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/Target/TargetLibraryInfo.h"
78 #include "llvm/Support/CommandLine.h"
79 #include "llvm/Support/ConstantRange.h"
80 #include "llvm/Support/Debug.h"
81 #include "llvm/Support/ErrorHandling.h"
82 #include "llvm/Support/GetElementPtrTypeIterator.h"
83 #include "llvm/Support/InstIterator.h"
84 #include "llvm/Support/MathExtras.h"
85 #include "llvm/Support/raw_ostream.h"
86 #include "llvm/ADT/Statistic.h"
87 #include "llvm/ADT/STLExtras.h"
88 #include "llvm/ADT/SmallPtrSet.h"
92 STATISTIC(NumArrayLenItCounts,
93 "Number of trip counts computed with array length");
94 STATISTIC(NumTripCountsComputed,
95 "Number of loops with predictable loop counts");
96 STATISTIC(NumTripCountsNotComputed,
97 "Number of loops without predictable loop counts");
98 STATISTIC(NumBruteForceTripCountsComputed,
99 "Number of loops with trip counts computed by force");
101 static cl::opt<unsigned>
102 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
103 cl::desc("Maximum number of iterations SCEV will "
104 "symbolically execute a constant "
108 INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
109 "Scalar Evolution Analysis", false, true)
110 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
111 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
112 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
113 INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
114 "Scalar Evolution Analysis", false, true)
115 char ScalarEvolution::ID = 0;
117 //===----------------------------------------------------------------------===//
118 // SCEV class definitions
119 //===----------------------------------------------------------------------===//
121 //===----------------------------------------------------------------------===//
122 // Implementation of the SCEV class.
125 void SCEV::dump() const {
130 void SCEV::print(raw_ostream &OS) const {
131 switch (getSCEVType()) {
133 WriteAsOperand(OS, cast<SCEVConstant>(this)->getValue(), false);
136 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(this);
137 const SCEV *Op = Trunc->getOperand();
138 OS << "(trunc " << *Op->getType() << " " << *Op << " to "
139 << *Trunc->getType() << ")";
143 const SCEVZeroExtendExpr *ZExt = cast<SCEVZeroExtendExpr>(this);
144 const SCEV *Op = ZExt->getOperand();
145 OS << "(zext " << *Op->getType() << " " << *Op << " to "
146 << *ZExt->getType() << ")";
150 const SCEVSignExtendExpr *SExt = cast<SCEVSignExtendExpr>(this);
151 const SCEV *Op = SExt->getOperand();
152 OS << "(sext " << *Op->getType() << " " << *Op << " to "
153 << *SExt->getType() << ")";
157 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(this);
158 OS << "{" << *AR->getOperand(0);
159 for (unsigned i = 1, e = AR->getNumOperands(); i != e; ++i)
160 OS << ",+," << *AR->getOperand(i);
162 if (AR->getNoWrapFlags(FlagNUW))
164 if (AR->getNoWrapFlags(FlagNSW))
166 if (AR->getNoWrapFlags(FlagNW) &&
167 !AR->getNoWrapFlags((NoWrapFlags)(FlagNUW | FlagNSW)))
169 WriteAsOperand(OS, AR->getLoop()->getHeader(), /*PrintType=*/false);
177 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(this);
178 const char *OpStr = 0;
179 switch (NAry->getSCEVType()) {
180 case scAddExpr: OpStr = " + "; break;
181 case scMulExpr: OpStr = " * "; break;
182 case scUMaxExpr: OpStr = " umax "; break;
183 case scSMaxExpr: OpStr = " smax "; break;
186 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
189 if (llvm::next(I) != E)
193 switch (NAry->getSCEVType()) {
196 if (NAry->getNoWrapFlags(FlagNUW))
198 if (NAry->getNoWrapFlags(FlagNSW))
204 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(this);
205 OS << "(" << *UDiv->getLHS() << " /u " << *UDiv->getRHS() << ")";
209 const SCEVUnknown *U = cast<SCEVUnknown>(this);
211 if (U->isSizeOf(AllocTy)) {
212 OS << "sizeof(" << *AllocTy << ")";
215 if (U->isAlignOf(AllocTy)) {
216 OS << "alignof(" << *AllocTy << ")";
222 if (U->isOffsetOf(CTy, FieldNo)) {
223 OS << "offsetof(" << *CTy << ", ";
224 WriteAsOperand(OS, FieldNo, false);
229 // Otherwise just print it normally.
