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 // If the input value is a chrec scev, truncate the chrec's operands.
823 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
824 SmallVector<const SCEV *, 4> Operands;
825 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
826 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
827 return getAddRecExpr(Operands, AddRec->getLoop());
830 // As a special case, fold trunc(undef) to undef. We don't want to
831 // know too much about SCEVUnknowns, but this special case is handy
833 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
834 if (isa<UndefValue>(U->getValue()))
835 return getSCEV(UndefValue::get(Ty));
837 // The cast wasn't folded; create an explicit cast node. We can reuse
838 // the existing insert position since if we get here, we won't have
839 // made any changes which would invalidate it.
840 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
842 UniqueSCEVs.InsertNode(S, IP);
846 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
848 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
849 "This is not an extending conversion!");
850 assert(isSCEVable(Ty) &&
851 "This is not a conversion to a SCEVable type!");
852 Ty = getEffectiveSCEVType(Ty);
854 // Fold if the operand is constant.
855 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
857 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(),
858 getEffectiveSCEVType(Ty))));
860 // zext(zext(x)) --> zext(x)
861 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
862 return getZeroExtendExpr(SZ->getOperand(), Ty);
864 // Before doing any expensive analysis, check to see if we've already
865 // computed a SCEV for this Op and Ty.
867 ID.AddInteger(scZeroExtend);
871 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
873 // If the input value is a chrec scev, and we can prove that the value
874 // did not overflow the old, smaller, value, we can zero extend all of the
875 // operands (often constants). This allows analysis of something like
876 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
877 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
878 if (AR->isAffine()) {
879 const SCEV *Start = AR->getStart();
880 const SCEV *Step = AR->getStepRecurrence(*this);
881 unsigned BitWidth = getTypeSizeInBits(AR->getType());
882 const Loop *L = AR->getLoop();
884 // If we have special knowledge that this addrec won't overflow,
885 // we don't need to do any further analysis.
886 if (AR->hasNoUnsignedWrap())
887 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
888 getZeroExtendExpr(Step, Ty),
891 // Check whether the backedge-taken count is SCEVCouldNotCompute.
892 // Note that this serves two purposes: It filters out loops that are
893 // simply not analyzable, and it covers the case where this code is
894 // being called from within backedge-taken count analysis, such that
895 // attempting to ask for the backedge-taken count would likely result
896 // in infinite recursion. In the later case, the analysis code will
897 // cope with a conservative value, and it will take care to purge
898 // that value once it has finished.
899 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
900 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
901 // Manually compute the final value for AR, checking for
904 // Check whether the backedge-taken count can be losslessly casted to
905 // the addrec's type. The count is always unsigned.
906 const SCEV *CastedMaxBECount =
907 getTruncateOrZeroExtend(MaxBECount, Start->getType());
908 const SCEV *RecastedMaxBECount =
909 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
910 if (MaxBECount == RecastedMaxBECount) {
911 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
912 // Check whether Start+Step*MaxBECount has no unsigned overflow.
913 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
914 const SCEV *Add = getAddExpr(Start, ZMul);
915 const SCEV *OperandExtendedAdd =
916 getAddExpr(getZeroExtendExpr(Start, WideTy),
917 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
918 getZeroExtendExpr(Step, WideTy)));
919 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
920 // Return the expression with the addrec on the outside.
921 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
922 getZeroExtendExpr(Step, Ty),
925 // Similar to above, only this time treat the step value as signed.
926 // This covers loops that count down.
927 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
928 Add = getAddExpr(Start, SMul);
930 getAddExpr(getZeroExtendExpr(Start, WideTy),
931 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
932 getSignExtendExpr(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 getSignExtendExpr(Step, Ty),
940 // If the backedge is guarded by a comparison with the pre-inc value
941 // the addrec is safe. Also, if the entry is guarded by a comparison
942 // with the start value and the backedge is guarded by a comparison
943 // with the post-inc value, the addrec is safe.
944 if (isKnownPositive(Step)) {
945 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
946 getUnsignedRange(Step).getUnsignedMax());
947 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
948 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
949 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
950 AR->getPostIncExpr(*this), N)))
951 // Return the expression with the addrec on the outside.
952 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
953 getZeroExtendExpr(Step, Ty),
955 } else if (isKnownNegative(Step)) {
956 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
957 getSignedRange(Step).getSignedMin());
958 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
959 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
960 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
961 AR->getPostIncExpr(*this), N)))
962 // Return the expression with the addrec on the outside.
963 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
964 getSignExtendExpr(Step, Ty),
970 // The cast wasn't folded; create an explicit cast node.
971 // Recompute the insert position, as it may have been invalidated.
972 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
973 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
975 UniqueSCEVs.InsertNode(S, IP);
979 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
981 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
982 "This is not an extending conversion!");
983 assert(isSCEVable(Ty) &&
984 "This is not a conversion to a SCEVable type!");
985 Ty = getEffectiveSCEVType(Ty);
987 // Fold if the operand is constant.
988 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
990 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(),
991 getEffectiveSCEVType(Ty))));
993 // sext(sext(x)) --> sext(x)
994 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
995 return getSignExtendExpr(SS->getOperand(), Ty);
997 // sext(zext(x)) --> zext(x)
998 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
999 return getZeroExtendExpr(SZ->getOperand(), Ty);
1001 // Before doing any expensive analysis, check to see if we've already
1002 // computed a SCEV for this Op and Ty.
1003 FoldingSetNodeID ID;
1004 ID.AddInteger(scSignExtend);
1008 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1010 // If the input value is a chrec scev, and we can prove that the value
1011 // did not overflow the old, smaller, value, we can sign extend all of the
1012 // operands (often constants). This allows analysis of something like
1013 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1014 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1015 if (AR->isAffine()) {
1016 const SCEV *Start = AR->getStart();
1017 const SCEV *Step = AR->getStepRecurrence(*this);
1018 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1019 const Loop *L = AR->getLoop();
1021 // If we have special knowledge that this addrec won't overflow,
1022 // we don't need to do any further analysis.
1023 if (AR->hasNoSignedWrap())
1024 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1025 getSignExtendExpr(Step, Ty),
1028 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1029 // Note that this serves two purposes: It filters out loops that are
1030 // simply not analyzable, and it covers the case where this code is
1031 // being called from within backedge-taken count analysis, such that
1032 // attempting to ask for the backedge-taken count would likely result
1033 // in infinite recursion. In the later case, the analysis code will
1034 // cope with a conservative value, and it will take care to purge
1035 // that value once it has finished.
1036 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1037 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1038 // Manually compute the final value for AR, checking for
1041 // Check whether the backedge-taken count can be losslessly casted to
1042 // the addrec's type. The count is always unsigned.
1043 const SCEV *CastedMaxBECount =
1044 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1045 const SCEV *RecastedMaxBECount =
1046 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1047 if (MaxBECount == RecastedMaxBECount) {
1048 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1049 // Check whether Start+Step*MaxBECount has no signed overflow.
1050 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1051 const SCEV *Add = getAddExpr(Start, SMul);
1052 const SCEV *OperandExtendedAdd =
1053 getAddExpr(getSignExtendExpr(Start, WideTy),
1054 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1055 getSignExtendExpr(Step, WideTy)));
1056 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1057 // Return the expression with the addrec on the outside.
1058 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1059 getSignExtendExpr(Step, Ty),
1062 // Similar to above, only this time treat the step value as unsigned.
1063 // This covers loops that count up with an unsigned step.
1064 const SCEV *UMul = getMulExpr(CastedMaxBECount, Step);
1065 Add = getAddExpr(Start, UMul);
1066 OperandExtendedAdd =
1067 getAddExpr(getSignExtendExpr(Start, WideTy),
1068 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1069 getZeroExtendExpr(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 getZeroExtendExpr(Step, Ty),
1077 // If the backedge is guarded by a comparison with the pre-inc value
1078 // the addrec is safe. Also, if the entry is guarded by a comparison
1079 // with the start value and the backedge is guarded by a comparison
1080 // with the post-inc value, the addrec is safe.
1081 if (isKnownPositive(Step)) {
1082 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1083 getSignedRange(Step).getSignedMax());
1084 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1085 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1086 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1087 AR->getPostIncExpr(*this), N)))
1088 // Return the expression with the addrec on the outside.
1089 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1090 getSignExtendExpr(Step, Ty),
1092 } else if (isKnownNegative(Step)) {
1093 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1094 getSignedRange(Step).getSignedMin());
1095 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1096 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1097 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1098 AR->getPostIncExpr(*this), N)))
1099 // Return the expression with the addrec on the outside.
1100 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1101 getSignExtendExpr(Step, Ty),
1107 // The cast wasn't folded; create an explicit cast node.
1108 // Recompute the insert position, as it may have been invalidated.
1109 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1110 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1112 UniqueSCEVs.InsertNode(S, IP);
1116 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1117 /// unspecified bits out to the given type.
1119 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1121 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1122 "This is not an extending conversion!");
1123 assert(isSCEVable(Ty) &&
1124 "This is not a conversion to a SCEVable type!");
1125 Ty = getEffectiveSCEVType(Ty);
1127 // Sign-extend negative constants.
1128 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1129 if (SC->getValue()->getValue().isNegative())
1130 return getSignExtendExpr(Op, Ty);
1132 // Peel off a truncate cast.
1133 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1134 const SCEV *NewOp = T->getOperand();
1135 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1136 return getAnyExtendExpr(NewOp, Ty);
1137 return getTruncateOrNoop(NewOp, Ty);
1140 // Next try a zext cast. If the cast is folded, use it.
1141 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1142 if (!isa<SCEVZeroExtendExpr>(ZExt))
1145 // Next try a sext cast. If the cast is folded, use it.
1146 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1147 if (!isa<SCEVSignExtendExpr>(SExt))
1150 // Force the cast to be folded into the operands of an addrec.
1151 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1152 SmallVector<const SCEV *, 4> Ops;
1153 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1155 Ops.push_back(getAnyExtendExpr(*I, Ty));
1156 return getAddRecExpr(Ops, AR->getLoop());
1159 // As a special case, fold anyext(undef) to undef. We don't want to
1160 // know too much about SCEVUnknowns, but this special case is handy
1162 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
1163 if (isa<UndefValue>(U->getValue()))
1164 return getSCEV(UndefValue::get(Ty));
1166 // If the expression is obviously signed, use the sext cast value.
1167 if (isa<SCEVSMaxExpr>(Op))
1170 // Absent any other information, use the zext cast value.
1174 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1175 /// a list of operands to be added under the given scale, update the given
1176 /// map. This is a helper function for getAddRecExpr. As an example of
1177 /// what it does, given a sequence of operands that would form an add
1178 /// expression like this:
1180 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1182 /// where A and B are constants, update the map with these values:
1184 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1186 /// and add 13 + A*B*29 to AccumulatedConstant.
1187 /// This will allow getAddRecExpr to produce this:
1189 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1191 /// This form often exposes folding opportunities that are hidden in
1192 /// the original operand list.
1194 /// Return true iff it appears that any interesting folding opportunities
1195 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1196 /// the common case where no interesting opportunities are present, and
1197 /// is also used as a check to avoid infinite recursion.
1200 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1201 SmallVector<const SCEV *, 8> &NewOps,
1202 APInt &AccumulatedConstant,
1203 const SCEV *const *Ops, size_t NumOperands,
1205 ScalarEvolution &SE) {
1206 bool Interesting = false;
1208 // Iterate over the add operands. They are sorted, with constants first.
1210 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1212 // Pull a buried constant out to the outside.
1213 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1215 AccumulatedConstant += Scale * C->getValue()->getValue();
1218 // Next comes everything else. We're especially interested in multiplies
1219 // here, but they're in the middle, so just visit the rest with one loop.
1220 for (; i != NumOperands; ++i) {
1221 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1222 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1224 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1225 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1226 // A multiplication of a constant with another add; recurse.
1227 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1229 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1230 Add->op_begin(), Add->getNumOperands(),
1233 // A multiplication of a constant with some other value. Update
1235 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1236 const SCEV *Key = SE.getMulExpr(MulOps);
1237 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1238 M.insert(std::make_pair(Key, NewScale));
1240 NewOps.push_back(Pair.first->first);
1242 Pair.first->second += NewScale;
1243 // The map already had an entry for this value, which may indicate
1244 // a folding opportunity.
1249 // An ordinary operand. Update the map.
1250 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1251 M.insert(std::make_pair(Ops[i], Scale));
1253 NewOps.push_back(Pair.first->first);
1255 Pair.first->second += Scale;
1256 // The map already had an entry for this value, which may indicate
1257 // a folding opportunity.
1267 struct APIntCompare {
1268 bool operator()(const APInt &LHS, const APInt &RHS) const {
1269 return LHS.ult(RHS);
1274 /// getAddExpr - Get a canonical add expression, or something simpler if
1276 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1277 bool HasNUW, bool HasNSW) {
1278 assert(!Ops.empty() && "Cannot get empty add!");
1279 if (Ops.size() == 1) return Ops[0];
1281 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1282 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1283 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1284 "SCEVAddExpr operand types don't match!");
1287 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1288 if (!HasNUW && HasNSW) {
1290 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1291 E = Ops.end(); I != E; ++I)
1292 if (!isKnownNonNegative(*I)) {
1296 if (All) HasNUW = true;
1299 // Sort by complexity, this groups all similar expression types together.
1300 GroupByComplexity(Ops, LI);
1302 // If there are any constants, fold them together.
1304 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1306 assert(Idx < Ops.size());
1307 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1308 // We found two constants, fold them together!
1309 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1310 RHSC->getValue()->getValue());
1311 if (Ops.size() == 2) return Ops[0];
1312 Ops.erase(Ops.begin()+1); // Erase the folded element
1313 LHSC = cast<SCEVConstant>(Ops[0]);
1316 // If we are left with a constant zero being added, strip it off.
1317 if (LHSC->getValue()->isZero()) {
1318 Ops.erase(Ops.begin());
1322 if (Ops.size() == 1) return Ops[0];
1325 // Okay, check to see if the same value occurs in the operand list more than
1326 // once. If so, merge them together into an multiply expression. Since we
1327 // sorted the list, these values are required to be adjacent.
1328 const Type *Ty = Ops[0]->getType();
1329 bool FoundMatch = false;
1330 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1331 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1332 // Scan ahead to count how many equal operands there are.
1334 while (i+Count != e && Ops[i+Count] == Ops[i])
1336 // Merge the values into a multiply.
1337 const SCEV *Scale = getConstant(Ty, Count);
1338 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1339 if (Ops.size() == Count)
1342 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1343 --i; e -= Count - 1;
1347 return getAddExpr(Ops, HasNUW, HasNSW);
1349 // Check for truncates. If all the operands are truncated from the same
1350 // type, see if factoring out the truncate would permit the result to be
1351 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1352 // if the contents of the resulting outer trunc fold to something simple.
1353 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1354 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1355 const Type *DstType = Trunc->getType();
1356 const Type *SrcType = Trunc->getOperand()->getType();
1357 SmallVector<const SCEV *, 8> LargeOps;
1359 // Check all the operands to see if they can be represented in the
1360 // source type of the truncate.
1361 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1362 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1363 if (T->getOperand()->getType() != SrcType) {
1367 LargeOps.push_back(T->getOperand());
1368 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1369 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1370 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1371 SmallVector<const SCEV *, 8> LargeMulOps;
1372 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1373 if (const SCEVTruncateExpr *T =
1374 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1375 if (T->getOperand()->getType() != SrcType) {
1379 LargeMulOps.push_back(T->getOperand());
1380 } else if (const SCEVConstant *C =
1381 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1382 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1389 LargeOps.push_back(getMulExpr(LargeMulOps));
1396 // Evaluate the expression in the larger type.
1397 const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
1398 // If it folds to something simple, use it. Otherwise, don't.
1399 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1400 return getTruncateExpr(Fold, DstType);
1404 // Skip past any other cast SCEVs.
1405 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1408 // If there are add operands they would be next.
1409 if (Idx < Ops.size()) {
1410 bool DeletedAdd = false;
1411 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1412 // If we have an add, expand the add operands onto the end of the operands
1414 Ops.erase(Ops.begin()+Idx);
1415 Ops.append(Add->op_begin(), Add->op_end());
1419 // If we deleted at least one add, we added operands to the end of the list,
1420 // and they are not necessarily sorted. Recurse to resort and resimplify
1421 // any operands we just acquired.
1423 return getAddExpr(Ops);
1426 // Skip over the add expression until we get to a multiply.
1427 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1430 // Check to see if there are any folding opportunities present with
1431 // operands multiplied by constant values.
1432 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1433 uint64_t BitWidth = getTypeSizeInBits(Ty);
1434 DenseMap<const SCEV *, APInt> M;
1435 SmallVector<const SCEV *, 8> NewOps;
1436 APInt AccumulatedConstant(BitWidth, 0);
1437 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1438 Ops.data(), Ops.size(),
1439 APInt(BitWidth, 1), *this)) {
1440 // Some interesting folding opportunity is present, so its worthwhile to
1441 // re-generate the operands list. Group the operands by constant scale,
1442 // to avoid multiplying by the same constant scale multiple times.
1443 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1444 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1445 E = NewOps.end(); I != E; ++I)
1446 MulOpLists[M.find(*I)->second].push_back(*I);
1447 // Re-generate the operands list.
1449 if (AccumulatedConstant != 0)
1450 Ops.push_back(getConstant(AccumulatedConstant));
1451 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1452 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1454 Ops.push_back(getMulExpr(getConstant(I->first),
1455 getAddExpr(I->second)));
1457 return getConstant(Ty, 0);
1458 if (Ops.size() == 1)
1460 return getAddExpr(Ops);
1464 // If we are adding something to a multiply expression, make sure the
1465 // something is not already an operand of the multiply. If so, merge it into
1467 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1468 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1469 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1470 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1471 if (isa<SCEVConstant>(MulOpSCEV))
1473 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1474 if (MulOpSCEV == Ops[AddOp]) {
1475 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1476 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1477 if (Mul->getNumOperands() != 2) {
1478 // If the multiply has more than two operands, we must get the
1480 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1481 Mul->op_begin()+MulOp);
1482 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1483 InnerMul = getMulExpr(MulOps);
1485 const SCEV *One = getConstant(Ty, 1);
1486 const SCEV *AddOne = getAddExpr(One, InnerMul);
1487 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1488 if (Ops.size() == 2) return OuterMul;
1490 Ops.erase(Ops.begin()+AddOp);
1491 Ops.erase(Ops.begin()+Idx-1);
1493 Ops.erase(Ops.begin()+Idx);
1494 Ops.erase(Ops.begin()+AddOp-1);
1496 Ops.push_back(OuterMul);
1497 return getAddExpr(Ops);
1500 // Check this multiply against other multiplies being added together.
