1 //===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the implementation of the scalar evolution analysis
11 // engine, which is used primarily to analyze expressions involving induction
12 // variables in loops.
14 // There are several aspects to this library. First is the representation of
15 // scalar expressions, which are represented as subclasses of the SCEV class.
16 // These classes are used to represent certain types of subexpressions that we
17 // can handle. We only create one SCEV of a particular shape, so
18 // pointer-comparisons for equality are legal.
20 // One important aspect of the SCEV objects is that they are never cyclic, even
21 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
22 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
23 // recurrence) then we represent it directly as a recurrence node, otherwise we
24 // represent it as a SCEVUnknown node.
26 // In addition to being able to represent expressions of various types, we also
27 // have folders that are used to build the *canonical* representation for a
28 // particular expression. These folders are capable of using a variety of
29 // rewrite rules to simplify the expressions.
31 // Once the folders are defined, we can implement the more interesting
32 // higher-level code, such as the code that recognizes PHI nodes of various
33 // types, computes the execution count of a loop, etc.
35 // TODO: We should use these routines and value representations to implement
36 // dependence analysis!
38 //===----------------------------------------------------------------------===//
40 // There are several good references for the techniques used in this analysis.
42 // Chains of recurrences -- a method to expedite the evaluation
43 // of closed-form functions
44 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
46 // On computational properties of chains of recurrences
49 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
50 // Robert A. van Engelen
52 // Efficient Symbolic Analysis for Optimizing Compilers
53 // Robert A. van Engelen
55 // Using the chains of recurrences algebra for data dependence testing and
56 // induction variable substitution
57 // MS Thesis, Johnie Birch
59 //===----------------------------------------------------------------------===//
61 #define DEBUG_TYPE "scalar-evolution"
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/Constants.h"
64 #include "llvm/DerivedTypes.h"
65 #include "llvm/GlobalVariable.h"
66 #include "llvm/GlobalAlias.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/LLVMContext.h"
69 #include "llvm/Operator.h"
70 #include "llvm/Analysis/ConstantFolding.h"
71 #include "llvm/Analysis/Dominators.h"
72 #include "llvm/Analysis/InstructionSimplify.h"
73 #include "llvm/Analysis/LoopInfo.h"
74 #include "llvm/Analysis/ValueTracking.h"
75 #include "llvm/Assembly/Writer.h"
76 #include "llvm/Target/TargetData.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/ConstantRange.h"
79 #include "llvm/Support/Debug.h"
80 #include "llvm/Support/ErrorHandling.h"
81 #include "llvm/Support/GetElementPtrTypeIterator.h"
82 #include "llvm/Support/InstIterator.h"
83 #include "llvm/Support/MathExtras.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/ADT/Statistic.h"
86 #include "llvm/ADT/STLExtras.h"
87 #include "llvm/ADT/SmallPtrSet.h"
91 STATISTIC(NumArrayLenItCounts,
92 "Number of trip counts computed with array length");
93 STATISTIC(NumTripCountsComputed,
94 "Number of loops with predictable loop counts");
95 STATISTIC(NumTripCountsNotComputed,
96 "Number of loops without predictable loop counts");
97 STATISTIC(NumBruteForceTripCountsComputed,
98 "Number of loops with trip counts computed by force");
100 static cl::opt<unsigned>
101 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
102 cl::desc("Maximum number of iterations SCEV will "
103 "symbolically execute a constant "
107 INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
108 "Scalar Evolution Analysis", false, true)
109 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
110 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
111 INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
112 "Scalar Evolution Analysis", false, true)
113 char ScalarEvolution::ID = 0;
115 //===----------------------------------------------------------------------===//
116 // SCEV class definitions
117 //===----------------------------------------------------------------------===//
119 //===----------------------------------------------------------------------===//
120 // Implementation of the SCEV class.
123 void SCEV::dump() const {
128 void SCEV::print(raw_ostream &OS) const {
129 switch (getSCEVType()) {
131 WriteAsOperand(OS, cast<SCEVConstant>(this)->getValue(), false);
134 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(this);
135 const SCEV *Op = Trunc->getOperand();
136 OS << "(trunc " << *Op->getType() << " " << *Op << " to "
137 << *Trunc->getType() << ")";
141 const SCEVZeroExtendExpr *ZExt = cast<SCEVZeroExtendExpr>(this);
142 const SCEV *Op = ZExt->getOperand();
143 OS << "(zext " << *Op->getType() << " " << *Op << " to "
144 << *ZExt->getType() << ")";
148 const SCEVSignExtendExpr *SExt = cast<SCEVSignExtendExpr>(this);
149 const SCEV *Op = SExt->getOperand();
150 OS << "(sext " << *Op->getType() << " " << *Op << " to "
151 << *SExt->getType() << ")";
155 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(this);
156 OS << "{" << *AR->getOperand(0);
157 for (unsigned i = 1, e = AR->getNumOperands(); i != e; ++i)
158 OS << ",+," << *AR->getOperand(i);
160 if (AR->hasNoUnsignedWrap())
162 if (AR->hasNoSignedWrap())
164 WriteAsOperand(OS, AR->getLoop()->getHeader(), /*PrintType=*/false);
172 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(this);
173 const char *OpStr = 0;
174 switch (NAry->getSCEVType()) {
175 case scAddExpr: OpStr = " + "; break;
176 case scMulExpr: OpStr = " * "; break;
177 case scUMaxExpr: OpStr = " umax "; break;
178 case scSMaxExpr: OpStr = " smax "; break;
181 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
184 if (llvm::next(I) != E)
191 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(this);
192 OS << "(" << *UDiv->getLHS() << " /u " << *UDiv->getRHS() << ")";
196 const SCEVUnknown *U = cast<SCEVUnknown>(this);
198 if (U->isSizeOf(AllocTy)) {
199 OS << "sizeof(" << *AllocTy << ")";
202 if (U->isAlignOf(AllocTy)) {
203 OS << "alignof(" << *AllocTy << ")";
209 if (U->isOffsetOf(CTy, FieldNo)) {
210 OS << "offsetof(" << *CTy << ", ";
211 WriteAsOperand(OS, FieldNo, false);
216 // Otherwise just print it normally.
217 WriteAsOperand(OS, U->getValue(), false);
220 case scCouldNotCompute:
221 OS << "***COULDNOTCOMPUTE***";
225 llvm_unreachable("Unknown SCEV kind!");
228 const Type *SCEV::getType() const {
229 switch (getSCEVType()) {
231 return cast<SCEVConstant>(this)->getType();
235 return cast<SCEVCastExpr>(this)->getType();
240 return cast<SCEVNAryExpr>(this)->getType();
242 return cast<SCEVAddExpr>(this)->getType();
244 return cast<SCEVUDivExpr>(this)->getType();
246 return cast<SCEVUnknown>(this)->getType();
247 case scCouldNotCompute:
248 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
252 llvm_unreachable("Unknown SCEV kind!");
256 bool SCEV::isZero() const {
257 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
258 return SC->getValue()->isZero();
262 bool SCEV::isOne() const {
263 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
264 return SC->getValue()->isOne();
268 bool SCEV::isAllOnesValue() const {
269 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
270 return SC->getValue()->isAllOnesValue();
274 SCEVCouldNotCompute::SCEVCouldNotCompute() :
275 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
277 bool SCEVCouldNotCompute::classof(const SCEV *S) {
278 return S->getSCEVType() == scCouldNotCompute;
281 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
283 ID.AddInteger(scConstant);
286 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
287 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V);
288 UniqueSCEVs.InsertNode(S, IP);
292 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
293 return getConstant(ConstantInt::get(getContext(), Val));
297 ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
298 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
299 return getConstant(ConstantInt::get(ITy, V, isSigned));
302 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
303 unsigned SCEVTy, const SCEV *op, const Type *ty)
304 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
306 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
307 const SCEV *op, const Type *ty)
308 : SCEVCastExpr(ID, scTruncate, op, ty) {
309 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
310 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
311 "Cannot truncate non-integer value!");
314 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
315 const SCEV *op, const Type *ty)
316 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
317 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
318 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
319 "Cannot zero extend non-integer value!");
322 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
323 const SCEV *op, const Type *ty)
324 : SCEVCastExpr(ID, scSignExtend, op, ty) {
325 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
326 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
327 "Cannot sign extend non-integer value!");
330 void SCEVUnknown::deleted() {
331 // Clear this SCEVUnknown from various maps.
332 SE->forgetMemoizedResults(this);
334 // Remove this SCEVUnknown from the uniquing map.
335 SE->UniqueSCEVs.RemoveNode(this);
337 // Release the value.
341 void SCEVUnknown::allUsesReplacedWith(Value *New) {
342 // Clear this SCEVUnknown from various maps.
343 SE->forgetMemoizedResults(this);
345 // Remove this SCEVUnknown from the uniquing map.
346 SE->UniqueSCEVs.RemoveNode(this);
348 // Update this SCEVUnknown to point to the new value. This is needed
349 // because there may still be outstanding SCEVs which still point to
354 bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
355 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
356 if (VCE->getOpcode() == Instruction::PtrToInt)
357 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
358 if (CE->getOpcode() == Instruction::GetElementPtr &&
359 CE->getOperand(0)->isNullValue() &&
360 CE->getNumOperands() == 2)
361 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
363 AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
371 bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
372 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
373 if (VCE->getOpcode() == Instruction::PtrToInt)
374 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
375 if (CE->getOpcode() == Instruction::GetElementPtr &&
376 CE->getOperand(0)->isNullValue()) {
378 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
379 if (const StructType *STy = dyn_cast<StructType>(Ty))
380 if (!STy->isPacked() &&
381 CE->getNumOperands() == 3 &&
382 CE->getOperand(1)->isNullValue()) {
383 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
385 STy->getNumElements() == 2 &&
386 STy->getElementType(0)->isIntegerTy(1)) {
387 AllocTy = STy->getElementType(1);
396 bool SCEVUnknown::isOffsetOf(const Type *&CTy, Constant *&FieldNo) const {
397 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
398 if (VCE->getOpcode() == Instruction::PtrToInt)
399 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
400 if (CE->getOpcode() == Instruction::GetElementPtr &&
401 CE->getNumOperands() == 3 &&
402 CE->getOperand(0)->isNullValue() &&
403 CE->getOperand(1)->isNullValue()) {
405 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
406 // Ignore vector types here so that ScalarEvolutionExpander doesn't
407 // emit getelementptrs that index into vectors.
408 if (Ty->isStructTy() || Ty->isArrayTy()) {
410 FieldNo = CE->getOperand(2);
418 //===----------------------------------------------------------------------===//
420 //===----------------------------------------------------------------------===//
423 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
424 /// than the complexity of the RHS. This comparator is used to canonicalize
426 class SCEVComplexityCompare {
427 const LoopInfo *const LI;
429 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
431 // Return true or false if LHS is less than, or at least RHS, respectively.
432 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
433 return compare(LHS, RHS) < 0;
436 // Return negative, zero, or positive, if LHS is less than, equal to, or
437 // greater than RHS, respectively. A three-way result allows recursive
438 // comparisons to be more efficient.
439 int compare(const SCEV *LHS, const SCEV *RHS) const {
440 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
444 // Primarily, sort the SCEVs by their getSCEVType().
445 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
447 return (int)LType - (int)RType;
449 // Aside from the getSCEVType() ordering, the particular ordering
450 // isn't very important except that it's beneficial to be consistent,
451 // so that (a + b) and (b + a) don't end up as different expressions.
454 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS);
455 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
457 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
458 // not as complete as it could be.
459 const Value *LV = LU->getValue(), *RV = RU->getValue();
461 // Order pointer values after integer values. This helps SCEVExpander
463 bool LIsPointer = LV->getType()->isPointerTy(),
464 RIsPointer = RV->getType()->isPointerTy();
465 if (LIsPointer != RIsPointer)
466 return (int)LIsPointer - (int)RIsPointer;
468 // Compare getValueID values.
469 unsigned LID = LV->getValueID(),
470 RID = RV->getValueID();
472 return (int)LID - (int)RID;
474 // Sort arguments by their position.
475 if (const Argument *LA = dyn_cast<Argument>(LV)) {
476 const Argument *RA = cast<Argument>(RV);
477 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo();
478 return (int)LArgNo - (int)RArgNo;
481 // For instructions, compare their loop depth, and their operand
482 // count. This is pretty loose.
483 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) {
484 const Instruction *RInst = cast<Instruction>(RV);
486 // Compare loop depths.
487 const BasicBlock *LParent = LInst->getParent(),
488 *RParent = RInst->getParent();
489 if (LParent != RParent) {
490 unsigned LDepth = LI->getLoopDepth(LParent),
491 RDepth = LI->getLoopDepth(RParent);
492 if (LDepth != RDepth)
493 return (int)LDepth - (int)RDepth;
496 // Compare the number of operands.
497 unsigned LNumOps = LInst->getNumOperands(),
498 RNumOps = RInst->getNumOperands();
499 return (int)LNumOps - (int)RNumOps;
506 const SCEVConstant *LC = cast<SCEVConstant>(LHS);
507 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
509 // Compare constant values.
510 const APInt &LA = LC->getValue()->getValue();
511 const APInt &RA = RC->getValue()->getValue();
512 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth();
513 if (LBitWidth != RBitWidth)
514 return (int)LBitWidth - (int)RBitWidth;
515 return LA.ult(RA) ? -1 : 1;
519 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS);
520 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
522 // Compare addrec loop depths.
523 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop();
524 if (LLoop != RLoop) {
525 unsigned LDepth = LLoop->getLoopDepth(),
526 RDepth = RLoop->getLoopDepth();
527 if (LDepth != RDepth)
528 return (int)LDepth - (int)RDepth;
531 // Addrec complexity grows with operand count.
532 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands();
533 if (LNumOps != RNumOps)
534 return (int)LNumOps - (int)RNumOps;
536 // Lexicographically compare.
537 for (unsigned i = 0; i != LNumOps; ++i) {
538 long X = compare(LA->getOperand(i), RA->getOperand(i));
550 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS);
551 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
553 // Lexicographically compare n-ary expressions.
554 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
555 for (unsigned i = 0; i != LNumOps; ++i) {
558 long X = compare(LC->getOperand(i), RC->getOperand(i));
562 return (int)LNumOps - (int)RNumOps;
566 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
567 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
569 // Lexicographically compare udiv expressions.
570 long X = compare(LC->getLHS(), RC->getLHS());
573 return compare(LC->getRHS(), RC->getRHS());
579 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
580 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
582 // Compare cast expressions by operand.
583 return compare(LC->getOperand(), RC->getOperand());
590 llvm_unreachable("Unknown SCEV kind!");
596 /// GroupByComplexity - Given a list of SCEV objects, order them by their
597 /// complexity, and group objects of the same complexity together by value.
598 /// When this routine is finished, we know that any duplicates in the vector are
599 /// consecutive and that complexity is monotonically increasing.
601 /// Note that we go take special precautions to ensure that we get deterministic
602 /// results from this routine. In other words, we don't want the results of
603 /// this to depend on where the addresses of various SCEV objects happened to
606 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
608 if (Ops.size() < 2) return; // Noop
609 if (Ops.size() == 2) {
610 // This is the common case, which also happens to be trivially simple.
612 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
613 if (SCEVComplexityCompare(LI)(RHS, LHS))
618 // Do the rough sort by complexity.
619 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
621 // Now that we are sorted by complexity, group elements of the same
622 // complexity. Note that this is, at worst, N^2, but the vector is likely to
623 // be extremely short in practice. Note that we take this approach because we
624 // do not want to depend on the addresses of the objects we are grouping.
625 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
626 const SCEV *S = Ops[i];
627 unsigned Complexity = S->getSCEVType();
629 // If there are any objects of the same complexity and same value as this
631 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
632 if (Ops[j] == S) { // Found a duplicate.
633 // Move it to immediately after i'th element.
634 std::swap(Ops[i+1], Ops[j]);
635 ++i; // no need to rescan it.
636 if (i == e-2) return; // Done!
644 //===----------------------------------------------------------------------===//
645 // Simple SCEV method implementations
646 //===----------------------------------------------------------------------===//
648 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
650 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
652 const Type* ResultTy) {
653 // Handle the simplest case efficiently.
655 return SE.getTruncateOrZeroExtend(It, ResultTy);
657 // We are using the following formula for BC(It, K):
659 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
661 // Suppose, W is the bitwidth of the return value. We must be prepared for
662 // overflow. Hence, we must assure that the result of our computation is
663 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
664 // safe in modular arithmetic.
666 // However, this code doesn't use exactly that formula; the formula it uses
667 // is something like the following, where T is the number of factors of 2 in
668 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
671 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
673 // This formula is trivially equivalent to the previous formula. However,
674 // this formula can be implemented much more efficiently. The trick is that
675 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
676 // arithmetic. To do exact division in modular arithmetic, all we have
677 // to do is multiply by the inverse. Therefore, this step can be done at
680 // The next issue is how to safely do the division by 2^T. The way this
681 // is done is by doing the multiplication step at a width of at least W + T
682 // bits. This way, the bottom W+T bits of the product are accurate. Then,
683 // when we perform the division by 2^T (which is equivalent to a right shift
684 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
685 // truncated out after the division by 2^T.
687 // In comparison to just directly using the first formula, this technique
688 // is much more efficient; using the first formula requires W * K bits,
689 // but this formula less than W + K bits. Also, the first formula requires
690 // a division step, whereas this formula only requires multiplies and shifts.
692 // It doesn't matter whether the subtraction step is done in the calculation
693 // width or the input iteration count's width; if the subtraction overflows,
694 // the result must be zero anyway. We prefer here to do it in the width of
695 // the induction variable because it helps a lot for certain cases; CodeGen
696 // isn't smart enough to ignore the overflow, which leads to much less
697 // efficient code if the width of the subtraction is wider than the native
700 // (It's possible to not widen at all by pulling out factors of 2 before
701 // the multiplication; for example, K=2 can be calculated as
702 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
703 // extra arithmetic, so it's not an obvious win, and it gets
704 // much more complicated for K > 3.)
706 // Protection from insane SCEVs; this bound is conservative,
707 // but it probably doesn't matter.
709 return SE.getCouldNotCompute();
711 unsigned W = SE.getTypeSizeInBits(ResultTy);
713 // Calculate K! / 2^T and T; we divide out the factors of two before
714 // multiplying for calculating K! / 2^T to avoid overflow.
715 // Other overflow doesn't matter because we only care about the bottom
716 // W bits of the result.
717 APInt OddFactorial(W, 1);
719 for (unsigned i = 3; i <= K; ++i) {
721 unsigned TwoFactors = Mult.countTrailingZeros();
723 Mult = Mult.lshr(TwoFactors);
724 OddFactorial *= Mult;
727 // We need at least W + T bits for the multiplication step
728 unsigned CalculationBits = W + T;
730 // Calculate 2^T, at width T+W.
731 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
733 // Calculate the multiplicative inverse of K! / 2^T;
734 // this multiplication factor will perform the exact division by
736 APInt Mod = APInt::getSignedMinValue(W+1);
737 APInt MultiplyFactor = OddFactorial.zext(W+1);
738 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
739 MultiplyFactor = MultiplyFactor.trunc(W);
741 // Calculate the product, at width T+W
742 const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
744 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
745 for (unsigned i = 1; i != K; ++i) {
746 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
747 Dividend = SE.getMulExpr(Dividend,
748 SE.getTruncateOrZeroExtend(S, CalculationTy));
752 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
754 // Truncate the result, and divide by K! / 2^T.
756 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
757 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
760 /// evaluateAtIteration - Return the value of this chain of recurrences at
761 /// the specified iteration number. We can evaluate this recurrence by
762 /// multiplying each element in the chain by the binomial coefficient
763 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
765 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
767 /// where BC(It, k) stands for binomial coefficient.
769 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
770 ScalarEvolution &SE) const {
771 const SCEV *Result = getStart();
772 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
773 // The computation is correct in the face of overflow provided that the
774 // multiplication is performed _after_ the evaluation of the binomial
776 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
777 if (isa<SCEVCouldNotCompute>(Coeff))
780 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
785 //===----------------------------------------------------------------------===//
786 // SCEV Expression folder implementations
787 //===----------------------------------------------------------------------===//
789 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
791 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
792 "This is not a truncating conversion!");
793 assert(isSCEVable(Ty) &&
794 "This is not a conversion to a SCEVable type!");
795 Ty = getEffectiveSCEVType(Ty);
798 ID.AddInteger(scTruncate);
802 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
804 // Fold if the operand is constant.
805 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
807 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(),
808 getEffectiveSCEVType(Ty))));
810 // trunc(trunc(x)) --> trunc(x)
811 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
812 return getTruncateExpr(ST->getOperand(), Ty);
814 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
815 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
816 return getTruncateOrSignExtend(SS->getOperand(), Ty);
818 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
819 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
820 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
822 // trunc(x1+x2+...+xN) --> trunc(x1)+trunc(x2)+...+trunc(xN) if we can
823 // eliminate all the truncates.
824 if (const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Op)) {
825 SmallVector<const SCEV *, 4> Operands;
826 bool hasTrunc = false;
827 for (unsigned i = 0, e = SA->getNumOperands(); i != e && !hasTrunc; ++i) {
828 const SCEV *S = getTruncateExpr(SA->getOperand(i), Ty);
829 hasTrunc = isa<SCEVTruncateExpr>(S);
830 Operands.push_back(S);
833 return getAddExpr(Operands, false, false);
836 // trunc(x1*x2*...*xN) --> trunc(x1)*trunc(x2)*...*trunc(xN) if we can
837 // eliminate all the truncates.