230 WriteAsOperand(OS, U->getValue(), false);
233 case scCouldNotCompute:
234 OS << "***COULDNOTCOMPUTE***";
238 llvm_unreachable("Unknown SCEV kind!");
241 Type *SCEV::getType() const {
242 switch (getSCEVType()) {
244 return cast<SCEVConstant>(this)->getType();
248 return cast<SCEVCastExpr>(this)->getType();
253 return cast<SCEVNAryExpr>(this)->getType();
255 return cast<SCEVAddExpr>(this)->getType();
257 return cast<SCEVUDivExpr>(this)->getType();
259 return cast<SCEVUnknown>(this)->getType();
260 case scCouldNotCompute:
261 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
263 llvm_unreachable("Unknown SCEV kind!");
267 bool SCEV::isZero() const {
268 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
269 return SC->getValue()->isZero();
273 bool SCEV::isOne() const {
274 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
275 return SC->getValue()->isOne();
279 bool SCEV::isAllOnesValue() const {
280 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
281 return SC->getValue()->isAllOnesValue();
285 /// isNonConstantNegative - Return true if the specified scev is negated, but
287 bool SCEV::isNonConstantNegative() const {
288 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(this);
289 if (!Mul) return false;
291 // If there is a constant factor, it will be first.
292 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
293 if (!SC) return false;
295 // Return true if the value is negative, this matches things like (-42 * V).
296 return SC->getValue()->getValue().isNegative();
299 SCEVCouldNotCompute::SCEVCouldNotCompute() :
300 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
302 bool SCEVCouldNotCompute::classof(const SCEV *S) {
303 return S->getSCEVType() == scCouldNotCompute;
306 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
308 ID.AddInteger(scConstant);
311 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
312 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V);
313 UniqueSCEVs.InsertNode(S, IP);
317 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
318 return getConstant(ConstantInt::get(getContext(), Val));
322 ScalarEvolution::getConstant(Type *Ty, uint64_t V, bool isSigned) {
323 IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
324 return getConstant(ConstantInt::get(ITy, V, isSigned));
327 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
328 unsigned SCEVTy, const SCEV *op, Type *ty)
329 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
331 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
332 const SCEV *op, Type *ty)
333 : SCEVCastExpr(ID, scTruncate, op, ty) {
334 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
335 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
336 "Cannot truncate non-integer value!");
339 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
340 const SCEV *op, Type *ty)
341 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
342 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
343 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
344 "Cannot zero extend non-integer value!");
347 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
348 const SCEV *op, Type *ty)
349 : SCEVCastExpr(ID, scSignExtend, op, ty) {
350 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
351 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
352 "Cannot sign extend non-integer value!");
355 void SCEVUnknown::deleted() {
356 // Clear this SCEVUnknown from various maps.
357 SE->forgetMemoizedResults(this);
359 // Remove this SCEVUnknown from the uniquing map.
360 SE->UniqueSCEVs.RemoveNode(this);
362 // Release the value.
366 void SCEVUnknown::allUsesReplacedWith(Value *New) {
367 // Clear this SCEVUnknown from various maps.
368 SE->forgetMemoizedResults(this);
370 // Remove this SCEVUnknown from the uniquing map.