1501 for (unsigned OtherMulIdx = Idx+1;
1502 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1504 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1505 // If MulOp occurs in OtherMul, we can fold the two multiplies
1507 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1508 OMulOp != e; ++OMulOp)
1509 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1510 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1511 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1512 if (Mul->getNumOperands() != 2) {
1513 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1514 Mul->op_begin()+MulOp);
1515 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1516 InnerMul1 = getMulExpr(MulOps);
1518 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1519 if (OtherMul->getNumOperands() != 2) {
1520 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1521 OtherMul->op_begin()+OMulOp);
1522 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1523 InnerMul2 = getMulExpr(MulOps);
1525 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1526 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1527 if (Ops.size() == 2) return OuterMul;
1528 Ops.erase(Ops.begin()+Idx);
1529 Ops.erase(Ops.begin()+OtherMulIdx-1);
1530 Ops.push_back(OuterMul);
1531 return getAddExpr(Ops);
1537 // If there are any add recurrences in the operands list, see if any other
1538 // added values are loop invariant. If so, we can fold them into the
1540 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1543 // Scan over all recurrences, trying to fold loop invariants into them.
1544 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1545 // Scan all of the other operands to this add and add them to the vector if
1546 // they are loop invariant w.r.t. the recurrence.
1547 SmallVector<const SCEV *, 8> LIOps;
1548 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1549 const Loop *AddRecLoop = AddRec->getLoop();
1550 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1551 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1552 LIOps.push_back(Ops[i]);
1553 Ops.erase(Ops.begin()+i);
1557 // If we found some loop invariants, fold them into the recurrence.
1558 if (!LIOps.empty()) {
1559 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1560 LIOps.push_back(AddRec->getStart());
1562 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1564 AddRecOps[0] = getAddExpr(LIOps);
1566 // Build the new addrec. Propagate the NUW and NSW flags if both the
1567 // outer add and the inner addrec are guaranteed to have no overflow.
1568 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop,
1569 HasNUW && AddRec->hasNoUnsignedWrap(),
1570 HasNSW && AddRec->hasNoSignedWrap());
1572 // If all of the other operands were loop invariant, we are done.
1573 if (Ops.size() == 1) return NewRec;
1575 // Otherwise, add the folded AddRec by the non-liv parts.
1576 for (unsigned i = 0;; ++i)
1577 if (Ops[i] == AddRec) {
1581 return getAddExpr(Ops);
1584 // Okay, if there weren't any loop invariants to be folded, check to see if
1585 // there are multiple AddRec's with the same loop induction variable being
1586 // added together. If so, we can fold them.
1587 for (unsigned OtherIdx = Idx+1;
1588 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1590 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1591 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1592 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1594 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1596 if (const SCEVAddRecExpr *OtherAddRec =
1597 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1598 if (OtherAddRec->getLoop() == AddRecLoop) {
1599 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1601 if (i >= AddRecOps.size()) {
1602 AddRecOps.append(OtherAddRec->op_begin()+i,
1603 OtherAddRec->op_end());
1606 AddRecOps[i] = getAddExpr(AddRecOps[i],
1607 OtherAddRec->getOperand(i));
1609 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1611 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop);
1612 return getAddExpr(Ops);
1615 // Otherwise couldn't fold anything into this recurrence. Move onto the
1619 // Okay, it looks like we really DO need an add expr. Check to see if we
1620 // already have one, otherwise create a new one.
1621 FoldingSetNodeID ID;
1622 ID.AddInteger(scAddExpr);
1623 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1624 ID.AddPointer(Ops[i]);
1627 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1629 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1630 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1631 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1633 UniqueSCEVs.InsertNode(S, IP);
1635 if (HasNUW) S->setHasNoUnsignedWrap(true);
1636 if (HasNSW) S->setHasNoSignedWrap(true);
1640 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1642 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1643 bool HasNUW, bool HasNSW) {
1644 assert(!Ops.empty() && "Cannot get empty mul!");
1645 if (Ops.size() == 1) return Ops[0];
1647 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1648 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1649 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1650 "SCEVMulExpr operand types don't match!");
1653 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1654 if (!HasNUW && HasNSW) {
1656 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1657 E = Ops.end(); I != E; ++I)
1658 if (!isKnownNonNegative(*I)) {
1662 if (All) HasNUW = true;
1665 // Sort by complexity, this groups all similar expression types together.
1666 GroupByComplexity(Ops, LI);
1668 // If there are any constants, fold them together.
1670 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1672 // C1*(C2+V) -> C1*C2 + C1*V
1673 if (Ops.size() == 2)
1674 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1675 if (Add->getNumOperands() == 2 &&
1676 isa<SCEVConstant>(Add->getOperand(0)))
1677 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1678 getMulExpr(LHSC, Add->getOperand(1)));
1681 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1682 // We found two constants, fold them together!
1683 ConstantInt *Fold = ConstantInt::get(getContext(),
1684 LHSC->getValue()->getValue() *
1685 RHSC->getValue()->getValue());
1686 Ops[0] = getConstant(Fold);
1687 Ops.erase(Ops.begin()+1); // Erase the folded element
1688 if (Ops.size() == 1) return Ops[0];
1689 LHSC = cast<SCEVConstant>(Ops[0]);
1692 // If we are left with a constant one being multiplied, strip it off.
1693 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1694 Ops.erase(Ops.begin());
1696 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1697 // If we have a multiply of zero, it will always be zero.
1699 } else if (Ops[0]->isAllOnesValue()) {
1700 // If we have a mul by -1 of an add, try distributing the -1 among the
1702 if (Ops.size() == 2)
1703 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1704 SmallVector<const SCEV *, 4> NewOps;
1705 bool AnyFolded = false;
1706 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1708 const SCEV *Mul = getMulExpr(Ops[0], *I);
1709 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1710 NewOps.push_back(Mul);
1713 return getAddExpr(NewOps);
1717 if (Ops.size() == 1)
1721 // Skip over the add expression until we get to a multiply.
1722 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1725 // If there are mul operands inline them all into this expression.
1726 if (Idx < Ops.size()) {
1727 bool DeletedMul = false;
1728 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1729 // If we have an mul, expand the mul operands onto the end of the operands
1731 Ops.erase(Ops.begin()+Idx);
1732 Ops.append(Mul->op_begin(), Mul->op_end());
1736 // If we deleted at least one mul, we added operands to the end of the list,
1737 // and they are not necessarily sorted. Recurse to resort and resimplify
1738 // any operands we just acquired.
1740 return getMulExpr(Ops);
1743 // If there are any add recurrences in the operands list, see if any other
1744 // added values are loop invariant. If so, we can fold them into the
1746 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1749 // Scan over all recurrences, trying to fold loop invariants into them.
1750 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1751 // Scan all of the other operands to this mul and add them to the vector if
1752 // they are loop invariant w.r.t. the recurrence.
1753 SmallVector<const SCEV *, 8> LIOps;
1754 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1755 const Loop *AddRecLoop = AddRec->getLoop();
1756 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1757 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1758 LIOps.push_back(Ops[i]);
1759 Ops.erase(Ops.begin()+i);
1763 // If we found some loop invariants, fold them into the recurrence.
1764 if (!LIOps.empty()) {
1765 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1766 SmallVector<const SCEV *, 4> NewOps;
1767 NewOps.reserve(AddRec->getNumOperands());
1768 const SCEV *Scale = getMulExpr(LIOps);
1769 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1770 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1772 // Build the new addrec. Propagate the NUW and NSW flags if both the
1773 // outer mul and the inner addrec are guaranteed to have no overflow.
1774 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop,
1775 HasNUW && AddRec->hasNoUnsignedWrap(),
1776 HasNSW && AddRec->hasNoSignedWrap());
1778 // If all of the other operands were loop invariant, we are done.
1779 if (Ops.size() == 1) return NewRec;
1781 // Otherwise, multiply the folded AddRec by the non-liv parts.
1782 for (unsigned i = 0;; ++i)
1783 if (Ops[i] == AddRec) {
1787 return getMulExpr(Ops);
1790 // Okay, if there weren't any loop invariants to be folded, check to see if
1791 // there are multiple AddRec's with the same loop induction variable being
1792 // multiplied together. If so, we can fold them.
1793 for (unsigned OtherIdx = Idx+1;
1794 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1796 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1797 // F * G, where F = {A,+,B}<L> and G = {C,+,D}<L> -->
1798 // {A*C,+,F*D + G*B + B*D}<L>
1799 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1801 if (const SCEVAddRecExpr *OtherAddRec =
1802 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1803 if (OtherAddRec->getLoop() == AddRecLoop) {
1804 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1805 const SCEV *NewStart = getMulExpr(F->getStart(), G->getStart());
1806 const SCEV *B = F->getStepRecurrence(*this);
1807 const SCEV *D = G->getStepRecurrence(*this);
1808 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1811 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1813 if (Ops.size() == 2) return NewAddRec;
1814 Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
1815 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1817 return getMulExpr(Ops);
1820 // Otherwise couldn't fold anything into this recurrence. Move onto the
1824 // Okay, it looks like we really DO need an mul expr. Check to see if we
1825 // already have one, otherwise create a new one.
1826 FoldingSetNodeID ID;
1827 ID.AddInteger(scMulExpr);
1828 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1829 ID.AddPointer(Ops[i]);
1832 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1834 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1835 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1836 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
1838 UniqueSCEVs.InsertNode(S, IP);
1840 if (HasNUW) S->setHasNoUnsignedWrap(true);
1841 if (HasNSW) S->setHasNoSignedWrap(true);
1845 /// getUDivExpr - Get a canonical unsigned division expression, or something
1846 /// simpler if possible.
1847 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1849 assert(getEffectiveSCEVType(LHS->getType()) ==
1850 getEffectiveSCEVType(RHS->getType()) &&
1851 "SCEVUDivExpr operand types don't match!");
1853 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1854 if (RHSC->getValue()->equalsInt(1))
1855 return LHS; // X udiv 1 --> x
1856 // If the denominator is zero, the result of the udiv is undefined. Don't
1857 // try to analyze it, because the resolution chosen here may differ from
1858 // the resolution chosen in other parts of the compiler.
1859 if (!RHSC->getValue()->isZero()) {
1860 // Determine if the division can be folded into the operands of
1862 // TODO: Generalize this to non-constants by using known-bits information.
1863 const Type *Ty = LHS->getType();
1864 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1865 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
1866 // For non-power-of-two values, effectively round the value up to the
1867 // nearest power of two.
1868 if (!RHSC->getValue()->getValue().isPowerOf2())
1870 const IntegerType *ExtTy =
1871 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1872 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1873 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1874 if (const SCEVConstant *Step =
1875 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1876 if (!Step->getValue()->getValue()
1877 .urem(RHSC->getValue()->getValue()) &&
1878 getZeroExtendExpr(AR, ExtTy) ==
1879 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1880 getZeroExtendExpr(Step, ExtTy),
1882 SmallVector<const SCEV *, 4> Operands;
1883 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1884 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1885 return getAddRecExpr(Operands, AR->getLoop());
1887 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1888 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1889 SmallVector<const SCEV *, 4> Operands;
1890 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1891 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1892 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1893 // Find an operand that's safely divisible.
1894 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1895 const SCEV *Op = M->getOperand(i);
1896 const SCEV *Div = getUDivExpr(Op, RHSC);
1897 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1898 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
1901 return getMulExpr(Operands);
1905 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1906 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1907 SmallVector<const SCEV *, 4> Operands;
1908 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1909 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1910 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1912 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1913 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
1914 if (isa<SCEVUDivExpr>(Op) ||
1915 getMulExpr(Op, RHS) != A->getOperand(i))
1917 Operands.push_back(Op);
1919 if (Operands.size() == A->getNumOperands())
1920 return getAddExpr(Operands);
1924 // Fold if both operands are constant.
1925 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1926 Constant *LHSCV = LHSC->getValue();
1927 Constant *RHSCV = RHSC->getValue();
1928 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
1934 FoldingSetNodeID ID;
1935 ID.AddInteger(scUDivExpr);
1939 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1940 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
1942 UniqueSCEVs.InsertNode(S, IP);
1947 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1948 /// Simplify the expression as much as possible.
1949 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
1950 const SCEV *Step, const Loop *L,
1951 bool HasNUW, bool HasNSW) {
1952 SmallVector<const SCEV *, 4> Operands;
1953 Operands.push_back(Start);
1954 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1955 if (StepChrec->getLoop() == L) {
1956 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
1957 return getAddRecExpr(Operands, L);
1960 Operands.push_back(Step);
1961 return getAddRecExpr(Operands, L, HasNUW, HasNSW);
1964 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1965 /// Simplify the expression as much as possible.
1967 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
1969 bool HasNUW, bool HasNSW) {
1970 if (Operands.size() == 1) return Operands[0];
1972 const Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
1973 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
1974 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
1975 "SCEVAddRecExpr operand types don't match!");
1976 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1977 assert(isLoopInvariant(Operands[i], L) &&
1978 "SCEVAddRecExpr operand is not loop-invariant!");
1981 if (Operands.back()->isZero()) {
1982 Operands.pop_back();
1983 return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
1986 // It's tempting to want to call getMaxBackedgeTakenCount count here and
1987 // use that information to infer NUW and NSW flags. However, computing a
1988 // BE count requires calling getAddRecExpr, so we may not yet have a
1989 // meaningful BE count at this point (and if we don't, we'd be stuck
1990 // with a SCEVCouldNotCompute as the cached BE count).
1992 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1993 if (!HasNUW && HasNSW) {
1995 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
1996 E = Operands.end(); I != E; ++I)
1997 if (!isKnownNonNegative(*I)) {
2001 if (All) HasNUW = true;
2004 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2005 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2006 const Loop *NestedLoop = NestedAR->getLoop();
2007 if (L->contains(NestedLoop) ?
2008 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2009 (!NestedLoop->contains(L) &&
2010 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2011 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2012 NestedAR->op_end());
2013 Operands[0] = NestedAR->getStart();
2014 // AddRecs require their operands be loop-invariant with respect to their
2015 // loops. Don't perform this transformation if it would break this
2017 bool AllInvariant = true;
2018 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2019 if (!isLoopInvariant(Operands[i], L)) {
2020 AllInvariant = false;
2024 NestedOperands[0] = getAddRecExpr(Operands, L);
2025 AllInvariant = true;
2026 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2027 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2028 AllInvariant = false;
2032 // Ok, both add recurrences are valid after the transformation.
2033 return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
2035 // Reset Operands to its original state.
2036 Operands[0] = NestedAR;
2040 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2041 // already have one, otherwise create a new one.
2042 FoldingSetNodeID ID;
2043 ID.AddInteger(scAddRecExpr);
2044 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2045 ID.AddPointer(Operands[i]);
2049 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2051 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2052 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2053 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2054 O, Operands.size(), L);
2055 UniqueSCEVs.InsertNode(S, IP);
2057 if (HasNUW) S->setHasNoUnsignedWrap(true);
2058 if (HasNSW) S->setHasNoSignedWrap(true);
2062 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2064 SmallVector<const SCEV *, 2> Ops;
2067 return getSMaxExpr(Ops);
2071 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2072 assert(!Ops.empty() && "Cannot get empty smax!");
2073 if (Ops.size() == 1) return Ops[0];
2075 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2076 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2077 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2078 "SCEVSMaxExpr operand types don't match!");
2081 // Sort by complexity, this groups all similar expression types together.
2082 GroupByComplexity(Ops, LI);
2084 // If there are any constants, fold them together.
2086 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2088 assert(Idx < Ops.size());
2089 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2090 // We found two constants, fold them together!
2091 ConstantInt *Fold = ConstantInt::get(getContext(),
2092 APIntOps::smax(LHSC->getValue()->getValue(),
2093 RHSC->getValue()->getValue()));
2094 Ops[0] = getConstant(Fold);
2095 Ops.erase(Ops.begin()+1); // Erase the folded element
2096 if (Ops.size() == 1) return Ops[0];
2097 LHSC = cast<SCEVConstant>(Ops[0]);
2100 // If we are left with a constant minimum-int, strip it off.
2101 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2102 Ops.erase(Ops.begin());
2104 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2105 // If we have an smax with a constant maximum-int, it will always be
2110 if (Ops.size() == 1) return Ops[0];
2113 // Find the first SMax
2114 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2117 // Check to see if one of the operands is an SMax. If so, expand its operands
2118 // onto our operand list, and recurse to simplify.
2119 if (Idx < Ops.size()) {
2120 bool DeletedSMax = false;
2121 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2122 Ops.erase(Ops.begin()+Idx);
2123 Ops.append(SMax->op_begin(), SMax->op_end());
2128 return getSMaxExpr(Ops);
2131 // Okay, check to see if the same value occurs in the operand list twice. If
2132 // so, delete one. Since we sorted the list, these values are required to
2134 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2135 // X smax Y smax Y --> X smax Y
2136 // X smax Y --> X, if X is always greater than Y
2137 if (Ops[i] == Ops[i+1] ||
2138 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2139 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2141 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2142 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2146 if (Ops.size() == 1) return Ops[0];
2148 assert(!Ops.empty() && "Reduced smax down to nothing!");
2150 // Okay, it looks like we really DO need an smax expr. Check to see if we
2151 // already have one, otherwise create a new one.
2152 FoldingSetNodeID ID;
2153 ID.AddInteger(scSMaxExpr);
2154 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2155 ID.AddPointer(Ops[i]);
2157 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2158 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2159 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2160 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2162 UniqueSCEVs.InsertNode(S, IP);
2166 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2168 SmallVector<const SCEV *, 2> Ops;
2171 return getUMaxExpr(Ops);
2175 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2176 assert(!Ops.empty() && "Cannot get empty umax!");
2177 if (Ops.size() == 1) return Ops[0];
2179 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2180 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2181 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2182 "SCEVUMaxExpr operand types don't match!");
2185 // Sort by complexity, this groups all similar expression types together.
2186 GroupByComplexity(Ops, LI);
2188 // If there are any constants, fold them together.
2190 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2192 assert(Idx < Ops.size());
2193 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2194 // We found two constants, fold them together!
2195 ConstantInt *Fold = ConstantInt::get(getContext(),
2196 APIntOps::umax(LHSC->getValue()->getValue(),
2197 RHSC->getValue()->getValue()));
2198 Ops[0] = getConstant(Fold);
2199 Ops.erase(Ops.begin()+1); // Erase the folded element
2200 if (Ops.size() == 1) return Ops[0];
2201 LHSC = cast<SCEVConstant>(Ops[0]);
2204 // If we are left with a constant minimum-int, strip it off.
2205 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2206 Ops.erase(Ops.begin());
2208 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2209 // If we have an umax with a constant maximum-int, it will always be
2214 if (Ops.size() == 1) return Ops[0];
2217 // Find the first UMax
2218 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2221 // Check to see if one of the operands is a UMax. If so, expand its operands
2222 // onto our operand list, and recurse to simplify.