838 if (const SCEVMulExpr *SM = dyn_cast<SCEVMulExpr>(Op)) {
839 SmallVector<const SCEV *, 4> Operands;
840 bool hasTrunc = false;
841 for (unsigned i = 0, e = SM->getNumOperands(); i != e && !hasTrunc; ++i) {
842 const SCEV *S = getTruncateExpr(SM->getOperand(i), Ty);
843 hasTrunc = isa<SCEVTruncateExpr>(S);
844 Operands.push_back(S);
847 return getMulExpr(Operands, false, false);
850 // If the input value is a chrec scev, truncate the chrec's operands.
851 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
852 SmallVector<const SCEV *, 4> Operands;
853 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
854 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
855 return getAddRecExpr(Operands, AddRec->getLoop());
858 // As a special case, fold trunc(undef) to undef. We don't want to
859 // know too much about SCEVUnknowns, but this special case is handy
861 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
862 if (isa<UndefValue>(U->getValue()))
863 return getSCEV(UndefValue::get(Ty));
865 // The cast wasn't folded; create an explicit cast node. We can reuse
866 // the existing insert position since if we get here, we won't have
867 // made any changes which would invalidate it.
868 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
870 UniqueSCEVs.InsertNode(S, IP);
874 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
876 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
877 "This is not an extending conversion!");
878 assert(isSCEVable(Ty) &&
879 "This is not a conversion to a SCEVable type!");
880 Ty = getEffectiveSCEVType(Ty);
882 // Fold if the operand is constant.
883 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
885 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(),
886 getEffectiveSCEVType(Ty))));
888 // zext(zext(x)) --> zext(x)
889 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
890 return getZeroExtendExpr(SZ->getOperand(), Ty);
892 // Before doing any expensive analysis, check to see if we've already
893 // computed a SCEV for this Op and Ty.
895 ID.AddInteger(scZeroExtend);
899 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
901 // If the input value is a chrec scev, and we can prove that the value
902 // did not overflow the old, smaller, value, we can zero extend all of the
903 // operands (often constants). This allows analysis of something like
904 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
905 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
906 if (AR->isAffine()) {
907 const SCEV *Start = AR->getStart();
908 const SCEV *Step = AR->getStepRecurrence(*this);
909 unsigned BitWidth = getTypeSizeInBits(AR->getType());
910 const Loop *L = AR->getLoop();
912 // If we have special knowledge that this addrec won't overflow,
913 // we don't need to do any further analysis.
914 if (AR->hasNoUnsignedWrap())
915 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
916 getZeroExtendExpr(Step, Ty),
919 // Check whether the backedge-taken count is SCEVCouldNotCompute.
920 // Note that this serves two purposes: It filters out loops that are
921 // simply not analyzable, and it covers the case where this code is
922 // being called from within backedge-taken count analysis, such that
923 // attempting to ask for the backedge-taken count would likely result
924 // in infinite recursion. In the later case, the analysis code will
925 // cope with a conservative value, and it will take care to purge
926 // that value once it has finished.
927 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
928 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
929 // Manually compute the final value for AR, checking for
932 // Check whether the backedge-taken count can be losslessly casted to
933 // the addrec's type. The count is always unsigned.
934 const SCEV *CastedMaxBECount =
935 getTruncateOrZeroExtend(MaxBECount, Start->getType());
936 const SCEV *RecastedMaxBECount =
937 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
938 if (MaxBECount == RecastedMaxBECount) {
939 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
940 // Check whether Start+Step*MaxBECount has no unsigned overflow.
941 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
942 const SCEV *Add = getAddExpr(Start, ZMul);
943 const SCEV *OperandExtendedAdd =
944 getAddExpr(getZeroExtendExpr(Start, WideTy),
945 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
946 getZeroExtendExpr(Step, WideTy)));
947 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
948 // Return the expression with the addrec on the outside.
949 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
950 getZeroExtendExpr(Step, Ty),
953 // Similar to above, only this time treat the step value as signed.
954 // This covers loops that count down.
955 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
956 Add = getAddExpr(Start, SMul);
958 getAddExpr(getZeroExtendExpr(Start, WideTy),
959 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
960 getSignExtendExpr(Step, WideTy)));
961 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
962 // Return the expression with the addrec on the outside.
963 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
964 getSignExtendExpr(Step, Ty),
968 // If the backedge is guarded by a comparison with the pre-inc value
969 // the addrec is safe. Also, if the entry is guarded by a comparison
970 // with the start value and the backedge is guarded by a comparison
971 // with the post-inc value, the addrec is safe.
972 if (isKnownPositive(Step)) {
973 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
974 getUnsignedRange(Step).getUnsignedMax());
975 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
976 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
977 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
978 AR->getPostIncExpr(*this), N)))
979 // Return the expression with the addrec on the outside.
980 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
981 getZeroExtendExpr(Step, Ty),
983 } else if (isKnownNegative(Step)) {
984 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
985 getSignedRange(Step).getSignedMin());
986 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
987 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
988 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
989 AR->getPostIncExpr(*this), N)))
990 // Return the expression with the addrec on the outside.
991 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
992 getSignExtendExpr(Step, Ty),
998 // The cast wasn't folded; create an explicit cast node.
999 // Recompute the insert position, as it may have been invalidated.
1000 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1001 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1003 UniqueSCEVs.InsertNode(S, IP);
1007 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1009 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1010 "This is not an extending conversion!");
1011 assert(isSCEVable(Ty) &&
1012 "This is not a conversion to a SCEVable type!");
1013 Ty = getEffectiveSCEVType(Ty);
1015 // Fold if the operand is constant.
1016 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1018 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(),
1019 getEffectiveSCEVType(Ty))));
1021 // sext(sext(x)) --> sext(x)
1022 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1023 return getSignExtendExpr(SS->getOperand(), Ty);
1025 // sext(zext(x)) --> zext(x)
1026 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
1027 return getZeroExtendExpr(SZ->getOperand(), Ty);
1029 // Before doing any expensive analysis, check to see if we've already
1030 // computed a SCEV for this Op and Ty.
1031 FoldingSetNodeID ID;
1032 ID.AddInteger(scSignExtend);
1036 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1038 // If the input value is provably positive, build a zext instead.
1039 if (isKnownNonNegative(Op))
1040 return getZeroExtendExpr(Op, Ty);
1042 // If the input value is a chrec scev, and we can prove that the value
1043 // did not overflow the old, smaller, value, we can sign extend all of the
1044 // operands (often constants). This allows analysis of something like
1045 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1046 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1047 if (AR->isAffine()) {
1048 const SCEV *Start = AR->getStart();
1049 const SCEV *Step = AR->getStepRecurrence(*this);
1050 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1051 const Loop *L = AR->getLoop();
1053 // If we have special knowledge that this addrec won't overflow,
1054 // we don't need to do any further analysis.
1055 if (AR->hasNoSignedWrap())
1056 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1057 getSignExtendExpr(Step, Ty),
1060 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1061 // Note that this serves two purposes: It filters out loops that are
1062 // simply not analyzable, and it covers the case where this code is
1063 // being called from within backedge-taken count analysis, such that
1064 // attempting to ask for the backedge-taken count would likely result
1065 // in infinite recursion. In the later case, the analysis code will
1066 // cope with a conservative value, and it will take care to purge
1067 // that value once it has finished.
1068 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1069 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1070 // Manually compute the final value for AR, checking for
1073 // Check whether the backedge-taken count can be losslessly casted to
1074 // the addrec's type. The count is always unsigned.
1075 const SCEV *CastedMaxBECount =
1076 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1077 const SCEV *RecastedMaxBECount =
1078 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1079 if (MaxBECount == RecastedMaxBECount) {
1080 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1081 // Check whether Start+Step*MaxBECount has no signed overflow.
1082 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1083 const SCEV *Add = getAddExpr(Start, SMul);
1084 const SCEV *OperandExtendedAdd =
1085 getAddExpr(getSignExtendExpr(Start, WideTy),
1086 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1087 getSignExtendExpr(Step, WideTy)));
1088 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1089 // Return the expression with the addrec on the outside.
1090 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1091 getSignExtendExpr(Step, Ty),
1094 // Similar to above, only this time treat the step value as unsigned.
1095 // This covers loops that count up with an unsigned step.
1096 const SCEV *UMul = getMulExpr(CastedMaxBECount, Step);
1097 Add = getAddExpr(Start, UMul);
1098 OperandExtendedAdd =
1099 getAddExpr(getSignExtendExpr(Start, WideTy),
1100 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1101 getZeroExtendExpr(Step, WideTy)));
1102 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1103 // Return the expression with the addrec on the outside.
1104 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1105 getZeroExtendExpr(Step, Ty),
1109 // If the backedge is guarded by a comparison with the pre-inc value
1110 // the addrec is safe. Also, if the entry is guarded by a comparison
1111 // with the start value and the backedge is guarded by a comparison
1112 // with the post-inc value, the addrec is safe.
1113 if (isKnownPositive(Step)) {
1114 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1115 getSignedRange(Step).getSignedMax());
1116 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1117 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1118 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1119 AR->getPostIncExpr(*this), N)))
1120 // Return the expression with the addrec on the outside.
1121 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1122 getSignExtendExpr(Step, Ty),
1124 } else if (isKnownNegative(Step)) {
1125 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1126 getSignedRange(Step).getSignedMin());
1127 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1128 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1129 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1130 AR->getPostIncExpr(*this), N)))
1131 // Return the expression with the addrec on the outside.
1132 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1133 getSignExtendExpr(Step, Ty),
1139 // The cast wasn't folded; create an explicit cast node.
1140 // Recompute the insert position, as it may have been invalidated.
1141 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1142 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1144 UniqueSCEVs.InsertNode(S, IP);
1148 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1149 /// unspecified bits out to the given type.
1151 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1153 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1154 "This is not an extending conversion!");
1155 assert(isSCEVable(Ty) &&
1156 "This is not a conversion to a SCEVable type!");
1157 Ty = getEffectiveSCEVType(Ty);
1159 // Sign-extend negative constants.
1160 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1161 if (SC->getValue()->getValue().isNegative())
1162 return getSignExtendExpr(Op, Ty);
1164 // Peel off a truncate cast.
1165 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1166 const SCEV *NewOp = T->getOperand();
1167 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1168 return getAnyExtendExpr(NewOp, Ty);
1169 return getTruncateOrNoop(NewOp, Ty);
1172 // Next try a zext cast. If the cast is folded, use it.
1173 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1174 if (!isa<SCEVZeroExtendExpr>(ZExt))
1177 // Next try a sext cast. If the cast is folded, use it.
1178 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1179 if (!isa<SCEVSignExtendExpr>(SExt))
1182 // Force the cast to be folded into the operands of an addrec.
1183 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1184 SmallVector<const SCEV *, 4> Ops;
1185 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1187 Ops.push_back(getAnyExtendExpr(*I, Ty));
1188 return getAddRecExpr(Ops, AR->getLoop());
1191 // As a special case, fold anyext(undef) to undef. We don't want to
1192 // know too much about SCEVUnknowns, but this special case is handy
1194 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
1195 if (isa<UndefValue>(U->getValue()))
1196 return getSCEV(UndefValue::get(Ty));
1198 // If the expression is obviously signed, use the sext cast value.
1199 if (isa<SCEVSMaxExpr>(Op))
1202 // Absent any other information, use the zext cast value.
1206 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1207 /// a list of operands to be added under the given scale, update the given
1208 /// map. This is a helper function for getAddRecExpr. As an example of
1209 /// what it does, given a sequence of operands that would form an add
1210 /// expression like this:
1212 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1214 /// where A and B are constants, update the map with these values:
1216 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1218 /// and add 13 + A*B*29 to AccumulatedConstant.
1219 /// This will allow getAddRecExpr to produce this:
1221 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1223 /// This form often exposes folding opportunities that are hidden in
1224 /// the original operand list.
1226 /// Return true iff it appears that any interesting folding opportunities
1227 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1228 /// the common case where no interesting opportunities are present, and
1229 /// is also used as a check to avoid infinite recursion.
1232 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1233 SmallVector<const SCEV *, 8> &NewOps,
1234 APInt &AccumulatedConstant,
1235 const SCEV *const *Ops, size_t NumOperands,
1237 ScalarEvolution &SE) {
1238 bool Interesting = false;
1240 // Iterate over the add operands. They are sorted, with constants first.
1242 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1244 // Pull a buried constant out to the outside.
1245 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1247 AccumulatedConstant += Scale * C->getValue()->getValue();
1250 // Next comes everything else. We're especially interested in multiplies
1251 // here, but they're in the middle, so just visit the rest with one loop.
1252 for (; i != NumOperands; ++i) {
1253 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1254 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1256 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1257 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1258 // A multiplication of a constant with another add; recurse.
1259 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1261 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1262 Add->op_begin(), Add->getNumOperands(),
1265 // A multiplication of a constant with some other value. Update
1267 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1268 const SCEV *Key = SE.getMulExpr(MulOps);
1269 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1270 M.insert(std::make_pair(Key, NewScale));
1272 NewOps.push_back(Pair.first->first);
1274 Pair.first->second += NewScale;
1275 // The map already had an entry for this value, which may indicate
1276 // a folding opportunity.
1281 // An ordinary operand. Update the map.
1282 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1283 M.insert(std::make_pair(Ops[i], Scale));
1285 NewOps.push_back(Pair.first->first);
1287 Pair.first->second += Scale;
1288 // The map already had an entry for this value, which may indicate
1289 // a folding opportunity.
1299 struct APIntCompare {
1300 bool operator()(const APInt &LHS, const APInt &RHS) const {
1301 return LHS.ult(RHS);
1306 /// getAddExpr - Get a canonical add expression, or something simpler if
1308 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1309 bool HasNUW, bool HasNSW) {
1310 assert(!Ops.empty() && "Cannot get empty add!");
1311 if (Ops.size() == 1) return Ops[0];
1313 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1314 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1315 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1316 "SCEVAddExpr operand types don't match!");
1319 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1320 if (!HasNUW && HasNSW) {
1322 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1323 E = Ops.end(); I != E; ++I)
1324 if (!isKnownNonNegative(*I)) {
1328 if (All) HasNUW = true;
1331 // Sort by complexity, this groups all similar expression types together.
1332 GroupByComplexity(Ops, LI);
1334 // If there are any constants, fold them together.
1336 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1338 assert(Idx < Ops.size());
1339 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1340 // We found two constants, fold them together!
1341 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1342 RHSC->getValue()->getValue());
1343 if (Ops.size() == 2) return Ops[0];
1344 Ops.erase(Ops.begin()+1); // Erase the folded element
1345 LHSC = cast<SCEVConstant>(Ops[0]);
1348 // If we are left with a constant zero being added, strip it off.
1349 if (LHSC->getValue()->isZero()) {
1350 Ops.erase(Ops.begin());
1354 if (Ops.size() == 1) return Ops[0];
1357 // Okay, check to see if the same value occurs in the operand list more than
1358 // once. If so, merge them together into an multiply expression. Since we
1359 // sorted the list, these values are required to be adjacent.
1360 const Type *Ty = Ops[0]->getType();
1361 bool FoundMatch = false;
1362 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1363 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1364 // Scan ahead to count how many equal operands there are.
1366 while (i+Count != e && Ops[i+Count] == Ops[i])
1368 // Merge the values into a multiply.
1369 const SCEV *Scale = getConstant(Ty, Count);
1370 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1371 if (Ops.size() == Count)
1374 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1375 --i; e -= Count - 1;
1379 return getAddExpr(Ops, HasNUW, HasNSW);
1381 // Check for truncates. If all the operands are truncated from the same
1382 // type, see if factoring out the truncate would permit the result to be
1383 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1384 // if the contents of the resulting outer trunc fold to something simple.
1385 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1386 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1387 const Type *DstType = Trunc->getType();
1388 const Type *SrcType = Trunc->getOperand()->getType();
1389 SmallVector<const SCEV *, 8> LargeOps;
1391 // Check all the operands to see if they can be represented in the
1392 // source type of the truncate.
1393 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1394 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1395 if (T->getOperand()->getType() != SrcType) {
1399 LargeOps.push_back(T->getOperand());
1400 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1401 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1402 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1403 SmallVector<const SCEV *, 8> LargeMulOps;
1404 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1405 if (const SCEVTruncateExpr *T =
1406 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1407 if (T->getOperand()->getType() != SrcType) {
1411 LargeMulOps.push_back(T->getOperand());
1412 } else if (const SCEVConstant *C =
1413 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1414 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1421 LargeOps.push_back(getMulExpr(LargeMulOps));
1428 // Evaluate the expression in the larger type.
1429 const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
1430 // If it folds to something simple, use it. Otherwise, don't.
1431 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1432 return getTruncateExpr(Fold, DstType);
1436 // Skip past any other cast SCEVs.
1437 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1440 // If there are add operands they would be next.
1441 if (Idx < Ops.size()) {
1442 bool DeletedAdd = false;
1443 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1444 // If we have an add, expand the add operands onto the end of the operands
1446 Ops.erase(Ops.begin()+Idx);
1447 Ops.append(Add->op_begin(), Add->op_end());
1451 // If we deleted at least one add, we added operands to the end of the list,
1452 // and they are not necessarily sorted. Recurse to resort and resimplify
1453 // any operands we just acquired.
1455 return getAddExpr(Ops);
1458 // Skip over the add expression until we get to a multiply.
1459 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1462 // Check to see if there are any folding opportunities present with
1463 // operands multiplied by constant values.
1464 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1465 uint64_t BitWidth = getTypeSizeInBits(Ty);
1466 DenseMap<const SCEV *, APInt> M;
1467 SmallVector<const SCEV *, 8> NewOps;
1468 APInt AccumulatedConstant(BitWidth, 0);
1469 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1470 Ops.data(), Ops.size(),
1471 APInt(BitWidth, 1), *this)) {
1472 // Some interesting folding opportunity is present, so its worthwhile to
1473 // re-generate the operands list. Group the operands by constant scale,
1474 // to avoid multiplying by the same constant scale multiple times.
1475 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1476 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1477 E = NewOps.end(); I != E; ++I)
1478 MulOpLists[M.find(*I)->second].push_back(*I);
1479 // Re-generate the operands list.
1481 if (AccumulatedConstant != 0)
1482 Ops.push_back(getConstant(AccumulatedConstant));
1483 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1484 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1486 Ops.push_back(getMulExpr(getConstant(I->first),
1487 getAddExpr(I->second)));
1489 return getConstant(Ty, 0);
1490 if (Ops.size() == 1)
1492 return getAddExpr(Ops);
1496 // If we are adding something to a multiply expression, make sure the
1497 // something is not already an operand of the multiply. If so, merge it into
1499 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1500 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1501 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1502 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1503 if (isa<SCEVConstant>(MulOpSCEV))
1505 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1506 if (MulOpSCEV == Ops[AddOp]) {
1507 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1508 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1509 if (Mul->getNumOperands() != 2) {
1510 // If the multiply has more than two operands, we must get the
1512 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1513 Mul->op_begin()+MulOp);
1514 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1515 InnerMul = getMulExpr(MulOps);
1517 const SCEV *One = getConstant(Ty, 1);
1518 const SCEV *AddOne = getAddExpr(One, InnerMul);
1519 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1520 if (Ops.size() == 2) return OuterMul;
1522 Ops.erase(Ops.begin()+AddOp);
1523 Ops.erase(Ops.begin()+Idx-1);
1525 Ops.erase(Ops.begin()+Idx);
1526 Ops.erase(Ops.begin()+AddOp-1);
1528 Ops.push_back(OuterMul);
1529 return getAddExpr(Ops);
1532 // Check this multiply against other multiplies being added together.
1533 for (unsigned OtherMulIdx = Idx+1;
1534 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1536 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1537 // If MulOp occurs in OtherMul, we can fold the two multiplies
1539 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1540 OMulOp != e; ++OMulOp)
1541 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1542 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1543 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1544 if (Mul->getNumOperands() != 2) {
1545 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1546 Mul->op_begin()+MulOp);
1547 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1548 InnerMul1 = getMulExpr(MulOps);
1550 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1551 if (OtherMul->getNumOperands() != 2) {
1552 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1553 OtherMul->op_begin()+OMulOp);
1554 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1555 InnerMul2 = getMulExpr(MulOps);
1557 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1558 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1559 if (Ops.size() == 2) return OuterMul;
1560 Ops.erase(Ops.begin()+Idx);
1561 Ops.erase(Ops.begin()+OtherMulIdx-1);
1562 Ops.push_back(OuterMul);
1563 return getAddExpr(Ops);
1569 // If there are any add recurrences in the operands list, see if any other
1570 // added values are loop invariant. If so, we can fold them into the
1572 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1575 // Scan over all recurrences, trying to fold loop invariants into them.
1576 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1577 // Scan all of the other operands to this add and add them to the vector if
1578 // they are loop invariant w.r.t. the recurrence.
1579 SmallVector<const SCEV *, 8> LIOps;
1580 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1581 const Loop *AddRecLoop = AddRec->getLoop();
1582 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1583 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1584 LIOps.push_back(Ops[i]);
1585 Ops.erase(Ops.begin()+i);
1589 // If we found some loop invariants, fold them into the recurrence.
1590 if (!LIOps.empty()) {
1591 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1592 LIOps.push_back(AddRec->getStart());
1594 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1596 AddRecOps[0] = getAddExpr(LIOps);
1598 // Build the new addrec. Propagate the NUW and NSW flags if both the
1599 // outer add and the inner addrec are guaranteed to have no overflow.