371 SE->UniqueSCEVs.RemoveNode(this);
373 // Update this SCEVUnknown to point to the new value. This is needed
374 // because there may still be outstanding SCEVs which still point to
379 bool SCEVUnknown::isSizeOf(Type *&AllocTy) const {
380 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
381 if (VCE->getOpcode() == Instruction::PtrToInt)
382 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
383 if (CE->getOpcode() == Instruction::GetElementPtr &&
384 CE->getOperand(0)->isNullValue() &&
385 CE->getNumOperands() == 2)
386 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
388 AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
396 bool SCEVUnknown::isAlignOf(Type *&AllocTy) 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->getOperand(0)->isNullValue()) {
403 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
404 if (StructType *STy = dyn_cast<StructType>(Ty))
405 if (!STy->isPacked() &&
406 CE->getNumOperands() == 3 &&
407 CE->getOperand(1)->isNullValue()) {
408 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
410 STy->getNumElements() == 2 &&
411 STy->getElementType(0)->isIntegerTy(1)) {
412 AllocTy = STy->getElementType(1);
421 bool SCEVUnknown::isOffsetOf(Type *&CTy, Constant *&FieldNo) const {
422 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
423 if (VCE->getOpcode() == Instruction::PtrToInt)
424 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
425 if (CE->getOpcode() == Instruction::GetElementPtr &&
426 CE->getNumOperands() == 3 &&
427 CE->getOperand(0)->isNullValue() &&
428 CE->getOperand(1)->isNullValue()) {
430 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
431 // Ignore vector types here so that ScalarEvolutionExpander doesn't
432 // emit getelementptrs that index into vectors.
433 if (Ty->isStructTy() || Ty->isArrayTy()) {
435 FieldNo = CE->getOperand(2);
443 //===----------------------------------------------------------------------===//
445 //===----------------------------------------------------------------------===//
448 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
449 /// than the complexity of the RHS. This comparator is used to canonicalize
451 class SCEVComplexityCompare {
452 const LoopInfo *const LI;
454 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
456 // Return true or false if LHS is less than, or at least RHS, respectively.
457 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
458 return compare(LHS, RHS) < 0;
461 // Return negative, zero, or positive, if LHS is less than, equal to, or
462 // greater than RHS, respectively. A three-way result allows recursive
463 // comparisons to be more efficient.
464 int compare(const SCEV *LHS, const SCEV *RHS) const {
465 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
469 // Primarily, sort the SCEVs by their getSCEVType().
470 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
472 return (int)LType - (int)RType;
474 // Aside from the getSCEVType() ordering, the particular ordering
475 // isn't very important except that it's beneficial to be consistent,
476 // so that (a + b) and (b + a) don't end up as different expressions.
479 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS);
480 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
482 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
483 // not as complete as it could be.
484 const Value *LV = LU->getValue(), *RV = RU->getValue();
486 // Order pointer values after integer values. This helps SCEVExpander
488 bool LIsPointer = LV->getType()->isPointerTy(),
489 RIsPointer = RV->getType()->isPointerTy();
490 if (LIsPointer != RIsPointer)
491 return (int)LIsPointer - (int)RIsPointer;
493 // Compare getValueID values.
494 unsigned LID = LV->getValueID(),
495 RID = RV->getValueID();
497 return (int)LID - (int)RID;
499 // Sort arguments by their position.
500 if (const Argument *LA = dyn_cast<Argument>(LV)) {
501 const Argument *RA = cast<Argument>(RV);
502 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo();
503 return (int)LArgNo - (int)RArgNo;
506 // For instructions, compare their loop depth, and their operand
507 // count. This is pretty loose.
508 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) {
509 const Instruction *RInst = cast<Instruction>(RV);
511 // Compare loop depths.
512 const BasicBlock *LParent = LInst->getParent(),
513 *RParent = RInst->getParent();
514 if (LParent != RParent) {
515 unsigned LDepth = LI->getLoopDepth(LParent),
516 RDepth = LI->getLoopDepth(RParent);
517 if (LDepth != RDepth)
518 return (int)LDepth - (int)RDepth;
521 // Compare the number of operands.
