StartRange.add(MaxBECountRange.multiply(StepRange));
3192 // Check for overflow. This must be done with ConstantRange arithmetic
3193 // because we could be called from within the ScalarEvolution overflow
3195 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3196 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3197 ConstantRange ExtMaxBECountRange =
3198 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3199 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3200 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3202 return setSignedRange(AddRec, ConservativeResult);
3204 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3205 EndRange.getSignedMin());
3206 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3207 EndRange.getSignedMax());
3208 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3209 return setSignedRange(AddRec, ConservativeResult);
3210 return setSignedRange(AddRec,
3211 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3215 return setSignedRange(AddRec, ConservativeResult);
3218 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3219 // For a SCEVUnknown, ask ValueTracking.
3220 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3221 return setSignedRange(U, ConservativeResult);
3222 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3224 return setSignedRange(U, ConservativeResult);
3225 return setSignedRange(U, ConservativeResult.intersectWith(
3226 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3227 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3230 return setSignedRange(S, ConservativeResult);
3233 /// createSCEV - We know that there is no SCEV for the specified value.
3234 /// Analyze the expression.
3236 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3237 if (!isSCEVable(V->getType()))
3238 return getUnknown(V);
3240 unsigned Opcode = Instruction::UserOp1;
3241 if (Instruction *I = dyn_cast<Instruction>(V)) {
3242 Opcode = I->getOpcode();
3244 // Don't attempt to analyze instructions in blocks that aren't
3245 // reachable. Such instructions don't matter, and they aren't required
3246 // to obey basic rules for definitions dominating uses which this
3247 // analysis depends on.
3248 if (!DT->isReachableFromEntry(I->getParent()))
3249 return getUnknown(V);
3250 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3251 Opcode = CE->getOpcode();
3252 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3253 return getConstant(CI);
3254 else if (isa<ConstantPointerNull>(V))
3255 return getConstant(V->getType(), 0);
3256 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3257 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3259 return getUnknown(V);
3261 Operator *U = cast<Operator>(V);
3263 case Instruction::Add: {
3264 // The simple thing to do would be to just call getSCEV on both operands
3265 // and call getAddExpr with the result. However if we're looking at a
3266 // bunch of things all added together, this can be quite inefficient,
3267 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3268 // Instead, gather up all the operands and make a single getAddExpr call.
3269 // LLVM IR canonical form means we need only traverse the left operands.
3270 SmallVector<const SCEV *, 4> AddOps;
3271 AddOps.push_back(getSCEV(U->getOperand(1)));
3272 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3273 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3274 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3276 U = cast<Operator>(Op);
3277 const SCEV *Op1 = getSCEV(U->getOperand(1));
3278 if (Opcode == Instruction::Sub)
3279 AddOps.push_back(getNegativeSCEV(Op1));
3281 AddOps.push_back(Op1);
3283 AddOps.push_back(getSCEV(U->getOperand(0)));
3284 return getAddExpr(AddOps);
3286 case Instruction::Mul: {
3287 // See the Add code above.
3288 SmallVector<const SCEV *, 4> MulOps;
3289 MulOps.push_back(getSCEV(U->getOperand(1)));
3290 for (Value *Op = U->getOperand(0);
3291 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3292 Op = U->getOperand(0)) {
3293 U = cast<Operator>(Op);
3294 MulOps.push_back(getSCEV(U->getOperand(1)));
3296 MulOps.push_back(getSCEV(U->getOperand(0)));
3297 return getMulExpr(MulOps);
3299 case Instruction::UDiv:
3300 return getUDivExpr(getSCEV(U->getOperand(0)),
3301 getSCEV(U->getOperand(1)));
3302 case Instruction::Sub:
3303 return getMinusSCEV(getSCEV(U->getOperand(0)),
3304 getSCEV(U->getOperand(1)));
3305 case Instruction::And:
3306 // For an expression like x&255 that merely masks off the high bits,
3307 // use zext(trunc(x)) as the SCEV expression.
3308 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3309 if (CI->isNullValue())
3310 return getSCEV(U->getOperand(1));
3311 if (CI->isAllOnesValue())
3312 return getSCEV(U->getOperand(0));
3313 const APInt &A = CI->getValue();
3315 // Instcombine's ShrinkDemandedConstant may strip bits out of
3316 // constants, obscuring what would otherwise be a low-bits mask.
3317 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3318 // knew about to reconstruct a low-bits mask value.
3319 unsigned LZ = A.countLeadingZeros();
3320 unsigned BitWidth = A.getBitWidth();
3321 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
3322 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3323 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
3325 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3327 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3329 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3330 IntegerType::get(getContext(), BitWidth - LZ)),
3335 case Instruction::Or:
3336 // If the RHS of the Or is a constant, we may have something like:
3337 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3338 // optimizations will transparently handle this case.
3340 // In order for this transformation to be safe, the LHS must be of the
3341 // form X*(2^n) and the Or constant must be less than 2^n.
3342 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3343 const SCEV *LHS = getSCEV(U->getOperand(0));
3344 const APInt &CIVal = CI->getValue();
3345 if (GetMinTrailingZeros(LHS) >=
3346 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3347 // Build a plain add SCEV.
3348 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3349 // If the LHS of the add was an addrec and it has no-wrap flags,
3350 // transfer the no-wrap flags, since an or won't introduce a wrap.
3351 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3352 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3353 if (OldAR->hasNoUnsignedWrap())
3354 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3355 if (OldAR->hasNoSignedWrap())
3356 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3362 case Instruction::Xor:
3363 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3364 // If the RHS of the xor is a signbit, then this is just an add.
3365 // Instcombine turns add of signbit into xor as a strength reduction step.
3366 if (CI->getValue().isSignBit())
3367 return getAddExpr(getSCEV(U->getOperand(0)),
3368 getSCEV(U->getOperand(1)));
3370 // If the RHS of xor is -1, then this is a not operation.
3371 if (CI->isAllOnesValue())
3372 return getNotSCEV(getSCEV(U->getOperand(0)));
3374 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3375 // This is a variant of the check for xor with -1, and it handles
3376 // the case where instcombine has trimmed non-demanded bits out
3377 // of an xor with -1.
3378 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3379 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3380 if (BO->getOpcode() == Instruction::And &&
3381 LCI->getValue() == CI->getValue())
3382 if (const SCEVZeroExtendExpr *Z =
3383 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3384 const Type *UTy = U->getType();
3385 const SCEV *Z0 = Z->getOperand();
3386 const Type *Z0Ty = Z0->getType();
3387 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3389 // If C is a low-bits mask, the zero extend is serving to
3390 // mask off the high bits. Complement the operand and
3391 // re-apply the zext.
3392 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3393 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3395 // If C is a single bit, it may be in the sign-bit position
3396 // before the zero-extend. In this case, represent the xor
3397 // using an add, which is equivalent, and re-apply the zext.
3398 APInt Trunc = CI->getValue().trunc(Z0TySize);
3399 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3401 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3407 case Instruction::Shl:
3408 // Turn shift left of a constant amount into a multiply.
3409 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3410 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3412 // If the shift count is not less than the bitwidth, the result of
3413 // the shift is undefined. Don't try to analyze it, because the
3414 // resolution chosen here may differ from the resolution chosen in
3415 // other parts of the compiler.
3416 if (SA->getValue().uge(BitWidth))
3419 Constant *X = ConstantInt::get(getContext(),
3420 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3421 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3425 case Instruction::LShr:
3426 // Turn logical shift right of a constant into a unsigned divide.
3427 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3428 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3430 // If the shift count is not less than the bitwidth, the result of
3431 // the shift is undefined. Don't try to analyze it, because the
3432 // resolution chosen here may differ from the resolution chosen in
3433 // other parts of the compiler.
3434 if (SA->getValue().uge(BitWidth))
3437 Constant *X = ConstantInt::get(getContext(),
3438 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3439 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3443 case Instruction::AShr:
3444 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3445 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3446 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3447 if (L->getOpcode() == Instruction::Shl &&
3448 L->getOperand(1) == U->getOperand(1)) {
3449 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3451 // If the shift count is not less than the bitwidth, the result of
3452 // the shift is undefined. Don't try to analyze it, because the
3453 // resolution chosen here may differ from the resolution chosen in
3454 // other parts of the compiler.
3455 if (CI->getValue().uge(BitWidth))
3458 uint64_t Amt = BitWidth - CI->getZExtValue();
3459 if (Amt == BitWidth)
3460 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3462 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3463 IntegerType::get(getContext(),
3469 case Instruction::Trunc:
3470 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3472 case Instruction::ZExt:
3473 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3475 case Instruction::SExt:
3476 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3478 case Instruction::BitCast:
3479 // BitCasts are no-op casts so we just eliminate the cast.
3480 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3481 return getSCEV(U->getOperand(0));
3484 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3485 // lead to pointer expressions which cannot safely be expanded to GEPs,
3486 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3487 // simplifying integer expressions.
3489 case Instruction::GetElementPtr:
3490 return createNodeForGEP(cast<GEPOperator>(U));
3492 case Instruction::PHI:
3493 return createNodeForPHI(cast<PHINode>(U));
3495 case Instruction::Select:
3496 // This could be a smax or umax that was lowered earlier.
3497 // Try to recover it.
3498 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3499 Value *LHS = ICI->getOperand(0);
3500 Value *RHS = ICI->getOperand(1);
3501 switch (ICI->getPredicate()) {
3502 case ICmpInst::ICMP_SLT:
3503 case ICmpInst::ICMP_SLE:
3504 std::swap(LHS, RHS);
3506 case ICmpInst::ICMP_SGT:
3507 case ICmpInst::ICMP_SGE:
3508 // a >s b ? a+x : b+x -> smax(a, b)+x
3509 // a >s b ? b+x : a+x -> smin(a, b)+x
3510 if (LHS->getType() == U->getType()) {
3511 const SCEV *LS = getSCEV(LHS);
3512 const SCEV *RS = getSCEV(RHS);
3513 const SCEV *LA = getSCEV(U->getOperand(1));
3514 const SCEV *RA = getSCEV(U->getOperand(2));
3515 const SCEV *LDiff = getMinusSCEV(LA, LS);
3516 const SCEV *RDiff = getMinusSCEV(RA, RS);
3518 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3519 LDiff = getMinusSCEV(LA, RS);
3520 RDiff = getMinusSCEV(RA, LS);
3522 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3525 case ICmpInst::ICMP_ULT:
3526 case ICmpInst::ICMP_ULE:
3527 std::swap(LHS, RHS);
3529 case ICmpInst::ICMP_UGT:
3530 case ICmpInst::ICMP_UGE:
3531 // a >u b ? a+x : b+x -> umax(a, b)+x
3532 // a >u b ? b+x : a+x -> umin(a, b)+x
3533 if (LHS->getType() == U->getType()) {
3534 const SCEV *LS = getSCEV(LHS);
3535 const SCEV *RS = getSCEV(RHS);
3536 const SCEV *LA = getSCEV(U->getOperand(1));
3537 const SCEV *RA = getSCEV(U->getOperand(2));
3538 const SCEV *LDiff = getMinusSCEV(LA, LS);
3539 const SCEV *RDiff = getMinusSCEV(RA, RS);
3541 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3542 LDiff = getMinusSCEV(LA, RS);
3543 RDiff = getMinusSCEV(RA, LS);
3545 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3548 case ICmpInst::ICMP_NE:
3549 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3550 if (LHS->getType() == U->getType() &&
3551 isa<ConstantInt>(RHS) &&
3552 cast<ConstantInt>(RHS)->isZero()) {
3553 const SCEV *One = getConstant(LHS->getType(), 1);
3554 const SCEV *LS = getSCEV(LHS);
3555 const SCEV *LA = getSCEV(U->getOperand(1));
3556 const SCEV *RA = getSCEV(U->getOperand(2));
3557 const SCEV *LDiff = getMinusSCEV(LA, LS);
3558 const SCEV *RDiff = getMinusSCEV(RA, One);
3560 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3563 case ICmpInst::ICMP_EQ:
3564 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3565 if (LHS->getType() == U->getType() &&
3566 isa<ConstantInt>(RHS) &&
3567 cast<ConstantInt>(RHS)->isZero()) {
3568 const SCEV *One = getConstant(LHS->getType(), 1);
3569 const SCEV *LS = getSCEV(LHS);
3570 const SCEV *LA = getSCEV(U->getOperand(1));
3571 const SCEV *RA = getSCEV(U->getOperand(2));
3572 const SCEV *LDiff = getMinusSCEV(LA, One);
3573 const SCEV *RDiff = getMinusSCEV(RA, LS);
3575 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3583 default: // We cannot analyze this expression.