2223 if (Idx < Ops.size()) {
2224 bool DeletedUMax = false;
2225 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2226 Ops.erase(Ops.begin()+Idx);
2227 Ops.append(UMax->op_begin(), UMax->op_end());
2232 return getUMaxExpr(Ops);
2235 // Okay, check to see if the same value occurs in the operand list twice. If
2236 // so, delete one. Since we sorted the list, these values are required to
2238 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2239 // X umax Y umax Y --> X umax Y
2240 // X umax Y --> X, if X is always greater than Y
2241 if (Ops[i] == Ops[i+1] ||
2242 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2243 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2245 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2246 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2250 if (Ops.size() == 1) return Ops[0];
2252 assert(!Ops.empty() && "Reduced umax down to nothing!");
2254 // Okay, it looks like we really DO need a umax expr. Check to see if we
2255 // already have one, otherwise create a new one.
2256 FoldingSetNodeID ID;
2257 ID.AddInteger(scUMaxExpr);
2258 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2259 ID.AddPointer(Ops[i]);
2261 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2262 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2263 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2264 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2266 UniqueSCEVs.InsertNode(S, IP);
2270 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2272 // ~smax(~x, ~y) == smin(x, y).
2273 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2276 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2278 // ~umax(~x, ~y) == umin(x, y)
2279 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2282 const SCEV *ScalarEvolution::getSizeOfExpr(const Type *AllocTy) {
2283 // If we have TargetData, we can bypass creating a target-independent
2284 // constant expression and then folding it back into a ConstantInt.
2285 // This is just a compile-time optimization.
2287 return getConstant(TD->getIntPtrType(getContext()),
2288 TD->getTypeAllocSize(AllocTy));
2290 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2291 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2292 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2294 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2295 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2298 const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
2299 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2300 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2301 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2303 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2304 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2307 const SCEV *ScalarEvolution::getOffsetOfExpr(const StructType *STy,
2309 // If we have TargetData, we can bypass creating a target-independent
2310 // constant expression and then folding it back into a ConstantInt.
2311 // This is just a compile-time optimization.
2313 return getConstant(TD->getIntPtrType(getContext()),
2314 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2316 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2317 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2318 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2320 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2321 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2324 const SCEV *ScalarEvolution::getOffsetOfExpr(const Type *CTy,
2325 Constant *FieldNo) {
2326 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2327 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2328 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2330 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2331 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2334 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2335 // Don't attempt to do anything other than create a SCEVUnknown object
2336 // here. createSCEV only calls getUnknown after checking for all other
2337 // interesting possibilities, and any other code that calls getUnknown
2338 // is doing so in order to hide a value from SCEV canonicalization.
2340 FoldingSetNodeID ID;
2341 ID.AddInteger(scUnknown);
2344 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2345 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2346 "Stale SCEVUnknown in uniquing map!");
2349 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2351 FirstUnknown = cast<SCEVUnknown>(S);
2352 UniqueSCEVs.InsertNode(S, IP);
2356 //===----------------------------------------------------------------------===//
2357 // Basic SCEV Analysis and PHI Idiom Recognition Code
2360 /// isSCEVable - Test if values of the given type are analyzable within
2361 /// the SCEV framework. This primarily includes integer types, and it
2362 /// can optionally include pointer types if the ScalarEvolution class
2363 /// has access to target-specific information.
2364 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2365 // Integers and pointers are always SCEVable.
2366 return Ty->isIntegerTy() || Ty->isPointerTy();
2369 /// getTypeSizeInBits - Return the size in bits of the specified type,
2370 /// for which isSCEVable must return true.
2371 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2372 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2374 // If we have a TargetData, use it!
2376 return TD->getTypeSizeInBits(Ty);
2378 // Integer types have fixed sizes.
2379 if (Ty->isIntegerTy())
2380 return Ty->getPrimitiveSizeInBits();
2382 // The only other support type is pointer. Without TargetData, conservatively
2383 // assume pointers are 64-bit.
2384 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2388 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2389 /// the given type and which represents how SCEV will treat the given
2390 /// type, for which isSCEVable must return true. For pointer types,
2391 /// this is the pointer-sized integer type.
2392 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2393 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2395 if (Ty->isIntegerTy())
2398 // The only other support type is pointer.
2399 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2400 if (TD) return TD->getIntPtrType(getContext());
2402 // Without TargetData, conservatively assume pointers are 64-bit.
2403 return Type::getInt64Ty(getContext());
2406 const SCEV *ScalarEvolution::getCouldNotCompute() {
2407 return &CouldNotCompute;
2410 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2411 /// expression and create a new one.
2412 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2413 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2415 ValueExprMapType::const_iterator I = ValueExprMap.find(V);
2416 if (I != ValueExprMap.end()) return I->second;
2417 const SCEV *S = createSCEV(V);
2419 // The process of creating a SCEV for V may have caused other SCEVs
2420 // to have been created, so it's necessary to insert the new entry
2421 // from scratch, rather than trying to remember the insert position
2423 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2427 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2429 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2430 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2432 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2434 const Type *Ty = V->getType();
2435 Ty = getEffectiveSCEVType(Ty);
2436 return getMulExpr(V,
2437 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2440 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2441 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2442 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2444 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2446 const Type *Ty = V->getType();
2447 Ty = getEffectiveSCEVType(Ty);
2448 const SCEV *AllOnes =
2449 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2450 return getMinusSCEV(AllOnes, V);
2453 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1,
2454 /// and thus the HasNUW and HasNSW bits apply to the resultant add, not
2455 /// whether the sub would have overflowed.
2456 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2457 bool HasNUW, bool HasNSW) {
2458 // Fast path: X - X --> 0.
2460 return getConstant(LHS->getType(), 0);
2463 return getAddExpr(LHS, getNegativeSCEV(RHS), HasNUW, HasNSW);
2466 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2467 /// input value to the specified type. If the type must be extended, it is zero
2470 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, const Type *Ty) {
2471 const Type *SrcTy = V->getType();
2472 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2473 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2474 "Cannot truncate or zero extend with non-integer arguments!");
2475 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2476 return V; // No conversion
2477 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2478 return getTruncateExpr(V, Ty);
2479 return getZeroExtendExpr(V, Ty);
2482 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2483 /// input value to the specified type. If the type must be extended, it is sign
2486 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2488 const Type *SrcTy = V->getType();
2489 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2490 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2491 "Cannot truncate or zero extend with non-integer arguments!");
2492 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2493 return V; // No conversion
2494 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2495 return getTruncateExpr(V, Ty);
2496 return getSignExtendExpr(V, Ty);
2499 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2500 /// input value to the specified type. If the type must be extended, it is zero
2501 /// extended. The conversion must not be narrowing.
2503 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2504 const Type *SrcTy = V->getType();
2505 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2506 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2507 "Cannot noop or zero extend with non-integer arguments!");
2508 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2509 "getNoopOrZeroExtend cannot truncate!");
2510 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2511 return V; // No conversion
2512 return getZeroExtendExpr(V, Ty);
2515 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2516 /// input value to the specified type. If the type must be extended, it is sign
2517 /// extended. The conversion must not be narrowing.
2519 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
2520 const Type *SrcTy = V->getType();
2521 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2522 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2523 "Cannot noop or sign extend with non-integer arguments!");
2524 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2525 "getNoopOrSignExtend cannot truncate!");
2526 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2527 return V; // No conversion
2528 return getSignExtendExpr(V, Ty);
2531 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2532 /// the input value to the specified type. If the type must be extended,
2533 /// it is extended with unspecified bits. The conversion must not be
2536 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2537 const Type *SrcTy = V->getType();
2538 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2539 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2540 "Cannot noop or any extend with non-integer arguments!");
2541 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2542 "getNoopOrAnyExtend cannot truncate!");
2543 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2544 return V; // No conversion
2545 return getAnyExtendExpr(V, Ty);
2548 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2549 /// input value to the specified type. The conversion must not be widening.
2551 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2552 const Type *SrcTy = V->getType();
2553 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2554 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2555 "Cannot truncate or noop with non-integer arguments!");
2556 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2557 "getTruncateOrNoop cannot extend!");
2558 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2559 return V; // No conversion
2560 return getTruncateExpr(V, Ty);
2563 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2564 /// the types using zero-extension, and then perform a umax operation
2566 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2568 const SCEV *PromotedLHS = LHS;
2569 const SCEV *PromotedRHS = RHS;
2571 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2572 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2574 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2576 return getUMaxExpr(PromotedLHS, PromotedRHS);
2579 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2580 /// the types using zero-extension, and then perform a umin operation
2582 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2584 const SCEV *PromotedLHS = LHS;
2585 const SCEV *PromotedRHS = RHS;
2587 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2588 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2590 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2592 return getUMinExpr(PromotedLHS, PromotedRHS);
2595 /// PushDefUseChildren - Push users of the given Instruction
2596 /// onto the given Worklist.
2598 PushDefUseChildren(Instruction *I,
2599 SmallVectorImpl<Instruction *> &Worklist) {
2600 // Push the def-use children onto the Worklist stack.
2601 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2603 Worklist.push_back(cast<Instruction>(*UI));
2606 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2607 /// instructions that depend on the given instruction and removes them from
2608 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2611 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2612 SmallVector<Instruction *, 16> Worklist;
2613 PushDefUseChildren(PN, Worklist);
2615 SmallPtrSet<Instruction *, 8> Visited;
2617 while (!Worklist.empty()) {
2618 Instruction *I = Worklist.pop_back_val();
2619 if (!Visited.insert(I)) continue;
2621 ValueExprMapType::iterator It =
2622 ValueExprMap.find(static_cast<Value *>(I));
2623 if (It != ValueExprMap.end()) {
2624 const SCEV *Old = It->second;
2626 // Short-circuit the def-use traversal if the symbolic name
2627 // ceases to appear in expressions.
2628 if (Old != SymName && !hasOperand(Old, SymName))
2631 // SCEVUnknown for a PHI either means that it has an unrecognized
2632 // structure, it's a PHI that's in the progress of being computed
2633 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2634 // additional loop trip count information isn't going to change anything.
2635 // In the second case, createNodeForPHI will perform the necessary
2636 // updates on its own when it gets to that point. In the third, we do
2637 // want to forget the SCEVUnknown.
2638 if (!isa<PHINode>(I) ||
2639 !isa<SCEVUnknown>(Old) ||
2640 (I != PN && Old == SymName)) {
2641 forgetMemoizedResults(Old);
2642 ValueExprMap.erase(It);
2646 PushDefUseChildren(I, Worklist);
2650 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2651 /// a loop header, making it a potential recurrence, or it doesn't.
2653 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2654 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2655 if (L->getHeader() == PN->getParent()) {
2656 // The loop may have multiple entrances or multiple exits; we can analyze
2657 // this phi as an addrec if it has a unique entry value and a unique
2659 Value *BEValueV = 0, *StartValueV = 0;
2660 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2661 Value *V = PN->getIncomingValue(i);
2662 if (L->contains(PN->getIncomingBlock(i))) {
2665 } else if (BEValueV != V) {
2669 } else if (!StartValueV) {
2671 } else if (StartValueV != V) {
2676 if (BEValueV && StartValueV) {
2677 // While we are analyzing this PHI node, handle its value symbolically.
2678 const SCEV *SymbolicName = getUnknown(PN);
2679 assert(ValueExprMap.find(PN) == ValueExprMap.end() &&
2680 "PHI node already processed?");
2681 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2683 // Using this symbolic name for the PHI, analyze the value coming around
2685 const SCEV *BEValue = getSCEV(BEValueV);
2687 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2688 // has a special value for the first iteration of the loop.
2690 // If the value coming around the backedge is an add with the symbolic
2691 // value we just inserted, then we found a simple induction variable!
2692 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2693 // If there is a single occurrence of the symbolic value, replace it
2694 // with a recurrence.
2695 unsigned FoundIndex = Add->getNumOperands();
2696 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2697 if (Add->getOperand(i) == SymbolicName)
2698 if (FoundIndex == e) {
2703 if (FoundIndex != Add->getNumOperands()) {
2704 // Create an add with everything but the specified operand.
2705 SmallVector<const SCEV *, 8> Ops;
2706 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2707 if (i != FoundIndex)
2708 Ops.push_back(Add->getOperand(i));
2709 const SCEV *Accum = getAddExpr(Ops);
2711 // This is not a valid addrec if the step amount is varying each
2712 // loop iteration, but is not itself an addrec in this loop.
2713 if (isLoopInvariant(Accum, L) ||
2714 (isa<SCEVAddRecExpr>(Accum) &&
2715 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2716 bool HasNUW = false;
2717 bool HasNSW = false;
2719 // If the increment doesn't overflow, then neither the addrec nor
2720 // the post-increment will overflow.
2721 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
2722 if (OBO->hasNoUnsignedWrap())
2724 if (OBO->hasNoSignedWrap())
2726 } else if (const GEPOperator *GEP =
2727 dyn_cast<GEPOperator>(BEValueV)) {
2728 // If the increment is a GEP, then we know it won't perform an
2729 // unsigned overflow, because the address space cannot be
2731 HasNUW |= GEP->isInBounds();
2734 const SCEV *StartVal = getSCEV(StartValueV);
2735 const SCEV *PHISCEV =
2736 getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
2738 // Since the no-wrap flags are on the increment, they apply to the
2739 // post-incremented value as well.
2740 if (isLoopInvariant(Accum, L))
2741 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
2742 Accum, L, HasNUW, HasNSW);
2744 // Okay, for the entire analysis of this edge we assumed the PHI
2745 // to be symbolic. We now need to go back and purge all of the
2746 // entries for the scalars that use the symbolic expression.
2747 ForgetSymbolicName(PN, SymbolicName);
2748 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2752 } else if (const SCEVAddRecExpr *AddRec =
2753 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2754 // Otherwise, this could be a loop like this:
2755 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2756 // In this case, j = {1,+,1} and BEValue is j.
2757 // Because the other in-value of i (0) fits the evolution of BEValue
2758 // i really is an addrec evolution.
2759 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2760 const SCEV *StartVal = getSCEV(StartValueV);
2762 // If StartVal = j.start - j.stride, we can use StartVal as the
2763 // initial step of the addrec evolution.
2764 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2765 AddRec->getOperand(1))) {
2766 const SCEV *PHISCEV =
2767 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2769 // Okay, for the entire analysis of this edge we assumed the PHI
2770 // to be symbolic. We now need to go back and purge all of the
2771 // entries for the scalars that use the symbolic expression.
2772 ForgetSymbolicName(PN, SymbolicName);
2773 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2781 // If the PHI has a single incoming value, follow that value, unless the
2782 // PHI's incoming blocks are in a different loop, in which case doing so
2783 // risks breaking LCSSA form. Instcombine would normally zap these, but
2784 // it doesn't have DominatorTree information, so it may miss cases.
2785 if (Value *V = SimplifyInstruction(PN, TD, DT))
2786 if (LI->replacementPreservesLCSSAForm(PN, V))
2789 // If it's not a loop phi, we can't handle it yet.
2790 return getUnknown(PN);
2793 /// createNodeForGEP - Expand GEP instructions into add and multiply
2794 /// operations. This allows them to be analyzed by regular SCEV code.
2796 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
2798 // Don't blindly transfer the inbounds flag from the GEP instruction to the
2799 // Add expression, because the Instruction may be guarded by control flow
2800 // and the no-overflow bits may not be valid for the expression in any
2803 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2804 Value *Base = GEP->getOperand(0);
2805 // Don't attempt to analyze GEPs over unsized objects.
2806 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2807 return getUnknown(GEP);
2808 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
2809 gep_type_iterator GTI = gep_type_begin(GEP);
2810 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
2814 // Compute the (potentially symbolic) offset in bytes for this index.
2815 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2816 // For a struct, add the member offset.
2817 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2818 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
2820 // Add the field offset to the running total offset.
2821 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
2823 // For an array, add the element offset, explicitly scaled.
2824 const SCEV *ElementSize = getSizeOfExpr(*GTI);
2825 const SCEV *IndexS = getSCEV(Index);
2826 // Getelementptr indices are signed.
2827 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
2829 // Multiply the index by the element size to compute the element offset.
2830 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize);
2832 // Add the element offset to the running total offset.
2833 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2837 // Get the SCEV for the GEP base.
2838 const SCEV *BaseS = getSCEV(Base);
2840 // Add the total offset from all the GEP indices to the base.
2841 return getAddExpr(BaseS, TotalOffset);
2844 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2845 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2846 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2847 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2849 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2850 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2851 return C->getValue()->getValue().countTrailingZeros();
2853 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2854 return std::min(GetMinTrailingZeros(T->getOperand()),
2855 (uint32_t)getTypeSizeInBits(T->getType()));
2857 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2858 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2859 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2860 getTypeSizeInBits(E->getType()) : OpRes;
2863 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2864 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2865 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2866 getTypeSizeInBits(E->getType()) : OpRes;
2869 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2870 // The result is the min of all operands results.
2871 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2872 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2873 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2877 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2878 // The result is the sum of all operands results.
2879 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2880 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2881 for (unsigned i = 1, e = M->getNumOperands();
2882 SumOpRes != BitWidth && i != e; ++i)
2883 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2888 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2889 // The result is the min of all operands results.
2890 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2891 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2892 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2896 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2897 // The result is the min of all operands results.
2898 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2899 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2900 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2904 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2905 // The result is the min of all operands results.
2906 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2907 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2908 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2912 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2913 // For a SCEVUnknown, ask ValueTracking.
2914 unsigned BitWidth = getTypeSizeInBits(U->getType());
2915 APInt Mask = APInt::getAllOnesValue(BitWidth);
2916 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2917 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
2918 return Zeros.countTrailingOnes();
2925 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
2928 ScalarEvolution::getUnsignedRange(const SCEV *S) {
2929 // See if we've computed this range already.
2930 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
2931 if (I != UnsignedRanges.end())
2934 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2935 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
2937 unsigned BitWidth = getTypeSizeInBits(S->getType());
2938 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
2940 // If the value has known zeros, the maximum unsigned value will have those
2941 // known zeros as well.