522 unsigned LNumOps = LInst->getNumOperands(),
523 RNumOps = RInst->getNumOperands();
524 return (int)LNumOps - (int)RNumOps;
531 const SCEVConstant *LC = cast<SCEVConstant>(LHS);
532 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
534 // Compare constant values.
535 const APInt &LA = LC->getValue()->getValue();
536 const APInt &RA = RC->getValue()->getValue();
537 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth();
538 if (LBitWidth != RBitWidth)
539 return (int)LBitWidth - (int)RBitWidth;
540 return LA.ult(RA) ? -1 : 1;
544 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS);
545 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
547 // Compare addrec loop depths.
548 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop();
549 if (LLoop != RLoop) {
550 unsigned LDepth = LLoop->getLoopDepth(),
551 RDepth = RLoop->getLoopDepth();
552 if (LDepth != RDepth)
553 return (int)LDepth - (int)RDepth;
556 // Addrec complexity grows with operand count.
557 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands();
558 if (LNumOps != RNumOps)
559 return (int)LNumOps - (int)RNumOps;
561 // Lexicographically compare.
562 for (unsigned i = 0; i != LNumOps; ++i) {
563 long X = compare(LA->getOperand(i), RA->getOperand(i));
575 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS);
576 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
578 // Lexicographically compare n-ary expressions.
579 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
580 for (unsigned i = 0; i != LNumOps; ++i) {
583 long X = compare(LC->getOperand(i), RC->getOperand(i));
587 return (int)LNumOps - (int)RNumOps;
591 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
592 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
594 // Lexicographically compare udiv expressions.
595 long X = compare(LC->getLHS(), RC->getLHS());
598 return compare(LC->getRHS(), RC->getRHS());
604 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
605 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
607 // Compare cast expressions by operand.
608 return compare(LC->getOperand(), RC->getOperand());
612 llvm_unreachable("Unknown SCEV kind!");
618 /// GroupByComplexity - Given a list of SCEV objects, order them by their
619 /// complexity, and group objects of the same complexity together by value.
620 /// When this routine is finished, we know that any duplicates in the vector are
621 /// consecutive and that complexity is monotonically increasing.
623 /// Note that we go take special precautions to ensure that we get deterministic
624 /// results from this routine. In other words, we don't want the results of
625 /// this to depend on where the addresses of various SCEV objects happened to
628 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
630 if (Ops.size() < 2) return; // Noop
631 if (Ops.size() == 2) {
632 // This is the common case, which also happens to be trivially simple.
634 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
635 if (SCEVComplexityCompare(LI)(RHS, LHS))
640 // Do the rough sort by complexity.
641 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
643 // Now that we are sorted by complexity, group elements of the same
644 // complexity. Note that this is, at worst, N^2, but the vector is likely to
645 // be extremely short in practice. Note that we take this approach because we
646 // do not want to depend on the addresses of the objects we are grouping.
647 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
648 const SCEV *S = Ops[i];
649 unsigned Complexity = S->getSCEVType();
651 // If there are any objects of the same complexity and same value as this
653 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
654 if (Ops[j] == S) { // Found a duplicate.
655 // Move it to immediately after i'th element.
656 std::swap(Ops[i+1], Ops[j]);
657 ++i; // no need to rescan it.
658 if (i == e-2) return; // Done!
666 //===----------------------------------------------------------------------===//
667 // Simple SCEV method implementations
668 //===----------------------------------------------------------------------===//
670 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
672 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
675 // Handle the simplest case efficiently.
677 return SE.getTruncateOrZeroExtend(It, ResultTy);
679 // We are using the following formula for BC(It, K):
681 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
683 // Suppose, W is the bitwidth of the return value. We must be prepared for
684 // overflow. Hence, we must assure that the result of our computation is
685 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
686 // safe in modular arithmetic.