3587 return getUnknown(V);
3592 //===----------------------------------------------------------------------===//
3593 // Iteration Count Computation Code
3596 /// getBackedgeTakenCount - If the specified loop has a predictable
3597 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3598 /// object. The backedge-taken count is the number of times the loop header
3599 /// will be branched to from within the loop. This is one less than the
3600 /// trip count of the loop, since it doesn't count the first iteration,
3601 /// when the header is branched to from outside the loop.
3603 /// Note that it is not valid to call this method on a loop without a
3604 /// loop-invariant backedge-taken count (see
3605 /// hasLoopInvariantBackedgeTakenCount).
3607 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3608 return getBackedgeTakenInfo(L).Exact;
3611 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3612 /// return the least SCEV value that is known never to be less than the
3613 /// actual backedge taken count.
3614 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3615 return getBackedgeTakenInfo(L).Max;
3618 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3619 /// onto the given Worklist.
3621 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3622 BasicBlock *Header = L->getHeader();
3624 // Push all Loop-header PHIs onto the Worklist stack.
3625 for (BasicBlock::iterator I = Header->begin();
3626 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3627 Worklist.push_back(PN);
3630 const ScalarEvolution::BackedgeTakenInfo &
3631 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3632 // Initially insert a CouldNotCompute for this loop. If the insertion
3633 // succeeds, proceed to actually compute a backedge-taken count and
3634 // update the value. The temporary CouldNotCompute value tells SCEV
3635 // code elsewhere that it shouldn't attempt to request a new
3636 // backedge-taken count, which could result in infinite recursion.
3637 std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3638 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3640 return Pair.first->second;
3642 BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
3643 if (BECount.Exact != getCouldNotCompute()) {
3644 assert(isLoopInvariant(BECount.Exact, L) &&
3645 isLoopInvariant(BECount.Max, L) &&
3646 "Computed backedge-taken count isn't loop invariant for loop!");
3647 ++NumTripCountsComputed;
3649 // Update the value in the map.
3650 Pair.first->second = BECount;
3652 if (BECount.Max != getCouldNotCompute())
3653 // Update the value in the map.
3654 Pair.first->second = BECount;
3655 if (isa<PHINode>(L->getHeader()->begin()))
3656 // Only count loops that have phi nodes as not being computable.
3657 ++NumTripCountsNotComputed;
3660 // Now that we know more about the trip count for this loop, forget any
3661 // existing SCEV values for PHI nodes in this loop since they are only
3662 // conservative estimates made without the benefit of trip count
3663 // information. This is similar to the code in forgetLoop, except that
3664 // it handles SCEVUnknown PHI nodes specially.
3665 if (BECount.hasAnyInfo()) {
3666 SmallVector<Instruction *, 16> Worklist;
3667 PushLoopPHIs(L, Worklist);
3669 SmallPtrSet<Instruction *, 8> Visited;
3670 while (!Worklist.empty()) {
3671 Instruction *I = Worklist.pop_back_val();
3672 if (!Visited.insert(I)) continue;
3674 ValueExprMapType::iterator It =
3675 ValueExprMap.find(static_cast<Value *>(I));
3676 if (It != ValueExprMap.end()) {
3677 const SCEV *Old = It->second;
3679 // SCEVUnknown for a PHI either means that it has an unrecognized
3680 // structure, or it's a PHI that's in the progress of being computed
3681 // by createNodeForPHI. In the former case, additional loop trip
3682 // count information isn't going to change anything. In the later
3683 // case, createNodeForPHI will perform the necessary updates on its
3684 // own when it gets to that point.
3685 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
3686 forgetMemoizedResults(Old);
3687 ValueExprMap.erase(It);
3689 if (PHINode *PN = dyn_cast<PHINode>(I))
3690 ConstantEvolutionLoopExitValue.erase(PN);
3693 PushDefUseChildren(I, Worklist);
3696 return Pair.first->second;
3699 /// forgetLoop - This method should be called by the client when it has
3700 /// changed a loop in a way that may effect ScalarEvolution's ability to
3701 /// compute a trip count, or if the loop is deleted.
3702 void ScalarEvolution::forgetLoop(const Loop *L) {
3703 // Drop any stored trip count value.
3704 BackedgeTakenCounts.erase(L);
3706 // Drop information about expressions based on loop-header PHIs.
3707 SmallVector<Instruction *, 16> Worklist;
3708 PushLoopPHIs(L, Worklist);
3710 SmallPtrSet<Instruction *, 8> Visited;
3711 while (!Worklist.empty()) {
3712 Instruction *I = Worklist.pop_back_val();
3713 if (!Visited.insert(I)) continue;
3715 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3716 if (It != ValueExprMap.end()) {
3717 forgetMemoizedResults(It->second);
3718 ValueExprMap.erase(It);
3719 if (PHINode *PN = dyn_cast<PHINode>(I))
3720 ConstantEvolutionLoopExitValue.erase(PN);
3723 PushDefUseChildren(I, Worklist);
3726 // Forget all contained loops too, to avoid dangling entries in the
3727 // ValuesAtScopes map.
3728 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
3732 /// forgetValue - This method should be called by the client when it has
3733 /// changed a value in a way that may effect its value, or which may
3734 /// disconnect it from a def-use chain linking it to a loop.
3735 void ScalarEvolution::forgetValue(Value *V) {
3736 Instruction *I = dyn_cast<Instruction>(V);
3739 // Drop information about expressions based on loop-header PHIs.
3740 SmallVector<Instruction *, 16> Worklist;
3741 Worklist.push_back(I);
3743 SmallPtrSet<Instruction *, 8> Visited;
3744 while (!Worklist.empty()) {
3745 I = Worklist.pop_back_val();
3746 if (!Visited.insert(I)) continue;
3748 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3749 if (It != ValueExprMap.end()) {
3750 forgetMemoizedResults(It->second);
3751 ValueExprMap.erase(It);
3752 if (PHINode *PN = dyn_cast<PHINode>(I))
3753 ConstantEvolutionLoopExitValue.erase(PN);
3756 PushDefUseChildren(I, Worklist);
3760 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3761 /// of the specified loop will execute.
3762 ScalarEvolution::BackedgeTakenInfo
3763 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3764 SmallVector<BasicBlock *, 8> ExitingBlocks;
3765 L->getExitingBlocks(ExitingBlocks);
3767 // Examine all exits and pick the most conservative values.
3768 const SCEV *BECount = getCouldNotCompute();
3769 const SCEV *MaxBECount = getCouldNotCompute();
3770 bool CouldNotComputeBECount = false;
3771 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3772 BackedgeTakenInfo NewBTI =
3773 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3775 if (NewBTI.Exact == getCouldNotCompute()) {
3776 // We couldn't compute an exact value for this exit, so
3777 // we won't be able to compute an exact value for the loop.
3778 CouldNotComputeBECount = true;
3779 BECount = getCouldNotCompute();
3780 } else if (!CouldNotComputeBECount) {
3781 if (BECount == getCouldNotCompute())
3782 BECount = NewBTI.Exact;
3784 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3786 if (MaxBECount == getCouldNotCompute())
3787 MaxBECount = NewBTI.Max;
3788 else if (NewBTI.Max != getCouldNotCompute())
3789 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3792 return BackedgeTakenInfo(BECount, MaxBECount);
3795 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3796 /// of the specified loop will execute if it exits via the specified block.
3797 ScalarEvolution::BackedgeTakenInfo
3798 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3799 BasicBlock *ExitingBlock) {
3801 // Okay, we've chosen an exiting block. See what condition causes us to
3802 // exit at this block.
3804 // FIXME: we should be able to handle switch instructions (with a single exit)
3805 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3806 if (ExitBr == 0) return getCouldNotCompute();
3807 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3809 // At this point, we know we have a conditional branch that determines whether
3810 // the loop is exited. However, we don't know if the branch is executed each
3811 // time through the loop. If not, then the execution count of the branch will
3812 // not be equal to the trip count of the loop.
3814 // Currently we check for this by checking to see if the Exit branch goes to
3815 // the loop header. If so, we know it will always execute the same number of
3816 // times as the loop. We also handle the case where the exit block *is* the
3817 // loop header. This is common for un-rotated loops.
3819 // If both of those tests fail, walk up the unique predecessor chain to the
3820 // header, stopping if there is an edge that doesn't exit the loop. If the
3821 // header is reached, the execution count of the branch will be equal to the
3822 // trip count of the loop.
3824 // More extensive analysis could be done to handle more cases here.
3826 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3827 ExitBr->getSuccessor(1) != L->getHeader() &&
3828 ExitBr->getParent() != L->getHeader()) {
3829 // The simple checks failed, try climbing the unique predecessor chain
3830 // up to the header.
3832 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3833 BasicBlock *Pred = BB->getUniquePredecessor();
3835 return getCouldNotCompute();
3836 TerminatorInst *PredTerm = Pred->getTerminator();
3837 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3838 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3841 // If the predecessor has a successor that isn't BB and isn't
3842 // outside the loop, assume the worst.
3843 if (L->contains(PredSucc))
3844 return getCouldNotCompute();
3846 if (Pred == L->getHeader()) {
3853 return getCouldNotCompute();
3856 // Proceed to the next level to examine the exit condition expression.
3857 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3858 ExitBr->getSuccessor(0),
3859 ExitBr->getSuccessor(1));
3862 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3863 /// backedge of the specified loop will execute if its exit condition
3864 /// were a conditional branch of ExitCond, TBB, and FBB.
3865 ScalarEvolution::BackedgeTakenInfo
3866 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3870 // Check if the controlling expression for this loop is an And or Or.
3871 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3872 if (BO->getOpcode() == Instruction::And) {
3873 // Recurse on the operands of the and.