2942 uint32_t TZ = GetMinTrailingZeros(S);
2944 ConservativeResult =
2945 ConstantRange(APInt::getMinValue(BitWidth),
2946 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
2948 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2949 ConstantRange X = getUnsignedRange(Add->getOperand(0));
2950 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2951 X = X.add(getUnsignedRange(Add->getOperand(i)));
2952 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
2955 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2956 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
2957 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2958 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
2959 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
2962 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2963 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
2964 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2965 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
2966 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
2969 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2970 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
2971 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2972 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
2973 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
2976 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2977 ConstantRange X = getUnsignedRange(UDiv->getLHS());
2978 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
2979 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
2982 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2983 ConstantRange X = getUnsignedRange(ZExt->getOperand());
2984 return setUnsignedRange(ZExt,
2985 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
2988 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2989 ConstantRange X = getUnsignedRange(SExt->getOperand());
2990 return setUnsignedRange(SExt,
2991 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
2994 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
2995 ConstantRange X = getUnsignedRange(Trunc->getOperand());
2996 return setUnsignedRange(Trunc,
2997 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3000 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3001 // If there's no unsigned wrap, the value will never be less than its
3003 if (AddRec->hasNoUnsignedWrap())
3004 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3005 if (!C->getValue()->isZero())
3006 ConservativeResult =
3007 ConservativeResult.intersectWith(
3008 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3010 // TODO: non-affine addrec
3011 if (AddRec->isAffine()) {
3012 const Type *Ty = AddRec->getType();
3013 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3014 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3015 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3016 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3018 const SCEV *Start = AddRec->getStart();
3019 const SCEV *Step = AddRec->getStepRecurrence(*this);
3021 ConstantRange StartRange = getUnsignedRange(Start);
3022 ConstantRange StepRange = getSignedRange(Step);
3023 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3024 ConstantRange EndRange =
3025 StartRange.add(MaxBECountRange.multiply(StepRange));
3027 // Check for overflow. This must be done with ConstantRange arithmetic
3028 // because we could be called from within the ScalarEvolution overflow
3030 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3031 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3032 ConstantRange ExtMaxBECountRange =
3033 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3034 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3035 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3037 return setUnsignedRange(AddRec, ConservativeResult);
3039 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3040 EndRange.getUnsignedMin());
3041 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3042 EndRange.getUnsignedMax());
3043 if (Min.isMinValue() && Max.isMaxValue())
3044 return setUnsignedRange(AddRec, ConservativeResult);
3045 return setUnsignedRange(AddRec,
3046 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3050 return setUnsignedRange(AddRec, ConservativeResult);
3053 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3054 // For a SCEVUnknown, ask ValueTracking.
3055 APInt Mask = APInt::getAllOnesValue(BitWidth);
3056 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3057 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
3058 if (Ones == ~Zeros + 1)
3059 return setUnsignedRange(U, ConservativeResult);
3060 return setUnsignedRange(U,
3061 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3064 return setUnsignedRange(S, ConservativeResult);
3067 /// getSignedRange - Determine the signed range for a particular SCEV.
3070 ScalarEvolution::getSignedRange(const SCEV *S) {
3071 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3072 if (I != SignedRanges.end())
3075 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3076 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3078 unsigned BitWidth = getTypeSizeInBits(S->getType());
3079 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3081 // If the value has known zeros, the maximum signed value will have those
3082 // known zeros as well.
3083 uint32_t TZ = GetMinTrailingZeros(S);
3085 ConservativeResult =
3086 ConstantRange(APInt::getSignedMinValue(BitWidth),
3087 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3089 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3090 ConstantRange X = getSignedRange(Add->getOperand(0));
3091 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3092 X = X.add(getSignedRange(Add->getOperand(i)));
3093 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3096 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3097 ConstantRange X = getSignedRange(Mul->getOperand(0));
3098 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3099 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3100 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3103 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3104 ConstantRange X = getSignedRange(SMax->getOperand(0));
3105 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3106 X = X.smax(getSignedRange(SMax->getOperand(i)));
3107 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3110 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3111 ConstantRange X = getSignedRange(UMax->getOperand(0));
3112 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3113 X = X.umax(getSignedRange(UMax->getOperand(i)));
3114 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3117 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3118 ConstantRange X = getSignedRange(UDiv->getLHS());
3119 ConstantRange Y = getSignedRange(UDiv->getRHS());
3120 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3123 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3124 ConstantRange X = getSignedRange(ZExt->getOperand());
3125 return setSignedRange(ZExt,
3126 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3129 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3130 ConstantRange X = getSignedRange(SExt->getOperand());
3131 return setSignedRange(SExt,
3132 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3135 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3136 ConstantRange X = getSignedRange(Trunc->getOperand());
3137 return setSignedRange(Trunc,
3138 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3141 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3142 // If there's no signed wrap, and all the operands have the same sign or
3143 // zero, the value won't ever change sign.
3144 if (AddRec->hasNoSignedWrap()) {
3145 bool AllNonNeg = true;
3146 bool AllNonPos = true;
3147 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3148 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3149 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3152 ConservativeResult = ConservativeResult.intersectWith(
3153 ConstantRange(APInt(BitWidth, 0),
3154 APInt::getSignedMinValue(BitWidth)));
3156 ConservativeResult = ConservativeResult.intersectWith(
3157 ConstantRange(APInt::getSignedMinValue(BitWidth),
3158 APInt(BitWidth, 1)));
3161 // TODO: non-affine addrec
3162 if (AddRec->isAffine()) {
3163 const Type *Ty = AddRec->getType();
3164 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3165 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3166 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3167 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3169 const SCEV *Start = AddRec->getStart();
3170 const SCEV *Step = AddRec->getStepRecurrence(*this);
3172 ConstantRange StartRange = getSignedRange(Start);
3173 ConstantRange StepRange = getSignedRange(Step);
3174 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3175 ConstantRange EndRange =
3176 StartRange.add(MaxBECountRange.multiply(StepRange));
3178 // Check for overflow. This must be done with ConstantRange arithmetic
3179 // because we could be called from within the ScalarEvolution overflow
3181 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3182 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3183 ConstantRange ExtMaxBECountRange =
3184 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3185 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3186 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3188 return setSignedRange(AddRec, ConservativeResult);
3190 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3191 EndRange.getSignedMin());
3192 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3193 EndRange.getSignedMax());
3194 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3195 return setSignedRange(AddRec, ConservativeResult);
3196 return setSignedRange(AddRec,
3197 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3201 return setSignedRange(AddRec, ConservativeResult);
3204 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3205 // For a SCEVUnknown, ask ValueTracking.
3206 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3207 return setSignedRange(U, ConservativeResult);
3208 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3210 return setSignedRange(U, ConservativeResult);
3211 return setSignedRange(U, ConservativeResult.intersectWith(
3212 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3213 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3216 return setSignedRange(S, ConservativeResult);
3219 /// createSCEV - We know that there is no SCEV for the specified value.
3220 /// Analyze the expression.
3222 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3223 if (!isSCEVable(V->getType()))
3224 return getUnknown(V);
3226 unsigned Opcode = Instruction::UserOp1;
3227 if (Instruction *I = dyn_cast<Instruction>(V)) {
3228 Opcode = I->getOpcode();
3230 // Don't attempt to analyze instructions in blocks that aren't
3231 // reachable. Such instructions don't matter, and they aren't required
3232 // to obey basic rules for definitions dominating uses which this
3233 // analysis depends on.
3234 if (!DT->isReachableFromEntry(I->getParent()))
3235 return getUnknown(V);
3236 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3237 Opcode = CE->getOpcode();
3238 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3239 return getConstant(CI);
3240 else if (isa<ConstantPointerNull>(V))
3241 return getConstant(V->getType(), 0);
3242 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3243 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3245 return getUnknown(V);
3247 Operator *U = cast<Operator>(V);
3249 case Instruction::Add: {
3250 // The simple thing to do would be to just call getSCEV on both operands
3251 // and call getAddExpr with the result. However if we're looking at a
3252 // bunch of things all added together, this can be quite inefficient,
3253 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3254 // Instead, gather up all the operands and make a single getAddExpr call.
3255 // LLVM IR canonical form means we need only traverse the left operands.
3256 SmallVector<const SCEV *, 4> AddOps;
3257 AddOps.push_back(getSCEV(U->getOperand(1)));
3258 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3259 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3260 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3262 U = cast<Operator>(Op);
3263 const SCEV *Op1 = getSCEV(U->getOperand(1));
3264 if (Opcode == Instruction::Sub)
3265 AddOps.push_back(getNegativeSCEV(Op1));
3267 AddOps.push_back(Op1);
3269 AddOps.push_back(getSCEV(U->getOperand(0)));
3270 return getAddExpr(AddOps);
3272 case Instruction::Mul: {
3273 // See the Add code above.
3274 SmallVector<const SCEV *, 4> MulOps;
3275 MulOps.push_back(getSCEV(U->getOperand(1)));
3276 for (Value *Op = U->getOperand(0);
3277 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3278 Op = U->getOperand(0)) {
3279 U = cast<Operator>(Op);
3280 MulOps.push_back(getSCEV(U->getOperand(1)));
3282 MulOps.push_back(getSCEV(U->getOperand(0)));
3283 return getMulExpr(MulOps);
3285 case Instruction::UDiv:
3286 return getUDivExpr(getSCEV(U->getOperand(0)),
3287 getSCEV(U->getOperand(1)));
3288 case Instruction::Sub:
3289 return getMinusSCEV(getSCEV(U->getOperand(0)),
3290 getSCEV(U->getOperand(1)));
3291 case Instruction::And:
3292 // For an expression like x&255 that merely masks off the high bits,
3293 // use zext(trunc(x)) as the SCEV expression.
3294 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3295 if (CI->isNullValue())
3296 return getSCEV(U->getOperand(1));
3297 if (CI->isAllOnesValue())
3298 return getSCEV(U->getOperand(0));
3299 const APInt &A = CI->getValue();
3301 // Instcombine's ShrinkDemandedConstant may strip bits out of
3302 // constants, obscuring what would otherwise be a low-bits mask.
3303 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3304 // knew about to reconstruct a low-bits mask value.
3305 unsigned LZ = A.countLeadingZeros();
3306 unsigned BitWidth = A.getBitWidth();
3307 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
3308 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3309 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
3311 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3313 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3315 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3316 IntegerType::get(getContext(), BitWidth - LZ)),
3321 case Instruction::Or:
3322 // If the RHS of the Or is a constant, we may have something like:
3323 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3324 // optimizations will transparently handle this case.
3326 // In order for this transformation to be safe, the LHS must be of the
3327 // form X*(2^n) and the Or constant must be less than 2^n.
3328 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3329 const SCEV *LHS = getSCEV(U->getOperand(0));
3330 const APInt &CIVal = CI->getValue();
3331 if (GetMinTrailingZeros(LHS) >=
3332 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3333 // Build a plain add SCEV.
3334 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3335 // If the LHS of the add was an addrec and it has no-wrap flags,
3336 // transfer the no-wrap flags, since an or won't introduce a wrap.
3337 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3338 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3339 if (OldAR->hasNoUnsignedWrap())
3340 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3341 if (OldAR->hasNoSignedWrap())
3342 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3348 case Instruction::Xor:
3349 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3350 // If the RHS of the xor is a signbit, then this is just an add.
3351 // Instcombine turns add of signbit into xor as a strength reduction step.
3352 if (CI->getValue().isSignBit())
3353 return getAddExpr(getSCEV(U->getOperand(0)),
3354 getSCEV(U->getOperand(1)));
3356 // If the RHS of xor is -1, then this is a not operation.
3357 if (CI->isAllOnesValue())
3358 return getNotSCEV(getSCEV(U->getOperand(0)));
3360 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3361 // This is a variant of the check for xor with -1, and it handles
3362 // the case where instcombine has trimmed non-demanded bits out
3363 // of an xor with -1.
3364 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3365 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3366 if (BO->getOpcode() == Instruction::And &&
3367 LCI->getValue() == CI->getValue())
3368 if (const SCEVZeroExtendExpr *Z =
3369 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3370 const Type *UTy = U->getType();
3371 const SCEV *Z0 = Z->getOperand();
3372 const Type *Z0Ty = Z0->getType();
3373 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3375 // If C is a low-bits mask, the zero extend is serving to
3376 // mask off the high bits. Complement the operand and
3377 // re-apply the zext.
3378 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3379 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3381 // If C is a single bit, it may be in the sign-bit position
3382 // before the zero-extend. In this case, represent the xor
3383 // using an add, which is equivalent, and re-apply the zext.
3384 APInt Trunc = CI->getValue().trunc(Z0TySize);
3385 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3387 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3393 case Instruction::Shl:
3394 // Turn shift left of a constant amount into a multiply.
3395 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3396 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3398 // If the shift count is not less than the bitwidth, the result of
3399 // the shift is undefined. Don't try to analyze it, because the
3400 // resolution chosen here may differ from the resolution chosen in
3401 // other parts of the compiler.
3402 if (SA->getValue().uge(BitWidth))
3405 Constant *X = ConstantInt::get(getContext(),
3406 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3407 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3411 case Instruction::LShr:
3412 // Turn logical shift right of a constant into a unsigned divide.
3413 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3414 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3416 // If the shift count is not less than the bitwidth, the result of
3417 // the shift is undefined. Don't try to analyze it, because the
3418 // resolution chosen here may differ from the resolution chosen in
3419 // other parts of the compiler.
3420 if (SA->getValue().uge(BitWidth))
3423 Constant *X = ConstantInt::get(getContext(),
3424 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3425 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3429 case Instruction::AShr:
3430 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3431 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3432 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3433 if (L->getOpcode() == Instruction::Shl &&
3434 L->getOperand(1) == U->getOperand(1)) {
3435 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3437 // If the shift count is not less than the bitwidth, the result of
3438 // the shift is undefined. Don't try to analyze it, because the
3439 // resolution chosen here may differ from the resolution chosen in
3440 // other parts of the compiler.
3441 if (CI->getValue().uge(BitWidth))
3444 uint64_t Amt = BitWidth - CI->getZExtValue();
3445 if (Amt == BitWidth)
3446 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3448 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3449 IntegerType::get(getContext(),
3455 case Instruction::Trunc:
3456 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3458 case Instruction::ZExt:
3459 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3461 case Instruction::SExt:
3462 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3464 case Instruction::BitCast:
3465 // BitCasts are no-op casts so we just eliminate the cast.
3466 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3467 return getSCEV(U->getOperand(0));
3470 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3471 // lead to pointer expressions which cannot safely be expanded to GEPs,
3472 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3473 // simplifying integer expressions.
3475 case Instruction::GetElementPtr:
3476 return createNodeForGEP(cast<GEPOperator>(U));
3478 case Instruction::PHI:
3479 return createNodeForPHI(cast<PHINode>(U));
3481 case Instruction::Select:
3482 // This could be a smax or umax that was lowered earlier.
3483 // Try to recover it.
3484 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3485 Value *LHS = ICI->getOperand(0);
3486 Value *RHS = ICI->getOperand(1);
3487 switch (ICI->getPredicate()) {
3488 case ICmpInst::ICMP_SLT:
3489 case ICmpInst::ICMP_SLE:
3490 std::swap(LHS, RHS);
3492 case ICmpInst::ICMP_SGT:
3493 case ICmpInst::ICMP_SGE:
3494 // a >s b ? a+x : b+x -> smax(a, b)+x
3495 // a >s b ? b+x : a+x -> smin(a, b)+x
3496 if (LHS->getType() == U->getType()) {
3497 const SCEV *LS = getSCEV(LHS);
3498 const SCEV *RS = getSCEV(RHS);
3499 const SCEV *LA = getSCEV(U->getOperand(1));
3500 const SCEV *RA = getSCEV(U->getOperand(2));
3501 const SCEV *LDiff = getMinusSCEV(LA, LS);
3502 const SCEV *RDiff = getMinusSCEV(RA, RS);
3504 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3505 LDiff = getMinusSCEV(LA, RS);
3506 RDiff = getMinusSCEV(RA, LS);
3508 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3511 case ICmpInst::ICMP_ULT:
3512 case ICmpInst::ICMP_ULE:
3513 std::swap(LHS, RHS);
3515 case ICmpInst::ICMP_UGT:
3516 case ICmpInst::ICMP_UGE:
3517 // a >u b ? a+x : b+x -> umax(a, b)+x
3518 // a >u b ? b+x : a+x -> umin(a, b)+x
3519 if (LHS->getType() == U->getType()) {
3520 const SCEV *LS = getSCEV(LHS);
3521 const SCEV *RS = getSCEV(RHS);
3522 const SCEV *LA = getSCEV(U->getOperand(1));
3523 const SCEV *RA = getSCEV(U->getOperand(2));
3524 const SCEV *LDiff = getMinusSCEV(LA, LS);
3525 const SCEV *RDiff = getMinusSCEV(RA, RS);
3527 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3528 LDiff = getMinusSCEV(LA, RS);
3529 RDiff = getMinusSCEV(RA, LS);
3531 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3534 case ICmpInst::ICMP_NE:
3535 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3536 if (LHS->getType() == U->getType() &&
3537 isa<ConstantInt>(RHS) &&
3538 cast<ConstantInt>(RHS)->isZero()) {
3539 const SCEV *One = getConstant(LHS->getType(), 1);
3540 const SCEV *LS = getSCEV(LHS);
3541 const SCEV *LA = getSCEV(U->getOperand(1));
3542 const SCEV *RA = getSCEV(U->getOperand(2));
3543 const SCEV *LDiff = getMinusSCEV(LA, LS);
3544 const SCEV *RDiff = getMinusSCEV(RA, One);
3546 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3549 case ICmpInst::ICMP_EQ:
3550 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3551 if (LHS->getType() == U->getType() &&
3552 isa<ConstantInt>(RHS) &&
3553 cast<ConstantInt>(RHS)->isZero()) {
3554 const SCEV *One = getConstant(LHS->getType(), 1);
3555 const SCEV *LS = getSCEV(LHS);
3556 const SCEV *LA = getSCEV(U->getOperand(1));
3557 const SCEV *RA = getSCEV(U->getOperand(2));
3558 const SCEV *LDiff = getMinusSCEV(LA, One);
3559 const SCEV *RDiff = getMinusSCEV(RA, LS);
3561 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3569 default: // We cannot analyze this expression.
3573 return getUnknown(V);
3578 //===----------------------------------------------------------------------===//
3579 // Iteration Count Computation Code
3582 /// getBackedgeTakenCount - If the specified loop has a predictable
3583 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3584 /// object. The backedge-taken count is the number of times the loop header
3585 /// will be branched to from within the loop. This is one less than the
3586 /// trip count of the loop, since it doesn't count the first iteration,
3587 /// when the header is branched to from outside the loop.
3589 /// Note that it is not valid to call this method on a loop without a
3590 /// loop-invariant backedge-taken count (see
3591 /// hasLoopInvariantBackedgeTakenCount).
3593 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3594 return getBackedgeTakenInfo(L).Exact;
3597 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3598 /// return the least SCEV value that is known never to be less than the
3599 /// actual backedge taken count.
3600 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3601 return getBackedgeTakenInfo(L).Max;
3604 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3605 /// onto the given Worklist.
3607 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3608 BasicBlock *Header = L->getHeader();
3610 // Push all Loop-header PHIs onto the Worklist stack.
3611 for (BasicBlock::iterator I = Header->begin();
3612 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3613 Worklist.push_back(PN);
3616 const ScalarEvolution::BackedgeTakenInfo &
3617 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3618 // Initially insert a CouldNotCompute for this loop. If the insertion
3619 // succeeds, proceed to actually compute a backedge-taken count and
3620 // update the value. The temporary CouldNotCompute value tells SCEV
3621 // code elsewhere that it shouldn't attempt to request a new
3622 // backedge-taken count, which could result in infinite recursion.