1600 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop,
1601 HasNUW && AddRec->hasNoUnsignedWrap(),
1602 HasNSW && AddRec->hasNoSignedWrap());
1604 // If all of the other operands were loop invariant, we are done.
1605 if (Ops.size() == 1) return NewRec;
1607 // Otherwise, add the folded AddRec by the non-liv parts.
1608 for (unsigned i = 0;; ++i)
1609 if (Ops[i] == AddRec) {
1613 return getAddExpr(Ops);
1616 // Okay, if there weren't any loop invariants to be folded, check to see if
1617 // there are multiple AddRec's with the same loop induction variable being
1618 // added together. If so, we can fold them.
1619 for (unsigned OtherIdx = Idx+1;
1620 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1622 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1623 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1624 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1626 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1628 if (const SCEVAddRecExpr *OtherAddRec =
1629 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1630 if (OtherAddRec->getLoop() == AddRecLoop) {
1631 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1633 if (i >= AddRecOps.size()) {
1634 AddRecOps.append(OtherAddRec->op_begin()+i,
1635 OtherAddRec->op_end());
1638 AddRecOps[i] = getAddExpr(AddRecOps[i],
1639 OtherAddRec->getOperand(i));
1641 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1643 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop);
1644 return getAddExpr(Ops);
1647 // Otherwise couldn't fold anything into this recurrence. Move onto the
1651 // Okay, it looks like we really DO need an add expr. Check to see if we
1652 // already have one, otherwise create a new one.
1653 FoldingSetNodeID ID;
1654 ID.AddInteger(scAddExpr);
1655 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1656 ID.AddPointer(Ops[i]);
1659 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1661 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1662 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1663 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1665 UniqueSCEVs.InsertNode(S, IP);
1667 if (HasNUW) S->setHasNoUnsignedWrap(true);
1668 if (HasNSW) S->setHasNoSignedWrap(true);
1672 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1674 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1675 bool HasNUW, bool HasNSW) {
1676 assert(!Ops.empty() && "Cannot get empty mul!");
1677 if (Ops.size() == 1) return Ops[0];
1679 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1680 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1681 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1682 "SCEVMulExpr operand types don't match!");
1685 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1686 if (!HasNUW && HasNSW) {
1688 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1689 E = Ops.end(); I != E; ++I)
1690 if (!isKnownNonNegative(*I)) {
1694 if (All) HasNUW = true;
1697 // Sort by complexity, this groups all similar expression types together.
1698 GroupByComplexity(Ops, LI);
1700 // If there are any constants, fold them together.
1702 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1704 // C1*(C2+V) -> C1*C2 + C1*V
1705 if (Ops.size() == 2)
1706 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1707 if (Add->getNumOperands() == 2 &&
1708 isa<SCEVConstant>(Add->getOperand(0)))
1709 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1710 getMulExpr(LHSC, Add->getOperand(1)));
1713 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1714 // We found two constants, fold them together!
1715 ConstantInt *Fold = ConstantInt::get(getContext(),
1716 LHSC->getValue()->getValue() *
1717 RHSC->getValue()->getValue());
1718 Ops[0] = getConstant(Fold);
1719 Ops.erase(Ops.begin()+1); // Erase the folded element
1720 if (Ops.size() == 1) return Ops[0];
1721 LHSC = cast<SCEVConstant>(Ops[0]);
1724 // If we are left with a constant one being multiplied, strip it off.
1725 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1726 Ops.erase(Ops.begin());
1728 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1729 // If we have a multiply of zero, it will always be zero.
1731 } else if (Ops[0]->isAllOnesValue()) {
1732 // If we have a mul by -1 of an add, try distributing the -1 among the
1734 if (Ops.size() == 2)
1735 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1736 SmallVector<const SCEV *, 4> NewOps;
1737 bool AnyFolded = false;
1738 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1740 const SCEV *Mul = getMulExpr(Ops[0], *I);
1741 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1742 NewOps.push_back(Mul);
1745 return getAddExpr(NewOps);
1749 if (Ops.size() == 1)
1753 // Skip over the add expression until we get to a multiply.
1754 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1757 // If there are mul operands inline them all into this expression.
1758 if (Idx < Ops.size()) {
1759 bool DeletedMul = false;
1760 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1761 // If we have an mul, expand the mul operands onto the end of the operands
1763 Ops.erase(Ops.begin()+Idx);
1764 Ops.append(Mul->op_begin(), Mul->op_end());
1768 // If we deleted at least one mul, we added operands to the end of the list,
1769 // and they are not necessarily sorted. Recurse to resort and resimplify
1770 // any operands we just acquired.
1772 return getMulExpr(Ops);
1775 // If there are any add recurrences in the operands list, see if any other
1776 // added values are loop invariant. If so, we can fold them into the
1778 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1781 // Scan over all recurrences, trying to fold loop invariants into them.
1782 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1783 // Scan all of the other operands to this mul and add them to the vector if
1784 // they are loop invariant w.r.t. the recurrence.
1785 SmallVector<const SCEV *, 8> LIOps;
1786 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1787 const Loop *AddRecLoop = AddRec->getLoop();
1788 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1789 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1790 LIOps.push_back(Ops[i]);
1791 Ops.erase(Ops.begin()+i);
1795 // If we found some loop invariants, fold them into the recurrence.
1796 if (!LIOps.empty()) {
1797 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1798 SmallVector<const SCEV *, 4> NewOps;
1799 NewOps.reserve(AddRec->getNumOperands());
1800 const SCEV *Scale = getMulExpr(LIOps);
1801 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1802 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1804 // Build the new addrec. Propagate the NUW and NSW flags if both the
1805 // outer mul and the inner addrec are guaranteed to have no overflow.
1806 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop,
1807 HasNUW && AddRec->hasNoUnsignedWrap(),
1808 HasNSW && AddRec->hasNoSignedWrap());
1810 // If all of the other operands were loop invariant, we are done.
1811 if (Ops.size() == 1) return NewRec;
1813 // Otherwise, multiply the folded AddRec by the non-liv parts.
1814 for (unsigned i = 0;; ++i)
1815 if (Ops[i] == AddRec) {
1819 return getMulExpr(Ops);
1822 // Okay, if there weren't any loop invariants to be folded, check to see if
1823 // there are multiple AddRec's with the same loop induction variable being
1824 // multiplied together. If so, we can fold them.
1825 for (unsigned OtherIdx = Idx+1;
1826 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1828 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1829 // F * G, where F = {A,+,B}<L> and G = {C,+,D}<L> -->
1830 // {A*C,+,F*D + G*B + B*D}<L>
1831 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1833 if (const SCEVAddRecExpr *OtherAddRec =
1834 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1835 if (OtherAddRec->getLoop() == AddRecLoop) {
1836 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1837 const SCEV *NewStart = getMulExpr(F->getStart(), G->getStart());
1838 const SCEV *B = F->getStepRecurrence(*this);
1839 const SCEV *D = G->getStepRecurrence(*this);
1840 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1843 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1845 if (Ops.size() == 2) return NewAddRec;
1846 Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
1847 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1849 return getMulExpr(Ops);
1852 // Otherwise couldn't fold anything into this recurrence. Move onto the
1856 // Okay, it looks like we really DO need an mul expr. Check to see if we
1857 // already have one, otherwise create a new one.
1858 FoldingSetNodeID ID;
1859 ID.AddInteger(scMulExpr);
1860 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1861 ID.AddPointer(Ops[i]);
1864 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1866 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1867 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1868 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
1870 UniqueSCEVs.InsertNode(S, IP);
1872 if (HasNUW) S->setHasNoUnsignedWrap(true);
1873 if (HasNSW) S->setHasNoSignedWrap(true);
1877 /// getUDivExpr - Get a canonical unsigned division expression, or something
1878 /// simpler if possible.
1879 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1881 assert(getEffectiveSCEVType(LHS->getType()) ==
1882 getEffectiveSCEVType(RHS->getType()) &&
1883 "SCEVUDivExpr operand types don't match!");
1885 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1886 if (RHSC->getValue()->equalsInt(1))
1887 return LHS; // X udiv 1 --> x
1888 // If the denominator is zero, the result of the udiv is undefined. Don't
1889 // try to analyze it, because the resolution chosen here may differ from
1890 // the resolution chosen in other parts of the compiler.
1891 if (!RHSC->getValue()->isZero()) {
1892 // Determine if the division can be folded into the operands of
1894 // TODO: Generalize this to non-constants by using known-bits information.
1895 const Type *Ty = LHS->getType();
1896 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1897 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
1898 // For non-power-of-two values, effectively round the value up to the
1899 // nearest power of two.
1900 if (!RHSC->getValue()->getValue().isPowerOf2())
1902 const IntegerType *ExtTy =
1903 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1904 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1905 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1906 if (const SCEVConstant *Step =
1907 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1908 if (!Step->getValue()->getValue()
1909 .urem(RHSC->getValue()->getValue()) &&
1910 getZeroExtendExpr(AR, ExtTy) ==
1911 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1912 getZeroExtendExpr(Step, ExtTy),
1914 SmallVector<const SCEV *, 4> Operands;
1915 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1916 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1917 return getAddRecExpr(Operands, AR->getLoop());
1919 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1920 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1921 SmallVector<const SCEV *, 4> Operands;
1922 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1923 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1924 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1925 // Find an operand that's safely divisible.
1926 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1927 const SCEV *Op = M->getOperand(i);
1928 const SCEV *Div = getUDivExpr(Op, RHSC);
1929 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1930 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
1933 return getMulExpr(Operands);
1937 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1938 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1939 SmallVector<const SCEV *, 4> Operands;
1940 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1941 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1942 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1944 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1945 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
1946 if (isa<SCEVUDivExpr>(Op) ||
1947 getMulExpr(Op, RHS) != A->getOperand(i))
1949 Operands.push_back(Op);
1951 if (Operands.size() == A->getNumOperands())
1952 return getAddExpr(Operands);
1956 // Fold if both operands are constant.
1957 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1958 Constant *LHSCV = LHSC->getValue();
1959 Constant *RHSCV = RHSC->getValue();
1960 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
1966 FoldingSetNodeID ID;
1967 ID.AddInteger(scUDivExpr);
1971 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1972 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
1974 UniqueSCEVs.InsertNode(S, IP);
1979 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1980 /// Simplify the expression as much as possible.
1981 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
1982 const SCEV *Step, const Loop *L,
1983 bool HasNUW, bool HasNSW) {
1984 SmallVector<const SCEV *, 4> Operands;
1985 Operands.push_back(Start);
1986 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1987 if (StepChrec->getLoop() == L) {
1988 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
1989 return getAddRecExpr(Operands, L);
1992 Operands.push_back(Step);
1993 return getAddRecExpr(Operands, L, HasNUW, HasNSW);
1996 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1997 /// Simplify the expression as much as possible.
1999 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2001 bool HasNUW, bool HasNSW) {
2002 if (Operands.size() == 1) return Operands[0];
2004 const Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2005 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2006 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2007 "SCEVAddRecExpr operand types don't match!");
2008 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2009 assert(isLoopInvariant(Operands[i], L) &&
2010 "SCEVAddRecExpr operand is not loop-invariant!");
2013 if (Operands.back()->isZero()) {
2014 Operands.pop_back();
2015 return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
2018 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2019 // use that information to infer NUW and NSW flags. However, computing a
2020 // BE count requires calling getAddRecExpr, so we may not yet have a
2021 // meaningful BE count at this point (and if we don't, we'd be stuck
2022 // with a SCEVCouldNotCompute as the cached BE count).
2024 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
2025 if (!HasNUW && HasNSW) {
2027 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2028 E = Operands.end(); I != E; ++I)
2029 if (!isKnownNonNegative(*I)) {
2033 if (All) HasNUW = true;
2036 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2037 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2038 const Loop *NestedLoop = NestedAR->getLoop();
2039 if (L->contains(NestedLoop) ?
2040 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2041 (!NestedLoop->contains(L) &&
2042 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2043 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2044 NestedAR->op_end());
2045 Operands[0] = NestedAR->getStart();
2046 // AddRecs require their operands be loop-invariant with respect to their
2047 // loops. Don't perform this transformation if it would break this
2049 bool AllInvariant = true;
2050 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2051 if (!isLoopInvariant(Operands[i], L)) {
2052 AllInvariant = false;
2056 NestedOperands[0] = getAddRecExpr(Operands, L);
2057 AllInvariant = true;
2058 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2059 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2060 AllInvariant = false;
2064 // Ok, both add recurrences are valid after the transformation.
2065 return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
2067 // Reset Operands to its original state.
2068 Operands[0] = NestedAR;
2072 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2073 // already have one, otherwise create a new one.
2074 FoldingSetNodeID ID;
2075 ID.AddInteger(scAddRecExpr);
2076 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2077 ID.AddPointer(Operands[i]);
2081 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2083 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2084 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2085 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2086 O, Operands.size(), L);
2087 UniqueSCEVs.InsertNode(S, IP);
2089 if (HasNUW) S->setHasNoUnsignedWrap(true);
2090 if (HasNSW) S->setHasNoSignedWrap(true);
2094 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2096 SmallVector<const SCEV *, 2> Ops;
2099 return getSMaxExpr(Ops);
2103 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2104 assert(!Ops.empty() && "Cannot get empty smax!");
2105 if (Ops.size() == 1) return Ops[0];
2107 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2108 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2109 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2110 "SCEVSMaxExpr operand types don't match!");
2113 // Sort by complexity, this groups all similar expression types together.
2114 GroupByComplexity(Ops, LI);
2116 // If there are any constants, fold them together.
2118 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2120 assert(Idx < Ops.size());
2121 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2122 // We found two constants, fold them together!
2123 ConstantInt *Fold = ConstantInt::get(getContext(),
2124 APIntOps::smax(LHSC->getValue()->getValue(),
2125 RHSC->getValue()->getValue()));
2126 Ops[0] = getConstant(Fold);
2127 Ops.erase(Ops.begin()+1); // Erase the folded element
2128 if (Ops.size() == 1) return Ops[0];
2129 LHSC = cast<SCEVConstant>(Ops[0]);
2132 // If we are left with a constant minimum-int, strip it off.
2133 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2134 Ops.erase(Ops.begin());
2136 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2137 // If we have an smax with a constant maximum-int, it will always be
2142 if (Ops.size() == 1) return Ops[0];
2145 // Find the first SMax
2146 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2149 // Check to see if one of the operands is an SMax. If so, expand its operands
2150 // onto our operand list, and recurse to simplify.
2151 if (Idx < Ops.size()) {
2152 bool DeletedSMax = false;
2153 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2154 Ops.erase(Ops.begin()+Idx);
2155 Ops.append(SMax->op_begin(), SMax->op_end());
2160 return getSMaxExpr(Ops);
2163 // Okay, check to see if the same value occurs in the operand list twice. If
2164 // so, delete one. Since we sorted the list, these values are required to
2166 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2167 // X smax Y smax Y --> X smax Y
2168 // X smax Y --> X, if X is always greater than Y
2169 if (Ops[i] == Ops[i+1] ||
2170 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2171 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2173 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2174 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2178 if (Ops.size() == 1) return Ops[0];
2180 assert(!Ops.empty() && "Reduced smax down to nothing!");
2182 // Okay, it looks like we really DO need an smax expr. Check to see if we
2183 // already have one, otherwise create a new one.
2184 FoldingSetNodeID ID;
2185 ID.AddInteger(scSMaxExpr);
2186 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2187 ID.AddPointer(Ops[i]);
2189 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2190 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2191 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2192 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2194 UniqueSCEVs.InsertNode(S, IP);
2198 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2200 SmallVector<const SCEV *, 2> Ops;
2203 return getUMaxExpr(Ops);
2207 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2208 assert(!Ops.empty() && "Cannot get empty umax!");
2209 if (Ops.size() == 1) return Ops[0];
2211 const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2212 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2213 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2214 "SCEVUMaxExpr operand types don't match!");
2217 // Sort by complexity, this groups all similar expression types together.
2218 GroupByComplexity(Ops, LI);
2220 // If there are any constants, fold them together.
2222 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2224 assert(Idx < Ops.size());
2225 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2226 // We found two constants, fold them together!
2227 ConstantInt *Fold = ConstantInt::get(getContext(),
2228 APIntOps::umax(LHSC->getValue()->getValue(),
2229 RHSC->getValue()->getValue()));
2230 Ops[0] = getConstant(Fold);
2231 Ops.erase(Ops.begin()+1); // Erase the folded element
2232 if (Ops.size() == 1) return Ops[0];
2233 LHSC = cast<SCEVConstant>(Ops[0]);
2236 // If we are left with a constant minimum-int, strip it off.
2237 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2238 Ops.erase(Ops.begin());
2240 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2241 // If we have an umax with a constant maximum-int, it will always be
2246 if (Ops.size() == 1) return Ops[0];
2249 // Find the first UMax
2250 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2253 // Check to see if one of the operands is a UMax. If so, expand its operands
2254 // onto our operand list, and recurse to simplify.
2255 if (Idx < Ops.size()) {
2256 bool DeletedUMax = false;
2257 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2258 Ops.erase(Ops.begin()+Idx);
2259 Ops.append(UMax->op_begin(), UMax->op_end());
2264 return getUMaxExpr(Ops);
2267 // Okay, check to see if the same value occurs in the operand list twice. If
2268 // so, delete one. Since we sorted the list, these values are required to
2270 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2271 // X umax Y umax Y --> X umax Y
2272 // X umax Y --> X, if X is always greater than Y
2273 if (Ops[i] == Ops[i+1] ||
2274 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2275 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2277 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2278 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2282 if (Ops.size() == 1) return Ops[0];
2284 assert(!Ops.empty() && "Reduced umax down to nothing!");
2286 // Okay, it looks like we really DO need a umax expr. Check to see if we
2287 // already have one, otherwise create a new one.
2288 FoldingSetNodeID ID;
2289 ID.AddInteger(scUMaxExpr);
2290 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2291 ID.AddPointer(Ops[i]);
2293 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2294 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2295 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2296 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2298 UniqueSCEVs.InsertNode(S, IP);
2302 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2304 // ~smax(~x, ~y) == smin(x, y).
2305 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2308 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2310 // ~umax(~x, ~y) == umin(x, y)
2311 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2314 const SCEV *ScalarEvolution::getSizeOfExpr(const Type *AllocTy) {
2315 // If we have TargetData, we can bypass creating a target-independent
2316 // constant expression and then folding it back into a ConstantInt.
2317 // This is just a compile-time optimization.
2319 return getConstant(TD->getIntPtrType(getContext()),
2320 TD->getTypeAllocSize(AllocTy));
2322 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2323 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2324 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2326 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2327 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2330 const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
2331 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2332 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2333 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2335 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2336 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2339 const SCEV *ScalarEvolution::getOffsetOfExpr(const StructType *STy,
2341 // If we have TargetData, we can bypass creating a target-independent
2342 // constant expression and then folding it back into a ConstantInt.
2343 // This is just a compile-time optimization.
2345 return getConstant(TD->getIntPtrType(getContext()),
2346 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2348 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2349 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2350 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2352 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2353 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2356 const SCEV *ScalarEvolution::getOffsetOfExpr(const Type *CTy,
2357 Constant *FieldNo) {
2358 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2359 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2360 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2362 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2363 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2366 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2367 // Don't attempt to do anything other than create a SCEVUnknown object
2368 // here. createSCEV only calls getUnknown after checking for all other
2369 // interesting possibilities, and any other code that calls getUnknown
2370 // is doing so in order to hide a value from SCEV canonicalization.
2372 FoldingSetNodeID ID;
2373 ID.AddInteger(scUnknown);
2376 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2377 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2378 "Stale SCEVUnknown in uniquing map!");
2381 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2383 FirstUnknown = cast<SCEVUnknown>(S);
2384 UniqueSCEVs.InsertNode(S, IP);
2388 //===----------------------------------------------------------------------===//
2389 // Basic SCEV Analysis and PHI Idiom Recognition Code
2392 /// isSCEVable - Test if values of the given type are analyzable within
2393 /// the SCEV framework. This primarily includes integer types, and it
2394 /// can optionally include pointer types if the ScalarEvolution class
2395 /// has access to target-specific information.
2396 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2397 // Integers and pointers are always SCEVable.
2398 return Ty->isIntegerTy() || Ty->isPointerTy();
2401 /// getTypeSizeInBits - Return the size in bits of the specified type,
2402 /// for which isSCEVable must return true.
2403 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2404 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2406 // If we have a TargetData, use it!
2408 return TD->getTypeSizeInBits(Ty);
2410 // Integer types have fixed sizes.
2411 if (Ty->isIntegerTy())
2412 return Ty->getPrimitiveSizeInBits();
2414 // The only other support type is pointer. Without TargetData, conservatively
2415 // assume pointers are 64-bit.
2416 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2420 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2421 /// the given type and which represents how SCEV will treat the given
2422 /// type, for which isSCEVable must return true. For pointer types,
2423 /// this is the pointer-sized integer type.
2424 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2425 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2427 if (Ty->isIntegerTy())
2430 // The only other support type is pointer.
2431 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2432 if (TD) return TD->getIntPtrType(getContext());
2434 // Without TargetData, conservatively assume pointers are 64-bit.
2435 return Type::getInt64Ty(getContext());
2438 const SCEV *ScalarEvolution::getCouldNotCompute() {
2439 return &CouldNotCompute;
2442 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2443 /// expression and create a new one.