688 // However, this code doesn't use exactly that formula; the formula it uses
689 // is something like the following, where T is the number of factors of 2 in
690 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
693 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
695 // This formula is trivially equivalent to the previous formula. However,
696 // this formula can be implemented much more efficiently. The trick is that
697 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
698 // arithmetic. To do exact division in modular arithmetic, all we have
699 // to do is multiply by the inverse. Therefore, this step can be done at
702 // The next issue is how to safely do the division by 2^T. The way this
703 // is done is by doing the multiplication step at a width of at least W + T
704 // bits. This way, the bottom W+T bits of the product are accurate. Then,
705 // when we perform the division by 2^T (which is equivalent to a right shift
706 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
707 // truncated out after the division by 2^T.
709 // In comparison to just directly using the first formula, this technique
710 // is much more efficient; using the first formula requires W * K bits,
711 // but this formula less than W + K bits. Also, the first formula requires
712 // a division step, whereas this formula only requires multiplies and shifts.
714 // It doesn't matter whether the subtraction step is done in the calculation
715 // width or the input iteration count's width; if the subtraction overflows,
716 // the result must be zero anyway. We prefer here to do it in the width of
717 // the induction variable because it helps a lot for certain cases; CodeGen
718 // isn't smart enough to ignore the overflow, which leads to much less
719 // efficient code if the width of the subtraction is wider than the native
722 // (It's possible to not widen at all by pulling out factors of 2 before
723 // the multiplication; for example, K=2 can be calculated as
724 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
725 // extra arithmetic, so it's not an obvious win, and it gets
726 // much more complicated for K > 3.)
728 // Protection from insane SCEVs; this bound is conservative,
729 // but it probably doesn't matter.
731 return SE.getCouldNotCompute();
733 unsigned W = SE.getTypeSizeInBits(ResultTy);
735 // Calculate K! / 2^T and T; we divide out the factors of two before
736 // multiplying for calculating K! / 2^T to avoid overflow.
737 // Other overflow doesn't matter because we only care about the bottom
738 // W bits of the result.
739 APInt OddFactorial(W, 1);
741 for (unsigned i = 3; i <= K; ++i) {
743 unsigned TwoFactors = Mult.countTrailingZeros();
745 Mult = Mult.lshr(TwoFactors);
746 OddFactorial *= Mult;
749 // We need at least W + T bits for the multiplication step
750 unsigned CalculationBits = W + T;
752 // Calculate 2^T, at width T+W.
753 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
755 // Calculate the multiplicative inverse of K! / 2^T;
756 // this multiplication factor will perform the exact division by
758 APInt Mod = APInt::getSignedMinValue(W+1);
759 APInt MultiplyFactor = OddFactorial.zext(W+1);
760 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
761 MultiplyFactor = MultiplyFactor.trunc(W);
763 // Calculate the product, at width T+W
764 IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
766 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
767 for (unsigned i = 1; i != K; ++i) {
768 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
769 Dividend = SE.getMulExpr(Dividend,
770 SE.getTruncateOrZeroExtend(S, CalculationTy));
774 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
776 // Truncate the result, and divide by K! / 2^T.
778 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
779 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
782 /// evaluateAtIteration - Return the value of this chain of recurrences at
783 /// the specified iteration number. We can evaluate this recurrence by
784 /// multiplying each element in the chain by the binomial coefficient
785 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
787 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
789 /// where BC(It, k) stands for binomial coefficient.
791 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
792 ScalarEvolution &SE) const {
793 const SCEV *Result = getStart();
794 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
795 // The computation is correct in the face of overflow provided that the
796 // multiplication is performed _after_ the evaluation of the binomial
798 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
799 if (isa<SCEVCouldNotCompute>(Coeff))
802 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
807 //===----------------------------------------------------------------------===//
808 // SCEV Expression folder implementations
809 //===----------------------------------------------------------------------===//
811 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
813 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
814 "This is not a truncating conversion!");
815 assert(isSCEVable(Ty) &&
816 "This is not a conversion to a SCEVable type!");
817 Ty = getEffectiveSCEVType(Ty);
820 ID.AddInteger(scTruncate);
824 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
826 // Fold if the operand is constant.