3623 std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3624 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3626 return Pair.first->second;
3628 BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
3629 if (BECount.Exact != getCouldNotCompute()) {
3630 assert(isLoopInvariant(BECount.Exact, L) &&
3631 isLoopInvariant(BECount.Max, L) &&
3632 "Computed backedge-taken count isn't loop invariant for loop!");
3633 ++NumTripCountsComputed;
3635 // Update the value in the map.
3636 Pair.first->second = BECount;
3638 if (BECount.Max != getCouldNotCompute())
3639 // Update the value in the map.
3640 Pair.first->second = BECount;
3641 if (isa<PHINode>(L->getHeader()->begin()))
3642 // Only count loops that have phi nodes as not being computable.
3643 ++NumTripCountsNotComputed;
3646 // Now that we know more about the trip count for this loop, forget any
3647 // existing SCEV values for PHI nodes in this loop since they are only
3648 // conservative estimates made without the benefit of trip count
3649 // information. This is similar to the code in forgetLoop, except that
3650 // it handles SCEVUnknown PHI nodes specially.
3651 if (BECount.hasAnyInfo()) {
3652 SmallVector<Instruction *, 16> Worklist;
3653 PushLoopPHIs(L, Worklist);
3655 SmallPtrSet<Instruction *, 8> Visited;
3656 while (!Worklist.empty()) {
3657 Instruction *I = Worklist.pop_back_val();
3658 if (!Visited.insert(I)) continue;
3660 ValueExprMapType::iterator It =
3661 ValueExprMap.find(static_cast<Value *>(I));
3662 if (It != ValueExprMap.end()) {
3663 const SCEV *Old = It->second;
3665 // SCEVUnknown for a PHI either means that it has an unrecognized
3666 // structure, or it's a PHI that's in the progress of being computed
3667 // by createNodeForPHI. In the former case, additional loop trip
3668 // count information isn't going to change anything. In the later
3669 // case, createNodeForPHI will perform the necessary updates on its
3670 // own when it gets to that point.
3671 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
3672 forgetMemoizedResults(Old);
3673 ValueExprMap.erase(It);
3675 if (PHINode *PN = dyn_cast<PHINode>(I))
3676 ConstantEvolutionLoopExitValue.erase(PN);
3679 PushDefUseChildren(I, Worklist);
3682 return Pair.first->second;
3685 /// forgetLoop - This method should be called by the client when it has
3686 /// changed a loop in a way that may effect ScalarEvolution's ability to
3687 /// compute a trip count, or if the loop is deleted.
3688 void ScalarEvolution::forgetLoop(const Loop *L) {
3689 // Drop any stored trip count value.
3690 BackedgeTakenCounts.erase(L);
3692 // Drop information about expressions based on loop-header PHIs.
3693 SmallVector<Instruction *, 16> Worklist;
3694 PushLoopPHIs(L, Worklist);
3696 SmallPtrSet<Instruction *, 8> Visited;
3697 while (!Worklist.empty()) {
3698 Instruction *I = Worklist.pop_back_val();
3699 if (!Visited.insert(I)) continue;
3701 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3702 if (It != ValueExprMap.end()) {
3703 forgetMemoizedResults(It->second);
3704 ValueExprMap.erase(It);
3705 if (PHINode *PN = dyn_cast<PHINode>(I))
3706 ConstantEvolutionLoopExitValue.erase(PN);
3709 PushDefUseChildren(I, Worklist);
3712 // Forget all contained loops too, to avoid dangling entries in the
3713 // ValuesAtScopes map.
3714 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
3718 /// forgetValue - This method should be called by the client when it has
3719 /// changed a value in a way that may effect its value, or which may
3720 /// disconnect it from a def-use chain linking it to a loop.
3721 void ScalarEvolution::forgetValue(Value *V) {
3722 Instruction *I = dyn_cast<Instruction>(V);
3725 // Drop information about expressions based on loop-header PHIs.
3726 SmallVector<Instruction *, 16> Worklist;
3727 Worklist.push_back(I);
3729 SmallPtrSet<Instruction *, 8> Visited;
3730 while (!Worklist.empty()) {
3731 I = Worklist.pop_back_val();
3732 if (!Visited.insert(I)) continue;
3734 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3735 if (It != ValueExprMap.end()) {
3736 forgetMemoizedResults(It->second);
3737 ValueExprMap.erase(It);
3738 if (PHINode *PN = dyn_cast<PHINode>(I))
3739 ConstantEvolutionLoopExitValue.erase(PN);
3742 PushDefUseChildren(I, Worklist);
3746 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3747 /// of the specified loop will execute.
3748 ScalarEvolution::BackedgeTakenInfo
3749 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3750 SmallVector<BasicBlock *, 8> ExitingBlocks;
3751 L->getExitingBlocks(ExitingBlocks);
3753 // Examine all exits and pick the most conservative values.
3754 const SCEV *BECount = getCouldNotCompute();
3755 const SCEV *MaxBECount = getCouldNotCompute();
3756 bool CouldNotComputeBECount = false;
3757 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3758 BackedgeTakenInfo NewBTI =
3759 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3761 if (NewBTI.Exact == getCouldNotCompute()) {
3762 // We couldn't compute an exact value for this exit, so
3763 // we won't be able to compute an exact value for the loop.
3764 CouldNotComputeBECount = true;
3765 BECount = getCouldNotCompute();
3766 } else if (!CouldNotComputeBECount) {
3767 if (BECount == getCouldNotCompute())
3768 BECount = NewBTI.Exact;
3770 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3772 if (MaxBECount == getCouldNotCompute())
3773 MaxBECount = NewBTI.Max;
3774 else if (NewBTI.Max != getCouldNotCompute())
3775 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3778 return BackedgeTakenInfo(BECount, MaxBECount);
3781 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3782 /// of the specified loop will execute if it exits via the specified block.
3783 ScalarEvolution::BackedgeTakenInfo
3784 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3785 BasicBlock *ExitingBlock) {
3787 // Okay, we've chosen an exiting block. See what condition causes us to
3788 // exit at this block.
3790 // FIXME: we should be able to handle switch instructions (with a single exit)
3791 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3792 if (ExitBr == 0) return getCouldNotCompute();
3793 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3795 // At this point, we know we have a conditional branch that determines whether
3796 // the loop is exited. However, we don't know if the branch is executed each
3797 // time through the loop. If not, then the execution count of the branch will
3798 // not be equal to the trip count of the loop.
3800 // Currently we check for this by checking to see if the Exit branch goes to
3801 // the loop header. If so, we know it will always execute the same number of
3802 // times as the loop. We also handle the case where the exit block *is* the
3803 // loop header. This is common for un-rotated loops.
3805 // If both of those tests fail, walk up the unique predecessor chain to the
3806 // header, stopping if there is an edge that doesn't exit the loop. If the
3807 // header is reached, the execution count of the branch will be equal to the
3808 // trip count of the loop.
3810 // More extensive analysis could be done to handle more cases here.
3812 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3813 ExitBr->getSuccessor(1) != L->getHeader() &&
3814 ExitBr->getParent() != L->getHeader()) {
3815 // The simple checks failed, try climbing the unique predecessor chain
3816 // up to the header.
3818 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3819 BasicBlock *Pred = BB->getUniquePredecessor();
3821 return getCouldNotCompute();
3822 TerminatorInst *PredTerm = Pred->getTerminator();
3823 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3824 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3827 // If the predecessor has a successor that isn't BB and isn't
3828 // outside the loop, assume the worst.
3829 if (L->contains(PredSucc))
3830 return getCouldNotCompute();
3832 if (Pred == L->getHeader()) {
3839 return getCouldNotCompute();
3842 // Proceed to the next level to examine the exit condition expression.
3843 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3844 ExitBr->getSuccessor(0),
3845 ExitBr->getSuccessor(1));
3848 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3849 /// backedge of the specified loop will execute if its exit condition
3850 /// were a conditional branch of ExitCond, TBB, and FBB.
3851 ScalarEvolution::BackedgeTakenInfo
3852 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3856 // Check if the controlling expression for this loop is an And or Or.
3857 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3858 if (BO->getOpcode() == Instruction::And) {
3859 // Recurse on the operands of the and.
3860 BackedgeTakenInfo BTI0 =
3861 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3862 BackedgeTakenInfo BTI1 =
3863 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3864 const SCEV *BECount = getCouldNotCompute();
3865 const SCEV *MaxBECount = getCouldNotCompute();
3866 if (L->contains(TBB)) {
3867 // Both conditions must be true for the loop to continue executing.
3868 // Choose the less conservative count.
3869 if (BTI0.Exact == getCouldNotCompute() ||
3870 BTI1.Exact == getCouldNotCompute())
3871 BECount = getCouldNotCompute();
3873 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3874 if (BTI0.Max == getCouldNotCompute())
3875 MaxBECount = BTI1.Max;
3876 else if (BTI1.Max == getCouldNotCompute())
3877 MaxBECount = BTI0.Max;
3879 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3881 // Both conditions must be true at the same time for the loop to exit.
3882 // For now, be conservative.
3883 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3884 if (BTI0.Max == BTI1.Max)
3885 MaxBECount = BTI0.Max;
3886 if (BTI0.Exact == BTI1.Exact)
3887 BECount = BTI0.Exact;
3890 return BackedgeTakenInfo(BECount, MaxBECount);
3892 if (BO->getOpcode() == Instruction::Or) {
3893 // Recurse on the operands of the or.
3894 BackedgeTakenInfo BTI0 =
3895 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3896 BackedgeTakenInfo BTI1 =
3897 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3898 const SCEV *BECount = getCouldNotCompute();
3899 const SCEV *MaxBECount = getCouldNotCompute();
3900 if (L->contains(FBB)) {
3901 // Both conditions must be false for the loop to continue executing.
3902 // Choose the less conservative count.
3903 if (BTI0.Exact == getCouldNotCompute() ||
3904 BTI1.Exact == getCouldNotCompute())
3905 BECount = getCouldNotCompute();
3907 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3908 if (BTI0.Max == getCouldNotCompute())
3909 MaxBECount = BTI1.Max;
3910 else if (BTI1.Max == getCouldNotCompute())
3911 MaxBECount = BTI0.Max;
3913 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3915 // Both conditions must be false at the same time for the loop to exit.
3916 // For now, be conservative.
3917 assert(L->contains(TBB) && "Loop block has no successor in loop!");
3918 if (BTI0.Max == BTI1.Max)
3919 MaxBECount = BTI0.Max;
3920 if (BTI0.Exact == BTI1.Exact)
3921 BECount = BTI0.Exact;
3924 return BackedgeTakenInfo(BECount, MaxBECount);
3928 // With an icmp, it may be feasible to compute an exact backedge-taken count.
3929 // Proceed to the next level to examine the icmp.
3930 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
3931 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
3933 // Check for a constant condition. These are normally stripped out by
3934 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
3935 // preserve the CFG and is temporarily leaving constant conditions
3937 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
3938 if (L->contains(FBB) == !CI->getZExtValue())
3939 // The backedge is always taken.
3940 return getCouldNotCompute();
3942 // The backedge is never taken.
3943 return getConstant(CI->getType(), 0);
3946 // If it's not an integer or pointer comparison then compute it the hard way.
3947 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3950 static const SCEVAddRecExpr *
3951 isSimpleUnwrappingAddRec(const SCEV *S, const Loop *L) {
3952 const SCEVAddRecExpr *SA = dyn_cast<SCEVAddRecExpr>(S);
3954 // The SCEV must be an addrec of this loop.
3955 if (!SA || SA->getLoop() != L || !SA->isAffine())
3958 // The SCEV must be known to not wrap in some way to be interesting.
3959 if (!SA->hasNoUnsignedWrap() && !SA->hasNoSignedWrap())
3962 // The stride must be a constant so that we know if it is striding up or down.
3963 if (!isa<SCEVConstant>(SA->getOperand(1)))
3968 /// getMinusSCEVForExitTest - When considering an exit test for a loop with a
3969 /// "x != y" exit test, we turn this into a computation that evaluates x-y != 0,
3970 /// and this function returns the expression to use for x-y. We know and take
3971 /// advantage of the fact that this subtraction is only being used in a
3972 /// comparison by zero context.
3974 static const SCEV *getMinusSCEVForExitTest(const SCEV *LHS, const SCEV *RHS,
3975 const Loop *L, ScalarEvolution &SE) {
3976 // If either LHS or RHS is an AddRec SCEV (of this loop) that is known to not
3977 // wrap (either NSW or NUW), then we know that the value will either become
3978 // the other one (and thus the loop terminates), that the loop will terminate
3979 // through some other exit condition first, or that the loop has undefined
3980 // behavior. This information is useful when the addrec has a stride that is
3981 // != 1 or -1, because it means we can't "miss" the exit value.
3983 // In any of these three cases, it is safe to turn the exit condition into a
3984 // "counting down" AddRec (to zero) by subtracting the two inputs as normal,
3985 // but since we know that the "end cannot be missed" we can force the
3986 // resulting AddRec to be a NUW addrec. Since it is counting down, this means
3987 // that the AddRec *cannot* pass zero.
3989 // See if LHS and RHS are addrec's we can handle.
3990 const SCEVAddRecExpr *LHSA = isSimpleUnwrappingAddRec(LHS, L);
3991 const SCEVAddRecExpr *RHSA = isSimpleUnwrappingAddRec(RHS, L);
3993 // If neither addrec is interesting, just return a minus.
3994 if (RHSA == 0 && LHSA == 0)
3995 return SE.getMinusSCEV(LHS, RHS);
3997 // If only one of LHS and RHS are an AddRec of this loop, make sure it is LHS.
3998 if (RHSA && LHSA == 0) {
3999 // Safe because a-b === b-a for comparisons against zero.
4000 std::swap(LHS, RHS);
4001 std::swap(LHSA, RHSA);
4004 // Handle the case when only one is advancing in a non-overflowing way.
4006 // If RHS is loop varying, then we can't predict when LHS will cross it.
4007 if (!SE.isLoopInvariant(RHS, L))
4008 return SE.getMinusSCEV(LHS, RHS);
4010 // If LHS has a positive stride, then we compute RHS-LHS, because the loop
4011 // is counting up until it crosses RHS (which must be larger than LHS). If
4012 // it is negative, we compute LHS-RHS because we're counting down to RHS.
4013 const ConstantInt *Stride =
4014 cast<SCEVConstant>(LHSA->getOperand(1))->getValue();
4015 if (Stride->getValue().isNegative())
4016 std::swap(LHS, RHS);
4018 return SE.getMinusSCEV(RHS, LHS, true /*HasNUW*/);
4021 // If both LHS and RHS are interesting, we have something like:
4023 const ConstantInt *LHSStride =
4024 cast<SCEVConstant>(LHSA->getOperand(1))->getValue();
4025 const ConstantInt *RHSStride =
4026 cast<SCEVConstant>(RHSA->getOperand(1))->getValue();
4028 // If the strides are equal, then this is just a (complex) loop invariant
4029 // comparison of a and b.
4030 if (LHSStride == RHSStride)
4031 return SE.getMinusSCEV(LHSA->getStart(), RHSA->getStart());
4033 // If the signs of the strides differ, then the negative stride is counting
4034 // down to the positive stride.
4035 if (LHSStride->getValue().isNegative() != RHSStride->getValue().isNegative()){
4036 if (RHSStride->getValue().isNegative())
4037 std::swap(LHS, RHS);
4039 // If LHS's stride is smaller than RHS's stride, then "b" must be less than
4040 // "a" and "b" is RHS is counting up (catching up) to LHS. This is true
4041 // whether the strides are positive or negative.
4042 if (RHSStride->getValue().slt(LHSStride->getValue()))
4043 std::swap(LHS, RHS);
4046 return SE.getMinusSCEV(LHS, RHS, true /*HasNUW*/);
4049 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
4050 /// backedge of the specified loop will execute if its exit condition
4051 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4052 ScalarEvolution::BackedgeTakenInfo
4053 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
4058 // If the condition was exit on true, convert the condition to exit on false
4059 ICmpInst::Predicate Cond;
4060 if (!L->contains(FBB))
4061 Cond = ExitCond->getPredicate();
4063 Cond = ExitCond->getInversePredicate();
4065 // Handle common loops like: for (X = "string"; *X; ++X)
4066 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4067 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4068 BackedgeTakenInfo ItCnt =
4069 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
4070 if (ItCnt.hasAnyInfo())
4074 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4075 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4077 // Try to evaluate any dependencies out of the loop.
4078 LHS = getSCEVAtScope(LHS, L);
4079 RHS = getSCEVAtScope(RHS, L);
4081 // At this point, we would like to compute how many iterations of the
4082 // loop the predicate will return true for these inputs.
4083 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4084 // If there is a loop-invariant, force it into the RHS.
4085 std::swap(LHS, RHS);
4086 Cond = ICmpInst::getSwappedPredicate(Cond);
4089 // Simplify the operands before analyzing them.
4090 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4092 // If we have a comparison of a chrec against a constant, try to use value
4093 // ranges to answer this query.
4094 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4095 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4096 if (AddRec->getLoop() == L) {
4097 // Form the constant range.
4098 ConstantRange CompRange(
4099 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4101 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4102 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4106 case ICmpInst::ICMP_NE: { // while (X != Y)
4107 // Convert to: while (X-Y != 0)
4108 BackedgeTakenInfo BTI = HowFarToZero(getMinusSCEVForExitTest(LHS, RHS, L,
4110 if (BTI.hasAnyInfo()) return BTI;
4113 case ICmpInst::ICMP_EQ: { // while (X == Y)
4114 // Convert to: while (X-Y == 0)
4115 BackedgeTakenInfo BTI = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4116 if (BTI.hasAnyInfo()) return BTI;
4119 case ICmpInst::ICMP_SLT: {
4120 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
4121 if (BTI.hasAnyInfo()) return BTI;
4124 case ICmpInst::ICMP_SGT: {
4125 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4126 getNotSCEV(RHS), L, true);
4127 if (BTI.hasAnyInfo()) return BTI;
4130 case ICmpInst::ICMP_ULT: {
4131 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
4132 if (BTI.hasAnyInfo()) return BTI;
4135 case ICmpInst::ICMP_UGT: {
4136 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4137 getNotSCEV(RHS), L, false);
4138 if (BTI.hasAnyInfo()) return BTI;
4143 dbgs() << "ComputeBackedgeTakenCount ";
4144 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4145 dbgs() << "[unsigned] ";
4146 dbgs() << *LHS << " "
4147 << Instruction::getOpcodeName(Instruction::ICmp)
4148 << " " << *RHS << "\n";
4153 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
4156 static ConstantInt *
4157 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4158 ScalarEvolution &SE) {
4159 const SCEV *InVal = SE.getConstant(C);
4160 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4161 assert(isa<SCEVConstant>(Val) &&
4162 "Evaluation of SCEV at constant didn't fold correctly?");
4163 return cast<SCEVConstant>(Val)->getValue();
4166 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
4167 /// and a GEP expression (missing the pointer index) indexing into it, return
4168 /// the addressed element of the initializer or null if the index expression is
4171 GetAddressedElementFromGlobal(GlobalVariable *GV,
4172 const std::vector<ConstantInt*> &Indices) {
4173 Constant *Init = GV->getInitializer();
4174 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
4175 uint64_t Idx = Indices[i]->getZExtValue();
4176 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
4177 assert(Idx < CS->getNumOperands() && "Bad struct index!");
4178 Init = cast<Constant>(CS->getOperand(Idx));
4179 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
4180 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
4181 Init = cast<Constant>(CA->getOperand(Idx));
4182 } else if (isa<ConstantAggregateZero>(Init)) {
4183 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
4184 assert(Idx < STy->getNumElements() && "Bad struct index!");
4185 Init = Constant::getNullValue(STy->getElementType(Idx));
4186 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
4187 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
4188 Init = Constant::getNullValue(ATy->getElementType());
4190 llvm_unreachable("Unknown constant aggregate type!");
4194 return 0; // Unknown initializer type
4200 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
4201 /// 'icmp op load X, cst', try to see if we can compute the backedge
4202 /// execution count.