3874 BackedgeTakenInfo BTI0 =
3875 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3876 BackedgeTakenInfo BTI1 =
3877 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3878 const SCEV *BECount = getCouldNotCompute();
3879 const SCEV *MaxBECount = getCouldNotCompute();
3880 if (L->contains(TBB)) {
3881 // Both conditions must be true for the loop to continue executing.
3882 // Choose the less conservative count.
3883 if (BTI0.Exact == getCouldNotCompute() ||
3884 BTI1.Exact == getCouldNotCompute())
3885 BECount = getCouldNotCompute();
3887 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3888 if (BTI0.Max == getCouldNotCompute())
3889 MaxBECount = BTI1.Max;
3890 else if (BTI1.Max == getCouldNotCompute())
3891 MaxBECount = BTI0.Max;
3893 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3895 // Both conditions must be true at the same time for the loop to exit.
3896 // For now, be conservative.
3897 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3898 if (BTI0.Max == BTI1.Max)
3899 MaxBECount = BTI0.Max;
3900 if (BTI0.Exact == BTI1.Exact)
3901 BECount = BTI0.Exact;
3904 return BackedgeTakenInfo(BECount, MaxBECount);
3906 if (BO->getOpcode() == Instruction::Or) {
3907 // Recurse on the operands of the or.
3908 BackedgeTakenInfo BTI0 =
3909 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3910 BackedgeTakenInfo BTI1 =
3911 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3912 const SCEV *BECount = getCouldNotCompute();
3913 const SCEV *MaxBECount = getCouldNotCompute();
3914 if (L->contains(FBB)) {
3915 // Both conditions must be false for the loop to continue executing.
3916 // Choose the less conservative count.
3917 if (BTI0.Exact == getCouldNotCompute() ||
3918 BTI1.Exact == getCouldNotCompute())
3919 BECount = getCouldNotCompute();
3921 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3922 if (BTI0.Max == getCouldNotCompute())
3923 MaxBECount = BTI1.Max;
3924 else if (BTI1.Max == getCouldNotCompute())
3925 MaxBECount = BTI0.Max;
3927 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3929 // Both conditions must be false at the same time for the loop to exit.
3930 // For now, be conservative.
3931 assert(L->contains(TBB) && "Loop block has no successor in loop!");
3932 if (BTI0.Max == BTI1.Max)
3933 MaxBECount = BTI0.Max;
3934 if (BTI0.Exact == BTI1.Exact)
3935 BECount = BTI0.Exact;
3938 return BackedgeTakenInfo(BECount, MaxBECount);
3942 // With an icmp, it may be feasible to compute an exact backedge-taken count.
3943 // Proceed to the next level to examine the icmp.
3944 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
3945 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
3947 // Check for a constant condition. These are normally stripped out by
3948 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
3949 // preserve the CFG and is temporarily leaving constant conditions
3951 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
3952 if (L->contains(FBB) == !CI->getZExtValue())
3953 // The backedge is always taken.
3954 return getCouldNotCompute();
3956 // The backedge is never taken.
3957 return getConstant(CI->getType(), 0);
3960 // If it's not an integer or pointer comparison then compute it the hard way.
3961 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3964 static const SCEVAddRecExpr *
3965 isSimpleUnwrappingAddRec(const SCEV *S, const Loop *L) {
3966 const SCEVAddRecExpr *SA = dyn_cast<SCEVAddRecExpr>(S);
3968 // The SCEV must be an addrec of this loop.
3969 if (!SA || SA->getLoop() != L || !SA->isAffine())
3972 // The SCEV must be known to not wrap in some way to be interesting.
3973 if (!SA->hasNoUnsignedWrap() && !SA->hasNoSignedWrap())
3976 // The stride must be a constant so that we know if it is striding up or down.
3977 if (!isa<SCEVConstant>(SA->getOperand(1)))
3982 /// getMinusSCEVForExitTest - When considering an exit test for a loop with a
3983 /// "x != y" exit test, we turn this into a computation that evaluates x-y != 0,
3984 /// and this function returns the expression to use for x-y. We know and take
3985 /// advantage of the fact that this subtraction is only being used in a
3986 /// comparison by zero context.
3988 static const SCEV *getMinusSCEVForExitTest(const SCEV *LHS, const SCEV *RHS,
3989 const Loop *L, ScalarEvolution &SE) {
3990 // If either LHS or RHS is an AddRec SCEV (of this loop) that is known to not
3991 // wrap (either NSW or NUW), then we know that the value will either become
3992 // the other one (and thus the loop terminates), that the loop will terminate
3993 // through some other exit condition first, or that the loop has undefined
3994 // behavior. This information is useful when the addrec has a stride that is
3995 // != 1 or -1, because it means we can't "miss" the exit value.
3997 // In any of these three cases, it is safe to turn the exit condition into a
3998 // "counting down" AddRec (to zero) by subtracting the two inputs as normal,
3999 // but since we know that the "end cannot be missed" we can force the
4000 // resulting AddRec to be a NUW addrec. Since it is counting down, this means
4001 // that the AddRec *cannot* pass zero.
4003 // See if LHS and RHS are addrec's we can handle.
4004 const SCEVAddRecExpr *LHSA = isSimpleUnwrappingAddRec(LHS, L);
4005 const SCEVAddRecExpr *RHSA = isSimpleUnwrappingAddRec(RHS, L);
4007 // If neither addrec is interesting, just return a minus.
4008 if (RHSA == 0 && LHSA == 0)
4009 return SE.getMinusSCEV(LHS, RHS);
4011 // If only one of LHS and RHS are an AddRec of this loop, make sure it is LHS.
4012 if (RHSA && LHSA == 0) {
4013 // Safe because a-b === b-a for comparisons against zero.
4014 std::swap(LHS, RHS);
4015 std::swap(LHSA, RHSA);
4018 // Handle the case when only one is advancing in a non-overflowing way.
4020 // If RHS is loop varying, then we can't predict when LHS will cross it.
4021 if (!SE.isLoopInvariant(RHS, L))
4022 return SE.getMinusSCEV(LHS, RHS);
4024 // If LHS has a positive stride, then we compute RHS-LHS, because the loop
4025 // is counting up until it crosses RHS (which must be larger than LHS). If
4026 // it is negative, we compute LHS-RHS because we're counting down to RHS.
4027 const ConstantInt *Stride =
4028 cast<SCEVConstant>(LHSA->getOperand(1))->getValue();
4029 if (Stride->getValue().isNegative())
4030 std::swap(LHS, RHS);
4032 return SE.getMinusSCEV(RHS, LHS, true /*HasNUW*/);
4035 // If both LHS and RHS are interesting, we have something like:
4037 const ConstantInt *LHSStride =
4038 cast<SCEVConstant>(LHSA->getOperand(1))->getValue();
4039 const ConstantInt *RHSStride =
4040 cast<SCEVConstant>(RHSA->getOperand(1))->getValue();
4042 // If the strides are equal, then this is just a (complex) loop invariant
4043 // comparison of a and b.
4044 if (LHSStride == RHSStride)
4045 return SE.getMinusSCEV(LHSA->getStart(), RHSA->getStart());
4047 // If the signs of the strides differ, then the negative stride is counting
4048 // down to the positive stride.
4049 if (LHSStride->getValue().isNegative() != RHSStride->getValue().isNegative()){
4050 if (RHSStride->getValue().isNegative())
4051 std::swap(LHS, RHS);
4053 // If LHS's stride is smaller than RHS's stride, then "b" must be less than
4054 // "a" and "b" is RHS is counting up (catching up) to LHS. This is true
4055 // whether the strides are positive or negative.
4056 if (RHSStride->getValue().slt(LHSStride->getValue()))
4057 std::swap(LHS, RHS);
4060 return SE.getMinusSCEV(LHS, RHS, true /*HasNUW*/);
4063 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
4064 /// backedge of the specified loop will execute if its exit condition
4065 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4066 ScalarEvolution::BackedgeTakenInfo
4067 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
4072 // If the condition was exit on true, convert the condition to exit on false
4073 ICmpInst::Predicate Cond;
4074 if (!L->contains(FBB))
4075 Cond = ExitCond->getPredicate();
4077 Cond = ExitCond->getInversePredicate();
4079 // Handle common loops like: for (X = "string"; *X; ++X)
4080 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4081 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4082 BackedgeTakenInfo ItCnt =
4083 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
4084 if (ItCnt.hasAnyInfo())
4088 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4089 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4091 // Try to evaluate any dependencies out of the loop.
4092 LHS = getSCEVAtScope(LHS, L);
4093 RHS = getSCEVAtScope(RHS, L);
4095 // At this point, we would like to compute how many iterations of the
4096 // loop the predicate will return true for these inputs.
4097 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4098 // If there is a loop-invariant, force it into the RHS.
4099 std::swap(LHS, RHS);
4100 Cond = ICmpInst::getSwappedPredicate(Cond);
4103 // Simplify the operands before analyzing them.
4104 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4106 // If we have a comparison of a chrec against a constant, try to use value
4107 // ranges to answer this query.
4108 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4109 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4110 if (AddRec->getLoop() == L) {
4111 // Form the constant range.
4112 ConstantRange CompRange(
4113 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4115 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4116 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4120 case ICmpInst::ICMP_NE: { // while (X != Y)
4121 // Convert to: while (X-Y != 0)
4122 BackedgeTakenInfo BTI = HowFarToZero(getMinusSCEVForExitTest(LHS, RHS, L,
4124 if (BTI.hasAnyInfo()) return BTI;
4127 case ICmpInst::ICMP_EQ: { // while (X == Y)
4128 // Convert to: while (X-Y == 0)
4129 BackedgeTakenInfo BTI = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4130 if (BTI.hasAnyInfo()) return BTI;
4133 case ICmpInst::ICMP_SLT: {
4134 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
4135 if (BTI.hasAnyInfo()) return BTI;
4138 case ICmpInst::ICMP_SGT: {
4139 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4140 getNotSCEV(RHS), L, true);
4141 if (BTI.hasAnyInfo()) return BTI;
4144 case ICmpInst::ICMP_ULT: {
4145 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
4146 if (BTI.hasAnyInfo()) return BTI;
4149 case ICmpInst::ICMP_UGT: {
4150 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4151 getNotSCEV(RHS), L, false);
4152 if (BTI.hasAnyInfo()) return BTI;
4157 dbgs() << "ComputeBackedgeTakenCount ";
4158 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4159 dbgs() << "[unsigned] ";
4160 dbgs() << *LHS << " "
4161 << Instruction::getOpcodeName(Instruction::ICmp)
4162 << " " << *RHS << "\n";
4167 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
4170 static ConstantInt *
4171 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4172 ScalarEvolution &SE) {
4173 const SCEV *InVal = SE.getConstant(C);
4174 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4175 assert(isa<SCEVConstant>(Val) &&
4176 "Evaluation of SCEV at constant didn't fold correctly?");
4177 return cast<SCEVConstant>(Val)->getValue();
4180 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
4181 /// and a GEP expression (missing the pointer index) indexing into it, return
4182 /// the addressed element of the initializer or null if the index expression is
4185 GetAddressedElementFromGlobal(GlobalVariable *GV,
4186 const std::vector<ConstantInt*> &Indices) {
4187 Constant *Init = GV->getInitializer();
4188 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
4189 uint64_t Idx = Indices[i]->getZExtValue();
4190 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
4191 assert(Idx < CS->getNumOperands() && "Bad struct index!");
4192 Init = cast<Constant>(CS->getOperand(Idx));
4193 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
4194 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
4195 Init = cast<Constant>(CA->getOperand(Idx));
4196 } else if (isa<ConstantAggregateZero>(Init)) {
4197 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
4198 assert(Idx < STy->getNumElements() && "Bad struct index!");
4199 Init = Constant::getNullValue(STy->getElementType(Idx));
4200 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
4201 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
4202 Init = Constant::getNullValue(ATy->getElementType());
4204 llvm_unreachable("Unknown constant aggregate type!");
4208 return 0; // Unknown initializer type
4214 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
4215 /// 'icmp op load X, cst', try to see if we can compute the backedge
4216 /// execution count.