2444 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2445 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2447 ValueExprMapType::const_iterator I = ValueExprMap.find(V);
2448 if (I != ValueExprMap.end()) return I->second;
2449 const SCEV *S = createSCEV(V);
2451 // The process of creating a SCEV for V may have caused other SCEVs
2452 // to have been created, so it's necessary to insert the new entry
2453 // from scratch, rather than trying to remember the insert position
2455 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2459 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2461 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2462 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2464 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2466 const Type *Ty = V->getType();
2467 Ty = getEffectiveSCEVType(Ty);
2468 return getMulExpr(V,
2469 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2472 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2473 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2474 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2476 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2478 const Type *Ty = V->getType();
2479 Ty = getEffectiveSCEVType(Ty);
2480 const SCEV *AllOnes =
2481 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2482 return getMinusSCEV(AllOnes, V);
2485 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1,
2486 /// and thus the HasNUW and HasNSW bits apply to the resultant add, not
2487 /// whether the sub would have overflowed.
2488 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2489 bool HasNUW, bool HasNSW) {
2490 // Fast path: X - X --> 0.
2492 return getConstant(LHS->getType(), 0);
2495 return getAddExpr(LHS, getNegativeSCEV(RHS), HasNUW, HasNSW);
2498 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2499 /// input value to the specified type. If the type must be extended, it is zero
2502 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, const Type *Ty) {
2503 const Type *SrcTy = V->getType();
2504 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2505 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2506 "Cannot truncate or zero extend with non-integer arguments!");
2507 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2508 return V; // No conversion
2509 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2510 return getTruncateExpr(V, Ty);
2511 return getZeroExtendExpr(V, Ty);
2514 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2515 /// input value to the specified type. If the type must be extended, it is sign
2518 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2520 const Type *SrcTy = V->getType();
2521 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2522 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2523 "Cannot truncate or zero extend with non-integer arguments!");
2524 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2525 return V; // No conversion
2526 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2527 return getTruncateExpr(V, Ty);
2528 return getSignExtendExpr(V, Ty);
2531 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2532 /// input value to the specified type. If the type must be extended, it is zero
2533 /// extended. The conversion must not be narrowing.
2535 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2536 const Type *SrcTy = V->getType();
2537 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2538 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2539 "Cannot noop or zero extend with non-integer arguments!");
2540 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2541 "getNoopOrZeroExtend cannot truncate!");
2542 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2543 return V; // No conversion
2544 return getZeroExtendExpr(V, Ty);
2547 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2548 /// input value to the specified type. If the type must be extended, it is sign
2549 /// extended. The conversion must not be narrowing.
2551 ScalarEvolution::getNoopOrSignExtend(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 noop or sign extend with non-integer arguments!");
2556 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2557 "getNoopOrSignExtend cannot truncate!");
2558 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2559 return V; // No conversion
2560 return getSignExtendExpr(V, Ty);
2563 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2564 /// the input value to the specified type. If the type must be extended,
2565 /// it is extended with unspecified bits. The conversion must not be
2568 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2569 const Type *SrcTy = V->getType();
2570 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2571 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2572 "Cannot noop or any extend with non-integer arguments!");
2573 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2574 "getNoopOrAnyExtend cannot truncate!");
2575 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2576 return V; // No conversion
2577 return getAnyExtendExpr(V, Ty);
2580 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2581 /// input value to the specified type. The conversion must not be widening.
2583 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2584 const Type *SrcTy = V->getType();
2585 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2586 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2587 "Cannot truncate or noop with non-integer arguments!");
2588 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2589 "getTruncateOrNoop cannot extend!");
2590 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2591 return V; // No conversion
2592 return getTruncateExpr(V, Ty);
2595 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2596 /// the types using zero-extension, and then perform a umax operation
2598 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2600 const SCEV *PromotedLHS = LHS;
2601 const SCEV *PromotedRHS = RHS;
2603 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2604 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2606 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2608 return getUMaxExpr(PromotedLHS, PromotedRHS);
2611 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2612 /// the types using zero-extension, and then perform a umin operation
2614 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2616 const SCEV *PromotedLHS = LHS;
2617 const SCEV *PromotedRHS = RHS;
2619 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2620 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2622 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2624 return getUMinExpr(PromotedLHS, PromotedRHS);
2627 /// PushDefUseChildren - Push users of the given Instruction
2628 /// onto the given Worklist.
2630 PushDefUseChildren(Instruction *I,
2631 SmallVectorImpl<Instruction *> &Worklist) {
2632 // Push the def-use children onto the Worklist stack.
2633 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2635 Worklist.push_back(cast<Instruction>(*UI));
2638 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2639 /// instructions that depend on the given instruction and removes them from
2640 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2643 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2644 SmallVector<Instruction *, 16> Worklist;
2645 PushDefUseChildren(PN, Worklist);
2647 SmallPtrSet<Instruction *, 8> Visited;
2649 while (!Worklist.empty()) {
2650 Instruction *I = Worklist.pop_back_val();
2651 if (!Visited.insert(I)) continue;
2653 ValueExprMapType::iterator It =
2654 ValueExprMap.find(static_cast<Value *>(I));
2655 if (It != ValueExprMap.end()) {
2656 const SCEV *Old = It->second;
2658 // Short-circuit the def-use traversal if the symbolic name
2659 // ceases to appear in expressions.
2660 if (Old != SymName && !hasOperand(Old, SymName))
2663 // SCEVUnknown for a PHI either means that it has an unrecognized
2664 // structure, it's a PHI that's in the progress of being computed
2665 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2666 // additional loop trip count information isn't going to change anything.
2667 // In the second case, createNodeForPHI will perform the necessary
2668 // updates on its own when it gets to that point. In the third, we do
2669 // want to forget the SCEVUnknown.
2670 if (!isa<PHINode>(I) ||
2671 !isa<SCEVUnknown>(Old) ||
2672 (I != PN && Old == SymName)) {
2673 forgetMemoizedResults(Old);
2674 ValueExprMap.erase(It);
2678 PushDefUseChildren(I, Worklist);
2682 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2683 /// a loop header, making it a potential recurrence, or it doesn't.
2685 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2686 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2687 if (L->getHeader() == PN->getParent()) {
2688 // The loop may have multiple entrances or multiple exits; we can analyze
2689 // this phi as an addrec if it has a unique entry value and a unique
2691 Value *BEValueV = 0, *StartValueV = 0;
2692 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2693 Value *V = PN->getIncomingValue(i);
2694 if (L->contains(PN->getIncomingBlock(i))) {
2697 } else if (BEValueV != V) {
2701 } else if (!StartValueV) {
2703 } else if (StartValueV != V) {
2708 if (BEValueV && StartValueV) {
2709 // While we are analyzing this PHI node, handle its value symbolically.
2710 const SCEV *SymbolicName = getUnknown(PN);
2711 assert(ValueExprMap.find(PN) == ValueExprMap.end() &&
2712 "PHI node already processed?");
2713 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2715 // Using this symbolic name for the PHI, analyze the value coming around
2717 const SCEV *BEValue = getSCEV(BEValueV);
2719 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2720 // has a special value for the first iteration of the loop.
2722 // If the value coming around the backedge is an add with the symbolic
2723 // value we just inserted, then we found a simple induction variable!
2724 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2725 // If there is a single occurrence of the symbolic value, replace it
2726 // with a recurrence.
2727 unsigned FoundIndex = Add->getNumOperands();
2728 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2729 if (Add->getOperand(i) == SymbolicName)
2730 if (FoundIndex == e) {
2735 if (FoundIndex != Add->getNumOperands()) {
2736 // Create an add with everything but the specified operand.
2737 SmallVector<const SCEV *, 8> Ops;
2738 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2739 if (i != FoundIndex)
2740 Ops.push_back(Add->getOperand(i));
2741 const SCEV *Accum = getAddExpr(Ops);
2743 // This is not a valid addrec if the step amount is varying each
2744 // loop iteration, but is not itself an addrec in this loop.
2745 if (isLoopInvariant(Accum, L) ||
2746 (isa<SCEVAddRecExpr>(Accum) &&
2747 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2748 bool HasNUW = false;
2749 bool HasNSW = false;
2751 // If the increment doesn't overflow, then neither the addrec nor
2752 // the post-increment will overflow.
2753 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
2754 if (OBO->hasNoUnsignedWrap())
2756 if (OBO->hasNoSignedWrap())
2758 } else if (const GEPOperator *GEP =
2759 dyn_cast<GEPOperator>(BEValueV)) {
2760 // If the increment is a GEP, then we know it won't perform an
2761 // unsigned overflow, because the address space cannot be
2763 HasNUW |= GEP->isInBounds();
2766 const SCEV *StartVal = getSCEV(StartValueV);
2767 const SCEV *PHISCEV =
2768 getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
2770 // Since the no-wrap flags are on the increment, they apply to the
2771 // post-incremented value as well.
2772 if (isLoopInvariant(Accum, L))
2773 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
2774 Accum, L, HasNUW, HasNSW);
2776 // Okay, for the entire analysis of this edge we assumed the PHI
2777 // to be symbolic. We now need to go back and purge all of the
2778 // entries for the scalars that use the symbolic expression.
2779 ForgetSymbolicName(PN, SymbolicName);
2780 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2784 } else if (const SCEVAddRecExpr *AddRec =
2785 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2786 // Otherwise, this could be a loop like this:
2787 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2788 // In this case, j = {1,+,1} and BEValue is j.
2789 // Because the other in-value of i (0) fits the evolution of BEValue
2790 // i really is an addrec evolution.
2791 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2792 const SCEV *StartVal = getSCEV(StartValueV);
2794 // If StartVal = j.start - j.stride, we can use StartVal as the
2795 // initial step of the addrec evolution.
2796 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2797 AddRec->getOperand(1))) {
2798 const SCEV *PHISCEV =
2799 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2801 // Okay, for the entire analysis of this edge we assumed the PHI
2802 // to be symbolic. We now need to go back and purge all of the
2803 // entries for the scalars that use the symbolic expression.
2804 ForgetSymbolicName(PN, SymbolicName);
2805 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
2813 // If the PHI has a single incoming value, follow that value, unless the
2814 // PHI's incoming blocks are in a different loop, in which case doing so
2815 // risks breaking LCSSA form. Instcombine would normally zap these, but
2816 // it doesn't have DominatorTree information, so it may miss cases.
2817 if (Value *V = SimplifyInstruction(PN, TD, DT))
2818 if (LI->replacementPreservesLCSSAForm(PN, V))
2821 // If it's not a loop phi, we can't handle it yet.
2822 return getUnknown(PN);
2825 /// createNodeForGEP - Expand GEP instructions into add and multiply
2826 /// operations. This allows them to be analyzed by regular SCEV code.
2828 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
2830 // Don't blindly transfer the inbounds flag from the GEP instruction to the
2831 // Add expression, because the Instruction may be guarded by control flow
2832 // and the no-overflow bits may not be valid for the expression in any
2835 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2836 Value *Base = GEP->getOperand(0);
2837 // Don't attempt to analyze GEPs over unsized objects.
2838 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2839 return getUnknown(GEP);
2840 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
2841 gep_type_iterator GTI = gep_type_begin(GEP);
2842 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
2846 // Compute the (potentially symbolic) offset in bytes for this index.
2847 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2848 // For a struct, add the member offset.
2849 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2850 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
2852 // Add the field offset to the running total offset.
2853 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
2855 // For an array, add the element offset, explicitly scaled.
2856 const SCEV *ElementSize = getSizeOfExpr(*GTI);
2857 const SCEV *IndexS = getSCEV(Index);
2858 // Getelementptr indices are signed.
2859 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
2861 // Multiply the index by the element size to compute the element offset.
2862 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize);
2864 // Add the element offset to the running total offset.
2865 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2869 // Get the SCEV for the GEP base.
2870 const SCEV *BaseS = getSCEV(Base);
2872 // Add the total offset from all the GEP indices to the base.
2873 return getAddExpr(BaseS, TotalOffset);
2876 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2877 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2878 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2879 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2881 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2882 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2883 return C->getValue()->getValue().countTrailingZeros();
2885 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2886 return std::min(GetMinTrailingZeros(T->getOperand()),
2887 (uint32_t)getTypeSizeInBits(T->getType()));
2889 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2890 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2891 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2892 getTypeSizeInBits(E->getType()) : OpRes;
2895 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2896 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2897 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2898 getTypeSizeInBits(E->getType()) : OpRes;
2901 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2902 // The result is the min of all operands results.
2903 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2904 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2905 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2909 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2910 // The result is the sum of all operands results.
2911 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2912 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2913 for (unsigned i = 1, e = M->getNumOperands();
2914 SumOpRes != BitWidth && i != e; ++i)
2915 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2920 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2921 // The result is the min of all operands results.
2922 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2923 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2924 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2928 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2929 // The result is the min of all operands results.
2930 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2931 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2932 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2936 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2937 // The result is the min of all operands results.
2938 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2939 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2940 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2944 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2945 // For a SCEVUnknown, ask ValueTracking.
2946 unsigned BitWidth = getTypeSizeInBits(U->getType());
2947 APInt Mask = APInt::getAllOnesValue(BitWidth);
2948 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2949 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
2950 return Zeros.countTrailingOnes();
2957 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
2960 ScalarEvolution::getUnsignedRange(const SCEV *S) {
2961 // See if we've computed this range already.
2962 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
2963 if (I != UnsignedRanges.end())
2966 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2967 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
2969 unsigned BitWidth = getTypeSizeInBits(S->getType());
2970 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
2972 // If the value has known zeros, the maximum unsigned value will have those
2973 // known zeros as well.
2974 uint32_t TZ = GetMinTrailingZeros(S);
2976 ConservativeResult =
2977 ConstantRange(APInt::getMinValue(BitWidth),
2978 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
2980 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2981 ConstantRange X = getUnsignedRange(Add->getOperand(0));
2982 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2983 X = X.add(getUnsignedRange(Add->getOperand(i)));
2984 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
2987 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2988 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
2989 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2990 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
2991 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
2994 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2995 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
2996 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2997 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
2998 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
3001 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3002 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3003 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3004 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3005 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
3008 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3009 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3010 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3011 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3014 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3015 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3016 return setUnsignedRange(ZExt,
3017 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3020 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3021 ConstantRange X = getUnsignedRange(SExt->getOperand());
3022 return setUnsignedRange(SExt,
3023 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3026 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3027 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3028 return setUnsignedRange(Trunc,
3029 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3032 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3033 // If there's no unsigned wrap, the value will never be less than its
3035 if (AddRec->hasNoUnsignedWrap())
3036 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3037 if (!C->getValue()->isZero())
3038 ConservativeResult =
3039 ConservativeResult.intersectWith(
3040 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3042 // TODO: non-affine addrec
3043 if (AddRec->isAffine()) {
3044 const Type *Ty = AddRec->getType();
3045 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3046 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3047 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3048 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3050 const SCEV *Start = AddRec->getStart();
3051 const SCEV *Step = AddRec->getStepRecurrence(*this);
3053 ConstantRange StartRange = getUnsignedRange(Start);
3054 ConstantRange StepRange = getSignedRange(Step);
3055 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3056 ConstantRange EndRange =
3057 StartRange.add(MaxBECountRange.multiply(StepRange));
3059 // Check for overflow. This must be done with ConstantRange arithmetic
3060 // because we could be called from within the ScalarEvolution overflow
3062 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3063 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3064 ConstantRange ExtMaxBECountRange =
3065 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3066 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3067 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3069 return setUnsignedRange(AddRec, ConservativeResult);
3071 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3072 EndRange.getUnsignedMin());
3073 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3074 EndRange.getUnsignedMax());
3075 if (Min.isMinValue() && Max.isMaxValue())
3076 return setUnsignedRange(AddRec, ConservativeResult);
3077 return setUnsignedRange(AddRec,
3078 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3082 return setUnsignedRange(AddRec, ConservativeResult);
3085 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3086 // For a SCEVUnknown, ask ValueTracking.
3087 APInt Mask = APInt::getAllOnesValue(BitWidth);
3088 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3089 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
3090 if (Ones == ~Zeros + 1)
3091 return setUnsignedRange(U, ConservativeResult);
3092 return setUnsignedRange(U,
3093 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3096 return setUnsignedRange(S, ConservativeResult);
3099 /// getSignedRange - Determine the signed range for a particular SCEV.
3102 ScalarEvolution::getSignedRange(const SCEV *S) {
3103 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3104 if (I != SignedRanges.end())
3107 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3108 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3110 unsigned BitWidth = getTypeSizeInBits(S->getType());
3111 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3113 // If the value has known zeros, the maximum signed value will have those
3114 // known zeros as well.
3115 uint32_t TZ = GetMinTrailingZeros(S);
3117 ConservativeResult =
3118 ConstantRange(APInt::getSignedMinValue(BitWidth),
3119 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3121 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3122 ConstantRange X = getSignedRange(Add->getOperand(0));
3123 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3124 X = X.add(getSignedRange(Add->getOperand(i)));
3125 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3128 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3129 ConstantRange X = getSignedRange(Mul->getOperand(0));
3130 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3131 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3132 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3135 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3136 ConstantRange X = getSignedRange(SMax->getOperand(0));
3137 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3138 X = X.smax(getSignedRange(SMax->getOperand(i)));
3139 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3142 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3143 ConstantRange X = getSignedRange(UMax->getOperand(0));
3144 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3145 X = X.umax(getSignedRange(UMax->getOperand(i)));
3146 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3149 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3150 ConstantRange X = getSignedRange(UDiv->getLHS());
3151 ConstantRange Y = getSignedRange(UDiv->getRHS());
3152 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3155 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3156 ConstantRange X = getSignedRange(ZExt->getOperand());
3157 return setSignedRange(ZExt,
3158 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3161 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3162 ConstantRange X = getSignedRange(SExt->getOperand());
3163 return setSignedRange(SExt,
3164 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3167 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3168 ConstantRange X = getSignedRange(Trunc->getOperand());
3169 return setSignedRange(Trunc,
3170 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3173 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3174 // If there's no signed wrap, and all the operands have the same sign or
3175 // zero, the value won't ever change sign.
3176 if (AddRec->hasNoSignedWrap()) {
3177 bool AllNonNeg = true;
3178 bool AllNonPos = true;
3179 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3180 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3181 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3184 ConservativeResult = ConservativeResult.intersectWith(
3185 ConstantRange(APInt(BitWidth, 0),
3186 APInt::getSignedMinValue(BitWidth)));
3188 ConservativeResult = ConservativeResult.intersectWith(
3189 ConstantRange(APInt::getSignedMinValue(BitWidth),
3190 APInt(BitWidth, 1)));
3193 // TODO: non-affine addrec
3194 if (AddRec->isAffine()) {
3195 const Type *Ty = AddRec->getType();
3196 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3197 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3198 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3199 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3201 const SCEV *Start = AddRec->getStart();
3202 const SCEV *Step = AddRec->getStepRecurrence(*this);
3204 ConstantRange StartRange = getSignedRange(Start);
3205 ConstantRange StepRange = getSignedRange(Step);
3206 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3207 ConstantRange EndRange =
3208 StartRange.add(MaxBECountRange.multiply(StepRange));
3210 // Check for overflow. This must be done with ConstantRange arithmetic
3211 // because we could be called from within the ScalarEvolution overflow
3213 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3214 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3215 ConstantRange ExtMaxBECountRange =
3216 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3217 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3218 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3220 return setSignedRange(AddRec, ConservativeResult);
3222 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3223 EndRange.getSignedMin());
3224 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3225 EndRange.getSignedMax());
3226 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3227 return setSignedRange(AddRec, ConservativeResult);
3228 return setSignedRange(AddRec,
3229 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3233 return setSignedRange(AddRec, ConservativeResult);
3236 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3237 // For a SCEVUnknown, ask ValueTracking.
3238 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3239 return setSignedRange(U, ConservativeResult);
3240 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3242 return setSignedRange(U, ConservativeResult);
3243 return setSignedRange(U, ConservativeResult.intersectWith(
3244 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3245 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3248 return setSignedRange(S, ConservativeResult);
3251 /// createSCEV - We know that there is no SCEV for the specified value.
3252 /// Analyze the expression.
3254 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3255 if (!isSCEVable(V->getType()))
3256 return getUnknown(V);
3258 unsigned Opcode = Instruction::UserOp1;
3259 if (Instruction *I = dyn_cast<Instruction>(V)) {
3260 Opcode = I->getOpcode();
3262 // Don't attempt to analyze instructions in blocks that aren't
3263 // reachable. Such instructions don't matter, and they aren't required
3264 // to obey basic rules for definitions dominating uses which this
3265 // analysis depends on.
3266 if (!DT->isReachableFromEntry(I->getParent()))
3267 return getUnknown(V);
3268 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3269 Opcode = CE->getOpcode();
3270 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3271 return getConstant(CI);
3272 else if (isa<ConstantPointerNull>(V))
3273 return getConstant(V->getType(), 0);
3274 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3275 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3277 return getUnknown(V);
3279 Operator *U = cast<Operator>(V);
3281 case Instruction::Add: {
3282 // The simple thing to do would be to just call getSCEV on both operands
3283 // and call getAddExpr with the result. However if we're looking at a
3284 // bunch of things all added together, this can be quite inefficient,
3285 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3286 // Instead, gather up all the operands and make a single getAddExpr call.
3287 // LLVM IR canonical form means we need only traverse the left operands.
3288 SmallVector<const SCEV *, 4> AddOps;
3289 AddOps.push_back(getSCEV(U->getOperand(1)));
3290 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3291 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3292 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3294 U = cast<Operator>(Op);
3295 const SCEV *Op1 = getSCEV(U->getOperand(1));
3296 if (Opcode == Instruction::Sub)
3297 AddOps.push_back(getNegativeSCEV(Op1));
3299 AddOps.push_back(Op1);
3301 AddOps.push_back(getSCEV(U->getOperand(0)));
3302 return getAddExpr(AddOps);
3304 case Instruction::Mul: {
3305 // See the Add code above.