827 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
829 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
831 // trunc(trunc(x)) --> trunc(x)
832 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
833 return getTruncateExpr(ST->getOperand(), Ty);
835 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
836 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
837 return getTruncateOrSignExtend(SS->getOperand(), Ty);
839 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
840 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
841 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
843 // trunc(x1+x2+...+xN) --> trunc(x1)+trunc(x2)+...+trunc(xN) if we can
844 // eliminate all the truncates.
845 if (const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Op)) {
846 SmallVector<const SCEV *, 4> Operands;
847 bool hasTrunc = false;
848 for (unsigned i = 0, e = SA->getNumOperands(); i != e && !hasTrunc; ++i) {
849 const SCEV *S = getTruncateExpr(SA->getOperand(i), Ty);
850 hasTrunc = isa<SCEVTruncateExpr>(S);
851 Operands.push_back(S);
854 return getAddExpr(Operands);
855 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
858 // trunc(x1*x2*...*xN) --> trunc(x1)*trunc(x2)*...*trunc(xN) if we can
859 // eliminate all the truncates.
860 if (const SCEVMulExpr *SM = dyn_cast<SCEVMulExpr>(Op)) {
861 SmallVector<const SCEV *, 4> Operands;
862 bool hasTrunc = false;
863 for (unsigned i = 0, e = SM->getNumOperands(); i != e && !hasTrunc; ++i) {
864 const SCEV *S = getTruncateExpr(SM->getOperand(i), Ty);
865 hasTrunc = isa<SCEVTruncateExpr>(S);
866 Operands.push_back(S);
869 return getMulExpr(Operands);
870 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
873 // If the input value is a chrec scev, truncate the chrec's operands.
874 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
875 SmallVector<const SCEV *, 4> Operands;
876 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
877 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
878 return getAddRecExpr(Operands, AddRec->getLoop(), SCEV::FlagAnyWrap);
881 // The cast wasn't folded; create an explicit cast node. We can reuse
882 // the existing insert position since if we get here, we won't have
883 // made any changes which would invalidate it.
884 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
886 UniqueSCEVs.InsertNode(S, IP);
890 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
892 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
893 "This is not an extending conversion!");
894 assert(isSCEVable(Ty) &&
895 "This is not a conversion to a SCEVable type!");
896 Ty = getEffectiveSCEVType(Ty);
898 // Fold if the operand is constant.
899 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
901 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(), Ty)));
903 // zext(zext(x)) --> zext(x)
904 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
905 return getZeroExtendExpr(SZ->getOperand(), Ty);
907 // Before doing any expensive analysis, check to see if we've already
908 // computed a SCEV for this Op and Ty.
910 ID.AddInteger(scZeroExtend);
914 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
916 // zext(trunc(x)) --> zext(x) or x or trunc(x)
917 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
918 // It's possible the bits taken off by the truncate were all zero bits. If
919 // so, we should be able to simplify this further.
920 const SCEV *X = ST->getOperand();
921 ConstantRange CR = getUnsignedRange(X);
922 unsigned TruncBits = getTypeSizeInBits(ST->getType());
923 unsigned NewBits = getTypeSizeInBits(Ty);
924 if (CR.truncate(TruncBits).zeroExtend(NewBits).contains(
925 CR.zextOrTrunc(NewBits)))
926 return getTruncateOrZeroExtend(X, Ty);
929 // If the input value is a chrec scev, and we can prove that the value
930 // did not overflow the old, smaller, value, we can zero extend all of the
931 // operands (often constants). This allows analysis of something like
932 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
933 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
934 if (AR->isAffine()) {
935 const SCEV *Start = AR->getStart();
936 const SCEV *Step = AR->getStepRecurrence(*this);
937 unsigned BitWidth = getTypeSizeInBits(AR->getType());
938 const Loop *L = AR->getLoop();
940 // If we have special knowledge that this addrec won't overflow,
941 // we don't need to do any further analysis.