4217 ScalarEvolution::BackedgeTakenInfo
4218 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
4222 ICmpInst::Predicate predicate) {
4223 if (LI->isVolatile()) return getCouldNotCompute();
4225 // Check to see if the loaded pointer is a getelementptr of a global.
4226 // TODO: Use SCEV instead of manually grubbing with GEPs.
4227 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4228 if (!GEP) return getCouldNotCompute();
4230 // Make sure that it is really a constant global we are gepping, with an
4231 // initializer, and make sure the first IDX is really 0.
4232 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4233 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4234 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4235 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4236 return getCouldNotCompute();
4238 // Okay, we allow one non-constant index into the GEP instruction.
4240 std::vector<ConstantInt*> Indexes;
4241 unsigned VarIdxNum = 0;
4242 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4243 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4244 Indexes.push_back(CI);
4245 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4246 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4247 VarIdx = GEP->getOperand(i);
4249 Indexes.push_back(0);
4252 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4253 // Check to see if X is a loop variant variable value now.
4254 const SCEV *Idx = getSCEV(VarIdx);
4255 Idx = getSCEVAtScope(Idx, L);
4257 // We can only recognize very limited forms of loop index expressions, in
4258 // particular, only affine AddRec's like {C1,+,C2}.
4259 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4260 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4261 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4262 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4263 return getCouldNotCompute();
4265 unsigned MaxSteps = MaxBruteForceIterations;
4266 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4267 ConstantInt *ItCst = ConstantInt::get(
4268 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4269 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4271 // Form the GEP offset.
4272 Indexes[VarIdxNum] = Val;
4274 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
4275 if (Result == 0) break; // Cannot compute!
4277 // Evaluate the condition for this iteration.
4278 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4279 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4280 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4282 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4283 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4286 ++NumArrayLenItCounts;
4287 return getConstant(ItCst); // Found terminating iteration!
4290 return getCouldNotCompute();
4294 /// CanConstantFold - Return true if we can constant fold an instruction of the
4295 /// specified type, assuming that all operands were constants.
4296 static bool CanConstantFold(const Instruction *I) {
4297 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4298 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
4301 if (const CallInst *CI = dyn_cast<CallInst>(I))
4302 if (const Function *F = CI->getCalledFunction())
4303 return canConstantFoldCallTo(F);
4307 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4308 /// in the loop that V is derived from. We allow arbitrary operations along the
4309 /// way, but the operands of an operation must either be constants or a value
4310 /// derived from a constant PHI. If this expression does not fit with these
4311 /// constraints, return null.
4312 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4313 // If this is not an instruction, or if this is an instruction outside of the
4314 // loop, it can't be derived from a loop PHI.
4315 Instruction *I = dyn_cast<Instruction>(V);
4316 if (I == 0 || !L->contains(I)) return 0;
4318 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4319 if (L->getHeader() == I->getParent())
4322 // We don't currently keep track of the control flow needed to evaluate
4323 // PHIs, so we cannot handle PHIs inside of loops.
4327 // If we won't be able to constant fold this expression even if the operands
4328 // are constants, return early.
4329 if (!CanConstantFold(I)) return 0;
4331 // Otherwise, we can evaluate this instruction if all of its operands are
4332 // constant or derived from a PHI node themselves.
4334 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
4335 if (!isa<Constant>(I->getOperand(Op))) {
4336 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
4337 if (P == 0) return 0; // Not evolving from PHI
4341 return 0; // Evolving from multiple different PHIs.
4344 // This is a expression evolving from a constant PHI!
4348 /// EvaluateExpression - Given an expression that passes the
4349 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4350 /// in the loop has the value PHIVal. If we can't fold this expression for some
4351 /// reason, return null.
4352 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
4353 const TargetData *TD) {
4354 if (isa<PHINode>(V)) return PHIVal;
4355 if (Constant *C = dyn_cast<Constant>(V)) return C;
4356 Instruction *I = cast<Instruction>(V);
4358 std::vector<Constant*> Operands(I->getNumOperands());
4360 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4361 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
4362 if (Operands[i] == 0) return 0;
4365 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4366 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4368 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4369 &Operands[0], Operands.size(), TD);
4372 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4373 /// in the header of its containing loop, we know the loop executes a
4374 /// constant number of times, and the PHI node is just a recurrence
4375 /// involving constants, fold it.
4377 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4380 std::map<PHINode*, Constant*>::const_iterator I =
4381 ConstantEvolutionLoopExitValue.find(PN);
4382 if (I != ConstantEvolutionLoopExitValue.end())
4385 if (BEs.ugt(MaxBruteForceIterations))
4386 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4388 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4390 // Since the loop is canonicalized, the PHI node must have two entries. One
4391 // entry must be a constant (coming in from outside of the loop), and the
4392 // second must be derived from the same PHI.
4393 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4394 Constant *StartCST =
4395 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4397 return RetVal = 0; // Must be a constant.
4399 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4400 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4401 !isa<Constant>(BEValue))
4402 return RetVal = 0; // Not derived from same PHI.
4404 // Execute the loop symbolically to determine the exit value.
4405 if (BEs.getActiveBits() >= 32)
4406 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4408 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4409 unsigned IterationNum = 0;
4410 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
4411 if (IterationNum == NumIterations)
4412 return RetVal = PHIVal; // Got exit value!
4414 // Compute the value of the PHI node for the next iteration.
4415 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4416 if (NextPHI == PHIVal)
4417 return RetVal = NextPHI; // Stopped evolving!
4419 return 0; // Couldn't evaluate!
4424 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
4425 /// constant number of times (the condition evolves only from constants),
4426 /// try to evaluate a few iterations of the loop until we get the exit
4427 /// condition gets a value of ExitWhen (true or false). If we cannot
4428 /// evaluate the trip count of the loop, return getCouldNotCompute().
4430 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
4433 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4434 if (PN == 0) return getCouldNotCompute();
4436 // If the loop is canonicalized, the PHI will have exactly two entries.
4437 // That's the only form we support here.
4438 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4440 // One entry must be a constant (coming in from outside of the loop), and the
4441 // second must be derived from the same PHI.
4442 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4443 Constant *StartCST =
4444 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4445 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
4447 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4448 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4449 !isa<Constant>(BEValue))
4450 return getCouldNotCompute(); // Not derived from same PHI.
4452 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4453 // the loop symbolically to determine when the condition gets a value of
4455 unsigned IterationNum = 0;
4456 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4457 for (Constant *PHIVal = StartCST;
4458 IterationNum != MaxIterations; ++IterationNum) {
4459 ConstantInt *CondVal =
4460 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
4462 // Couldn't symbolically evaluate.
4463 if (!CondVal) return getCouldNotCompute();
4465 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4466 ++NumBruteForceTripCountsComputed;
4467 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4470 // Compute the value of the PHI node for the next iteration.
4471 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4472 if (NextPHI == 0 || NextPHI == PHIVal)
4473 return getCouldNotCompute();// Couldn't evaluate or not making progress...
4477 // Too many iterations were needed to evaluate.
4478 return getCouldNotCompute();
4481 /// getSCEVAtScope - Return a SCEV expression for the specified value
4482 /// at the specified scope in the program. The L value specifies a loop
4483 /// nest to evaluate the expression at, where null is the top-level or a
4484 /// specified loop is immediately inside of the loop.
4486 /// This method can be used to compute the exit value for a variable defined
4487 /// in a loop by querying what the value will hold in the parent loop.
4489 /// In the case that a relevant loop exit value cannot be computed, the
4490 /// original value V is returned.
4491 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4492 // Check to see if we've folded this expression at this loop before.
4493 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4494 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4495 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4497 return Pair.first->second ? Pair.first->second : V;
4499 // Otherwise compute it.
4500 const SCEV *C = computeSCEVAtScope(V, L);
4501 ValuesAtScopes[V][L] = C;
4505 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
4506 if (isa<SCEVConstant>(V)) return V;
4508 // If this instruction is evolved from a constant-evolving PHI, compute the
4509 // exit value from the loop without using SCEVs.
4510 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
4511 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
4512 const Loop *LI = (*this->LI)[I->getParent()];
4513 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
4514 if (PHINode *PN = dyn_cast<PHINode>(I))
4515 if (PN->getParent() == LI->getHeader()) {
4516 // Okay, there is no closed form solution for the PHI node. Check
4517 // to see if the loop that contains it has a known backedge-taken
4518 // count. If so, we may be able to force computation of the exit
4520 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
4521 if (const SCEVConstant *BTCC =
4522 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
4523 // Okay, we know how many times the containing loop executes. If
4524 // this is a constant evolving PHI node, get the final value at
4525 // the specified iteration number.
4526 Constant *RV = getConstantEvolutionLoopExitValue(PN,
4527 BTCC->getValue()->getValue(),
4529 if (RV) return getSCEV(RV);
4533 // Okay, this is an expression that we cannot symbolically evaluate
4534 // into a SCEV. Check to see if it's possible to symbolically evaluate
4535 // the arguments into constants, and if so, try to constant propagate the
4536 // result. This is particularly useful for computing loop exit values.
4537 if (CanConstantFold(I)) {
4538 SmallVector<Constant *, 4> Operands;
4539 bool MadeImprovement = false;
4540 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4541 Value *Op = I->getOperand(i);
4542 if (Constant *C = dyn_cast<Constant>(Op)) {
4543 Operands.push_back(C);
4547 // If any of the operands is non-constant and if they are
4548 // non-integer and non-pointer, don't even try to analyze them
4549 // with scev techniques.
4550 if (!isSCEVable(Op->getType()))
4553 const SCEV *OrigV = getSCEV(Op);
4554 const SCEV *OpV = getSCEVAtScope(OrigV, L);
4555 MadeImprovement |= OrigV != OpV;
4558 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
4560 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV))
4561 C = dyn_cast<Constant>(SU->getValue());
4563 if (C->getType() != Op->getType())
4564 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4568 Operands.push_back(C);
4571 // Check to see if getSCEVAtScope actually made an improvement.
4572 if (MadeImprovement) {
4574 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4575 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4576 Operands[0], Operands[1], TD);
4578 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4579 &Operands[0], Operands.size(), TD);
4586 // This is some other type of SCEVUnknown, just return it.
4590 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4591 // Avoid performing the look-up in the common case where the specified
4592 // expression has no loop-variant portions.
4593 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4594 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4595 if (OpAtScope != Comm->getOperand(i)) {
4596 // Okay, at least one of these operands is loop variant but might be
4597 // foldable. Build a new instance of the folded commutative expression.