4203 ScalarEvolution::BackedgeTakenInfo
4204 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
4208 ICmpInst::Predicate predicate) {
4209 if (LI->isVolatile()) return getCouldNotCompute();
4211 // Check to see if the loaded pointer is a getelementptr of a global.
4212 // TODO: Use SCEV instead of manually grubbing with GEPs.
4213 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4214 if (!GEP) return getCouldNotCompute();
4216 // Make sure that it is really a constant global we are gepping, with an
4217 // initializer, and make sure the first IDX is really 0.
4218 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4219 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4220 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4221 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4222 return getCouldNotCompute();
4224 // Okay, we allow one non-constant index into the GEP instruction.
4226 std::vector<ConstantInt*> Indexes;
4227 unsigned VarIdxNum = 0;
4228 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4229 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4230 Indexes.push_back(CI);
4231 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4232 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4233 VarIdx = GEP->getOperand(i);
4235 Indexes.push_back(0);
4238 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4239 // Check to see if X is a loop variant variable value now.
4240 const SCEV *Idx = getSCEV(VarIdx);
4241 Idx = getSCEVAtScope(Idx, L);
4243 // We can only recognize very limited forms of loop index expressions, in
4244 // particular, only affine AddRec's like {C1,+,C2}.
4245 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4246 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4247 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4248 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4249 return getCouldNotCompute();
4251 unsigned MaxSteps = MaxBruteForceIterations;
4252 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4253 ConstantInt *ItCst = ConstantInt::get(
4254 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4255 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4257 // Form the GEP offset.
4258 Indexes[VarIdxNum] = Val;
4260 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
4261 if (Result == 0) break; // Cannot compute!
4263 // Evaluate the condition for this iteration.
4264 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4265 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4266 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4268 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4269 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4272 ++NumArrayLenItCounts;
4273 return getConstant(ItCst); // Found terminating iteration!
4276 return getCouldNotCompute();
4280 /// CanConstantFold - Return true if we can constant fold an instruction of the
4281 /// specified type, assuming that all operands were constants.
4282 static bool CanConstantFold(const Instruction *I) {
4283 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4284 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
4287 if (const CallInst *CI = dyn_cast<CallInst>(I))
4288 if (const Function *F = CI->getCalledFunction())
4289 return canConstantFoldCallTo(F);
4293 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4294 /// in the loop that V is derived from. We allow arbitrary operations along the
4295 /// way, but the operands of an operation must either be constants or a value
4296 /// derived from a constant PHI. If this expression does not fit with these
4297 /// constraints, return null.
4298 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4299 // If this is not an instruction, or if this is an instruction outside of the
4300 // loop, it can't be derived from a loop PHI.
4301 Instruction *I = dyn_cast<Instruction>(V);
4302 if (I == 0 || !L->contains(I)) return 0;
4304 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4305 if (L->getHeader() == I->getParent())
4308 // We don't currently keep track of the control flow needed to evaluate
4309 // PHIs, so we cannot handle PHIs inside of loops.
4313 // If we won't be able to constant fold this expression even if the operands
4314 // are constants, return early.
4315 if (!CanConstantFold(I)) return 0;
4317 // Otherwise, we can evaluate this instruction if all of its operands are
4318 // constant or derived from a PHI node themselves.
4320 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
4321 if (!isa<Constant>(I->getOperand(Op))) {
4322 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
4323 if (P == 0) return 0; // Not evolving from PHI
4327 return 0; // Evolving from multiple different PHIs.
4330 // This is a expression evolving from a constant PHI!
4334 /// EvaluateExpression - Given an expression that passes the
4335 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4336 /// in the loop has the value PHIVal. If we can't fold this expression for some
4337 /// reason, return null.
4338 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
4339 const TargetData *TD) {
4340 if (isa<PHINode>(V)) return PHIVal;
4341 if (Constant *C = dyn_cast<Constant>(V)) return C;
4342 Instruction *I = cast<Instruction>(V);
4344 std::vector<Constant*> Operands(I->getNumOperands());
4346 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4347 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
4348 if (Operands[i] == 0) return 0;
4351 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4352 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4354 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4355 &Operands[0], Operands.size(), TD);
4358 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4359 /// in the header of its containing loop, we know the loop executes a
4360 /// constant number of times, and the PHI node is just a recurrence
4361 /// involving constants, fold it.
4363 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4366 std::map<PHINode*, Constant*>::const_iterator I =
4367 ConstantEvolutionLoopExitValue.find(PN);
4368 if (I != ConstantEvolutionLoopExitValue.end())
4371 if (BEs.ugt(MaxBruteForceIterations))
4372 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4374 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4376 // Since the loop is canonicalized, the PHI node must have two entries. One
4377 // entry must be a constant (coming in from outside of the loop), and the
4378 // second must be derived from the same PHI.
4379 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4380 Constant *StartCST =
4381 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4383 return RetVal = 0; // Must be a constant.
4385 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4386 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4387 !isa<Constant>(BEValue))
4388 return RetVal = 0; // Not derived from same PHI.
4390 // Execute the loop symbolically to determine the exit value.
4391 if (BEs.getActiveBits() >= 32)
4392 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4394 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4395 unsigned IterationNum = 0;
4396 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
4397 if (IterationNum == NumIterations)
4398 return RetVal = PHIVal; // Got exit value!
4400 // Compute the value of the PHI node for the next iteration.
4401 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4402 if (NextPHI == PHIVal)
4403 return RetVal = NextPHI; // Stopped evolving!
4405 return 0; // Couldn't evaluate!
4410 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
4411 /// constant number of times (the condition evolves only from constants),
4412 /// try to evaluate a few iterations of the loop until we get the exit
4413 /// condition gets a value of ExitWhen (true or false). If we cannot
4414 /// evaluate the trip count of the loop, return getCouldNotCompute().
4416 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
4419 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4420 if (PN == 0) return getCouldNotCompute();
4422 // If the loop is canonicalized, the PHI will have exactly two entries.
4423 // That's the only form we support here.
4424 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4426 // One entry must be a constant (coming in from outside of the loop), and the
4427 // second must be derived from the same PHI.
4428 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4429 Constant *StartCST =
4430 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4431 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
4433 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4434 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4435 !isa<Constant>(BEValue))
4436 return getCouldNotCompute(); // Not derived from same PHI.
4438 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4439 // the loop symbolically to determine when the condition gets a value of
4441 unsigned IterationNum = 0;
4442 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4443 for (Constant *PHIVal = StartCST;
4444 IterationNum != MaxIterations; ++IterationNum) {
4445 ConstantInt *CondVal =
4446 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
4448 // Couldn't symbolically evaluate.
4449 if (!CondVal) return getCouldNotCompute();
4451 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4452 ++NumBruteForceTripCountsComputed;
4453 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4456 // Compute the value of the PHI node for the next iteration.
4457 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4458 if (NextPHI == 0 || NextPHI == PHIVal)
4459 return getCouldNotCompute();// Couldn't evaluate or not making progress...
4463 // Too many iterations were needed to evaluate.
4464 return getCouldNotCompute();
4467 /// getSCEVAtScope - Return a SCEV expression for the specified value
4468 /// at the specified scope in the program. The L value specifies a loop
4469 /// nest to evaluate the expression at, where null is the top-level or a
4470 /// specified loop is immediately inside of the loop.
4472 /// This method can be used to compute the exit value for a variable defined
4473 /// in a loop by querying what the value will hold in the parent loop.
4475 /// In the case that a relevant loop exit value cannot be computed, the
4476 /// original value V is returned.
4477 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4478 // Check to see if we've folded this expression at this loop before.
4479 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4480 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4481 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4483 return Pair.first->second ? Pair.first->second : V;
4485 // Otherwise compute it.
4486 const SCEV *C = computeSCEVAtScope(V, L);
4487 ValuesAtScopes[V][L] = C;
4491 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
4492 if (isa<SCEVConstant>(V)) return V;
4494 // If this instruction is evolved from a constant-evolving PHI, compute the
4495 // exit value from the loop without using SCEVs.
4496 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
4497 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
4498 const Loop *LI = (*this->LI)[I->getParent()];
4499 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
4500 if (PHINode *PN = dyn_cast<PHINode>(I))
4501 if (PN->getParent() == LI->getHeader()) {
4502 // Okay, there is no closed form solution for the PHI node. Check
4503 // to see if the loop that contains it has a known backedge-taken
4504 // count. If so, we may be able to force computation of the exit
4506 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
4507 if (const SCEVConstant *BTCC =
4508 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
4509 // Okay, we know how many times the containing loop executes. If
4510 // this is a constant evolving PHI node, get the final value at
4511 // the specified iteration number.
4512 Constant *RV = getConstantEvolutionLoopExitValue(PN,
4513 BTCC->getValue()->getValue(),
4515 if (RV) return getSCEV(RV);
4519 // Okay, this is an expression that we cannot symbolically evaluate
4520 // into a SCEV. Check to see if it's possible to symbolically evaluate
4521 // the arguments into constants, and if so, try to constant propagate the
4522 // result. This is particularly useful for computing loop exit values.
4523 if (CanConstantFold(I)) {
4524 SmallVector<Constant *, 4> Operands;
4525 bool MadeImprovement = false;
4526 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4527 Value *Op = I->getOperand(i);
4528 if (Constant *C = dyn_cast<Constant>(Op)) {
4529 Operands.push_back(C);
4533 // If any of the operands is non-constant and if they are
4534 // non-integer and non-pointer, don't even try to analyze them
4535 // with scev techniques.
4536 if (!isSCEVable(Op->getType()))
4539 const SCEV *OrigV = getSCEV(Op);
4540 const SCEV *OpV = getSCEVAtScope(OrigV, L);
4541 MadeImprovement |= OrigV != OpV;
4544 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
4546 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV))
4547 C = dyn_cast<Constant>(SU->getValue());
4549 if (C->getType() != Op->getType())
4550 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4554 Operands.push_back(C);
4557 // Check to see if getSCEVAtScope actually made an improvement.
4558 if (MadeImprovement) {
4560 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4561 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4562 Operands[0], Operands[1], TD);
4564 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4565 &Operands[0], Operands.size(), TD);
4572 // This is some other type of SCEVUnknown, just return it.
4576 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4577 // Avoid performing the look-up in the common case where the specified
4578 // expression has no loop-variant portions.
4579 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4580 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4581 if (OpAtScope != Comm->getOperand(i)) {
4582 // Okay, at least one of these operands is loop variant but might be
4583 // foldable. Build a new instance of the folded commutative expression.
4584 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4585 Comm->op_begin()+i);
4586 NewOps.push_back(OpAtScope);
4588 for (++i; i != e; ++i) {
4589 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4590 NewOps.push_back(OpAtScope);
4592 if (isa<SCEVAddExpr>(Comm))
4593 return getAddExpr(NewOps);
4594 if (isa<SCEVMulExpr>(Comm))
4595 return getMulExpr(NewOps);
4596 if (isa<SCEVSMaxExpr>(Comm))
4597 return getSMaxExpr(NewOps);
4598 if (isa<SCEVUMaxExpr>(Comm))
4599 return getUMaxExpr(NewOps);
4600 llvm_unreachable("Unknown commutative SCEV type!");
4603 // If we got here, all operands are loop invariant.
4607 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4608 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4609 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4610 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4611 return Div; // must be loop invariant
4612 return getUDivExpr(LHS, RHS);
4615 // If this is a loop recurrence for a loop that does not contain L, then we
4616 // are dealing with the final value computed by the loop.
4617 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4618 // First, attempt to evaluate each operand.
4619 // Avoid performing the look-up in the common case where the specified
4620 // expression has no loop-variant portions.
4621 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
4622 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
4623 if (OpAtScope == AddRec->getOperand(i))
4626 // Okay, at least one of these operands is loop variant but might be
4627 // foldable. Build a new instance of the folded commutative expression.
4628 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
4629 AddRec->op_begin()+i);
4630 NewOps.push_back(OpAtScope);
4631 for (++i; i != e; ++i)
4632 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
4634 AddRec = cast<SCEVAddRecExpr>(getAddRecExpr(NewOps, AddRec->getLoop()));
4638 // If the scope is outside the addrec's loop, evaluate it by using the
4639 // loop exit value of the addrec.
4640 if (!AddRec->getLoop()->contains(L)) {
4641 // To evaluate this recurrence, we need to know how many times the AddRec
4642 // loop iterates. Compute this now.
4643 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4644 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4646 // Then, evaluate the AddRec.
4647 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4653 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4654 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4655 if (Op == Cast->getOperand())
4656 return Cast; // must be loop invariant
4657 return getZeroExtendExpr(Op, Cast->getType());
4660 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4661 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4662 if (Op == Cast->getOperand())
4663 return Cast; // must be loop invariant
4664 return getSignExtendExpr(Op, Cast->getType());
4667 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4668 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4669 if (Op == Cast->getOperand())
4670 return Cast; // must be loop invariant
4671 return getTruncateExpr(Op, Cast->getType());
4674 llvm_unreachable("Unknown SCEV type!");
4678 /// getSCEVAtScope - This is a convenience function which does
4679 /// getSCEVAtScope(getSCEV(V), L).
4680 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4681 return getSCEVAtScope(getSCEV(V), L);
4684 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4685 /// following equation:
4687 /// A * X = B (mod N)
4689 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4690 /// A and B isn't important.
4692 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4693 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4694 ScalarEvolution &SE) {
4695 uint32_t BW = A.getBitWidth();
4696 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4697 assert(A != 0 && "A must be non-zero.");
4701 // The gcd of A and N may have only one prime factor: 2. The number of
4702 // trailing zeros in A is its multiplicity
4703 uint32_t Mult2 = A.countTrailingZeros();
4706 // 2. Check if B is divisible by D.
4708 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4709 // is not less than multiplicity of this prime factor for D.
4710 if (B.countTrailingZeros() < Mult2)
4711 return SE.getCouldNotCompute();
4713 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4716 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4717 // bit width during computations.
4718 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4719 APInt Mod(BW + 1, 0);
4720 Mod.setBit(BW - Mult2); // Mod = N / D
4721 APInt I = AD.multiplicativeInverse(Mod);
4723 // 4. Compute the minimum unsigned root of the equation:
4724 // I * (B / D) mod (N / D)
4725 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4727 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4729 return SE.getConstant(Result.trunc(BW));
4732 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4733 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4734 /// might be the same) or two SCEVCouldNotCompute objects.
4736 static std::pair<const SCEV *,const SCEV *>
4737 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4738 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4739 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4740 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4741 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4743 // We currently can only solve this if the coefficients are constants.
4744 if (!LC || !MC || !NC) {
4745 const SCEV *CNC = SE.getCouldNotCompute();
4746 return std::make_pair(CNC, CNC);
4749 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4750 const APInt &L = LC->getValue()->getValue();
4751 const APInt &M = MC->getValue()->getValue();
4752 const APInt &N = NC->getValue()->getValue();
4753 APInt Two(BitWidth, 2);
4754 APInt Four(BitWidth, 4);
4757 using namespace APIntOps;
4759 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4760 // The B coefficient is M-N/2
4764 // The A coefficient is N/2
4765 APInt A(N.sdiv(Two));
4767 // Compute the B^2-4ac term.
4770 SqrtTerm -= Four * (A * C);
4772 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4773 // integer value or else APInt::sqrt() will assert.
4774 APInt SqrtVal(SqrtTerm.sqrt());
4776 // Compute the two solutions for the quadratic formula.
4777 // The divisions must be performed as signed divisions.
4779 APInt TwoA( A << 1 );
4780 if (TwoA.isMinValue()) {
4781 const SCEV *CNC = SE.getCouldNotCompute();
4782 return std::make_pair(CNC, CNC);
4785 LLVMContext &Context = SE.getContext();
4787 ConstantInt *Solution1 =
4788 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4789 ConstantInt *Solution2 =
4790 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4792 return std::make_pair(SE.getConstant(Solution1),
4793 SE.getConstant(Solution2));
4794 } // end APIntOps namespace
4797 /// HowFarToZero - Return the number of times a backedge comparing the specified
4798 /// value to zero will execute. If not computable, return CouldNotCompute.
4799 ScalarEvolution::BackedgeTakenInfo
4800 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4801 // If the value is a constant
4802 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4803 // If the value is already zero, the branch will execute zero times.
4804 if (C->getValue()->isZero()) return C;
4805 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4808 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4809 if (!AddRec || AddRec->getLoop() != L)
4810 return getCouldNotCompute();
4812 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4813 // the quadratic equation to solve it.
4814 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
4815 std::pair<const SCEV *,const SCEV *> Roots =
4816 SolveQuadraticEquation(AddRec, *this);
4817 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4818 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4821 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
4822 << " sol#2: " << *R2 << "\n";
4824 // Pick the smallest positive root value.
4825 if (ConstantInt *CB =
4826 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
4829 if (CB->getZExtValue() == false)
4830 std::swap(R1, R2); // R1 is the minimum root now.
4832 // We can only use this value if the chrec ends up with an exact zero
4833 // value at this index. When solving for "X*X != 5", for example, we
4834 // should not accept a root of 2.
4835 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4837 return R1; // We found a quadratic root!
4840 return getCouldNotCompute();
4843 // Otherwise we can only handle this if it is affine.
4844 if (!AddRec->isAffine())
4845 return getCouldNotCompute();
4847 // If this is an affine expression, the execution count of this branch is
4848 // the minimum unsigned root of the following equation:
4850 // Start + Step*N = 0 (mod 2^BW)
4854 // Step*N = -Start (mod 2^BW)
4856 // where BW is the common bit width of Start and Step.