942 if (AR->getNoWrapFlags(SCEV::FlagNUW))
943 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
944 getZeroExtendExpr(Step, Ty),
945 L, AR->getNoWrapFlags());
947 // Check whether the backedge-taken count is SCEVCouldNotCompute.
948 // Note that this serves two purposes: It filters out loops that are
949 // simply not analyzable, and it covers the case where this code is
950 // being called from within backedge-taken count analysis, such that
951 // attempting to ask for the backedge-taken count would likely result
952 // in infinite recursion. In the later case, the analysis code will
953 // cope with a conservative value, and it will take care to purge
954 // that value once it has finished.
955 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
956 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
957 // Manually compute the final value for AR, checking for
960 // Check whether the backedge-taken count can be losslessly casted to
961 // the addrec's type. The count is always unsigned.
962 const SCEV *CastedMaxBECount =
963 getTruncateOrZeroExtend(MaxBECount, Start->getType());
964 const SCEV *RecastedMaxBECount =
965 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
966 if (MaxBECount == RecastedMaxBECount) {
967 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
968 // Check whether Start+Step*MaxBECount has no unsigned overflow.
969 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
970 const SCEV *ZAdd = getZeroExtendExpr(getAddExpr(Start, ZMul), WideTy);
971 const SCEV *WideStart = getZeroExtendExpr(Start, WideTy);
972 const SCEV *WideMaxBECount =
973 getZeroExtendExpr(CastedMaxBECount, WideTy);
974 const SCEV *OperandExtendedAdd =
975 getAddExpr(WideStart,
976 getMulExpr(WideMaxBECount,
977 getZeroExtendExpr(Step, WideTy)));
978 if (ZAdd == OperandExtendedAdd) {
979 // Cache knowledge of AR NUW, which is propagated to this AddRec.
980 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
981 // Return the expression with the addrec on the outside.
982 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
983 getZeroExtendExpr(Step, Ty),
984 L, AR->getNoWrapFlags());
986 // Similar to above, only this time treat the step value as signed.
987 // This covers loops that count down.
989 getAddExpr(WideStart,
990 getMulExpr(WideMaxBECount,
991 getSignExtendExpr(Step, WideTy)));
992 if (ZAdd == OperandExtendedAdd) {
993 // Cache knowledge of AR NW, which is propagated to this AddRec.
994 // Negative step causes unsigned wrap, but it still can't self-wrap.
995 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
996 // Return the expression with the addrec on the outside.
997 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
998 getSignExtendExpr(Step, Ty),
999 L, AR->getNoWrapFlags());
1003 // If the backedge is guarded by a comparison with the pre-inc value
1004 // the addrec is safe. Also, if the entry is guarded by a comparison
1005 // with the start value and the backedge is guarded by a comparison
1006 // with the post-inc value, the addrec is safe.
1007 if (isKnownPositive(Step)) {
1008 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
1009 getUnsignedRange(Step).getUnsignedMax());
1010 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
1011 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
1012 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
1013 AR->getPostIncExpr(*this), N))) {
1014 // Cache knowledge of AR NUW, which is propagated to this AddRec.
1015 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
1016 // Return the expression with the addrec on the outside.
1017 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1018 getZeroExtendExpr(Step, Ty),
1019 L, AR->getNoWrapFlags());
1021 } else if (isKnownNegative(Step)) {
1022 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
1023 getSignedRange(Step).getSignedMin());
1024 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
1025 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
1026 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
1027 AR->getPostIncExpr(*this), N))) {
1028 // Cache knowledge of AR NW, which is propagated to this AddRec.
1029 // Negative step causes unsigned wrap, but it still can't self-wrap.
1030 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1031 // Return the expression with the addrec on the outside.
1032 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1033 getSignExtendExpr(Step, Ty),
1034 L, AR->getNoWrapFlags());
1040 // The cast wasn't folded; create an explicit cast node.