4598 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4599 Comm->op_begin()+i);
4600 NewOps.push_back(OpAtScope);
4602 for (++i; i != e; ++i) {
4603 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4604 NewOps.push_back(OpAtScope);
4606 if (isa<SCEVAddExpr>(Comm))
4607 return getAddExpr(NewOps);
4608 if (isa<SCEVMulExpr>(Comm))
4609 return getMulExpr(NewOps);
4610 if (isa<SCEVSMaxExpr>(Comm))
4611 return getSMaxExpr(NewOps);
4612 if (isa<SCEVUMaxExpr>(Comm))
4613 return getUMaxExpr(NewOps);
4614 llvm_unreachable("Unknown commutative SCEV type!");
4617 // If we got here, all operands are loop invariant.
4621 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4622 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4623 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4624 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4625 return Div; // must be loop invariant
4626 return getUDivExpr(LHS, RHS);
4629 // If this is a loop recurrence for a loop that does not contain L, then we
4630 // are dealing with the final value computed by the loop.
4631 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4632 // First, attempt to evaluate each operand.
4633 // Avoid performing the look-up in the common case where the specified
4634 // expression has no loop-variant portions.
4635 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
4636 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
4637 if (OpAtScope == AddRec->getOperand(i))
4640 // Okay, at least one of these operands is loop variant but might be
4641 // foldable. Build a new instance of the folded commutative expression.
4642 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
4643 AddRec->op_begin()+i);
4644 NewOps.push_back(OpAtScope);
4645 for (++i; i != e; ++i)
4646 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
4648 AddRec = cast<SCEVAddRecExpr>(getAddRecExpr(NewOps, AddRec->getLoop()));
4652 // If the scope is outside the addrec's loop, evaluate it by using the
4653 // loop exit value of the addrec.
4654 if (!AddRec->getLoop()->contains(L)) {
4655 // To evaluate this recurrence, we need to know how many times the AddRec
4656 // loop iterates. Compute this now.
4657 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4658 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4660 // Then, evaluate the AddRec.
4661 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4667 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4668 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4669 if (Op == Cast->getOperand())
4670 return Cast; // must be loop invariant
4671 return getZeroExtendExpr(Op, Cast->getType());
4674 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4675 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4676 if (Op == Cast->getOperand())
4677 return Cast; // must be loop invariant
4678 return getSignExtendExpr(Op, Cast->getType());
4681 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4682 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4683 if (Op == Cast->getOperand())
4684 return Cast; // must be loop invariant
4685 return getTruncateExpr(Op, Cast->getType());
4688 llvm_unreachable("Unknown SCEV type!");
4692 /// getSCEVAtScope - This is a convenience function which does
4693 /// getSCEVAtScope(getSCEV(V), L).
4694 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4695 return getSCEVAtScope(getSCEV(V), L);
4698 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4699 /// following equation:
4701 /// A * X = B (mod N)
4703 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4704 /// A and B isn't important.
4706 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4707 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4708 ScalarEvolution &SE) {
4709 uint32_t BW = A.getBitWidth();
4710 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4711 assert(A != 0 && "A must be non-zero.");
4715 // The gcd of A and N may have only one prime factor: 2. The number of
4716 // trailing zeros in A is its multiplicity
4717 uint32_t Mult2 = A.countTrailingZeros();
4720 // 2. Check if B is divisible by D.
4722 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4723 // is not less than multiplicity of this prime factor for D.
4724 if (B.countTrailingZeros() < Mult2)
4725 return SE.getCouldNotCompute();
4727 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4730 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4731 // bit width during computations.
4732 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4733 APInt Mod(BW + 1, 0);
4734 Mod.setBit(BW - Mult2); // Mod = N / D
4735 APInt I = AD.multiplicativeInverse(Mod);
4737 // 4. Compute the minimum unsigned root of the equation:
4738 // I * (B / D) mod (N / D)
4739 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4741 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4743 return SE.getConstant(Result.trunc(BW));
4746 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4747 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4748 /// might be the same) or two SCEVCouldNotCompute objects.
4750 static std::pair<const SCEV *,const SCEV *>
4751 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4752 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4753 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4754 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4755 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4757 // We currently can only solve this if the coefficients are constants.
4758 if (!LC || !MC || !NC) {
4759 const SCEV *CNC = SE.getCouldNotCompute();
4760 return std::make_pair(CNC, CNC);
4763 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4764 const APInt &L = LC->getValue()->getValue();
4765 const APInt &M = MC->getValue()->getValue();
4766 const APInt &N = NC->getValue()->getValue();
4767 APInt Two(BitWidth, 2);
4768 APInt Four(BitWidth, 4);
4771 using namespace APIntOps;
4773 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4774 // The B coefficient is M-N/2
4778 // The A coefficient is N/2
4779 APInt A(N.sdiv(Two));
4781 // Compute the B^2-4ac term.
4784 SqrtTerm -= Four * (A * C);
4786 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4787 // integer value or else APInt::sqrt() will assert.
4788 APInt SqrtVal(SqrtTerm.sqrt());
4790 // Compute the two solutions for the quadratic formula.
4791 // The divisions must be performed as signed divisions.
4793 APInt TwoA( A << 1 );
4794 if (TwoA.isMinValue()) {
4795 const SCEV *CNC = SE.getCouldNotCompute();
4796 return std::make_pair(CNC, CNC);
4799 LLVMContext &Context = SE.getContext();
4801 ConstantInt *Solution1 =
4802 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4803 ConstantInt *Solution2 =
4804 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4806 return std::make_pair(SE.getConstant(Solution1),
4807 SE.getConstant(Solution2));
4808 } // end APIntOps namespace
4811 /// HowFarToZero - Return the number of times a backedge comparing the specified
4812 /// value to zero will execute. If not computable, return CouldNotCompute.
4813 ScalarEvolution::BackedgeTakenInfo
4814 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4815 // If the value is a constant
4816 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4817 // If the value is already zero, the branch will execute zero times.
4818 if (C->getValue()->isZero()) return C;
4819 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4822 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4823 if (!AddRec || AddRec->getLoop() != L)
4824 return getCouldNotCompute();
4826 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4827 // the quadratic equation to solve it.
4828 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
4829 std::pair<const SCEV *,const SCEV *> Roots =
4830 SolveQuadraticEquation(AddRec, *this);
4831 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4832 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4835 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
4836 << " sol#2: " << *R2 << "\n";
4838 // Pick the smallest positive root value.
4839 if (ConstantInt *CB =
4840 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
4843 if (CB->getZExtValue() == false)
4844 std::swap(R1, R2); // R1 is the minimum root now.
4846 // We can only use this value if the chrec ends up with an exact zero
4847 // value at this index. When solving for "X*X != 5", for example, we
4848 // should not accept a root of 2.
4849 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4851 return R1; // We found a quadratic root!
4854 return getCouldNotCompute();
4857 // Otherwise we can only handle this if it is affine.
4858 if (!AddRec->isAffine())
4859 return getCouldNotCompute();
4861 // If this is an affine expression, the execution count of this branch is
4862 // the minimum unsigned root of the following equation:
4864 // Start + Step*N = 0 (mod 2^BW)
4868 // Step*N = -Start (mod 2^BW)
4870 // where BW is the common bit width of Start and Step.
4872 // Get the initial value for the loop.
4873 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
4874 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
4876 // If the AddRec is NUW, then (in an unsigned sense) it cannot be counting up
4877 // to wrap to 0, it must be counting down to equal 0. Also, while counting
4878 // down, it cannot "miss" 0 (which would cause it to wrap), regardless of what
4879 // the stride is. As such, NUW addrec's will always become zero in
4880 // "start / -stride" steps, and we know that the division is exact.
4881 if (AddRec->hasNoUnsignedWrap())
4882 // FIXME: We really want an "isexact" bit for udiv.
4883 return getUDivExpr(Start, getNegativeSCEV(Step));
4885 // For now we handle only constant steps.
4886 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
4888 return getCouldNotCompute();
4890 // First, handle unitary steps.
4891 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4892 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4894 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4895 return Start; // N = Start (as unsigned)
4897 // Then, try to solve the above equation provided that Start is constant.
4898 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4899 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4900 -StartC->getValue()->getValue(),
4902 return getCouldNotCompute();
4905 /// HowFarToNonZero - Return the number of times a backedge checking the
4906 /// specified value for nonzero will execute. If not computable, return
4908 ScalarEvolution::BackedgeTakenInfo
4909 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4910 // Loops that look like: while (X == 0) are very strange indeed. We don't
4911 // handle them yet except for the trivial case. This could be expanded in the
4912 // future as needed.
4914 // If the value is a constant, check to see if it is known to be non-zero
4915 // already. If so, the backedge will execute zero times.
4916 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4917 if (!C->getValue()->isNullValue())
4918 return getConstant(C->getType(), 0);
4919 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4922 // We could implement others, but I really doubt anyone writes loops like
4923 // this, and if they did, they would already be constant folded.
4924 return getCouldNotCompute();
4927 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4928 /// (which may not be an immediate predecessor) which has exactly one
4929 /// successor from which BB is reachable, or null if no such block is
4932 std::pair<BasicBlock *, BasicBlock *>
4933 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4934 // If the block has a unique predecessor, then there is no path from the
4935 // predecessor to the block that does not go through the direct edge
4936 // from the predecessor to the block.
4937 if (BasicBlock *Pred = BB->getSinglePredecessor())
4938 return std::make_pair(Pred, BB);
4940 // A loop's header is defined to be a block that dominates the loop.
4941 // If the header has a unique predecessor outside the loop, it must be
4942 // a block that has exactly one successor that can reach the loop.
4943 if (Loop *L = LI->getLoopFor(BB))
4944 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
4946 return std::pair<BasicBlock *, BasicBlock *>();
4949 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4950 /// testing whether two expressions are equal, however for the purposes of
4951 /// looking for a condition guarding a loop, it can be useful to be a little
4952 /// more general, since a front-end may have replicated the controlling
4955 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4956 // Quick check to see if they are the same SCEV.
4957 if (A == B) return true;
4959 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4960 // two different instructions with the same value. Check for this case.
4961 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4962 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4963 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4964 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4965 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4968 // Otherwise assume they may have a different value.
4972 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
4973 /// predicate Pred. Return true iff any changes were made.
4975 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
4976 const SCEV *&LHS, const SCEV *&RHS) {
4977 bool Changed = false;
4979 // Canonicalize a constant to the right side.
4980 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
4981 // Check for both operands constant.
4982 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
4983 if (ConstantExpr::getICmp(Pred,
4985 RHSC->getValue())->isNullValue())
4986 goto trivially_false;
4988 goto trivially_true;
4990 // Otherwise swap the operands to put the constant on the right.
4991 std::swap(LHS, RHS);
4992 Pred = ICmpInst::getSwappedPredicate(Pred);
4996 // If we're comparing an addrec with a value which is loop-invariant in the
4997 // addrec's loop, put the addrec on the left. Also make a dominance check,
4998 // as both operands could be addrecs loop-invariant in each other's loop.