3306 SmallVector<const SCEV *, 4> MulOps;
3307 MulOps.push_back(getSCEV(U->getOperand(1)));
3308 for (Value *Op = U->getOperand(0);
3309 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3310 Op = U->getOperand(0)) {
3311 U = cast<Operator>(Op);
3312 MulOps.push_back(getSCEV(U->getOperand(1)));
3314 MulOps.push_back(getSCEV(U->getOperand(0)));
3315 return getMulExpr(MulOps);
3317 case Instruction::UDiv:
3318 return getUDivExpr(getSCEV(U->getOperand(0)),
3319 getSCEV(U->getOperand(1)));
3320 case Instruction::Sub:
3321 return getMinusSCEV(getSCEV(U->getOperand(0)),
3322 getSCEV(U->getOperand(1)));
3323 case Instruction::And:
3324 // For an expression like x&255 that merely masks off the high bits,
3325 // use zext(trunc(x)) as the SCEV expression.
3326 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3327 if (CI->isNullValue())
3328 return getSCEV(U->getOperand(1));
3329 if (CI->isAllOnesValue())
3330 return getSCEV(U->getOperand(0));
3331 const APInt &A = CI->getValue();
3333 // Instcombine's ShrinkDemandedConstant may strip bits out of
3334 // constants, obscuring what would otherwise be a low-bits mask.
3335 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3336 // knew about to reconstruct a low-bits mask value.
3337 unsigned LZ = A.countLeadingZeros();
3338 unsigned BitWidth = A.getBitWidth();
3339 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
3340 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3341 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
3343 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3345 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3347 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3348 IntegerType::get(getContext(), BitWidth - LZ)),
3353 case Instruction::Or:
3354 // If the RHS of the Or is a constant, we may have something like:
3355 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3356 // optimizations will transparently handle this case.
3358 // In order for this transformation to be safe, the LHS must be of the
3359 // form X*(2^n) and the Or constant must be less than 2^n.
3360 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3361 const SCEV *LHS = getSCEV(U->getOperand(0));
3362 const APInt &CIVal = CI->getValue();
3363 if (GetMinTrailingZeros(LHS) >=
3364 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3365 // Build a plain add SCEV.
3366 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3367 // If the LHS of the add was an addrec and it has no-wrap flags,
3368 // transfer the no-wrap flags, since an or won't introduce a wrap.
3369 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3370 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3371 if (OldAR->hasNoUnsignedWrap())
3372 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3373 if (OldAR->hasNoSignedWrap())
3374 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3380 case Instruction::Xor:
3381 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3382 // If the RHS of the xor is a signbit, then this is just an add.
3383 // Instcombine turns add of signbit into xor as a strength reduction step.
3384 if (CI->getValue().isSignBit())
3385 return getAddExpr(getSCEV(U->getOperand(0)),
3386 getSCEV(U->getOperand(1)));
3388 // If the RHS of xor is -1, then this is a not operation.
3389 if (CI->isAllOnesValue())
3390 return getNotSCEV(getSCEV(U->getOperand(0)));
3392 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3393 // This is a variant of the check for xor with -1, and it handles
3394 // the case where instcombine has trimmed non-demanded bits out
3395 // of an xor with -1.
3396 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3397 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3398 if (BO->getOpcode() == Instruction::And &&
3399 LCI->getValue() == CI->getValue())
3400 if (const SCEVZeroExtendExpr *Z =
3401 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3402 const Type *UTy = U->getType();
3403 const SCEV *Z0 = Z->getOperand();
3404 const Type *Z0Ty = Z0->getType();
3405 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3407 // If C is a low-bits mask, the zero extend is serving to
3408 // mask off the high bits. Complement the operand and
3409 // re-apply the zext.
3410 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3411 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3413 // If C is a single bit, it may be in the sign-bit position
3414 // before the zero-extend. In this case, represent the xor
3415 // using an add, which is equivalent, and re-apply the zext.
3416 APInt Trunc = CI->getValue().trunc(Z0TySize);
3417 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3419 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3425 case Instruction::Shl:
3426 // Turn shift left of a constant amount into a multiply.
3427 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3428 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3430 // If the shift count is not less than the bitwidth, the result of
3431 // the shift is undefined. Don't try to analyze it, because the
3432 // resolution chosen here may differ from the resolution chosen in
3433 // other parts of the compiler.
3434 if (SA->getValue().uge(BitWidth))
3437 Constant *X = ConstantInt::get(getContext(),
3438 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3439 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3443 case Instruction::LShr:
3444 // Turn logical shift right of a constant into a unsigned divide.
3445 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3446 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3448 // If the shift count is not less than the bitwidth, the result of
3449 // the shift is undefined. Don't try to analyze it, because the
3450 // resolution chosen here may differ from the resolution chosen in
3451 // other parts of the compiler.
3452 if (SA->getValue().uge(BitWidth))
3455 Constant *X = ConstantInt::get(getContext(),
3456 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3457 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3461 case Instruction::AShr:
3462 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3463 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3464 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3465 if (L->getOpcode() == Instruction::Shl &&
3466 L->getOperand(1) == U->getOperand(1)) {
3467 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3469 // If the shift count is not less than the bitwidth, the result of
3470 // the shift is undefined. Don't try to analyze it, because the
3471 // resolution chosen here may differ from the resolution chosen in
3472 // other parts of the compiler.
3473 if (CI->getValue().uge(BitWidth))
3476 uint64_t Amt = BitWidth - CI->getZExtValue();
3477 if (Amt == BitWidth)
3478 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3480 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3481 IntegerType::get(getContext(),
3487 case Instruction::Trunc:
3488 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3490 case Instruction::ZExt:
3491 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3493 case Instruction::SExt:
3494 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3496 case Instruction::BitCast:
3497 // BitCasts are no-op casts so we just eliminate the cast.
3498 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3499 return getSCEV(U->getOperand(0));
3502 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3503 // lead to pointer expressions which cannot safely be expanded to GEPs,
3504 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3505 // simplifying integer expressions.
3507 case Instruction::GetElementPtr:
3508 return createNodeForGEP(cast<GEPOperator>(U));
3510 case Instruction::PHI:
3511 return createNodeForPHI(cast<PHINode>(U));
3513 case Instruction::Select:
3514 // This could be a smax or umax that was lowered earlier.
3515 // Try to recover it.
3516 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3517 Value *LHS = ICI->getOperand(0);
3518 Value *RHS = ICI->getOperand(1);
3519 switch (ICI->getPredicate()) {
3520 case ICmpInst::ICMP_SLT:
3521 case ICmpInst::ICMP_SLE:
3522 std::swap(LHS, RHS);
3524 case ICmpInst::ICMP_SGT:
3525 case ICmpInst::ICMP_SGE:
3526 // a >s b ? a+x : b+x -> smax(a, b)+x
3527 // a >s b ? b+x : a+x -> smin(a, b)+x
3528 if (LHS->getType() == U->getType()) {
3529 const SCEV *LS = getSCEV(LHS);
3530 const SCEV *RS = getSCEV(RHS);
3531 const SCEV *LA = getSCEV(U->getOperand(1));
3532 const SCEV *RA = getSCEV(U->getOperand(2));
3533 const SCEV *LDiff = getMinusSCEV(LA, LS);
3534 const SCEV *RDiff = getMinusSCEV(RA, RS);
3536 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3537 LDiff = getMinusSCEV(LA, RS);
3538 RDiff = getMinusSCEV(RA, LS);
3540 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3543 case ICmpInst::ICMP_ULT:
3544 case ICmpInst::ICMP_ULE:
3545 std::swap(LHS, RHS);
3547 case ICmpInst::ICMP_UGT:
3548 case ICmpInst::ICMP_UGE:
3549 // a >u b ? a+x : b+x -> umax(a, b)+x
3550 // a >u b ? b+x : a+x -> umin(a, b)+x
3551 if (LHS->getType() == U->getType()) {
3552 const SCEV *LS = getSCEV(LHS);
3553 const SCEV *RS = getSCEV(RHS);
3554 const SCEV *LA = getSCEV(U->getOperand(1));
3555 const SCEV *RA = getSCEV(U->getOperand(2));
3556 const SCEV *LDiff = getMinusSCEV(LA, LS);
3557 const SCEV *RDiff = getMinusSCEV(RA, RS);
3559 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3560 LDiff = getMinusSCEV(LA, RS);
3561 RDiff = getMinusSCEV(RA, LS);
3563 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3566 case ICmpInst::ICMP_NE:
3567 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3568 if (LHS->getType() == U->getType() &&
3569 isa<ConstantInt>(RHS) &&
3570 cast<ConstantInt>(RHS)->isZero()) {
3571 const SCEV *One = getConstant(LHS->getType(), 1);
3572 const SCEV *LS = getSCEV(LHS);
3573 const SCEV *LA = getSCEV(U->getOperand(1));
3574 const SCEV *RA = getSCEV(U->getOperand(2));
3575 const SCEV *LDiff = getMinusSCEV(LA, LS);
3576 const SCEV *RDiff = getMinusSCEV(RA, One);
3578 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3581 case ICmpInst::ICMP_EQ:
3582 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3583 if (LHS->getType() == U->getType() &&
3584 isa<ConstantInt>(RHS) &&
3585 cast<ConstantInt>(RHS)->isZero()) {
3586 const SCEV *One = getConstant(LHS->getType(), 1);
3587 const SCEV *LS = getSCEV(LHS);
3588 const SCEV *LA = getSCEV(U->getOperand(1));
3589 const SCEV *RA = getSCEV(U->getOperand(2));
3590 const SCEV *LDiff = getMinusSCEV(LA, One);
3591 const SCEV *RDiff = getMinusSCEV(RA, LS);
3593 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3601 default: // We cannot analyze this expression.
3605 return getUnknown(V);
3610 //===----------------------------------------------------------------------===//
3611 // Iteration Count Computation Code
3614 /// getBackedgeTakenCount - If the specified loop has a predictable
3615 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3616 /// object. The backedge-taken count is the number of times the loop header
3617 /// will be branched to from within the loop. This is one less than the
3618 /// trip count of the loop, since it doesn't count the first iteration,
3619 /// when the header is branched to from outside the loop.
3621 /// Note that it is not valid to call this method on a loop without a
3622 /// loop-invariant backedge-taken count (see
3623 /// hasLoopInvariantBackedgeTakenCount).
3625 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3626 return getBackedgeTakenInfo(L).Exact;
3629 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3630 /// return the least SCEV value that is known never to be less than the
3631 /// actual backedge taken count.
3632 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3633 return getBackedgeTakenInfo(L).Max;
3636 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3637 /// onto the given Worklist.
3639 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3640 BasicBlock *Header = L->getHeader();
3642 // Push all Loop-header PHIs onto the Worklist stack.
3643 for (BasicBlock::iterator I = Header->begin();
3644 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3645 Worklist.push_back(PN);
3648 const ScalarEvolution::BackedgeTakenInfo &
3649 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3650 // Initially insert a CouldNotCompute for this loop. If the insertion
3651 // succeeds, proceed to actually compute a backedge-taken count and
3652 // update the value. The temporary CouldNotCompute value tells SCEV
3653 // code elsewhere that it shouldn't attempt to request a new
3654 // backedge-taken count, which could result in infinite recursion.
3655 std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3656 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3658 return Pair.first->second;
3660 BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
3661 if (BECount.Exact != getCouldNotCompute()) {
3662 assert(isLoopInvariant(BECount.Exact, L) &&
3663 isLoopInvariant(BECount.Max, L) &&
3664 "Computed backedge-taken count isn't loop invariant for loop!");
3665 ++NumTripCountsComputed;
3667 // Update the value in the map.
3668 Pair.first->second = BECount;
3670 if (BECount.Max != getCouldNotCompute())
3671 // Update the value in the map.
3672 Pair.first->second = BECount;
3673 if (isa<PHINode>(L->getHeader()->begin()))
3674 // Only count loops that have phi nodes as not being computable.
3675 ++NumTripCountsNotComputed;
3678 // Now that we know more about the trip count for this loop, forget any
3679 // existing SCEV values for PHI nodes in this loop since they are only
3680 // conservative estimates made without the benefit of trip count
3681 // information. This is similar to the code in forgetLoop, except that
3682 // it handles SCEVUnknown PHI nodes specially.
3683 if (BECount.hasAnyInfo()) {
3684 SmallVector<Instruction *, 16> Worklist;
3685 PushLoopPHIs(L, Worklist);
3687 SmallPtrSet<Instruction *, 8> Visited;
3688 while (!Worklist.empty()) {
3689 Instruction *I = Worklist.pop_back_val();
3690 if (!Visited.insert(I)) continue;
3692 ValueExprMapType::iterator It =
3693 ValueExprMap.find(static_cast<Value *>(I));
3694 if (It != ValueExprMap.end()) {
3695 const SCEV *Old = It->second;
3697 // SCEVUnknown for a PHI either means that it has an unrecognized
3698 // structure, or it's a PHI that's in the progress of being computed
3699 // by createNodeForPHI. In the former case, additional loop trip
3700 // count information isn't going to change anything. In the later
3701 // case, createNodeForPHI will perform the necessary updates on its
3702 // own when it gets to that point.
3703 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
3704 forgetMemoizedResults(Old);
3705 ValueExprMap.erase(It);
3707 if (PHINode *PN = dyn_cast<PHINode>(I))
3708 ConstantEvolutionLoopExitValue.erase(PN);
3711 PushDefUseChildren(I, Worklist);
3714 return Pair.first->second;
3717 /// forgetLoop - This method should be called by the client when it has
3718 /// changed a loop in a way that may effect ScalarEvolution's ability to
3719 /// compute a trip count, or if the loop is deleted.
3720 void ScalarEvolution::forgetLoop(const Loop *L) {
3721 // Drop any stored trip count value.
3722 BackedgeTakenCounts.erase(L);
3724 // Drop information about expressions based on loop-header PHIs.
3725 SmallVector<Instruction *, 16> Worklist;
3726 PushLoopPHIs(L, Worklist);
3728 SmallPtrSet<Instruction *, 8> Visited;
3729 while (!Worklist.empty()) {
3730 Instruction *I = Worklist.pop_back_val();
3731 if (!Visited.insert(I)) continue;
3733 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3734 if (It != ValueExprMap.end()) {
3735 forgetMemoizedResults(It->second);
3736 ValueExprMap.erase(It);
3737 if (PHINode *PN = dyn_cast<PHINode>(I))
3738 ConstantEvolutionLoopExitValue.erase(PN);
3741 PushDefUseChildren(I, Worklist);
3744 // Forget all contained loops too, to avoid dangling entries in the
3745 // ValuesAtScopes map.
3746 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
3750 /// forgetValue - This method should be called by the client when it has
3751 /// changed a value in a way that may effect its value, or which may
3752 /// disconnect it from a def-use chain linking it to a loop.
3753 void ScalarEvolution::forgetValue(Value *V) {
3754 Instruction *I = dyn_cast<Instruction>(V);
3757 // Drop information about expressions based on loop-header PHIs.
3758 SmallVector<Instruction *, 16> Worklist;
3759 Worklist.push_back(I);
3761 SmallPtrSet<Instruction *, 8> Visited;
3762 while (!Worklist.empty()) {
3763 I = Worklist.pop_back_val();
3764 if (!Visited.insert(I)) continue;
3766 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
3767 if (It != ValueExprMap.end()) {
3768 forgetMemoizedResults(It->second);
3769 ValueExprMap.erase(It);
3770 if (PHINode *PN = dyn_cast<PHINode>(I))
3771 ConstantEvolutionLoopExitValue.erase(PN);
3774 PushDefUseChildren(I, Worklist);
3778 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3779 /// of the specified loop will execute.
3780 ScalarEvolution::BackedgeTakenInfo
3781 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3782 SmallVector<BasicBlock *, 8> ExitingBlocks;
3783 L->getExitingBlocks(ExitingBlocks);
3785 // Examine all exits and pick the most conservative values.
3786 const SCEV *BECount = getCouldNotCompute();
3787 const SCEV *MaxBECount = getCouldNotCompute();
3788 bool CouldNotComputeBECount = false;
3789 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3790 BackedgeTakenInfo NewBTI =
3791 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3793 if (NewBTI.Exact == getCouldNotCompute()) {
3794 // We couldn't compute an exact value for this exit, so
3795 // we won't be able to compute an exact value for the loop.
3796 CouldNotComputeBECount = true;
3797 BECount = getCouldNotCompute();
3798 } else if (!CouldNotComputeBECount) {
3799 if (BECount == getCouldNotCompute())
3800 BECount = NewBTI.Exact;
3802 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3804 if (MaxBECount == getCouldNotCompute())
3805 MaxBECount = NewBTI.Max;
3806 else if (NewBTI.Max != getCouldNotCompute())
3807 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3810 return BackedgeTakenInfo(BECount, MaxBECount);
3813 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3814 /// of the specified loop will execute if it exits via the specified block.
3815 ScalarEvolution::BackedgeTakenInfo
3816 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3817 BasicBlock *ExitingBlock) {
3819 // Okay, we've chosen an exiting block. See what condition causes us to
3820 // exit at this block.
3822 // FIXME: we should be able to handle switch instructions (with a single exit)
3823 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3824 if (ExitBr == 0) return getCouldNotCompute();
3825 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3827 // At this point, we know we have a conditional branch that determines whether
3828 // the loop is exited. However, we don't know if the branch is executed each
3829 // time through the loop. If not, then the execution count of the branch will
3830 // not be equal to the trip count of the loop.
3832 // Currently we check for this by checking to see if the Exit branch goes to
3833 // the loop header. If so, we know it will always execute the same number of
3834 // times as the loop. We also handle the case where the exit block *is* the
3835 // loop header. This is common for un-rotated loops.
3837 // If both of those tests fail, walk up the unique predecessor chain to the
3838 // header, stopping if there is an edge that doesn't exit the loop. If the
3839 // header is reached, the execution count of the branch will be equal to the
3840 // trip count of the loop.
3842 // More extensive analysis could be done to handle more cases here.
3844 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3845 ExitBr->getSuccessor(1) != L->getHeader() &&
3846 ExitBr->getParent() != L->getHeader()) {
3847 // The simple checks failed, try climbing the unique predecessor chain
3848 // up to the header.
3850 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3851 BasicBlock *Pred = BB->getUniquePredecessor();
3853 return getCouldNotCompute();
3854 TerminatorInst *PredTerm = Pred->getTerminator();
3855 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3856 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3859 // If the predecessor has a successor that isn't BB and isn't
3860 // outside the loop, assume the worst.
3861 if (L->contains(PredSucc))
3862 return getCouldNotCompute();
3864 if (Pred == L->getHeader()) {
3871 return getCouldNotCompute();
3874 // Proceed to the next level to examine the exit condition expression.
3875 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3876 ExitBr->getSuccessor(0),
3877 ExitBr->getSuccessor(1));
3880 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3881 /// backedge of the specified loop will execute if its exit condition
3882 /// were a conditional branch of ExitCond, TBB, and FBB.
3883 ScalarEvolution::BackedgeTakenInfo
3884 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3888 // Check if the controlling expression for this loop is an And or Or.
3889 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3890 if (BO->getOpcode() == Instruction::And) {
3891 // Recurse on the operands of the and.
3892 BackedgeTakenInfo BTI0 =
3893 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3894 BackedgeTakenInfo BTI1 =
3895 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3896 const SCEV *BECount = getCouldNotCompute();
3897 const SCEV *MaxBECount = getCouldNotCompute();
3898 if (L->contains(TBB)) {
3899 // Both conditions must be true for the loop to continue executing.
3900 // Choose the less conservative count.
3901 if (BTI0.Exact == getCouldNotCompute() ||
3902 BTI1.Exact == getCouldNotCompute())
3903 BECount = getCouldNotCompute();
3905 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3906 if (BTI0.Max == getCouldNotCompute())
3907 MaxBECount = BTI1.Max;
3908 else if (BTI1.Max == getCouldNotCompute())
3909 MaxBECount = BTI0.Max;
3911 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3913 // Both conditions must be true at the same time for the loop to exit.
3914 // For now, be conservative.
3915 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3916 if (BTI0.Max == BTI1.Max)
3917 MaxBECount = BTI0.Max;
3918 if (BTI0.Exact == BTI1.Exact)
3919 BECount = BTI0.Exact;
3922 return BackedgeTakenInfo(BECount, MaxBECount);
3924 if (BO->getOpcode() == Instruction::Or) {
3925 // Recurse on the operands of the or.
3926 BackedgeTakenInfo BTI0 =
3927 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3928 BackedgeTakenInfo BTI1 =
3929 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3930 const SCEV *BECount = getCouldNotCompute();
3931 const SCEV *MaxBECount = getCouldNotCompute();
3932 if (L->contains(FBB)) {
3933 // Both conditions must be false for the loop to continue executing.
3934 // Choose the less conservative count.
3935 if (BTI0.Exact == getCouldNotCompute() ||
3936 BTI1.Exact == getCouldNotCompute())
3937 BECount = getCouldNotCompute();
3939 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3940 if (BTI0.Max == getCouldNotCompute())
3941 MaxBECount = BTI1.Max;
3942 else if (BTI1.Max == getCouldNotCompute())
3943 MaxBECount = BTI0.Max;
3945 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3947 // Both conditions must be false at the same time for the loop to exit.
3948 // For now, be conservative.