4858 // Get the initial value for the loop.
4859 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
4860 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
4862 // If the AddRec is NUW, then (in an unsigned sense) it cannot be counting up
4863 // to wrap to 0, it must be counting down to equal 0. Also, while counting
4864 // down, it cannot "miss" 0 (which would cause it to wrap), regardless of what
4865 // the stride is. As such, NUW addrec's will always become zero in
4866 // "start / -stride" steps, and we know that the division is exact.
4867 if (AddRec->hasNoUnsignedWrap())
4868 // FIXME: We really want an "isexact" bit for udiv.
4869 return getUDivExpr(Start, getNegativeSCEV(Step));
4871 // For now we handle only constant steps.
4872 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
4874 return getCouldNotCompute();
4876 // First, handle unitary steps.
4877 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4878 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4880 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4881 return Start; // N = Start (as unsigned)
4883 // Then, try to solve the above equation provided that Start is constant.
4884 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4885 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4886 -StartC->getValue()->getValue(),
4888 return getCouldNotCompute();
4891 /// HowFarToNonZero - Return the number of times a backedge checking the
4892 /// specified value for nonzero will execute. If not computable, return
4894 ScalarEvolution::BackedgeTakenInfo
4895 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4896 // Loops that look like: while (X == 0) are very strange indeed. We don't
4897 // handle them yet except for the trivial case. This could be expanded in the
4898 // future as needed.
4900 // If the value is a constant, check to see if it is known to be non-zero
4901 // already. If so, the backedge will execute zero times.
4902 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4903 if (!C->getValue()->isNullValue())
4904 return getConstant(C->getType(), 0);
4905 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4908 // We could implement others, but I really doubt anyone writes loops like
4909 // this, and if they did, they would already be constant folded.
4910 return getCouldNotCompute();
4913 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4914 /// (which may not be an immediate predecessor) which has exactly one
4915 /// successor from which BB is reachable, or null if no such block is
4918 std::pair<BasicBlock *, BasicBlock *>
4919 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4920 // If the block has a unique predecessor, then there is no path from the
4921 // predecessor to the block that does not go through the direct edge
4922 // from the predecessor to the block.
4923 if (BasicBlock *Pred = BB->getSinglePredecessor())
4924 return std::make_pair(Pred, BB);
4926 // A loop's header is defined to be a block that dominates the loop.
4927 // If the header has a unique predecessor outside the loop, it must be
4928 // a block that has exactly one successor that can reach the loop.
4929 if (Loop *L = LI->getLoopFor(BB))
4930 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
4932 return std::pair<BasicBlock *, BasicBlock *>();
4935 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4936 /// testing whether two expressions are equal, however for the purposes of
4937 /// looking for a condition guarding a loop, it can be useful to be a little
4938 /// more general, since a front-end may have replicated the controlling
4941 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4942 // Quick check to see if they are the same SCEV.
4943 if (A == B) return true;
4945 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4946 // two different instructions with the same value. Check for this case.
4947 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4948 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4949 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4950 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4951 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4954 // Otherwise assume they may have a different value.
4958 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
4959 /// predicate Pred. Return true iff any changes were made.
4961 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
4962 const SCEV *&LHS, const SCEV *&RHS) {
4963 bool Changed = false;
4965 // Canonicalize a constant to the right side.
4966 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
4967 // Check for both operands constant.
4968 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
4969 if (ConstantExpr::getICmp(Pred,
4971 RHSC->getValue())->isNullValue())
4972 goto trivially_false;
4974 goto trivially_true;
4976 // Otherwise swap the operands to put the constant on the right.
4977 std::swap(LHS, RHS);
4978 Pred = ICmpInst::getSwappedPredicate(Pred);
4982 // If we're comparing an addrec with a value which is loop-invariant in the
4983 // addrec's loop, put the addrec on the left. Also make a dominance check,
4984 // as both operands could be addrecs loop-invariant in each other's loop.
4985 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
4986 const Loop *L = AR->getLoop();
4987 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
4988 std::swap(LHS, RHS);
4989 Pred = ICmpInst::getSwappedPredicate(Pred);
4994 // If there's a constant operand, canonicalize comparisons with boundary
4995 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
4996 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
4997 const APInt &RA = RC->getValue()->getValue();
4999 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5000 case ICmpInst::ICMP_EQ:
5001 case ICmpInst::ICMP_NE:
5003 case ICmpInst::ICMP_UGE:
5004 if ((RA - 1).isMinValue()) {
5005 Pred = ICmpInst::ICMP_NE;
5006 RHS = getConstant(RA - 1);
5010 if (RA.isMaxValue()) {
5011 Pred = ICmpInst::ICMP_EQ;
5015 if (RA.isMinValue()) goto trivially_true;
5017 Pred = ICmpInst::ICMP_UGT;
5018 RHS = getConstant(RA - 1);
5021 case ICmpInst::ICMP_ULE:
5022 if ((RA + 1).isMaxValue()) {
5023 Pred = ICmpInst::ICMP_NE;
5024 RHS = getConstant(RA + 1);
5028 if (RA.isMinValue()) {
5029 Pred = ICmpInst::ICMP_EQ;
5033 if (RA.isMaxValue()) goto trivially_true;
5035 Pred = ICmpInst::ICMP_ULT;
5036 RHS = getConstant(RA + 1);
5039 case ICmpInst::ICMP_SGE:
5040 if ((RA - 1).isMinSignedValue()) {
5041 Pred = ICmpInst::ICMP_NE;
5042 RHS = getConstant(RA - 1);
5046 if (RA.isMaxSignedValue()) {
5047 Pred = ICmpInst::ICMP_EQ;
5051 if (RA.isMinSignedValue()) goto trivially_true;
5053 Pred = ICmpInst::ICMP_SGT;
5054 RHS = getConstant(RA - 1);
5057 case ICmpInst::ICMP_SLE:
5058 if ((RA + 1).isMaxSignedValue()) {
5059 Pred = ICmpInst::ICMP_NE;
5060 RHS = getConstant(RA + 1);
5064 if (RA.isMinSignedValue()) {
5065 Pred = ICmpInst::ICMP_EQ;
5069 if (RA.isMaxSignedValue()) goto trivially_true;
5071 Pred = ICmpInst::ICMP_SLT;
5072 RHS = getConstant(RA + 1);
5075 case ICmpInst::ICMP_UGT:
5076 if (RA.isMinValue()) {
5077 Pred = ICmpInst::ICMP_NE;
5081 if ((RA + 1).isMaxValue()) {
5082 Pred = ICmpInst::ICMP_EQ;
5083 RHS = getConstant(RA + 1);
5087 if (RA.isMaxValue()) goto trivially_false;
5089 case ICmpInst::ICMP_ULT:
5090 if (RA.isMaxValue()) {
5091 Pred = ICmpInst::ICMP_NE;
5095 if ((RA - 1).isMinValue()) {
5096 Pred = ICmpInst::ICMP_EQ;
5097 RHS = getConstant(RA - 1);
5101 if (RA.isMinValue()) goto trivially_false;
5103 case ICmpInst::ICMP_SGT:
5104 if (RA.isMinSignedValue()) {
5105 Pred = ICmpInst::ICMP_NE;
5109 if ((RA + 1).isMaxSignedValue()) {
5110 Pred = ICmpInst::ICMP_EQ;
5111 RHS = getConstant(RA + 1);
5115 if (RA.isMaxSignedValue()) goto trivially_false;
5117 case ICmpInst::ICMP_SLT:
5118 if (RA.isMaxSignedValue()) {
5119 Pred = ICmpInst::ICMP_NE;
5123 if ((RA - 1).isMinSignedValue()) {
5124 Pred = ICmpInst::ICMP_EQ;
5125 RHS = getConstant(RA - 1);
5129 if (RA.isMinSignedValue()) goto trivially_false;
5134 // Check for obvious equality.
5135 if (HasSameValue(LHS, RHS)) {
5136 if (ICmpInst::isTrueWhenEqual(Pred))
5137 goto trivially_true;
5138 if (ICmpInst::isFalseWhenEqual(Pred))
5139 goto trivially_false;
5142 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5143 // adding or subtracting 1 from one of the operands.
5145 case ICmpInst::ICMP_SLE:
5146 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5147 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5148 /*HasNUW=*/false, /*HasNSW=*/true);
5149 Pred = ICmpInst::ICMP_SLT;
5151 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5152 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5153 /*HasNUW=*/false, /*HasNSW=*/true);
5154 Pred = ICmpInst::ICMP_SLT;
5158 case ICmpInst::ICMP_SGE:
5159 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5160 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5161 /*HasNUW=*/false, /*HasNSW=*/true);
5162 Pred = ICmpInst::ICMP_SGT;
5164 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5165 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5166 /*HasNUW=*/false, /*HasNSW=*/true);
5167 Pred = ICmpInst::ICMP_SGT;
5171 case ICmpInst::ICMP_ULE:
5172 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5173 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5174 /*HasNUW=*/true, /*HasNSW=*/false);
5175 Pred = ICmpInst::ICMP_ULT;
5177 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5178 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5179 /*HasNUW=*/true, /*HasNSW=*/false);
5180 Pred = ICmpInst::ICMP_ULT;
5184 case ICmpInst::ICMP_UGE:
5185 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5186 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5187 /*HasNUW=*/true, /*HasNSW=*/false);
5188 Pred = ICmpInst::ICMP_UGT;
5190 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5191 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5192 /*HasNUW=*/true, /*HasNSW=*/false);
5193 Pred = ICmpInst::ICMP_UGT;
5201 // TODO: More simplifications are possible here.
5207 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5208 Pred = ICmpInst::ICMP_EQ;
5213 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5214 Pred = ICmpInst::ICMP_NE;
5218 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5219 return getSignedRange(S).getSignedMax().isNegative();
5222 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5223 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5226 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5227 return !getSignedRange(S).getSignedMin().isNegative();
5230 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5231 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5234 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5235 return isKnownNegative(S) || isKnownPositive(S);
5238 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5239 const SCEV *LHS, const SCEV *RHS) {
5240 // Canonicalize the inputs first.
5241 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5243 // If LHS or RHS is an addrec, check to see if the condition is true in
5244 // every iteration of the loop.
5245 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5246 if (isLoopEntryGuardedByCond(
5247 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5248 isLoopBackedgeGuardedByCond(
5249 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5251 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5252 if (isLoopEntryGuardedByCond(
5253 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5254 isLoopBackedgeGuardedByCond(
5255 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5258 // Otherwise see what can be done with known constant ranges.
5259 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5263 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5264 const SCEV *LHS, const SCEV *RHS) {
5265 if (HasSameValue(LHS, RHS))
5266 return ICmpInst::isTrueWhenEqual(Pred);
5268 // This code is split out from isKnownPredicate because it is called from
5269 // within isLoopEntryGuardedByCond.
5272 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5274 case ICmpInst::ICMP_SGT:
5275 Pred = ICmpInst::ICMP_SLT;
5276 std::swap(LHS, RHS);
5277 case ICmpInst::ICMP_SLT: {
5278 ConstantRange LHSRange = getSignedRange(LHS);
5279 ConstantRange RHSRange = getSignedRange(RHS);
5280 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5282 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5286 case ICmpInst::ICMP_SGE:
5287 Pred = ICmpInst::ICMP_SLE;
5288 std::swap(LHS, RHS);
5289 case ICmpInst::ICMP_SLE: {
5290 ConstantRange LHSRange = getSignedRange(LHS);
5291 ConstantRange RHSRange = getSignedRange(RHS);
5292 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5294 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
5298 case ICmpInst::ICMP_UGT:
5299 Pred = ICmpInst::ICMP_ULT;
5300 std::swap(LHS, RHS);
5301 case ICmpInst::ICMP_ULT: {
5302 ConstantRange LHSRange = getUnsignedRange(LHS);
5303 ConstantRange RHSRange = getUnsignedRange(RHS);
5304 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
5306 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
5310 case ICmpInst::ICMP_UGE:
5311 Pred = ICmpInst::ICMP_ULE;
5312 std::swap(LHS, RHS);
5313 case ICmpInst::ICMP_ULE: {
5314 ConstantRange LHSRange = getUnsignedRange(LHS);
5315 ConstantRange RHSRange = getUnsignedRange(RHS);
5316 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
5318 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
5322 case ICmpInst::ICMP_NE: {
5323 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
5325 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
5328 const SCEV *Diff = getMinusSCEV(LHS, RHS);
5329 if (isKnownNonZero(Diff))
5333 case ICmpInst::ICMP_EQ:
5334 // The check at the top of the function catches the case where
5335 // the values are known to be equal.
5341 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
5342 /// protected by a conditional between LHS and RHS. This is used to
5343 /// to eliminate casts.
5345 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
5346 ICmpInst::Predicate Pred,
5347 const SCEV *LHS, const SCEV *RHS) {
5348 // Interpret a null as meaning no loop, where there is obviously no guard
5349 // (interprocedural conditions notwithstanding).
5350 if (!L) return true;
5352 BasicBlock *Latch = L->getLoopLatch();
5356 BranchInst *LoopContinuePredicate =
5357 dyn_cast<BranchInst>(Latch->getTerminator());
5358 if (!LoopContinuePredicate ||
5359 LoopContinuePredicate->isUnconditional())
5362 return isImpliedCond(Pred, LHS, RHS,
5363 LoopContinuePredicate->getCondition(),
5364 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
5367 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
5368 /// by a conditional between LHS and RHS. This is used to help avoid max
5369 /// expressions in loop trip counts, and to eliminate casts.
5371 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
5372 ICmpInst::Predicate Pred,
5373 const SCEV *LHS, const SCEV *RHS) {
5374 // Interpret a null as meaning no loop, where there is obviously no guard
5375 // (interprocedural conditions notwithstanding).
5376 if (!L) return false;
5378 // Starting at the loop predecessor, climb up the predecessor chain, as long
5379 // as there are predecessors that can be found that have unique successors
5380 // leading to the original header.
5381 for (std::pair<BasicBlock *, BasicBlock *>
5382 Pair(L->getLoopPredecessor(), L->getHeader());
5384 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
5386 BranchInst *LoopEntryPredicate =
5387 dyn_cast<BranchInst>(Pair.first->getTerminator());
5388 if (!LoopEntryPredicate ||
5389 LoopEntryPredicate->isUnconditional())
5392 if (isImpliedCond(Pred, LHS, RHS,
5393 LoopEntryPredicate->getCondition(),
5394 LoopEntryPredicate->getSuccessor(0) != Pair.second))
5401 /// isImpliedCond - Test whether the condition described by Pred, LHS,
5402 /// and RHS is true whenever the given Cond value evaluates to true.
5403 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
5404 const SCEV *LHS, const SCEV *RHS,
5405 Value *FoundCondValue,
5407 // Recursively handle And and Or conditions.
5408 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
5409 if (BO->getOpcode() == Instruction::And) {
5411 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5412 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5413 } else if (BO->getOpcode() == Instruction::Or) {
5415 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5416 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5420 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
5421 if (!ICI) return false;
5423 // Bail if the ICmp's operands' types are wider than the needed type
5424 // before attempting to call getSCEV on them. This avoids infinite
5425 // recursion, since the analysis of widening casts can require loop
5426 // exit condition information for overflow checking, which would
5428 if (getTypeSizeInBits(LHS->getType()) <
5429 getTypeSizeInBits(ICI->getOperand(0)->getType()))
5432 // Now that we found a conditional branch that dominates the loop, check to
5433 // see if it is the comparison we are looking for.
5434 ICmpInst::Predicate FoundPred;
5436 FoundPred = ICI->getInversePredicate();
5438 FoundPred = ICI->getPredicate();
5440 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
5441 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
5443 // Balance the types. The case where FoundLHS' type is wider than
5444 // LHS' type is checked for above.
5445 if (getTypeSizeInBits(LHS->getType()) >
5446 getTypeSizeInBits(FoundLHS->getType())) {
5447 if (CmpInst::isSigned(Pred)) {
5448 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
5449 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
5451 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
5452 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
5456 // Canonicalize the query to match the way instcombine will have
5457 // canonicalized the comparison.
5458 if (SimplifyICmpOperands(Pred, LHS, RHS))
5460 return CmpInst::isTrueWhenEqual(Pred);
5461 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
5462 if (FoundLHS == FoundRHS)
5463 return CmpInst::isFalseWhenEqual(Pred);
5465 // Check to see if we can make the LHS or RHS match.
5466 if (LHS == FoundRHS || RHS == FoundLHS) {
5467 if (isa<SCEVConstant>(RHS)) {
5468 std::swap(FoundLHS, FoundRHS);
5469 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
5471 std::swap(LHS, RHS);
5472 Pred = ICmpInst::getSwappedPredicate(Pred);
5476 // Check whether the found predicate is the same as the desired predicate.
5477 if (FoundPred == Pred)
5478 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
5480 // Check whether swapping the found predicate makes it the same as the
5481 // desired predicate.
5482 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
5483 if (isa<SCEVConstant>(RHS))
5484 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
5486 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
5487 RHS, LHS, FoundLHS, FoundRHS);
5490 // Check whether the actual condition is beyond sufficient.
5491 if (FoundPred == ICmpInst::ICMP_EQ)
5492 if (ICmpInst::isTrueWhenEqual(Pred))
5493 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
5495 if (Pred == ICmpInst::ICMP_NE)
5496 if (!ICmpInst::isTrueWhenEqual(FoundPred))
5497 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
5500 // Otherwise assume the worst.
5504 /// isImpliedCondOperands - Test whether the condition described by Pred,
5505 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
5506 /// and FoundRHS is true.
5507 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
5508 const SCEV *LHS, const SCEV *RHS,
5509 const SCEV *FoundLHS,
5510 const SCEV *FoundRHS) {
5511 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
5512 FoundLHS, FoundRHS) ||
5513 // ~x < ~y --> x > y
5514 isImpliedCondOperandsHelper(Pred, LHS, RHS,
5515 getNotSCEV(FoundRHS),
5516 getNotSCEV(FoundLHS));
5519 /// isImpliedCondOperandsHelper - Test whether the condition described by
5520 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
5521 /// FoundLHS, and FoundRHS is true.
5523 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
5524 const SCEV *LHS, const SCEV *RHS,
5525 const SCEV *FoundLHS,
5526 const SCEV *FoundRHS) {
5528 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5529 case ICmpInst::ICMP_EQ:
5530 case ICmpInst::ICMP_NE:
5531 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
5534 case ICmpInst::ICMP_SLT:
5535 case ICmpInst::ICMP_SLE:
5536 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
5537 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
5540 case ICmpInst::ICMP_SGT:
5541 case ICmpInst::ICMP_SGE:
5542 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
5543 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
5546 case ICmpInst::ICMP_ULT:
5547 case ICmpInst::ICMP_ULE:
5548 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
5549 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
5552 case ICmpInst::ICMP_UGT:
5553 case ICmpInst::ICMP_UGE:
5554 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
5555 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
5563 /// getBECount - Subtract the end and start values and divide by the step,
5564 /// rounding up, to get the number of times the backedge is executed. Return
5565 /// CouldNotCompute if an intermediate computation overflows.