4999 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
5000 const Loop *L = AR->getLoop();
5001 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
5002 std::swap(LHS, RHS);
5003 Pred = ICmpInst::getSwappedPredicate(Pred);
5008 // If there's a constant operand, canonicalize comparisons with boundary
5009 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
5010 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
5011 const APInt &RA = RC->getValue()->getValue();
5013 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5014 case ICmpInst::ICMP_EQ:
5015 case ICmpInst::ICMP_NE:
5017 case ICmpInst::ICMP_UGE:
5018 if ((RA - 1).isMinValue()) {
5019 Pred = ICmpInst::ICMP_NE;
5020 RHS = getConstant(RA - 1);
5024 if (RA.isMaxValue()) {
5025 Pred = ICmpInst::ICMP_EQ;
5029 if (RA.isMinValue()) goto trivially_true;
5031 Pred = ICmpInst::ICMP_UGT;
5032 RHS = getConstant(RA - 1);
5035 case ICmpInst::ICMP_ULE:
5036 if ((RA + 1).isMaxValue()) {
5037 Pred = ICmpInst::ICMP_NE;
5038 RHS = getConstant(RA + 1);
5042 if (RA.isMinValue()) {
5043 Pred = ICmpInst::ICMP_EQ;
5047 if (RA.isMaxValue()) goto trivially_true;
5049 Pred = ICmpInst::ICMP_ULT;
5050 RHS = getConstant(RA + 1);
5053 case ICmpInst::ICMP_SGE:
5054 if ((RA - 1).isMinSignedValue()) {
5055 Pred = ICmpInst::ICMP_NE;
5056 RHS = getConstant(RA - 1);
5060 if (RA.isMaxSignedValue()) {
5061 Pred = ICmpInst::ICMP_EQ;
5065 if (RA.isMinSignedValue()) goto trivially_true;
5067 Pred = ICmpInst::ICMP_SGT;
5068 RHS = getConstant(RA - 1);
5071 case ICmpInst::ICMP_SLE:
5072 if ((RA + 1).isMaxSignedValue()) {
5073 Pred = ICmpInst::ICMP_NE;
5074 RHS = getConstant(RA + 1);
5078 if (RA.isMinSignedValue()) {
5079 Pred = ICmpInst::ICMP_EQ;
5083 if (RA.isMaxSignedValue()) goto trivially_true;
5085 Pred = ICmpInst::ICMP_SLT;
5086 RHS = getConstant(RA + 1);
5089 case ICmpInst::ICMP_UGT:
5090 if (RA.isMinValue()) {
5091 Pred = ICmpInst::ICMP_NE;
5095 if ((RA + 1).isMaxValue()) {
5096 Pred = ICmpInst::ICMP_EQ;
5097 RHS = getConstant(RA + 1);
5101 if (RA.isMaxValue()) goto trivially_false;
5103 case ICmpInst::ICMP_ULT:
5104 if (RA.isMaxValue()) {
5105 Pred = ICmpInst::ICMP_NE;
5109 if ((RA - 1).isMinValue()) {
5110 Pred = ICmpInst::ICMP_EQ;
5111 RHS = getConstant(RA - 1);
5115 if (RA.isMinValue()) goto trivially_false;
5117 case ICmpInst::ICMP_SGT:
5118 if (RA.isMinSignedValue()) {
5119 Pred = ICmpInst::ICMP_NE;
5123 if ((RA + 1).isMaxSignedValue()) {
5124 Pred = ICmpInst::ICMP_EQ;
5125 RHS = getConstant(RA + 1);
5129 if (RA.isMaxSignedValue()) goto trivially_false;
5131 case ICmpInst::ICMP_SLT:
5132 if (RA.isMaxSignedValue()) {
5133 Pred = ICmpInst::ICMP_NE;
5137 if ((RA - 1).isMinSignedValue()) {
5138 Pred = ICmpInst::ICMP_EQ;
5139 RHS = getConstant(RA - 1);
5143 if (RA.isMinSignedValue()) goto trivially_false;
5148 // Check for obvious equality.
5149 if (HasSameValue(LHS, RHS)) {
5150 if (ICmpInst::isTrueWhenEqual(Pred))
5151 goto trivially_true;
5152 if (ICmpInst::isFalseWhenEqual(Pred))
5153 goto trivially_false;
5156 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5157 // adding or subtracting 1 from one of the operands.
5159 case ICmpInst::ICMP_SLE:
5160 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5161 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5162 /*HasNUW=*/false, /*HasNSW=*/true);
5163 Pred = ICmpInst::ICMP_SLT;
5165 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5166 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5167 /*HasNUW=*/false, /*HasNSW=*/true);
5168 Pred = ICmpInst::ICMP_SLT;
5172 case ICmpInst::ICMP_SGE:
5173 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5174 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5175 /*HasNUW=*/false, /*HasNSW=*/true);
5176 Pred = ICmpInst::ICMP_SGT;
5178 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5179 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5180 /*HasNUW=*/false, /*HasNSW=*/true);
5181 Pred = ICmpInst::ICMP_SGT;
5185 case ICmpInst::ICMP_ULE:
5186 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5187 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5188 /*HasNUW=*/true, /*HasNSW=*/false);
5189 Pred = ICmpInst::ICMP_ULT;
5191 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5192 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5193 /*HasNUW=*/true, /*HasNSW=*/false);
5194 Pred = ICmpInst::ICMP_ULT;
5198 case ICmpInst::ICMP_UGE:
5199 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5200 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5201 /*HasNUW=*/true, /*HasNSW=*/false);
5202 Pred = ICmpInst::ICMP_UGT;
5204 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5205 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5206 /*HasNUW=*/true, /*HasNSW=*/false);
5207 Pred = ICmpInst::ICMP_UGT;
5215 // TODO: More simplifications are possible here.
5221 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5222 Pred = ICmpInst::ICMP_EQ;
5227 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5228 Pred = ICmpInst::ICMP_NE;
5232 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5233 return getSignedRange(S).getSignedMax().isNegative();
5236 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5237 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5240 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5241 return !getSignedRange(S).getSignedMin().isNegative();
5244 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5245 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5248 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5249 return isKnownNegative(S) || isKnownPositive(S);
5252 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5253 const SCEV *LHS, const SCEV *RHS) {
5254 // Canonicalize the inputs first.
5255 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5257 // If LHS or RHS is an addrec, check to see if the condition is true in
5258 // every iteration of the loop.
5259 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5260 if (isLoopEntryGuardedByCond(
5261 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5262 isLoopBackedgeGuardedByCond(
5263 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5265 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5266 if (isLoopEntryGuardedByCond(
5267 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5268 isLoopBackedgeGuardedByCond(
5269 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5272 // Otherwise see what can be done with known constant ranges.
5273 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5277 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5278 const SCEV *LHS, const SCEV *RHS) {
5279 if (HasSameValue(LHS, RHS))
5280 return ICmpInst::isTrueWhenEqual(Pred);
5282 // This code is split out from isKnownPredicate because it is called from
5283 // within isLoopEntryGuardedByCond.
5286 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5288 case ICmpInst::ICMP_SGT:
5289 Pred = ICmpInst::ICMP_SLT;
5290 std::swap(LHS, RHS);
5291 case ICmpInst::ICMP_SLT: {
5292 ConstantRange LHSRange = getSignedRange(LHS);
5293 ConstantRange RHSRange = getSignedRange(RHS);
5294 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5296 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5300 case ICmpInst::ICMP_SGE:
5301 Pred = ICmpInst::ICMP_SLE;
5302 std::swap(LHS, RHS);
5303 case ICmpInst::ICMP_SLE: {
5304 ConstantRange LHSRange = getSignedRange(LHS);
5305 ConstantRange RHSRange = getSignedRange(RHS);
5306 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5308 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
5312 case ICmpInst::ICMP_UGT:
5313 Pred = ICmpInst::ICMP_ULT;
5314 std::swap(LHS, RHS);
5315 case ICmpInst::ICMP_ULT: {
5316 ConstantRange LHSRange = getUnsignedRange(LHS);
5317 ConstantRange RHSRange = getUnsignedRange(RHS);
5318 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
5320 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
5324 case ICmpInst::ICMP_UGE:
5325 Pred = ICmpInst::ICMP_ULE;
5326 std::swap(LHS, RHS);
5327 case ICmpInst::ICMP_ULE: {
5328 ConstantRange LHSRange = getUnsignedRange(LHS);
5329 ConstantRange RHSRange = getUnsignedRange(RHS);
5330 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
5332 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
5336 case ICmpInst::ICMP_NE: {
5337 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
5339 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
5342 const SCEV *Diff = getMinusSCEV(LHS, RHS);
5343 if (isKnownNonZero(Diff))
5347 case ICmpInst::ICMP_EQ:
5348 // The check at the top of the function catches the case where
5349 // the values are known to be equal.
5355 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
5356 /// protected by a conditional between LHS and RHS. This is used to
5357 /// to eliminate casts.
5359 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
5360 ICmpInst::Predicate Pred,
5361 const SCEV *LHS, const SCEV *RHS) {
5362 // Interpret a null as meaning no loop, where there is obviously no guard
5363 // (interprocedural conditions notwithstanding).
5364 if (!L) return true;
5366 BasicBlock *Latch = L->getLoopLatch();
5370 BranchInst *LoopContinuePredicate =
5371 dyn_cast<BranchInst>(Latch->getTerminator());
5372 if (!LoopContinuePredicate ||
5373 LoopContinuePredicate->isUnconditional())
5376 return isImpliedCond(Pred, LHS, RHS,
5377 LoopContinuePredicate->getCondition(),
5378 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
5381 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
5382 /// by a conditional between LHS and RHS. This is used to help avoid max
5383 /// expressions in loop trip counts, and to eliminate casts.
5385 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
5386 ICmpInst::Predicate Pred,
5387 const SCEV *LHS, const SCEV *RHS) {
5388 // Interpret a null as meaning no loop, where there is obviously no guard
5389 // (interprocedural conditions notwithstanding).
5390 if (!L) return false;
5392 // Starting at the loop predecessor, climb up the predecessor chain, as long
5393 // as there are predecessors that can be found that have unique successors
5394 // leading to the original header.
5395 for (std::pair<BasicBlock *, BasicBlock *>
5396 Pair(L->getLoopPredecessor(), L->getHeader());
5398 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
5400 BranchInst *LoopEntryPredicate =
5401 dyn_cast<BranchInst>(Pair.first->getTerminator());
5402 if (!LoopEntryPredicate ||
5403 LoopEntryPredicate->isUnconditional())
5406 if (isImpliedCond(Pred, LHS, RHS,
5407 LoopEntryPredicate->getCondition(),
5408 LoopEntryPredicate->getSuccessor(0) != Pair.second))
5415 /// isImpliedCond - Test whether the condition described by Pred, LHS,
5416 /// and RHS is true whenever the given Cond value evaluates to true.
5417 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
5418 const SCEV *LHS, const SCEV *RHS,
5419 Value *FoundCondValue,
5421 // Recursively handle And and Or conditions.
5422 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
5423 if (BO->getOpcode() == Instruction::And) {
5425 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5426 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5427 } else if (BO->getOpcode() == Instruction::Or) {
5429 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5430 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5434 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
5435 if (!ICI) return false;
5437 // Bail if the ICmp's operands' types are wider than the needed type
5438 // before attempting to call getSCEV on them. This avoids infinite
5439 // recursion, since the analysis of widening casts can require loop
5440 // exit condition information for overflow checking, which would
5442 if (getTypeSizeInBits(LHS->getType()) <
5443 getTypeSizeInBits(ICI->getOperand(0)->getType()))
5446 // Now that we found a conditional branch that dominates the loop, check to
5447 // see if it is the comparison we are looking for.