3949 assert(L->contains(TBB) && "Loop block has no successor in loop!");
3950 if (BTI0.Max == BTI1.Max)
3951 MaxBECount = BTI0.Max;
3952 if (BTI0.Exact == BTI1.Exact)
3953 BECount = BTI0.Exact;
3956 return BackedgeTakenInfo(BECount, MaxBECount);
3960 // With an icmp, it may be feasible to compute an exact backedge-taken count.
3961 // Proceed to the next level to examine the icmp.
3962 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
3963 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
3965 // Check for a constant condition. These are normally stripped out by
3966 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
3967 // preserve the CFG and is temporarily leaving constant conditions
3969 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
3970 if (L->contains(FBB) == !CI->getZExtValue())
3971 // The backedge is always taken.
3972 return getCouldNotCompute();
3974 // The backedge is never taken.
3975 return getConstant(CI->getType(), 0);
3978 // If it's not an integer or pointer comparison then compute it the hard way.
3979 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3982 static const SCEVAddRecExpr *
3983 isSimpleUnwrappingAddRec(const SCEV *S, const Loop *L) {
3984 const SCEVAddRecExpr *SA = dyn_cast<SCEVAddRecExpr>(S);
3986 // The SCEV must be an addrec of this loop.
3987 if (!SA || SA->getLoop() != L || !SA->isAffine())
3990 // The SCEV must be known to not wrap in some way to be interesting.
3991 if (!SA->hasNoUnsignedWrap() && !SA->hasNoSignedWrap())
3994 // The stride must be a constant so that we know if it is striding up or down.
3995 if (!isa<SCEVConstant>(SA->getOperand(1)))
4000 /// getMinusSCEVForExitTest - When considering an exit test for a loop with a
4001 /// "x != y" exit test, we turn this into a computation that evaluates x-y != 0,
4002 /// and this function returns the expression to use for x-y. We know and take
4003 /// advantage of the fact that this subtraction is only being used in a
4004 /// comparison by zero context.
4006 static const SCEV *getMinusSCEVForExitTest(const SCEV *LHS, const SCEV *RHS,
4007 const Loop *L, ScalarEvolution &SE) {
4008 // If either LHS or RHS is an AddRec SCEV (of this loop) that is known to not
4009 // wrap (either NSW or NUW), then we know that the value will either become
4010 // the other one (and thus the loop terminates), that the loop will terminate
4011 // through some other exit condition first, or that the loop has undefined
4012 // behavior. This information is useful when the addrec has a stride that is
4013 // != 1 or -1, because it means we can't "miss" the exit value.
4015 // In any of these three cases, it is safe to turn the exit condition into a
4016 // "counting down" AddRec (to zero) by subtracting the two inputs as normal,
4017 // but since we know that the "end cannot be missed" we can force the
4018 // resulting AddRec to be a NUW addrec. Since it is counting down, this means
4019 // that the AddRec *cannot* pass zero.
4021 // See if LHS and RHS are addrec's we can handle.
4022 const SCEVAddRecExpr *LHSA = isSimpleUnwrappingAddRec(LHS, L);
4023 const SCEVAddRecExpr *RHSA = isSimpleUnwrappingAddRec(RHS, L);
4025 // If neither addrec is interesting, just return a minus.
4026 if (RHSA == 0 && LHSA == 0)
4027 return SE.getMinusSCEV(LHS, RHS);
4029 // If only one of LHS and RHS are an AddRec of this loop, make sure it is LHS.
4030 if (RHSA && LHSA == 0) {
4031 // Safe because a-b === b-a for comparisons against zero.
4032 std::swap(LHS, RHS);
4033 std::swap(LHSA, RHSA);
4036 // Handle the case when only one is advancing in a non-overflowing way.
4038 // If RHS is loop varying, then we can't predict when LHS will cross it.
4039 if (!SE.isLoopInvariant(RHS, L))
4040 return SE.getMinusSCEV(LHS, RHS);
4042 // If LHS has a positive stride, then we compute RHS-LHS, because the loop
4043 // is counting up until it crosses RHS (which must be larger than LHS). If
4044 // it is negative, we compute LHS-RHS because we're counting down to RHS.
4045 const ConstantInt *Stride =
4046 cast<SCEVConstant>(LHSA->getOperand(1))->getValue();
4047 if (Stride->getValue().isNegative())
4048 std::swap(LHS, RHS);
4050 return SE.getMinusSCEV(RHS, LHS, true /*HasNUW*/);
4053 // If both LHS and RHS are interesting, we have something like:
4055 const ConstantInt *LHSStride =
4056 cast<SCEVConstant>(LHSA->getOperand(1))->getValue();
4057 const ConstantInt *RHSStride =
4058 cast<SCEVConstant>(RHSA->getOperand(1))->getValue();
4060 // If the strides are equal, then this is just a (complex) loop invariant
4061 // comparison of a and b.
4062 if (LHSStride == RHSStride)
4063 return SE.getMinusSCEV(LHSA->getStart(), RHSA->getStart());
4065 // If the signs of the strides differ, then the negative stride is counting
4066 // down to the positive stride.
4067 if (LHSStride->getValue().isNegative() != RHSStride->getValue().isNegative()){
4068 if (RHSStride->getValue().isNegative())
4069 std::swap(LHS, RHS);
4071 // If LHS's stride is smaller than RHS's stride, then "b" must be less than
4072 // "a" and "b" is RHS is counting up (catching up) to LHS. This is true
4073 // whether the strides are positive or negative.
4074 if (RHSStride->getValue().slt(LHSStride->getValue()))
4075 std::swap(LHS, RHS);
4078 return SE.getMinusSCEV(LHS, RHS, true /*HasNUW*/);
4081 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
4082 /// backedge of the specified loop will execute if its exit condition
4083 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4084 ScalarEvolution::BackedgeTakenInfo
4085 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
4090 // If the condition was exit on true, convert the condition to exit on false
4091 ICmpInst::Predicate Cond;
4092 if (!L->contains(FBB))
4093 Cond = ExitCond->getPredicate();
4095 Cond = ExitCond->getInversePredicate();
4097 // Handle common loops like: for (X = "string"; *X; ++X)
4098 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4099 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4100 BackedgeTakenInfo ItCnt =
4101 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
4102 if (ItCnt.hasAnyInfo())
4106 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4107 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4109 // Try to evaluate any dependencies out of the loop.
4110 LHS = getSCEVAtScope(LHS, L);
4111 RHS = getSCEVAtScope(RHS, L);
4113 // At this point, we would like to compute how many iterations of the
4114 // loop the predicate will return true for these inputs.
4115 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4116 // If there is a loop-invariant, force it into the RHS.
4117 std::swap(LHS, RHS);
4118 Cond = ICmpInst::getSwappedPredicate(Cond);
4121 // Simplify the operands before analyzing them.
4122 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4124 // If we have a comparison of a chrec against a constant, try to use value
4125 // ranges to answer this query.
4126 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4127 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4128 if (AddRec->getLoop() == L) {
4129 // Form the constant range.
4130 ConstantRange CompRange(
4131 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4133 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4134 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4138 case ICmpInst::ICMP_NE: { // while (X != Y)
4139 // Convert to: while (X-Y != 0)
4140 BackedgeTakenInfo BTI = HowFarToZero(getMinusSCEVForExitTest(LHS, RHS, L,
4142 if (BTI.hasAnyInfo()) return BTI;
4145 case ICmpInst::ICMP_EQ: { // while (X == Y)
4146 // Convert to: while (X-Y == 0)
4147 BackedgeTakenInfo BTI = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4148 if (BTI.hasAnyInfo()) return BTI;
4151 case ICmpInst::ICMP_SLT: {
4152 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
4153 if (BTI.hasAnyInfo()) return BTI;
4156 case ICmpInst::ICMP_SGT: {
4157 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4158 getNotSCEV(RHS), L, true);
4159 if (BTI.hasAnyInfo()) return BTI;
4162 case ICmpInst::ICMP_ULT: {
4163 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
4164 if (BTI.hasAnyInfo()) return BTI;
4167 case ICmpInst::ICMP_UGT: {
4168 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
4169 getNotSCEV(RHS), L, false);
4170 if (BTI.hasAnyInfo()) return BTI;
4175 dbgs() << "ComputeBackedgeTakenCount ";
4176 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4177 dbgs() << "[unsigned] ";
4178 dbgs() << *LHS << " "
4179 << Instruction::getOpcodeName(Instruction::ICmp)
4180 << " " << *RHS << "\n";
4185 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
4188 static ConstantInt *
4189 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4190 ScalarEvolution &SE) {
4191 const SCEV *InVal = SE.getConstant(C);
4192 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4193 assert(isa<SCEVConstant>(Val) &&
4194 "Evaluation of SCEV at constant didn't fold correctly?");
4195 return cast<SCEVConstant>(Val)->getValue();
4198 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
4199 /// and a GEP expression (missing the pointer index) indexing into it, return
4200 /// the addressed element of the initializer or null if the index expression is
4203 GetAddressedElementFromGlobal(GlobalVariable *GV,
4204 const std::vector<ConstantInt*> &Indices) {
4205 Constant *Init = GV->getInitializer();
4206 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
4207 uint64_t Idx = Indices[i]->getZExtValue();
4208 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
4209 assert(Idx < CS->getNumOperands() && "Bad struct index!");
4210 Init = cast<Constant>(CS->getOperand(Idx));
4211 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
4212 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
4213 Init = cast<Constant>(CA->getOperand(Idx));
4214 } else if (isa<ConstantAggregateZero>(Init)) {
4215 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
4216 assert(Idx < STy->getNumElements() && "Bad struct index!");
4217 Init = Constant::getNullValue(STy->getElementType(Idx));
4218 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
4219 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
4220 Init = Constant::getNullValue(ATy->getElementType());
4222 llvm_unreachable("Unknown constant aggregate type!");
4226 return 0; // Unknown initializer type
4232 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
4233 /// 'icmp op load X, cst', try to see if we can compute the backedge
4234 /// execution count.
4235 ScalarEvolution::BackedgeTakenInfo
4236 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
4240 ICmpInst::Predicate predicate) {
4241 if (LI->isVolatile()) return getCouldNotCompute();
4243 // Check to see if the loaded pointer is a getelementptr of a global.
4244 // TODO: Use SCEV instead of manually grubbing with GEPs.
4245 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4246 if (!GEP) return getCouldNotCompute();
4248 // Make sure that it is really a constant global we are gepping, with an
4249 // initializer, and make sure the first IDX is really 0.
4250 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4251 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4252 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4253 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4254 return getCouldNotCompute();
4256 // Okay, we allow one non-constant index into the GEP instruction.
4258 std::vector<ConstantInt*> Indexes;
4259 unsigned VarIdxNum = 0;
4260 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4261 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4262 Indexes.push_back(CI);
4263 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4264 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4265 VarIdx = GEP->getOperand(i);
4267 Indexes.push_back(0);
4270 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4271 // Check to see if X is a loop variant variable value now.
4272 const SCEV *Idx = getSCEV(VarIdx);
4273 Idx = getSCEVAtScope(Idx, L);
4275 // We can only recognize very limited forms of loop index expressions, in
4276 // particular, only affine AddRec's like {C1,+,C2}.
4277 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4278 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4279 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4280 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4281 return getCouldNotCompute();
4283 unsigned MaxSteps = MaxBruteForceIterations;
4284 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4285 ConstantInt *ItCst = ConstantInt::get(
4286 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4287 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4289 // Form the GEP offset.
4290 Indexes[VarIdxNum] = Val;
4292 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
4293 if (Result == 0) break; // Cannot compute!
4295 // Evaluate the condition for this iteration.
4296 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4297 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4298 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4300 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4301 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4304 ++NumArrayLenItCounts;
4305 return getConstant(ItCst); // Found terminating iteration!
4308 return getCouldNotCompute();
4312 /// CanConstantFold - Return true if we can constant fold an instruction of the
4313 /// specified type, assuming that all operands were constants.
4314 static bool CanConstantFold(const Instruction *I) {
4315 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4316 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
4319 if (const CallInst *CI = dyn_cast<CallInst>(I))
4320 if (const Function *F = CI->getCalledFunction())
4321 return canConstantFoldCallTo(F);
4325 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4326 /// in the loop that V is derived from. We allow arbitrary operations along the
4327 /// way, but the operands of an operation must either be constants or a value
4328 /// derived from a constant PHI. If this expression does not fit with these
4329 /// constraints, return null.
4330 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4331 // If this is not an instruction, or if this is an instruction outside of the
4332 // loop, it can't be derived from a loop PHI.
4333 Instruction *I = dyn_cast<Instruction>(V);
4334 if (I == 0 || !L->contains(I)) return 0;
4336 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4337 if (L->getHeader() == I->getParent())
4340 // We don't currently keep track of the control flow needed to evaluate
4341 // PHIs, so we cannot handle PHIs inside of loops.
4345 // If we won't be able to constant fold this expression even if the operands
4346 // are constants, return early.
4347 if (!CanConstantFold(I)) return 0;
4349 // Otherwise, we can evaluate this instruction if all of its operands are
4350 // constant or derived from a PHI node themselves.
4352 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
4353 if (!isa<Constant>(I->getOperand(Op))) {
4354 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
4355 if (P == 0) return 0; // Not evolving from PHI
4359 return 0; // Evolving from multiple different PHIs.
4362 // This is a expression evolving from a constant PHI!
4366 /// EvaluateExpression - Given an expression that passes the
4367 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4368 /// in the loop has the value PHIVal. If we can't fold this expression for some
4369 /// reason, return null.
4370 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
4371 const TargetData *TD) {
4372 if (isa<PHINode>(V)) return PHIVal;
4373 if (Constant *C = dyn_cast<Constant>(V)) return C;
4374 Instruction *I = cast<Instruction>(V);
4376 std::vector<Constant*> Operands(I->getNumOperands());
4378 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4379 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
4380 if (Operands[i] == 0) return 0;
4383 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4384 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4386 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4387 &Operands[0], Operands.size(), TD);
4390 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4391 /// in the header of its containing loop, we know the loop executes a
4392 /// constant number of times, and the PHI node is just a recurrence
4393 /// involving constants, fold it.
4395 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4398 std::map<PHINode*, Constant*>::const_iterator I =
4399 ConstantEvolutionLoopExitValue.find(PN);
4400 if (I != ConstantEvolutionLoopExitValue.end())
4403 if (BEs.ugt(MaxBruteForceIterations))
4404 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4406 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4408 // Since the loop is canonicalized, the PHI node must have two entries. One
4409 // entry must be a constant (coming in from outside of the loop), and the
4410 // second must be derived from the same PHI.
4411 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4412 Constant *StartCST =
4413 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4415 return RetVal = 0; // Must be a constant.
4417 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4418 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4419 !isa<Constant>(BEValue))
4420 return RetVal = 0; // Not derived from same PHI.
4422 // Execute the loop symbolically to determine the exit value.
4423 if (BEs.getActiveBits() >= 32)
4424 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4426 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4427 unsigned IterationNum = 0;
4428 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
4429 if (IterationNum == NumIterations)
4430 return RetVal = PHIVal; // Got exit value!
4432 // Compute the value of the PHI node for the next iteration.
4433 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4434 if (NextPHI == PHIVal)
4435 return RetVal = NextPHI; // Stopped evolving!
4437 return 0; // Couldn't evaluate!
4442 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
4443 /// constant number of times (the condition evolves only from constants),
4444 /// try to evaluate a few iterations of the loop until we get the exit
4445 /// condition gets a value of ExitWhen (true or false). If we cannot
4446 /// evaluate the trip count of the loop, return getCouldNotCompute().
4448 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
4451 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4452 if (PN == 0) return getCouldNotCompute();
4454 // If the loop is canonicalized, the PHI will have exactly two entries.
4455 // That's the only form we support here.
4456 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4458 // One entry must be a constant (coming in from outside of the loop), and the
4459 // second must be derived from the same PHI.
4460 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4461 Constant *StartCST =
4462 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4463 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
4465 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4466 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4467 !isa<Constant>(BEValue))
4468 return getCouldNotCompute(); // Not derived from same PHI.
4470 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4471 // the loop symbolically to determine when the condition gets a value of
4473 unsigned IterationNum = 0;
4474 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4475 for (Constant *PHIVal = StartCST;
4476 IterationNum != MaxIterations; ++IterationNum) {
4477 ConstantInt *CondVal =
4478 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
4480 // Couldn't symbolically evaluate.
4481 if (!CondVal) return getCouldNotCompute();
4483 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4484 ++NumBruteForceTripCountsComputed;
4485 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4488 // Compute the value of the PHI node for the next iteration.
4489 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4490 if (NextPHI == 0 || NextPHI == PHIVal)
4491 return getCouldNotCompute();// Couldn't evaluate or not making progress...
4495 // Too many iterations were needed to evaluate.
4496 return getCouldNotCompute();
4499 /// getSCEVAtScope - Return a SCEV expression for the specified value
4500 /// at the specified scope in the program. The L value specifies a loop
4501 /// nest to evaluate the expression at, where null is the top-level or a
4502 /// specified loop is immediately inside of the loop.
4504 /// This method can be used to compute the exit value for a variable defined
4505 /// in a loop by querying what the value will hold in the parent loop.
4507 /// In the case that a relevant loop exit value cannot be computed, the
4508 /// original value V is returned.
4509 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4510 // Check to see if we've folded this expression at this loop before.
4511 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4512 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4513 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4515 return Pair.first->second ? Pair.first->second : V;
4517 // Otherwise compute it.
4518 const SCEV *C = computeSCEVAtScope(V, L);
4519 ValuesAtScopes[V][L] = C;
4523 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
4524 if (isa<SCEVConstant>(V)) return V;
4526 // If this instruction is evolved from a constant-evolving PHI, compute the
4527 // exit value from the loop without using SCEVs.
4528 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
4529 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
4530 const Loop *LI = (*this->LI)[I->getParent()];
4531 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
4532 if (PHINode *PN = dyn_cast<PHINode>(I))
4533 if (PN->getParent() == LI->getHeader()) {
4534 // Okay, there is no closed form solution for the PHI node. Check
4535 // to see if the loop that contains it has a known backedge-taken
4536 // count. If so, we may be able to force computation of the exit
4538 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
4539 if (const SCEVConstant *BTCC =
4540 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
4541 // Okay, we know how many times the containing loop executes. If
4542 // this is a constant evolving PHI node, get the final value at
4543 // the specified iteration number.
4544 Constant *RV = getConstantEvolutionLoopExitValue(PN,
4545 BTCC->getValue()->getValue(),
4547 if (RV) return getSCEV(RV);
4551 // Okay, this is an expression that we cannot symbolically evaluate
4552 // into a SCEV. Check to see if it's possible to symbolically evaluate
4553 // the arguments into constants, and if so, try to constant propagate the
4554 // result. This is particularly useful for computing loop exit values.
4555 if (CanConstantFold(I)) {
4556 SmallVector<Constant *, 4> Operands;
4557 bool MadeImprovement = false;
4558 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4559 Value *Op = I->getOperand(i);
4560 if (Constant *C = dyn_cast<Constant>(Op)) {
4561 Operands.push_back(C);
4565 // If any of the operands is non-constant and if they are
4566 // non-integer and non-pointer, don't even try to analyze them
4567 // with scev techniques.
4568 if (!isSCEVable(Op->getType()))
4571 const SCEV *OrigV = getSCEV(Op);
4572 const SCEV *OpV = getSCEVAtScope(OrigV, L);
4573 MadeImprovement |= OrigV != OpV;
4576 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
4578 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV))
4579 C = dyn_cast<Constant>(SU->getValue());
4581 if (C->getType() != Op->getType())
4582 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4586 Operands.push_back(C);
4589 // Check to see if getSCEVAtScope actually made an improvement.
4590 if (MadeImprovement) {
4592 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4593 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4594 Operands[0], Operands[1], TD);
4596 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4597 &Operands[0], Operands.size(), TD);
4604 // This is some other type of SCEVUnknown, just return it.
4608 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4609 // Avoid performing the look-up in the common case where the specified
4610 // expression has no loop-variant portions.
4611 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4612 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4613 if (OpAtScope != Comm->getOperand(i)) {
4614 // Okay, at least one of these operands is loop variant but might be
4615 // foldable. Build a new instance of the folded commutative expression.
4616 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4617 Comm->op_begin()+i);
4618 NewOps.push_back(OpAtScope);
4620 for (++i; i != e; ++i) {
4621 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4622 NewOps.push_back(OpAtScope);
4624 if (isa<SCEVAddExpr>(Comm))
4625 return getAddExpr(NewOps);
4626 if (isa<SCEVMulExpr>(Comm))
4627 return getMulExpr(NewOps);
4628 if (isa<SCEVSMaxExpr>(Comm))
4629 return getSMaxExpr(NewOps);
4630 if (isa<SCEVUMaxExpr>(Comm))
4631 return getUMaxExpr(NewOps);
4632 llvm_unreachable("Unknown commutative SCEV type!");
4635 // If we got here, all operands are loop invariant.
4639 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4640 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4641 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4642 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4643 return Div; // must be loop invariant
4644 return getUDivExpr(LHS, RHS);
4647 // If this is a loop recurrence for a loop that does not contain L, then we
4648 // are dealing with the final value computed by the loop.
4649 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4650 // First, attempt to evaluate each operand.
4651 // Avoid performing the look-up in the common case where the specified
4652 // expression has no loop-variant portions.
4653 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
4654 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
4655 if (OpAtScope == AddRec->getOperand(i))
4658 // Okay, at least one of these operands is loop variant but might be
4659 // foldable. Build a new instance of the folded commutative expression.