5566 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
5570 assert(!isKnownNegative(Step) &&
5571 "This code doesn't handle negative strides yet!");
5573 const Type *Ty = Start->getType();
5574 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
5575 const SCEV *Diff = getMinusSCEV(End, Start);
5576 const SCEV *RoundUp = getAddExpr(Step, NegOne);
5578 // Add an adjustment to the difference between End and Start so that
5579 // the division will effectively round up.
5580 const SCEV *Add = getAddExpr(Diff, RoundUp);
5583 // Check Add for unsigned overflow.
5584 // TODO: More sophisticated things could be done here.
5585 const Type *WideTy = IntegerType::get(getContext(),
5586 getTypeSizeInBits(Ty) + 1);
5587 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
5588 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
5589 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
5590 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
5591 return getCouldNotCompute();
5594 return getUDivExpr(Add, Step);
5597 /// HowManyLessThans - Return the number of times a backedge containing the
5598 /// specified less-than comparison will execute. If not computable, return
5599 /// CouldNotCompute.
5600 ScalarEvolution::BackedgeTakenInfo
5601 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
5602 const Loop *L, bool isSigned) {
5603 // Only handle: "ADDREC < LoopInvariant".
5604 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute();
5606 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
5607 if (!AddRec || AddRec->getLoop() != L)
5608 return getCouldNotCompute();
5610 // Check to see if we have a flag which makes analysis easy.
5611 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
5612 AddRec->hasNoUnsignedWrap();
5614 if (AddRec->isAffine()) {
5615 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
5616 const SCEV *Step = AddRec->getStepRecurrence(*this);
5619 return getCouldNotCompute();
5620 if (Step->isOne()) {
5621 // With unit stride, the iteration never steps past the limit value.
5622 } else if (isKnownPositive(Step)) {
5623 // Test whether a positive iteration can step past the limit
5624 // value and past the maximum value for its type in a single step.
5625 // Note that it's not sufficient to check NoWrap here, because even
5626 // though the value after a wrap is undefined, it's not undefined
5627 // behavior, so if wrap does occur, the loop could either terminate or
5628 // loop infinitely, but in either case, the loop is guaranteed to
5629 // iterate at least until the iteration where the wrapping occurs.
5630 const SCEV *One = getConstant(Step->getType(), 1);
5632 APInt Max = APInt::getSignedMaxValue(BitWidth);
5633 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
5634 .slt(getSignedRange(RHS).getSignedMax()))
5635 return getCouldNotCompute();
5637 APInt Max = APInt::getMaxValue(BitWidth);
5638 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
5639 .ult(getUnsignedRange(RHS).getUnsignedMax()))
5640 return getCouldNotCompute();
5643 // TODO: Handle negative strides here and below.
5644 return getCouldNotCompute();
5646 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
5647 // m. So, we count the number of iterations in which {n,+,s} < m is true.
5648 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
5649 // treat m-n as signed nor unsigned due to overflow possibility.
5651 // First, we get the value of the LHS in the first iteration: n
5652 const SCEV *Start = AddRec->getOperand(0);
5654 // Determine the minimum constant start value.
5655 const SCEV *MinStart = getConstant(isSigned ?
5656 getSignedRange(Start).getSignedMin() :
5657 getUnsignedRange(Start).getUnsignedMin());
5659 // If we know that the condition is true in order to enter the loop,
5660 // then we know that it will run exactly (m-n)/s times. Otherwise, we
5661 // only know that it will execute (max(m,n)-n)/s times. In both cases,
5662 // the division must round up.
5663 const SCEV *End = RHS;
5664 if (!isLoopEntryGuardedByCond(L,
5665 isSigned ? ICmpInst::ICMP_SLT :
5667 getMinusSCEV(Start, Step), RHS))
5668 End = isSigned ? getSMaxExpr(RHS, Start)
5669 : getUMaxExpr(RHS, Start);
5671 // Determine the maximum constant end value.
5672 const SCEV *MaxEnd = getConstant(isSigned ?
5673 getSignedRange(End).getSignedMax() :
5674 getUnsignedRange(End).getUnsignedMax());
5676 // If MaxEnd is within a step of the maximum integer value in its type,
5677 // adjust it down to the minimum value which would produce the same effect.
5678 // This allows the subsequent ceiling division of (N+(step-1))/step to
5679 // compute the correct value.
5680 const SCEV *StepMinusOne = getMinusSCEV(Step,
5681 getConstant(Step->getType(), 1));
5684 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
5687 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
5690 // Finally, we subtract these two values and divide, rounding up, to get
5691 // the number of times the backedge is executed.
5692 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
5694 // The maximum backedge count is similar, except using the minimum start
5695 // value and the maximum end value.
5696 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
5698 return BackedgeTakenInfo(BECount, MaxBECount);
5701 return getCouldNotCompute();
5704 /// getNumIterationsInRange - Return the number of iterations of this loop that
5705 /// produce values in the specified constant range. Another way of looking at
5706 /// this is that it returns the first iteration number where the value is not in
5707 /// the condition, thus computing the exit count. If the iteration count can't
5708 /// be computed, an instance of SCEVCouldNotCompute is returned.
5709 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
5710 ScalarEvolution &SE) const {
5711 if (Range.isFullSet()) // Infinite loop.
5712 return SE.getCouldNotCompute();
5714 // If the start is a non-zero constant, shift the range to simplify things.
5715 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
5716 if (!SC->getValue()->isZero()) {
5717 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
5718 Operands[0] = SE.getConstant(SC->getType(), 0);
5719 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
5720 if (const SCEVAddRecExpr *ShiftedAddRec =
5721 dyn_cast<SCEVAddRecExpr>(Shifted))
5722 return ShiftedAddRec->getNumIterationsInRange(
5723 Range.subtract(SC->getValue()->getValue()), SE);
5724 // This is strange and shouldn't happen.
5725 return SE.getCouldNotCompute();
5728 // The only time we can solve this is when we have all constant indices.
5729 // Otherwise, we cannot determine the overflow conditions.
5730 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5731 if (!isa<SCEVConstant>(getOperand(i)))
5732 return SE.getCouldNotCompute();
5735 // Okay at this point we know that all elements of the chrec are constants and
5736 // that the start element is zero.
5738 // First check to see if the range contains zero. If not, the first
5740 unsigned BitWidth = SE.getTypeSizeInBits(getType());
5741 if (!Range.contains(APInt(BitWidth, 0)))
5742 return SE.getConstant(getType(), 0);
5745 // If this is an affine expression then we have this situation:
5746 // Solve {0,+,A} in Range === Ax in Range
5748 // We know that zero is in the range. If A is positive then we know that
5749 // the upper value of the range must be the first possible exit value.
5750 // If A is negative then the lower of the range is the last possible loop
5751 // value. Also note that we already checked for a full range.
5752 APInt One(BitWidth,1);
5753 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5754 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5756 // The exit value should be (End+A)/A.
5757 APInt ExitVal = (End + A).udiv(A);
5758 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5760 // Evaluate at the exit value. If we really did fall out of the valid
5761 // range, then we computed our trip count, otherwise wrap around or other
5762 // things must have happened.
5763 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5764 if (Range.contains(Val->getValue()))
5765 return SE.getCouldNotCompute(); // Something strange happened
5767 // Ensure that the previous value is in the range. This is a sanity check.
5768 assert(Range.contains(
5769 EvaluateConstantChrecAtConstant(this,
5770 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5771 "Linear scev computation is off in a bad way!");
5772 return SE.getConstant(ExitValue);
5773 } else if (isQuadratic()) {
5774 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5775 // quadratic equation to solve it. To do this, we must frame our problem in
5776 // terms of figuring out when zero is crossed, instead of when
5777 // Range.getUpper() is crossed.
5778 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5779 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5780 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5782 // Next, solve the constructed addrec
5783 std::pair<const SCEV *,const SCEV *> Roots =
5784 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5785 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5786 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5788 // Pick the smallest positive root value.
5789 if (ConstantInt *CB =
5790 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5791 R1->getValue(), R2->getValue()))) {
5792 if (CB->getZExtValue() == false)
5793 std::swap(R1, R2); // R1 is the minimum root now.
5795 // Make sure the root is not off by one. The returned iteration should
5796 // not be in the range, but the previous one should be. When solving
5797 // for "X*X < 5", for example, we should not return a root of 2.
5798 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5801 if (Range.contains(R1Val->getValue())) {
5802 // The next iteration must be out of the range...
5803 ConstantInt *NextVal =
5804 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5806 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5807 if (!Range.contains(R1Val->getValue()))
5808 return SE.getConstant(NextVal);
5809 return SE.getCouldNotCompute(); // Something strange happened
5812 // If R1 was not in the range, then it is a good return value. Make
5813 // sure that R1-1 WAS in the range though, just in case.
5814 ConstantInt *NextVal =
5815 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5816 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5817 if (Range.contains(R1Val->getValue()))
5819 return SE.getCouldNotCompute(); // Something strange happened
5824 return SE.getCouldNotCompute();
5829 //===----------------------------------------------------------------------===//
5830 // SCEVCallbackVH Class Implementation
5831 //===----------------------------------------------------------------------===//
5833 void ScalarEvolution::SCEVCallbackVH::deleted() {
5834 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5835 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5836 SE->ConstantEvolutionLoopExitValue.erase(PN);
5837 SE->ValueExprMap.erase(getValPtr());
5838 // this now dangles!
5841 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
5842 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5844 // Forget all the expressions associated with users of the old value,
5845 // so that future queries will recompute the expressions using the new
5847 Value *Old = getValPtr();
5848 SmallVector<User *, 16> Worklist;
5849 SmallPtrSet<User *, 8> Visited;
5850 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5852 Worklist.push_back(*UI);
5853 while (!Worklist.empty()) {
5854 User *U = Worklist.pop_back_val();
5855 // Deleting the Old value will cause this to dangle. Postpone
5856 // that until everything else is done.
5859 if (!Visited.insert(U))
5861 if (PHINode *PN = dyn_cast<PHINode>(U))
5862 SE->ConstantEvolutionLoopExitValue.erase(PN);
5863 SE->ValueExprMap.erase(U);
5864 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5866 Worklist.push_back(*UI);
5868 // Delete the Old value.
5869 if (PHINode *PN = dyn_cast<PHINode>(Old))
5870 SE->ConstantEvolutionLoopExitValue.erase(PN);
5871 SE->ValueExprMap.erase(Old);
5872 // this now dangles!
5875 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5876 : CallbackVH(V), SE(se) {}
5878 //===----------------------------------------------------------------------===//
5879 // ScalarEvolution Class Implementation
5880 //===----------------------------------------------------------------------===//
5882 ScalarEvolution::ScalarEvolution()
5883 : FunctionPass(ID), FirstUnknown(0) {
5884 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
5887 bool ScalarEvolution::runOnFunction(Function &F) {
5889 LI = &getAnalysis<LoopInfo>();
5890 TD = getAnalysisIfAvailable<TargetData>();
5891 DT = &getAnalysis<DominatorTree>();
5895 void ScalarEvolution::releaseMemory() {
5896 // Iterate through all the SCEVUnknown instances and call their
5897 // destructors, so that they release their references to their values.
5898 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
5902 ValueExprMap.clear();
5903 BackedgeTakenCounts.clear();
5904 ConstantEvolutionLoopExitValue.clear();
5905 ValuesAtScopes.clear();
5906 LoopDispositions.clear();
5907 BlockDispositions.clear();
5908 UnsignedRanges.clear();
5909 SignedRanges.clear();
5910 UniqueSCEVs.clear();
5911 SCEVAllocator.Reset();
5914 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5915 AU.setPreservesAll();
5916 AU.addRequiredTransitive<LoopInfo>();
5917 AU.addRequiredTransitive<DominatorTree>();
5920 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5921 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5924 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5926 // Print all inner loops first
5927 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5928 PrintLoopInfo(OS, SE, *I);
5931 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5934 SmallVector<BasicBlock *, 8> ExitBlocks;
5935 L->getExitBlocks(ExitBlocks);
5936 if (ExitBlocks.size() != 1)
5937 OS << "<multiple exits> ";
5939 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5940 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5942 OS << "Unpredictable backedge-taken count. ";
5947 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5950 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5951 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5953 OS << "Unpredictable max backedge-taken count. ";
5959 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
5960 // ScalarEvolution's implementation of the print method is to print
5961 // out SCEV values of all instructions that are interesting. Doing
5962 // this potentially causes it to create new SCEV objects though,
5963 // which technically conflicts with the const qualifier. This isn't
5964 // observable from outside the class though, so casting away the
5965 // const isn't dangerous.
5966 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
5968 OS << "Classifying expressions for: ";
5969 WriteAsOperand(OS, F, /*PrintType=*/false);
5971 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
5972 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
5975 const SCEV *SV = SE.getSCEV(&*I);
5978 const Loop *L = LI->getLoopFor((*I).getParent());
5980 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
5987 OS << "\t\t" "Exits: ";
5988 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
5989 if (!SE.isLoopInvariant(ExitValue, L)) {
5990 OS << "<<Unknown>>";
5999 OS << "Determining loop execution counts for: ";
6000 WriteAsOperand(OS, F, /*PrintType=*/false);
6002 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6003 PrintLoopInfo(OS, &SE, *I);
6006 ScalarEvolution::LoopDisposition
6007 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
6008 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S];
6009 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair =
6010 Values.insert(std::make_pair(L, LoopVariant));
6012 return Pair.first->second;
6014 LoopDisposition D = computeLoopDisposition(S, L);
6015 return LoopDispositions[S][L] = D;
6018 ScalarEvolution::LoopDisposition
6019 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
6020 switch (S->getSCEVType()) {
6022 return LoopInvariant;
6026 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
6027 case scAddRecExpr: {
6028 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6030 // If L is the addrec's loop, it's computable.
6031 if (AR->getLoop() == L)
6032 return LoopComputable;
6034 // Add recurrences are never invariant in the function-body (null loop).
6038 // This recurrence is variant w.r.t. L if L contains AR's loop.
6039 if (L->contains(AR->getLoop()))
6042 // This recurrence is invariant w.r.t. L if AR's loop contains L.
6043 if (AR->getLoop()->contains(L))
6044 return LoopInvariant;
6046 // This recurrence is variant w.r.t. L if any of its operands
6048 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
6050 if (!isLoopInvariant(*I, L))
6053 // Otherwise it's loop-invariant.
6054 return LoopInvariant;
6060 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6061 bool HasVarying = false;
6062 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6064 LoopDisposition D = getLoopDisposition(*I, L);
6065 if (D == LoopVariant)
6067 if (D == LoopComputable)
6070 return HasVarying ? LoopComputable : LoopInvariant;
6073 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6074 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
6075 if (LD == LoopVariant)
6077 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
6078 if (RD == LoopVariant)
6080 return (LD == LoopInvariant && RD == LoopInvariant) ?
6081 LoopInvariant : LoopComputable;
6084 // All non-instruction values are loop invariant. All instructions are loop
6085 // invariant if they are not contained in the specified loop.
6086 // Instructions are never considered invariant in the function body
6087 // (null loop) because they are defined within the "loop".
6088 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
6089 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
6090 return LoopInvariant;
6091 case scCouldNotCompute:
6092 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6096 llvm_unreachable("Unknown SCEV kind!");
6100 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
6101 return getLoopDisposition(S, L) == LoopInvariant;
6104 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
6105 return getLoopDisposition(S, L) == LoopComputable;
6108 ScalarEvolution::BlockDisposition
6109 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6110 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S];
6111 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool>
6112 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock));
6114 return Pair.first->second;
6116 BlockDisposition D = computeBlockDisposition(S, BB);
6117 return BlockDispositions[S][BB] = D;
6120 ScalarEvolution::BlockDisposition
6121 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6122 switch (S->getSCEVType()) {
6124 return ProperlyDominatesBlock;
6128 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
6129 case scAddRecExpr: {
6130 // This uses a "dominates" query instead of "properly dominates" query
6131 // to test for proper dominance too, because the instruction which
6132 // produces the addrec's value is a PHI, and a PHI effectively properly
6133 // dominates its entire containing block.
6134 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6135 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
6136 return DoesNotDominateBlock;
6138 // FALL THROUGH into SCEVNAryExpr handling.
6143 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6145 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6147 BlockDisposition D = getBlockDisposition(*I, BB);
6148 if (D == DoesNotDominateBlock)
6149 return DoesNotDominateBlock;
6150 if (D == DominatesBlock)
6153 return Proper ? ProperlyDominatesBlock : DominatesBlock;
6156 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6157 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6158 BlockDisposition LD = getBlockDisposition(LHS, BB);
6159 if (LD == DoesNotDominateBlock)
6160 return DoesNotDominateBlock;
6161 BlockDisposition RD = getBlockDisposition(RHS, BB);
6162 if (RD == DoesNotDominateBlock)
6163 return DoesNotDominateBlock;
6164 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
6165 ProperlyDominatesBlock : DominatesBlock;
6168 if (Instruction *I =
6169 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
6170 if (I->getParent() == BB)
6171 return DominatesBlock;
6172 if (DT->properlyDominates(I->getParent(), BB))
6173 return ProperlyDominatesBlock;
6174 return DoesNotDominateBlock;
6176 return ProperlyDominatesBlock;
6177 case scCouldNotCompute:
6178 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6179 return DoesNotDominateBlock;
6182 llvm_unreachable("Unknown SCEV kind!");
6183 return DoesNotDominateBlock;
6186 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
6187 return getBlockDisposition(S, BB) >= DominatesBlock;
6190 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
6191 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
6194 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
6195 switch (S->getSCEVType()) {
6200 case scSignExtend: {
6201 const SCEVCastExpr *Cast = cast<SCEVCastExpr>(S);
6202 const SCEV *CastOp = Cast->getOperand();
6203 return Op == CastOp || hasOperand(CastOp, Op);
6210 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6211 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6213 const SCEV *NAryOp = *I;
6214 if (NAryOp == Op || hasOperand(NAryOp, Op))
6220 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6221 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6222 return LHS == Op || hasOperand(LHS, Op) ||
6223 RHS == Op || hasOperand(RHS, Op);
6227 case scCouldNotCompute:
6228 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6232 llvm_unreachable("Unknown SCEV kind!");
6236 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
6237 ValuesAtScopes.erase(S);
6238 LoopDispositions.erase(S);
6239 BlockDispositions.erase(S);
6240 UnsignedRanges.erase(S);
6241 SignedRanges.erase(S);