5448 ICmpInst::Predicate FoundPred;
5450 FoundPred = ICI->getInversePredicate();
5452 FoundPred = ICI->getPredicate();
5454 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
5455 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
5457 // Balance the types. The case where FoundLHS' type is wider than
5458 // LHS' type is checked for above.
5459 if (getTypeSizeInBits(LHS->getType()) >
5460 getTypeSizeInBits(FoundLHS->getType())) {
5461 if (CmpInst::isSigned(Pred)) {
5462 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
5463 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
5465 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
5466 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
5470 // Canonicalize the query to match the way instcombine will have
5471 // canonicalized the comparison.
5472 if (SimplifyICmpOperands(Pred, LHS, RHS))
5474 return CmpInst::isTrueWhenEqual(Pred);
5475 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
5476 if (FoundLHS == FoundRHS)
5477 return CmpInst::isFalseWhenEqual(Pred);
5479 // Check to see if we can make the LHS or RHS match.
5480 if (LHS == FoundRHS || RHS == FoundLHS) {
5481 if (isa<SCEVConstant>(RHS)) {
5482 std::swap(FoundLHS, FoundRHS);
5483 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
5485 std::swap(LHS, RHS);
5486 Pred = ICmpInst::getSwappedPredicate(Pred);
5490 // Check whether the found predicate is the same as the desired predicate.
5491 if (FoundPred == Pred)
5492 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
5494 // Check whether swapping the found predicate makes it the same as the
5495 // desired predicate.
5496 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
5497 if (isa<SCEVConstant>(RHS))
5498 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
5500 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
5501 RHS, LHS, FoundLHS, FoundRHS);
5504 // Check whether the actual condition is beyond sufficient.
5505 if (FoundPred == ICmpInst::ICMP_EQ)
5506 if (ICmpInst::isTrueWhenEqual(Pred))
5507 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
5509 if (Pred == ICmpInst::ICMP_NE)
5510 if (!ICmpInst::isTrueWhenEqual(FoundPred))
5511 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
5514 // Otherwise assume the worst.
5518 /// isImpliedCondOperands - Test whether the condition described by Pred,
5519 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
5520 /// and FoundRHS is true.
5521 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
5522 const SCEV *LHS, const SCEV *RHS,
5523 const SCEV *FoundLHS,
5524 const SCEV *FoundRHS) {
5525 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
5526 FoundLHS, FoundRHS) ||
5527 // ~x < ~y --> x > y
5528 isImpliedCondOperandsHelper(Pred, LHS, RHS,
5529 getNotSCEV(FoundRHS),
5530 getNotSCEV(FoundLHS));
5533 /// isImpliedCondOperandsHelper - Test whether the condition described by
5534 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
5535 /// FoundLHS, and FoundRHS is true.
5537 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
5538 const SCEV *LHS, const SCEV *RHS,
5539 const SCEV *FoundLHS,
5540 const SCEV *FoundRHS) {
5542 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5543 case ICmpInst::ICMP_EQ:
5544 case ICmpInst::ICMP_NE:
5545 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
5548 case ICmpInst::ICMP_SLT:
5549 case ICmpInst::ICMP_SLE:
5550 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
5551 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
5554 case ICmpInst::ICMP_SGT:
5555 case ICmpInst::ICMP_SGE:
5556 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
5557 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
5560 case ICmpInst::ICMP_ULT:
5561 case ICmpInst::ICMP_ULE:
5562 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
5563 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
5566 case ICmpInst::ICMP_UGT:
5567 case ICmpInst::ICMP_UGE:
5568 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
5569 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
5577 /// getBECount - Subtract the end and start values and divide by the step,
5578 /// rounding up, to get the number of times the backedge is executed. Return
5579 /// CouldNotCompute if an intermediate computation overflows.
5580 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
5584 assert(!isKnownNegative(Step) &&
5585 "This code doesn't handle negative strides yet!");
5587 const Type *Ty = Start->getType();
5588 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
5589 const SCEV *Diff = getMinusSCEV(End, Start);
5590 const SCEV *RoundUp = getAddExpr(Step, NegOne);
5592 // Add an adjustment to the difference between End and Start so that
5593 // the division will effectively round up.
5594 const SCEV *Add = getAddExpr(Diff, RoundUp);
5597 // Check Add for unsigned overflow.
5598 // TODO: More sophisticated things could be done here.
5599 const Type *WideTy = IntegerType::get(getContext(),
5600 getTypeSizeInBits(Ty) + 1);
5601 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
5602 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
5603 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
5604 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
5605 return getCouldNotCompute();
5608 return getUDivExpr(Add, Step);
5611 /// HowManyLessThans - Return the number of times a backedge containing the
5612 /// specified less-than comparison will execute. If not computable, return
5613 /// CouldNotCompute.
5614 ScalarEvolution::BackedgeTakenInfo
5615 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
5616 const Loop *L, bool isSigned) {
5617 // Only handle: "ADDREC < LoopInvariant".
5618 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute();
5620 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
5621 if (!AddRec || AddRec->getLoop() != L)
5622 return getCouldNotCompute();
5624 // Check to see if we have a flag which makes analysis easy.
5625 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
5626 AddRec->hasNoUnsignedWrap();
5628 if (AddRec->isAffine()) {
5629 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
5630 const SCEV *Step = AddRec->getStepRecurrence(*this);
5633 return getCouldNotCompute();
5634 if (Step->isOne()) {
5635 // With unit stride, the iteration never steps past the limit value.
5636 } else if (isKnownPositive(Step)) {
5637 // Test whether a positive iteration can step past the limit
5638 // value and past the maximum value for its type in a single step.
5639 // Note that it's not sufficient to check NoWrap here, because even
5640 // though the value after a wrap is undefined, it's not undefined
5641 // behavior, so if wrap does occur, the loop could either terminate or
5642 // loop infinitely, but in either case, the loop is guaranteed to
5643 // iterate at least until the iteration where the wrapping occurs.
5644 const SCEV *One = getConstant(Step->getType(), 1);
5646 APInt Max = APInt::getSignedMaxValue(BitWidth);
5647 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
5648 .slt(getSignedRange(RHS).getSignedMax()))
5649 return getCouldNotCompute();
5651 APInt Max = APInt::getMaxValue(BitWidth);
5652 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
5653 .ult(getUnsignedRange(RHS).getUnsignedMax()))
5654 return getCouldNotCompute();
5657 // TODO: Handle negative strides here and below.
5658 return getCouldNotCompute();
5660 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
5661 // m. So, we count the number of iterations in which {n,+,s} < m is true.
5662 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
5663 // treat m-n as signed nor unsigned due to overflow possibility.
5665 // First, we get the value of the LHS in the first iteration: n
5666 const SCEV *Start = AddRec->getOperand(0);
5668 // Determine the minimum constant start value.
5669 const SCEV *MinStart = getConstant(isSigned ?
5670 getSignedRange(Start).getSignedMin() :
5671 getUnsignedRange(Start).getUnsignedMin());
5673 // If we know that the condition is true in order to enter the loop,
5674 // then we know that it will run exactly (m-n)/s times. Otherwise, we
5675 // only know that it will execute (max(m,n)-n)/s times. In both cases,
5676 // the division must round up.
5677 const SCEV *End = RHS;
5678 if (!isLoopEntryGuardedByCond(L,
5679 isSigned ? ICmpInst::ICMP_SLT :
5681 getMinusSCEV(Start, Step), RHS))
5682 End = isSigned ? getSMaxExpr(RHS, Start)
5683 : getUMaxExpr(RHS, Start);
5685 // Determine the maximum constant end value.
5686 const SCEV *MaxEnd = getConstant(isSigned ?
5687 getSignedRange(End).getSignedMax() :
5688 getUnsignedRange(End).getUnsignedMax());
5690 // If MaxEnd is within a step of the maximum integer value in its type,
5691 // adjust it down to the minimum value which would produce the same effect.
5692 // This allows the subsequent ceiling division of (N+(step-1))/step to
5693 // compute the correct value.
5694 const SCEV *StepMinusOne = getMinusSCEV(Step,
5695 getConstant(Step->getType(), 1));
5698 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
5701 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
5704 // Finally, we subtract these two values and divide, rounding up, to get
5705 // the number of times the backedge is executed.
5706 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
5708 // The maximum backedge count is similar, except using the minimum start
5709 // value and the maximum end value.
5710 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
5712 return BackedgeTakenInfo(BECount, MaxBECount);
5715 return getCouldNotCompute();
5718 /// getNumIterationsInRange - Return the number of iterations of this loop that
5719 /// produce values in the specified constant range. Another way of looking at
5720 /// this is that it returns the first iteration number where the value is not in
5721 /// the condition, thus computing the exit count. If the iteration count can't
5722 /// be computed, an instance of SCEVCouldNotCompute is returned.
5723 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
5724 ScalarEvolution &SE) const {
5725 if (Range.isFullSet()) // Infinite loop.
5726 return SE.getCouldNotCompute();
5728 // If the start is a non-zero constant, shift the range to simplify things.
5729 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
5730 if (!SC->getValue()->isZero()) {
5731 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
5732 Operands[0] = SE.getConstant(SC->getType(), 0);
5733 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
5734 if (const SCEVAddRecExpr *ShiftedAddRec =
5735 dyn_cast<SCEVAddRecExpr>(Shifted))
5736 return ShiftedAddRec->getNumIterationsInRange(
5737 Range.subtract(SC->getValue()->getValue()), SE);
5738 // This is strange and shouldn't happen.
5739 return SE.getCouldNotCompute();
5742 // The only time we can solve this is when we have all constant indices.
5743 // Otherwise, we cannot determine the overflow conditions.
5744 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5745 if (!isa<SCEVConstant>(getOperand(i)))
5746 return SE.getCouldNotCompute();
5749 // Okay at this point we know that all elements of the chrec are constants and
5750 // that the start element is zero.
5752 // First check to see if the range contains zero. If not, the first
5754 unsigned BitWidth = SE.getTypeSizeInBits(getType());
5755 if (!Range.contains(APInt(BitWidth, 0)))
5756 return SE.getConstant(getType(), 0);
5759 // If this is an affine expression then we have this situation:
5760 // Solve {0,+,A} in Range === Ax in Range
5762 // We know that zero is in the range. If A is positive then we know that
5763 // the upper value of the range must be the first possible exit value.
5764 // If A is negative then the lower of the range is the last possible loop
5765 // value. Also note that we already checked for a full range.
5766 APInt One(BitWidth,1);
5767 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5768 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5770 // The exit value should be (End+A)/A.
5771 APInt ExitVal = (End + A).udiv(A);
5772 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5774 // Evaluate at the exit value. If we really did fall out of the valid
5775 // range, then we computed our trip count, otherwise wrap around or other
5776 // things must have happened.
5777 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5778 if (Range.contains(Val->getValue()))
5779 return SE.getCouldNotCompute(); // Something strange happened
5781 // Ensure that the previous value is in the range. This is a sanity check.