4660 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
4661 AddRec->op_begin()+i);
4662 NewOps.push_back(OpAtScope);
4663 for (++i; i != e; ++i)
4664 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
4666 AddRec = cast<SCEVAddRecExpr>(getAddRecExpr(NewOps, AddRec->getLoop()));
4670 // If the scope is outside the addrec's loop, evaluate it by using the
4671 // loop exit value of the addrec.
4672 if (!AddRec->getLoop()->contains(L)) {
4673 // To evaluate this recurrence, we need to know how many times the AddRec
4674 // loop iterates. Compute this now.
4675 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4676 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4678 // Then, evaluate the AddRec.
4679 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4685 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4686 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4687 if (Op == Cast->getOperand())
4688 return Cast; // must be loop invariant
4689 return getZeroExtendExpr(Op, Cast->getType());
4692 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4693 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4694 if (Op == Cast->getOperand())
4695 return Cast; // must be loop invariant
4696 return getSignExtendExpr(Op, Cast->getType());
4699 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4700 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4701 if (Op == Cast->getOperand())
4702 return Cast; // must be loop invariant
4703 return getTruncateExpr(Op, Cast->getType());
4706 llvm_unreachable("Unknown SCEV type!");
4710 /// getSCEVAtScope - This is a convenience function which does
4711 /// getSCEVAtScope(getSCEV(V), L).
4712 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4713 return getSCEVAtScope(getSCEV(V), L);
4716 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4717 /// following equation:
4719 /// A * X = B (mod N)
4721 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4722 /// A and B isn't important.
4724 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4725 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4726 ScalarEvolution &SE) {
4727 uint32_t BW = A.getBitWidth();
4728 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4729 assert(A != 0 && "A must be non-zero.");
4733 // The gcd of A and N may have only one prime factor: 2. The number of
4734 // trailing zeros in A is its multiplicity
4735 uint32_t Mult2 = A.countTrailingZeros();
4738 // 2. Check if B is divisible by D.
4740 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4741 // is not less than multiplicity of this prime factor for D.
4742 if (B.countTrailingZeros() < Mult2)
4743 return SE.getCouldNotCompute();
4745 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4748 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4749 // bit width during computations.
4750 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4751 APInt Mod(BW + 1, 0);
4752 Mod.setBit(BW - Mult2); // Mod = N / D
4753 APInt I = AD.multiplicativeInverse(Mod);
4755 // 4. Compute the minimum unsigned root of the equation:
4756 // I * (B / D) mod (N / D)
4757 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4759 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4761 return SE.getConstant(Result.trunc(BW));
4764 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4765 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4766 /// might be the same) or two SCEVCouldNotCompute objects.
4768 static std::pair<const SCEV *,const SCEV *>
4769 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4770 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4771 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4772 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4773 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4775 // We currently can only solve this if the coefficients are constants.
4776 if (!LC || !MC || !NC) {
4777 const SCEV *CNC = SE.getCouldNotCompute();
4778 return std::make_pair(CNC, CNC);
4781 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4782 const APInt &L = LC->getValue()->getValue();
4783 const APInt &M = MC->getValue()->getValue();
4784 const APInt &N = NC->getValue()->getValue();
4785 APInt Two(BitWidth, 2);
4786 APInt Four(BitWidth, 4);
4789 using namespace APIntOps;
4791 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4792 // The B coefficient is M-N/2
4796 // The A coefficient is N/2
4797 APInt A(N.sdiv(Two));
4799 // Compute the B^2-4ac term.
4802 SqrtTerm -= Four * (A * C);
4804 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4805 // integer value or else APInt::sqrt() will assert.
4806 APInt SqrtVal(SqrtTerm.sqrt());
4808 // Compute the two solutions for the quadratic formula.
4809 // The divisions must be performed as signed divisions.
4811 APInt TwoA( A << 1 );
4812 if (TwoA.isMinValue()) {
4813 const SCEV *CNC = SE.getCouldNotCompute();
4814 return std::make_pair(CNC, CNC);
4817 LLVMContext &Context = SE.getContext();
4819 ConstantInt *Solution1 =
4820 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4821 ConstantInt *Solution2 =
4822 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4824 return std::make_pair(SE.getConstant(Solution1),
4825 SE.getConstant(Solution2));
4826 } // end APIntOps namespace
4829 /// HowFarToZero - Return the number of times a backedge comparing the specified
4830 /// value to zero will execute. If not computable, return CouldNotCompute.
4831 ScalarEvolution::BackedgeTakenInfo
4832 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4833 // If the value is a constant
4834 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4835 // If the value is already zero, the branch will execute zero times.
4836 if (C->getValue()->isZero()) return C;
4837 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4840 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4841 if (!AddRec || AddRec->getLoop() != L)
4842 return getCouldNotCompute();
4844 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4845 // the quadratic equation to solve it.
4846 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
4847 std::pair<const SCEV *,const SCEV *> Roots =
4848 SolveQuadraticEquation(AddRec, *this);
4849 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4850 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4853 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
4854 << " sol#2: " << *R2 << "\n";
4856 // Pick the smallest positive root value.
4857 if (ConstantInt *CB =
4858 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
4861 if (CB->getZExtValue() == false)
4862 std::swap(R1, R2); // R1 is the minimum root now.
4864 // We can only use this value if the chrec ends up with an exact zero
4865 // value at this index. When solving for "X*X != 5", for example, we
4866 // should not accept a root of 2.
4867 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4869 return R1; // We found a quadratic root!
4872 return getCouldNotCompute();
4875 // Otherwise we can only handle this if it is affine.
4876 if (!AddRec->isAffine())
4877 return getCouldNotCompute();
4879 // If this is an affine expression, the execution count of this branch is
4880 // the minimum unsigned root of the following equation:
4882 // Start + Step*N = 0 (mod 2^BW)
4886 // Step*N = -Start (mod 2^BW)
4888 // where BW is the common bit width of Start and Step.
4890 // Get the initial value for the loop.
4891 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
4892 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
4894 // If the AddRec is NUW, then (in an unsigned sense) it cannot be counting up
4895 // to wrap to 0, it must be counting down to equal 0. Also, while counting
4896 // down, it cannot "miss" 0 (which would cause it to wrap), regardless of what
4897 // the stride is. As such, NUW addrec's will always become zero in
4898 // "start / -stride" steps, and we know that the division is exact.
4899 if (AddRec->hasNoUnsignedWrap())
4900 // FIXME: We really want an "isexact" bit for udiv.
4901 return getUDivExpr(Start, getNegativeSCEV(Step));
4903 // For now we handle only constant steps.
4904 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
4906 return getCouldNotCompute();
4908 // First, handle unitary steps.
4909 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4910 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4912 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4913 return Start; // N = Start (as unsigned)
4915 // Then, try to solve the above equation provided that Start is constant.
4916 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4917 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4918 -StartC->getValue()->getValue(),
4920 return getCouldNotCompute();
4923 /// HowFarToNonZero - Return the number of times a backedge checking the
4924 /// specified value for nonzero will execute. If not computable, return
4926 ScalarEvolution::BackedgeTakenInfo
4927 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4928 // Loops that look like: while (X == 0) are very strange indeed. We don't
4929 // handle them yet except for the trivial case. This could be expanded in the
4930 // future as needed.
4932 // If the value is a constant, check to see if it is known to be non-zero
4933 // already. If so, the backedge will execute zero times.
4934 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4935 if (!C->getValue()->isNullValue())
4936 return getConstant(C->getType(), 0);
4937 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4940 // We could implement others, but I really doubt anyone writes loops like
4941 // this, and if they did, they would already be constant folded.
4942 return getCouldNotCompute();
4945 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4946 /// (which may not be an immediate predecessor) which has exactly one
4947 /// successor from which BB is reachable, or null if no such block is
4950 std::pair<BasicBlock *, BasicBlock *>
4951 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4952 // If the block has a unique predecessor, then there is no path from the
4953 // predecessor to the block that does not go through the direct edge
4954 // from the predecessor to the block.
4955 if (BasicBlock *Pred = BB->getSinglePredecessor())
4956 return std::make_pair(Pred, BB);
4958 // A loop's header is defined to be a block that dominates the loop.
4959 // If the header has a unique predecessor outside the loop, it must be
4960 // a block that has exactly one successor that can reach the loop.
4961 if (Loop *L = LI->getLoopFor(BB))
4962 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
4964 return std::pair<BasicBlock *, BasicBlock *>();
4967 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4968 /// testing whether two expressions are equal, however for the purposes of
4969 /// looking for a condition guarding a loop, it can be useful to be a little
4970 /// more general, since a front-end may have replicated the controlling
4973 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4974 // Quick check to see if they are the same SCEV.
4975 if (A == B) return true;
4977 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4978 // two different instructions with the same value. Check for this case.
4979 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4980 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4981 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4982 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4983 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4986 // Otherwise assume they may have a different value.
4990 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
4991 /// predicate Pred. Return true iff any changes were made.
4993 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
4994 const SCEV *&LHS, const SCEV *&RHS) {
4995 bool Changed = false;
4997 // Canonicalize a constant to the right side.
4998 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
4999 // Check for both operands constant.
5000 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
5001 if (ConstantExpr::getICmp(Pred,
5003 RHSC->getValue())->isNullValue())
5004 goto trivially_false;
5006 goto trivially_true;
5008 // Otherwise swap the operands to put the constant on the right.
5009 std::swap(LHS, RHS);
5010 Pred = ICmpInst::getSwappedPredicate(Pred);
5014 // If we're comparing an addrec with a value which is loop-invariant in the
5015 // addrec's loop, put the addrec on the left. Also make a dominance check,
5016 // as both operands could be addrecs loop-invariant in each other's loop.
5017 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
5018 const Loop *L = AR->getLoop();
5019 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
5020 std::swap(LHS, RHS);
5021 Pred = ICmpInst::getSwappedPredicate(Pred);
5026 // If there's a constant operand, canonicalize comparisons with boundary
5027 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
5028 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
5029 const APInt &RA = RC->getValue()->getValue();
5031 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5032 case ICmpInst::ICMP_EQ:
5033 case ICmpInst::ICMP_NE:
5035 case ICmpInst::ICMP_UGE:
5036 if ((RA - 1).isMinValue()) {
5037 Pred = ICmpInst::ICMP_NE;
5038 RHS = getConstant(RA - 1);
5042 if (RA.isMaxValue()) {
5043 Pred = ICmpInst::ICMP_EQ;
5047 if (RA.isMinValue()) goto trivially_true;
5049 Pred = ICmpInst::ICMP_UGT;
5050 RHS = getConstant(RA - 1);
5053 case ICmpInst::ICMP_ULE:
5054 if ((RA + 1).isMaxValue()) {
5055 Pred = ICmpInst::ICMP_NE;
5056 RHS = getConstant(RA + 1);
5060 if (RA.isMinValue()) {
5061 Pred = ICmpInst::ICMP_EQ;
5065 if (RA.isMaxValue()) goto trivially_true;
5067 Pred = ICmpInst::ICMP_ULT;
5068 RHS = getConstant(RA + 1);
5071 case ICmpInst::ICMP_SGE:
5072 if ((RA - 1).isMinSignedValue()) {
5073 Pred = ICmpInst::ICMP_NE;
5074 RHS = getConstant(RA - 1);
5078 if (RA.isMaxSignedValue()) {
5079 Pred = ICmpInst::ICMP_EQ;
5083 if (RA.isMinSignedValue()) goto trivially_true;
5085 Pred = ICmpInst::ICMP_SGT;
5086 RHS = getConstant(RA - 1);
5089 case ICmpInst::ICMP_SLE:
5090 if ((RA + 1).isMaxSignedValue()) {
5091 Pred = ICmpInst::ICMP_NE;
5092 RHS = getConstant(RA + 1);
5096 if (RA.isMinSignedValue()) {
5097 Pred = ICmpInst::ICMP_EQ;
5101 if (RA.isMaxSignedValue()) goto trivially_true;
5103 Pred = ICmpInst::ICMP_SLT;
5104 RHS = getConstant(RA + 1);
5107 case ICmpInst::ICMP_UGT:
5108 if (RA.isMinValue()) {
5109 Pred = ICmpInst::ICMP_NE;
5113 if ((RA + 1).isMaxValue()) {
5114 Pred = ICmpInst::ICMP_EQ;
5115 RHS = getConstant(RA + 1);
5119 if (RA.isMaxValue()) goto trivially_false;
5121 case ICmpInst::ICMP_ULT:
5122 if (RA.isMaxValue()) {
5123 Pred = ICmpInst::ICMP_NE;
5127 if ((RA - 1).isMinValue()) {
5128 Pred = ICmpInst::ICMP_EQ;
5129 RHS = getConstant(RA - 1);
5133 if (RA.isMinValue()) goto trivially_false;
5135 case ICmpInst::ICMP_SGT:
5136 if (RA.isMinSignedValue()) {
5137 Pred = ICmpInst::ICMP_NE;
5141 if ((RA + 1).isMaxSignedValue()) {
5142 Pred = ICmpInst::ICMP_EQ;
5143 RHS = getConstant(RA + 1);
5147 if (RA.isMaxSignedValue()) goto trivially_false;
5149 case ICmpInst::ICMP_SLT:
5150 if (RA.isMaxSignedValue()) {
5151 Pred = ICmpInst::ICMP_NE;
5155 if ((RA - 1).isMinSignedValue()) {
5156 Pred = ICmpInst::ICMP_EQ;
5157 RHS = getConstant(RA - 1);
5161 if (RA.isMinSignedValue()) goto trivially_false;
5166 // Check for obvious equality.
5167 if (HasSameValue(LHS, RHS)) {
5168 if (ICmpInst::isTrueWhenEqual(Pred))
5169 goto trivially_true;
5170 if (ICmpInst::isFalseWhenEqual(Pred))
5171 goto trivially_false;
5174 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5175 // adding or subtracting 1 from one of the operands.
5177 case ICmpInst::ICMP_SLE:
5178 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5179 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5180 /*HasNUW=*/false, /*HasNSW=*/true);
5181 Pred = ICmpInst::ICMP_SLT;
5183 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5184 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5185 /*HasNUW=*/false, /*HasNSW=*/true);
5186 Pred = ICmpInst::ICMP_SLT;
5190 case ICmpInst::ICMP_SGE:
5191 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5192 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5193 /*HasNUW=*/false, /*HasNSW=*/true);
5194 Pred = ICmpInst::ICMP_SGT;
5196 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5197 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5198 /*HasNUW=*/false, /*HasNSW=*/true);
5199 Pred = ICmpInst::ICMP_SGT;
5203 case ICmpInst::ICMP_ULE:
5204 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5205 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5206 /*HasNUW=*/true, /*HasNSW=*/false);
5207 Pred = ICmpInst::ICMP_ULT;
5209 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5210 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5211 /*HasNUW=*/true, /*HasNSW=*/false);
5212 Pred = ICmpInst::ICMP_ULT;
5216 case ICmpInst::ICMP_UGE:
5217 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5218 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5219 /*HasNUW=*/true, /*HasNSW=*/false);
5220 Pred = ICmpInst::ICMP_UGT;
5222 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5223 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5224 /*HasNUW=*/true, /*HasNSW=*/false);
5225 Pred = ICmpInst::ICMP_UGT;
5233 // TODO: More simplifications are possible here.
5239 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5240 Pred = ICmpInst::ICMP_EQ;
5245 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5246 Pred = ICmpInst::ICMP_NE;
5250 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5251 return getSignedRange(S).getSignedMax().isNegative();
5254 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5255 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5258 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5259 return !getSignedRange(S).getSignedMin().isNegative();
5262 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5263 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5266 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5267 return isKnownNegative(S) || isKnownPositive(S);
5270 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5271 const SCEV *LHS, const SCEV *RHS) {
5272 // Canonicalize the inputs first.
5273 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5275 // If LHS or RHS is an addrec, check to see if the condition is true in
5276 // every iteration of the loop.
5277 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5278 if (isLoopEntryGuardedByCond(
5279 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5280 isLoopBackedgeGuardedByCond(
5281 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5283 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5284 if (isLoopEntryGuardedByCond(
5285 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5286 isLoopBackedgeGuardedByCond(
5287 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5290 // Otherwise see what can be done with known constant ranges.
5291 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5295 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5296 const SCEV *LHS, const SCEV *RHS) {
5297 if (HasSameValue(LHS, RHS))
5298 return ICmpInst::isTrueWhenEqual(Pred);
5300 // This code is split out from isKnownPredicate because it is called from
5301 // within isLoopEntryGuardedByCond.
5304 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5306 case ICmpInst::ICMP_SGT:
5307 Pred = ICmpInst::ICMP_SLT;
5308 std::swap(LHS, RHS);
5309 case ICmpInst::ICMP_SLT: {
5310 ConstantRange LHSRange = getSignedRange(LHS);
5311 ConstantRange RHSRange = getSignedRange(RHS);
5312 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5314 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5318 case ICmpInst::ICMP_SGE:
5319 Pred = ICmpInst::ICMP_SLE;
5320 std::swap(LHS, RHS);
5321 case ICmpInst::ICMP_SLE: {
5322 ConstantRange LHSRange = getSignedRange(LHS);
5323 ConstantRange RHSRange = getSignedRange(RHS);
5324 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5326 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
5330 case ICmpInst::ICMP_UGT:
5331 Pred = ICmpInst::ICMP_ULT;
5332 std::swap(LHS, RHS);
5333 case ICmpInst::ICMP_ULT: {
5334 ConstantRange LHSRange = getUnsignedRange(LHS);
5335 ConstantRange RHSRange = getUnsignedRange(RHS);
5336 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
5338 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
5342 case ICmpInst::ICMP_UGE:
5343 Pred = ICmpInst::ICMP_ULE;
5344 std::swap(LHS, RHS);
5345 case ICmpInst::ICMP_ULE: {
5346 ConstantRange LHSRange = getUnsignedRange(LHS);
5347 ConstantRange RHSRange = getUnsignedRange(RHS);
5348 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
5350 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
5354 case ICmpInst::ICMP_NE: {
5355 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
5357 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
5360 const SCEV *Diff = getMinusSCEV(LHS, RHS);
5361 if (isKnownNonZero(Diff))
5365 case ICmpInst::ICMP_EQ:
5366 // The check at the top of the function catches the case where
5367 // the values are known to be equal.
5373 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
5374 /// protected by a conditional between LHS and RHS. This is used to
5375 /// to eliminate casts.
5377 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
5378 ICmpInst::Predicate Pred,
5379 const SCEV *LHS, const SCEV *RHS) {
5380 // Interpret a null as meaning no loop, where there is obviously no guard
5381 // (interprocedural conditions notwithstanding).
5382 if (!L) return true;
5384 BasicBlock *Latch = L->getLoopLatch();
5388 BranchInst *LoopContinuePredicate =
5389 dyn_cast<BranchInst>(Latch->getTerminator());
5390 if (!LoopContinuePredicate ||
5391 LoopContinuePredicate->isUnconditional())
5394 return isImpliedCond(Pred, LHS, RHS,
5395 LoopContinuePredicate->getCondition(),
5396 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
5399 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
5400 /// by a conditional between LHS and RHS. This is used to help avoid max
5401 /// expressions in loop trip counts, and to eliminate casts.
5403 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
5404 ICmpInst::Predicate Pred,
5405 const SCEV *LHS, const SCEV *RHS) {
5406 // Interpret a null as meaning no loop, where there is obviously no guard
5407 // (interprocedural conditions notwithstanding).
5408 if (!L) return false;
5410 // Starting at the loop predecessor, climb up the predecessor chain, as long
5411 // as there are predecessors that can be found that have unique successors
5412 // leading to the original header.
5413 for (std::pair<BasicBlock *, BasicBlock *>
5414 Pair(L->getLoopPredecessor(), L->getHeader());
5416 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
5418 BranchInst *LoopEntryPredicate =
5419 dyn_cast<BranchInst>(Pair.first->getTerminator());
5420 if (!LoopEntryPredicate ||
5421 LoopEntryPredicate->isUnconditional())
5424 if (isImpliedCond(Pred, LHS, RHS,
5425 LoopEntryPredicate->getCondition(),
5426 LoopEntryPredicate->getSuccessor(0) != Pair.second))
5433 /// isImpliedCond - Test whether the condition described by Pred, LHS,
5434 /// and RHS is true whenever the given Cond value evaluates to true.
5435 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
5436 const SCEV *LHS, const SCEV *RHS,
5437 Value *FoundCondValue,
5439 // Recursively handle And and Or conditions.
5440 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
5441 if (BO->getOpcode() == Instruction::And) {
5443 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5444 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5445 } else if (BO->getOpcode() == Instruction::Or) {
5447 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5448 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5452 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
5453 if (!ICI) return false;
5455 // Bail if the ICmp's operands' types are wider than the needed type
5456 // before attempting to call getSCEV on them. This avoids infinite
5457 // recursion, since the analysis of widening casts can require loop
5458 // exit condition information for overflow checking, which would
5460 if (getTypeSizeInBits(LHS->getType()) <
5461 getTypeSizeInBits(ICI->getOperand(0)->getType()))
5464 // Now that we found a conditional branch that dominates the loop, check to
5465 // see if it is the comparison we are looking for.
5466 ICmpInst::Predicate FoundPred;
5468 FoundPred = ICI->getInversePredicate();
5470 FoundPred = ICI->getPredicate();
5472 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
5473 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
5475 // Balance the types. The case where FoundLHS' type is wider than
5476 // LHS' type is checked for above.
5477 if (getTypeSizeInBits(LHS->getType()) >
5478 getTypeSizeInBits(FoundLHS->getType())) {
5479 if (CmpInst::isSigned(Pred)) {
5480 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
5481 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
5483 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
5484 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
5488 // Canonicalize the query to match the way instcombine will have
5489 // canonicalized the comparison.