5782 assert(Range.contains(
5783 EvaluateConstantChrecAtConstant(this,
5784 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5785 "Linear scev computation is off in a bad way!");
5786 return SE.getConstant(ExitValue);
5787 } else if (isQuadratic()) {
5788 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5789 // quadratic equation to solve it. To do this, we must frame our problem in
5790 // terms of figuring out when zero is crossed, instead of when
5791 // Range.getUpper() is crossed.
5792 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5793 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5794 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5796 // Next, solve the constructed addrec
5797 std::pair<const SCEV *,const SCEV *> Roots =
5798 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5799 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5800 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5802 // Pick the smallest positive root value.
5803 if (ConstantInt *CB =
5804 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5805 R1->getValue(), R2->getValue()))) {
5806 if (CB->getZExtValue() == false)
5807 std::swap(R1, R2); // R1 is the minimum root now.
5809 // Make sure the root is not off by one. The returned iteration should
5810 // not be in the range, but the previous one should be. When solving
5811 // for "X*X < 5", for example, we should not return a root of 2.
5812 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5815 if (Range.contains(R1Val->getValue())) {
5816 // The next iteration must be out of the range...
5817 ConstantInt *NextVal =
5818 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5820 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5821 if (!Range.contains(R1Val->getValue()))
5822 return SE.getConstant(NextVal);
5823 return SE.getCouldNotCompute(); // Something strange happened
5826 // If R1 was not in the range, then it is a good return value. Make
5827 // sure that R1-1 WAS in the range though, just in case.
5828 ConstantInt *NextVal =
5829 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5830 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5831 if (Range.contains(R1Val->getValue()))
5833 return SE.getCouldNotCompute(); // Something strange happened
5838 return SE.getCouldNotCompute();
5843 //===----------------------------------------------------------------------===//
5844 // SCEVCallbackVH Class Implementation
5845 //===----------------------------------------------------------------------===//
5847 void ScalarEvolution::SCEVCallbackVH::deleted() {
5848 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5849 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5850 SE->ConstantEvolutionLoopExitValue.erase(PN);
5851 SE->ValueExprMap.erase(getValPtr());
5852 // this now dangles!
5855 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
5856 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5858 // Forget all the expressions associated with users of the old value,
5859 // so that future queries will recompute the expressions using the new
5861 Value *Old = getValPtr();
5862 SmallVector<User *, 16> Worklist;
5863 SmallPtrSet<User *, 8> Visited;
5864 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5866 Worklist.push_back(*UI);
5867 while (!Worklist.empty()) {
5868 User *U = Worklist.pop_back_val();
5869 // Deleting the Old value will cause this to dangle. Postpone
5870 // that until everything else is done.
5873 if (!Visited.insert(U))
5875 if (PHINode *PN = dyn_cast<PHINode>(U))
5876 SE->ConstantEvolutionLoopExitValue.erase(PN);
5877 SE->ValueExprMap.erase(U);
5878 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5880 Worklist.push_back(*UI);
5882 // Delete the Old value.
5883 if (PHINode *PN = dyn_cast<PHINode>(Old))
5884 SE->ConstantEvolutionLoopExitValue.erase(PN);
5885 SE->ValueExprMap.erase(Old);
5886 // this now dangles!
5889 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5890 : CallbackVH(V), SE(se) {}
5892 //===----------------------------------------------------------------------===//
5893 // ScalarEvolution Class Implementation
5894 //===----------------------------------------------------------------------===//
5896 ScalarEvolution::ScalarEvolution()
5897 : FunctionPass(ID), FirstUnknown(0) {
5898 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
5901 bool ScalarEvolution::runOnFunction(Function &F) {
5903 LI = &getAnalysis<LoopInfo>();
5904 TD = getAnalysisIfAvailable<TargetData>();
5905 DT = &getAnalysis<DominatorTree>();
5909 void ScalarEvolution::releaseMemory() {
5910 // Iterate through all the SCEVUnknown instances and call their
5911 // destructors, so that they release their references to their values.
5912 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
5916 ValueExprMap.clear();
5917 BackedgeTakenCounts.clear();
5918 ConstantEvolutionLoopExitValue.clear();
5919 ValuesAtScopes.clear();
5920 LoopDispositions.clear();
5921 BlockDispositions.clear();
5922 UnsignedRanges.clear();
5923 SignedRanges.clear();
5924 UniqueSCEVs.clear();
5925 SCEVAllocator.Reset();
5928 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5929 AU.setPreservesAll();
5930 AU.addRequiredTransitive<LoopInfo>();
5931 AU.addRequiredTransitive<DominatorTree>();
5934 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5935 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5938 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5940 // Print all inner loops first
5941 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5942 PrintLoopInfo(OS, SE, *I);
5945 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5948 SmallVector<BasicBlock *, 8> ExitBlocks;
5949 L->getExitBlocks(ExitBlocks);
5950 if (ExitBlocks.size() != 1)
5951 OS << "<multiple exits> ";
5953 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5954 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5956 OS << "Unpredictable backedge-taken count. ";
5961 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5964 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5965 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5967 OS << "Unpredictable max backedge-taken count. ";
5973 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
5974 // ScalarEvolution's implementation of the print method is to print
5975 // out SCEV values of all instructions that are interesting. Doing
5976 // this potentially causes it to create new SCEV objects though,
5977 // which technically conflicts with the const qualifier. This isn't
5978 // observable from outside the class though, so casting away the
5979 // const isn't dangerous.
5980 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
5982 OS << "Classifying expressions for: ";
5983 WriteAsOperand(OS, F, /*PrintType=*/false);
5985 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
5986 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
5989 const SCEV *SV = SE.getSCEV(&*I);
5992 const Loop *L = LI->getLoopFor((*I).getParent());
5994 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
6001 OS << "\t\t" "Exits: ";
6002 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
6003 if (!SE.isLoopInvariant(ExitValue, L)) {
6004 OS << "<<Unknown>>";
6013 OS << "Determining loop execution counts for: ";
6014 WriteAsOperand(OS, F, /*PrintType=*/false);
6016 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6017 PrintLoopInfo(OS, &SE, *I);
6020 ScalarEvolution::LoopDisposition
6021 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
6022 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S];
6023 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair =
6024 Values.insert(std::make_pair(L, LoopVariant));
6026 return Pair.first->second;
6028 LoopDisposition D = computeLoopDisposition(S, L);
6029 return LoopDispositions[S][L] = D;
6032 ScalarEvolution::LoopDisposition
6033 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
6034 switch (S->getSCEVType()) {
6036 return LoopInvariant;
6040 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
6041 case scAddRecExpr: {
6042 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6044 // If L is the addrec's loop, it's computable.
6045 if (AR->getLoop() == L)
6046 return LoopComputable;
6048 // Add recurrences are never invariant in the function-body (null loop).
6052 // This recurrence is variant w.r.t. L if L contains AR's loop.
6053 if (L->contains(AR->getLoop()))
6056 // This recurrence is invariant w.r.t. L if AR's loop contains L.
6057 if (AR->getLoop()->contains(L))
6058 return LoopInvariant;
6060 // This recurrence is variant w.r.t. L if any of its operands
6062 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
6064 if (!isLoopInvariant(*I, L))
6067 // Otherwise it's loop-invariant.
6068 return LoopInvariant;
6074 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6075 bool HasVarying = false;
6076 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6078 LoopDisposition D = getLoopDisposition(*I, L);
6079 if (D == LoopVariant)
6081 if (D == LoopComputable)
6084 return HasVarying ? LoopComputable : LoopInvariant;
6087 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6088 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
6089 if (LD == LoopVariant)
6091 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
6092 if (RD == LoopVariant)
6094 return (LD == LoopInvariant && RD == LoopInvariant) ?
6095 LoopInvariant : LoopComputable;
6098 // All non-instruction values are loop invariant. All instructions are loop
6099 // invariant if they are not contained in the specified loop.
6100 // Instructions are never considered invariant in the function body
6101 // (null loop) because they are defined within the "loop".
6102 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
6103 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
6104 return LoopInvariant;
6105 case scCouldNotCompute:
6106 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6110 llvm_unreachable("Unknown SCEV kind!");
6114 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
6115 return getLoopDisposition(S, L) == LoopInvariant;
6118 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
6119 return getLoopDisposition(S, L) == LoopComputable;
6122 ScalarEvolution::BlockDisposition
6123 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6124 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S];
6125 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool>
6126 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock));
6128 return Pair.first->second;
6130 BlockDisposition D = computeBlockDisposition(S, BB);
6131 return BlockDispositions[S][BB] = D;
6134 ScalarEvolution::BlockDisposition
6135 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6136 switch (S->getSCEVType()) {
6138 return ProperlyDominatesBlock;
6142 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
6143 case scAddRecExpr: {
6144 // This uses a "dominates" query instead of "properly dominates" query
6145 // to test for proper dominance too, because the instruction which
6146 // produces the addrec's value is a PHI, and a PHI effectively properly
6147 // dominates its entire containing block.
6148 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6149 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
6150 return DoesNotDominateBlock;
6152 // FALL THROUGH into SCEVNAryExpr handling.
6157 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6159 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6161 BlockDisposition D = getBlockDisposition(*I, BB);
6162 if (D == DoesNotDominateBlock)
6163 return DoesNotDominateBlock;
6164 if (D == DominatesBlock)
6167 return Proper ? ProperlyDominatesBlock : DominatesBlock;
6170 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6171 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6172 BlockDisposition LD = getBlockDisposition(LHS, BB);
6173 if (LD == DoesNotDominateBlock)
6174 return DoesNotDominateBlock;
6175 BlockDisposition RD = getBlockDisposition(RHS, BB);
6176 if (RD == DoesNotDominateBlock)
6177 return DoesNotDominateBlock;
6178 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
6179 ProperlyDominatesBlock : DominatesBlock;
6182 if (Instruction *I =
6183 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
6184 if (I->getParent() == BB)
6185 return DominatesBlock;
6186 if (DT->properlyDominates(I->getParent(), BB))
6187 return ProperlyDominatesBlock;
6188 return DoesNotDominateBlock;
6190 return ProperlyDominatesBlock;
6191 case scCouldNotCompute:
6192 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6193 return DoesNotDominateBlock;
6196 llvm_unreachable("Unknown SCEV kind!");
6197 return DoesNotDominateBlock;
6200 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
6201 return getBlockDisposition(S, BB) >= DominatesBlock;
6204 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
6205 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
6208 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
6209 switch (S->getSCEVType()) {
6214 case scSignExtend: {
6215 const SCEVCastExpr *Cast = cast<SCEVCastExpr>(S);
6216 const SCEV *CastOp = Cast->getOperand();
6217 return Op == CastOp || hasOperand(CastOp, Op);
6224 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6225 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6227 const SCEV *NAryOp = *I;
6228 if (NAryOp == Op || hasOperand(NAryOp, Op))
6234 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6235 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6236 return LHS == Op || hasOperand(LHS, Op) ||
6237 RHS == Op || hasOperand(RHS, Op);
6241 case scCouldNotCompute:
6242 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6246 llvm_unreachable("Unknown SCEV kind!");
6250 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
6251 ValuesAtScopes.erase(S);
6252 LoopDispositions.erase(S);
6253 BlockDispositions.erase(S);
6254 UnsignedRanges.erase(S);
6255 SignedRanges.erase(S);