5490 if (SimplifyICmpOperands(Pred, LHS, RHS))
5492 return CmpInst::isTrueWhenEqual(Pred);
5493 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
5494 if (FoundLHS == FoundRHS)
5495 return CmpInst::isFalseWhenEqual(Pred);
5497 // Check to see if we can make the LHS or RHS match.
5498 if (LHS == FoundRHS || RHS == FoundLHS) {
5499 if (isa<SCEVConstant>(RHS)) {
5500 std::swap(FoundLHS, FoundRHS);
5501 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
5503 std::swap(LHS, RHS);
5504 Pred = ICmpInst::getSwappedPredicate(Pred);
5508 // Check whether the found predicate is the same as the desired predicate.
5509 if (FoundPred == Pred)
5510 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
5512 // Check whether swapping the found predicate makes it the same as the
5513 // desired predicate.
5514 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
5515 if (isa<SCEVConstant>(RHS))
5516 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
5518 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
5519 RHS, LHS, FoundLHS, FoundRHS);
5522 // Check whether the actual condition is beyond sufficient.
5523 if (FoundPred == ICmpInst::ICMP_EQ)
5524 if (ICmpInst::isTrueWhenEqual(Pred))
5525 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
5527 if (Pred == ICmpInst::ICMP_NE)
5528 if (!ICmpInst::isTrueWhenEqual(FoundPred))
5529 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
5532 // Otherwise assume the worst.
5536 /// isImpliedCondOperands - Test whether the condition described by Pred,
5537 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
5538 /// and FoundRHS is true.
5539 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
5540 const SCEV *LHS, const SCEV *RHS,
5541 const SCEV *FoundLHS,
5542 const SCEV *FoundRHS) {
5543 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
5544 FoundLHS, FoundRHS) ||
5545 // ~x < ~y --> x > y
5546 isImpliedCondOperandsHelper(Pred, LHS, RHS,
5547 getNotSCEV(FoundRHS),
5548 getNotSCEV(FoundLHS));
5551 /// isImpliedCondOperandsHelper - Test whether the condition described by
5552 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
5553 /// FoundLHS, and FoundRHS is true.
5555 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
5556 const SCEV *LHS, const SCEV *RHS,
5557 const SCEV *FoundLHS,
5558 const SCEV *FoundRHS) {
5560 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5561 case ICmpInst::ICMP_EQ:
5562 case ICmpInst::ICMP_NE:
5563 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
5566 case ICmpInst::ICMP_SLT:
5567 case ICmpInst::ICMP_SLE:
5568 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
5569 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
5572 case ICmpInst::ICMP_SGT:
5573 case ICmpInst::ICMP_SGE:
5574 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
5575 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
5578 case ICmpInst::ICMP_ULT:
5579 case ICmpInst::ICMP_ULE:
5580 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
5581 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
5584 case ICmpInst::ICMP_UGT:
5585 case ICmpInst::ICMP_UGE:
5586 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
5587 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
5595 /// getBECount - Subtract the end and start values and divide by the step,
5596 /// rounding up, to get the number of times the backedge is executed. Return
5597 /// CouldNotCompute if an intermediate computation overflows.
5598 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
5602 assert(!isKnownNegative(Step) &&
5603 "This code doesn't handle negative strides yet!");
5605 const Type *Ty = Start->getType();
5606 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
5607 const SCEV *Diff = getMinusSCEV(End, Start);
5608 const SCEV *RoundUp = getAddExpr(Step, NegOne);
5610 // Add an adjustment to the difference between End and Start so that
5611 // the division will effectively round up.
5612 const SCEV *Add = getAddExpr(Diff, RoundUp);
5615 // Check Add for unsigned overflow.
5616 // TODO: More sophisticated things could be done here.
5617 const Type *WideTy = IntegerType::get(getContext(),
5618 getTypeSizeInBits(Ty) + 1);
5619 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
5620 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
5621 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
5622 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
5623 return getCouldNotCompute();
5626 return getUDivExpr(Add, Step);
5629 /// HowManyLessThans - Return the number of times a backedge containing the
5630 /// specified less-than comparison will execute. If not computable, return
5631 /// CouldNotCompute.
5632 ScalarEvolution::BackedgeTakenInfo
5633 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
5634 const Loop *L, bool isSigned) {
5635 // Only handle: "ADDREC < LoopInvariant".
5636 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute();
5638 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
5639 if (!AddRec || AddRec->getLoop() != L)
5640 return getCouldNotCompute();
5642 // Check to see if we have a flag which makes analysis easy.
5643 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
5644 AddRec->hasNoUnsignedWrap();
5646 if (AddRec->isAffine()) {
5647 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
5648 const SCEV *Step = AddRec->getStepRecurrence(*this);
5651 return getCouldNotCompute();
5652 if (Step->isOne()) {
5653 // With unit stride, the iteration never steps past the limit value.
5654 } else if (isKnownPositive(Step)) {
5655 // Test whether a positive iteration can step past the limit
5656 // value and past the maximum value for its type in a single step.
5657 // Note that it's not sufficient to check NoWrap here, because even
5658 // though the value after a wrap is undefined, it's not undefined
5659 // behavior, so if wrap does occur, the loop could either terminate or
5660 // loop infinitely, but in either case, the loop is guaranteed to
5661 // iterate at least until the iteration where the wrapping occurs.
5662 const SCEV *One = getConstant(Step->getType(), 1);
5664 APInt Max = APInt::getSignedMaxValue(BitWidth);
5665 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
5666 .slt(getSignedRange(RHS).getSignedMax()))
5667 return getCouldNotCompute();
5669 APInt Max = APInt::getMaxValue(BitWidth);
5670 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
5671 .ult(getUnsignedRange(RHS).getUnsignedMax()))
5672 return getCouldNotCompute();
5675 // TODO: Handle negative strides here and below.
5676 return getCouldNotCompute();
5678 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
5679 // m. So, we count the number of iterations in which {n,+,s} < m is true.
5680 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
5681 // treat m-n as signed nor unsigned due to overflow possibility.
5683 // First, we get the value of the LHS in the first iteration: n
5684 const SCEV *Start = AddRec->getOperand(0);
5686 // Determine the minimum constant start value.
5687 const SCEV *MinStart = getConstant(isSigned ?
5688 getSignedRange(Start).getSignedMin() :
5689 getUnsignedRange(Start).getUnsignedMin());
5691 // If we know that the condition is true in order to enter the loop,
5692 // then we know that it will run exactly (m-n)/s times. Otherwise, we
5693 // only know that it will execute (max(m,n)-n)/s times. In both cases,
5694 // the division must round up.
5695 const SCEV *End = RHS;
5696 if (!isLoopEntryGuardedByCond(L,
5697 isSigned ? ICmpInst::ICMP_SLT :
5699 getMinusSCEV(Start, Step), RHS))
5700 End = isSigned ? getSMaxExpr(RHS, Start)
5701 : getUMaxExpr(RHS, Start);
5703 // Determine the maximum constant end value.
5704 const SCEV *MaxEnd = getConstant(isSigned ?
5705 getSignedRange(End).getSignedMax() :
5706 getUnsignedRange(End).getUnsignedMax());
5708 // If MaxEnd is within a step of the maximum integer value in its type,
5709 // adjust it down to the minimum value which would produce the same effect.
5710 // This allows the subsequent ceiling division of (N+(step-1))/step to
5711 // compute the correct value.
5712 const SCEV *StepMinusOne = getMinusSCEV(Step,
5713 getConstant(Step->getType(), 1));
5716 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
5719 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
5722 // Finally, we subtract these two values and divide, rounding up, to get
5723 // the number of times the backedge is executed.
5724 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
5726 // The maximum backedge count is similar, except using the minimum start
5727 // value and the maximum end value.
5728 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
5730 return BackedgeTakenInfo(BECount, MaxBECount);
5733 return getCouldNotCompute();
5736 /// getNumIterationsInRange - Return the number of iterations of this loop that
5737 /// produce values in the specified constant range. Another way of looking at
5738 /// this is that it returns the first iteration number where the value is not in
5739 /// the condition, thus computing the exit count. If the iteration count can't
5740 /// be computed, an instance of SCEVCouldNotCompute is returned.
5741 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
5742 ScalarEvolution &SE) const {
5743 if (Range.isFullSet()) // Infinite loop.
5744 return SE.getCouldNotCompute();
5746 // If the start is a non-zero constant, shift the range to simplify things.
5747 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
5748 if (!SC->getValue()->isZero()) {
5749 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
5750 Operands[0] = SE.getConstant(SC->getType(), 0);
5751 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
5752 if (const SCEVAddRecExpr *ShiftedAddRec =
5753 dyn_cast<SCEVAddRecExpr>(Shifted))
5754 return ShiftedAddRec->getNumIterationsInRange(
5755 Range.subtract(SC->getValue()->getValue()), SE);
5756 // This is strange and shouldn't happen.
5757 return SE.getCouldNotCompute();
5760 // The only time we can solve this is when we have all constant indices.
5761 // Otherwise, we cannot determine the overflow conditions.
5762 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5763 if (!isa<SCEVConstant>(getOperand(i)))
5764 return SE.getCouldNotCompute();
5767 // Okay at this point we know that all elements of the chrec are constants and
5768 // that the start element is zero.
5770 // First check to see if the range contains zero. If not, the first
5772 unsigned BitWidth = SE.getTypeSizeInBits(getType());
5773 if (!Range.contains(APInt(BitWidth, 0)))
5774 return SE.getConstant(getType(), 0);
5777 // If this is an affine expression then we have this situation:
5778 // Solve {0,+,A} in Range === Ax in Range
5780 // We know that zero is in the range. If A is positive then we know that
5781 // the upper value of the range must be the first possible exit value.
5782 // If A is negative then the lower of the range is the last possible loop
5783 // value. Also note that we already checked for a full range.
5784 APInt One(BitWidth,1);
5785 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5786 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5788 // The exit value should be (End+A)/A.
5789 APInt ExitVal = (End + A).udiv(A);
5790 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5792 // Evaluate at the exit value. If we really did fall out of the valid
5793 // range, then we computed our trip count, otherwise wrap around or other
5794 // things must have happened.
5795 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5796 if (Range.contains(Val->getValue()))
5797 return SE.getCouldNotCompute(); // Something strange happened
5799 // Ensure that the previous value is in the range. This is a sanity check.
5800 assert(Range.contains(
5801 EvaluateConstantChrecAtConstant(this,
5802 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5803 "Linear scev computation is off in a bad way!");
5804 return SE.getConstant(ExitValue);
5805 } else if (isQuadratic()) {
5806 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5807 // quadratic equation to solve it. To do this, we must frame our problem in
5808 // terms of figuring out when zero is crossed, instead of when
5809 // Range.getUpper() is crossed.
5810 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5811 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5812 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5814 // Next, solve the constructed addrec
5815 std::pair<const SCEV *,const SCEV *> Roots =
5816 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5817 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5818 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5820 // Pick the smallest positive root value.
5821 if (ConstantInt *CB =
5822 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5823 R1->getValue(), R2->getValue()))) {
5824 if (CB->getZExtValue() == false)
5825 std::swap(R1, R2); // R1 is the minimum root now.
5827 // Make sure the root is not off by one. The returned iteration should
5828 // not be in the range, but the previous one should be. When solving
5829 // for "X*X < 5", for example, we should not return a root of 2.
5830 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5833 if (Range.contains(R1Val->getValue())) {
5834 // The next iteration must be out of the range...
5835 ConstantInt *NextVal =
5836 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5838 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5839 if (!Range.contains(R1Val->getValue()))
5840 return SE.getConstant(NextVal);
5841 return SE.getCouldNotCompute(); // Something strange happened
5844 // If R1 was not in the range, then it is a good return value. Make
5845 // sure that R1-1 WAS in the range though, just in case.
5846 ConstantInt *NextVal =
5847 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5848 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5849 if (Range.contains(R1Val->getValue()))
5851 return SE.getCouldNotCompute(); // Something strange happened
5856 return SE.getCouldNotCompute();
5861 //===----------------------------------------------------------------------===//
5862 // SCEVCallbackVH Class Implementation
5863 //===----------------------------------------------------------------------===//
5865 void ScalarEvolution::SCEVCallbackVH::deleted() {
5866 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5867 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5868 SE->ConstantEvolutionLoopExitValue.erase(PN);
5869 SE->ValueExprMap.erase(getValPtr());
5870 // this now dangles!
5873 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
5874 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5876 // Forget all the expressions associated with users of the old value,
5877 // so that future queries will recompute the expressions using the new
5879 Value *Old = getValPtr();
5880 SmallVector<User *, 16> Worklist;
5881 SmallPtrSet<User *, 8> Visited;
5882 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5884 Worklist.push_back(*UI);
5885 while (!Worklist.empty()) {
5886 User *U = Worklist.pop_back_val();
5887 // Deleting the Old value will cause this to dangle. Postpone
5888 // that until everything else is done.
5891 if (!Visited.insert(U))
5893 if (PHINode *PN = dyn_cast<PHINode>(U))
5894 SE->ConstantEvolutionLoopExitValue.erase(PN);
5895 SE->ValueExprMap.erase(U);
5896 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5898 Worklist.push_back(*UI);
5900 // Delete the Old value.
5901 if (PHINode *PN = dyn_cast<PHINode>(Old))
5902 SE->ConstantEvolutionLoopExitValue.erase(PN);
5903 SE->ValueExprMap.erase(Old);
5904 // this now dangles!
5907 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5908 : CallbackVH(V), SE(se) {}
5910 //===----------------------------------------------------------------------===//
5911 // ScalarEvolution Class Implementation
5912 //===----------------------------------------------------------------------===//
5914 ScalarEvolution::ScalarEvolution()
5915 : FunctionPass(ID), FirstUnknown(0) {
5916 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
5919 bool ScalarEvolution::runOnFunction(Function &F) {
5921 LI = &getAnalysis<LoopInfo>();
5922 TD = getAnalysisIfAvailable<TargetData>();
5923 DT = &getAnalysis<DominatorTree>();
5927 void ScalarEvolution::releaseMemory() {
5928 // Iterate through all the SCEVUnknown instances and call their
5929 // destructors, so that they release their references to their values.
5930 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
5934 ValueExprMap.clear();
5935 BackedgeTakenCounts.clear();
5936 ConstantEvolutionLoopExitValue.clear();
5937 ValuesAtScopes.clear();
5938 LoopDispositions.clear();
5939 BlockDispositions.clear();
5940 UnsignedRanges.clear();
5941 SignedRanges.clear();
5942 UniqueSCEVs.clear();
5943 SCEVAllocator.Reset();
5946 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5947 AU.setPreservesAll();
5948 AU.addRequiredTransitive<LoopInfo>();
5949 AU.addRequiredTransitive<DominatorTree>();
5952 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5953 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5956 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5958 // Print all inner loops first
5959 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5960 PrintLoopInfo(OS, SE, *I);
5963 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5966 SmallVector<BasicBlock *, 8> ExitBlocks;
5967 L->getExitBlocks(ExitBlocks);
5968 if (ExitBlocks.size() != 1)
5969 OS << "<multiple exits> ";
5971 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5972 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5974 OS << "Unpredictable backedge-taken count. ";
5979 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5982 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5983 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5985 OS << "Unpredictable max backedge-taken count. ";
5991 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
5992 // ScalarEvolution's implementation of the print method is to print
5993 // out SCEV values of all instructions that are interesting. Doing
5994 // this potentially causes it to create new SCEV objects though,
5995 // which technically conflicts with the const qualifier. This isn't
5996 // observable from outside the class though, so casting away the
5997 // const isn't dangerous.
5998 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
6000 OS << "Classifying expressions for: ";
6001 WriteAsOperand(OS, F, /*PrintType=*/false);
6003 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
6004 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
6007 const SCEV *SV = SE.getSCEV(&*I);
6010 const Loop *L = LI->getLoopFor((*I).getParent());
6012 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
6019 OS << "\t\t" "Exits: ";
6020 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
6021 if (!SE.isLoopInvariant(ExitValue, L)) {
6022 OS << "<<Unknown>>";
6031 OS << "Determining loop execution counts for: ";
6032 WriteAsOperand(OS, F, /*PrintType=*/false);
6034 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6035 PrintLoopInfo(OS, &SE, *I);
6038 ScalarEvolution::LoopDisposition
6039 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
6040 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S];
6041 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair =
6042 Values.insert(std::make_pair(L, LoopVariant));
6044 return Pair.first->second;
6046 LoopDisposition D = computeLoopDisposition(S, L);
6047 return LoopDispositions[S][L] = D;
6050 ScalarEvolution::LoopDisposition
6051 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
6052 switch (S->getSCEVType()) {
6054 return LoopInvariant;
6058 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
6059 case scAddRecExpr: {
6060 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6062 // If L is the addrec's loop, it's computable.
6063 if (AR->getLoop() == L)
6064 return LoopComputable;
6066 // Add recurrences are never invariant in the function-body (null loop).
6070 // This recurrence is variant w.r.t. L if L contains AR's loop.
6071 if (L->contains(AR->getLoop()))
6074 // This recurrence is invariant w.r.t. L if AR's loop contains L.
6075 if (AR->getLoop()->contains(L))
6076 return LoopInvariant;
6078 // This recurrence is variant w.r.t. L if any of its operands
6080 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
6082 if (!isLoopInvariant(*I, L))
6085 // Otherwise it's loop-invariant.
6086 return LoopInvariant;
6092 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6093 bool HasVarying = false;
6094 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6096 LoopDisposition D = getLoopDisposition(*I, L);
6097 if (D == LoopVariant)
6099 if (D == LoopComputable)
6102 return HasVarying ? LoopComputable : LoopInvariant;
6105 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6106 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
6107 if (LD == LoopVariant)
6109 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
6110 if (RD == LoopVariant)
6112 return (LD == LoopInvariant && RD == LoopInvariant) ?
6113 LoopInvariant : LoopComputable;
6116 // All non-instruction values are loop invariant. All instructions are loop
6117 // invariant if they are not contained in the specified loop.
6118 // Instructions are never considered invariant in the function body
6119 // (null loop) because they are defined within the "loop".
6120 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
6121 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
6122 return LoopInvariant;
6123 case scCouldNotCompute:
6124 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6128 llvm_unreachable("Unknown SCEV kind!");
6132 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
6133 return getLoopDisposition(S, L) == LoopInvariant;
6136 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
6137 return getLoopDisposition(S, L) == LoopComputable;
6140 ScalarEvolution::BlockDisposition
6141 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6142 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S];
6143 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool>
6144 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock));
6146 return Pair.first->second;
6148 BlockDisposition D = computeBlockDisposition(S, BB);
6149 return BlockDispositions[S][BB] = D;
6152 ScalarEvolution::BlockDisposition
6153 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6154 switch (S->getSCEVType()) {
6156 return ProperlyDominatesBlock;
6160 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
6161 case scAddRecExpr: {
6162 // This uses a "dominates" query instead of "properly dominates" query
6163 // to test for proper dominance too, because the instruction which
6164 // produces the addrec's value is a PHI, and a PHI effectively properly
6165 // dominates its entire containing block.
6166 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6167 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
6168 return DoesNotDominateBlock;
6170 // FALL THROUGH into SCEVNAryExpr handling.
6175 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6177 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6179 BlockDisposition D = getBlockDisposition(*I, BB);
6180 if (D == DoesNotDominateBlock)
6181 return DoesNotDominateBlock;
6182 if (D == DominatesBlock)
6185 return Proper ? ProperlyDominatesBlock : DominatesBlock;
6188 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6189 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6190 BlockDisposition LD = getBlockDisposition(LHS, BB);
6191 if (LD == DoesNotDominateBlock)
6192 return DoesNotDominateBlock;
6193 BlockDisposition RD = getBlockDisposition(RHS, BB);
6194 if (RD == DoesNotDominateBlock)
6195 return DoesNotDominateBlock;
6196 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
6197 ProperlyDominatesBlock : DominatesBlock;
6200 if (Instruction *I =
6201 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
6202 if (I->getParent() == BB)
6203 return DominatesBlock;
6204 if (DT->properlyDominates(I->getParent(), BB))
6205 return ProperlyDominatesBlock;
6206 return DoesNotDominateBlock;
6208 return ProperlyDominatesBlock;
6209 case scCouldNotCompute:
6210 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6211 return DoesNotDominateBlock;
6214 llvm_unreachable("Unknown SCEV kind!");
6215 return DoesNotDominateBlock;
6218 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
6219 return getBlockDisposition(S, BB) >= DominatesBlock;
6222 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
6223 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
6226 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
6227 switch (S->getSCEVType()) {
6232 case scSignExtend: {
6233 const SCEVCastExpr *Cast = cast<SCEVCastExpr>(S);
6234 const SCEV *CastOp = Cast->getOperand();
6235 return Op == CastOp || hasOperand(CastOp, Op);
6242 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6243 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6245 const SCEV *NAryOp = *I;
6246 if (NAryOp == Op || hasOperand(NAryOp, Op))
6252 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6253 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6254 return LHS == Op || hasOperand(LHS, Op) ||
6255 RHS == Op || hasOperand(RHS, Op);
6259 case scCouldNotCompute:
6260 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6264 llvm_unreachable("Unknown SCEV kind!");
6268 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
6269 ValuesAtScopes.erase(S);
6270 LoopDispositions.erase(S);
6271 BlockDispositions.erase(S);
6272 UnsignedRanges.erase(S);
6273 SignedRanges.erase(S);