1041 // Recompute the insert position, as it may have been invalidated.
1042 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1043 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1045 UniqueSCEVs.InsertNode(S, IP);
1049 // Get the limit of a recurrence such that incrementing by Step cannot cause
1050 // signed overflow as long as the value of the recurrence within the loop does
1051 // not exceed this limit before incrementing.
1052 static const SCEV *getOverflowLimitForStep(const SCEV *Step,
1053 ICmpInst::Predicate *Pred,
1054 ScalarEvolution *SE) {
1055 unsigned BitWidth = SE->getTypeSizeInBits(Step->getType());
1056 if (SE->isKnownPositive(Step)) {
1057 *Pred = ICmpInst::ICMP_SLT;
1058 return SE->getConstant(APInt::getSignedMinValue(BitWidth) -
1059 SE->getSignedRange(Step).getSignedMax());
1061 if (SE->isKnownNegative(Step)) {
1062 *Pred = ICmpInst::ICMP_SGT;
1063 return SE->getConstant(APInt::getSignedMaxValue(BitWidth) -
1064 SE->getSignedRange(Step).getSignedMin());
1069 // The recurrence AR has been shown to have no signed wrap. Typically, if we can
1070 // prove NSW for AR, then we can just as easily prove NSW for its preincrement
1071 // or postincrement sibling. This allows normalizing a sign extended AddRec as
1072 // such: {sext(Step + Start),+,Step} => {(Step + sext(Start),+,Step} As a
1073 // result, the expression "Step + sext(PreIncAR)" is congruent with
1074 // "sext(PostIncAR)"
1075 static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR,
1077 ScalarEvolution *SE) {
1078 const Loop *L = AR->getLoop();
1079 const SCEV *Start = AR->getStart();
1080 const SCEV *Step = AR->getStepRecurrence(*SE);
1082 // Check for a simple looking step prior to loop entry.
1083 const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start);
1087 // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV
1088 // subtraction is expensive. For this purpose, perform a quick and dirty
1089 // difference, by checking for Step in the operand list.
1090 SmallVector<const SCEV *, 4> DiffOps;
1091 for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end();
1094 DiffOps.push_back(*I);
1096 if (DiffOps.size() == SA->getNumOperands())
1099 // This is a postinc AR. Check for overflow on the preinc recurrence using the
1100 // same three conditions that getSignExtendedExpr checks.
1102 // 1. NSW flags on the step increment.
1103 const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags());
1104 const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>(
1105 SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap));
1107 if (PreAR && PreAR->getNoWrapFlags(SCEV::FlagNSW))
1110 // 2. Direct overflow check on the step operation's expression.
1111 unsigned BitWidth = SE->getTypeSizeInBits(AR->getType());
1112 Type *WideTy = IntegerType::get(SE->getContext(), BitWidth * 2);
1113 const SCEV *OperandExtendedStart =
1114 SE->getAddExpr(SE->getSignExtendExpr(PreStart, WideTy),
1115 SE->getSignExtendExpr(Step, WideTy));
1116 if (SE->getSignExtendExpr(Start, WideTy) == OperandExtendedStart) {
1117 // Cache knowledge of PreAR NSW.
1119 const_cast<SCEVAddRecExpr *>(PreAR)->setNoWrapFlags(SCEV::FlagNSW);
1120 // FIXME: this optimization needs a unit test
1121 DEBUG(dbgs() << "SCEV: untested prestart overflow check\n");
1125 // 3. Loop precondition.
1126 ICmpInst::Predicate Pred;
1127 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, SE);
1129 if (OverflowLimit &&
1130 SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) {
1136 // Get the normalized sign-extended expression for this AddRec's Start.
1137 static const SCEV *getSignExtendAddRecStart(const SCEVAddRecExpr *AR,
1139 ScalarEvolution *SE) {
1140 const SCEV *PreStart = getPreStartForSignExtend(AR, Ty, SE);
1142 return SE->getSignExtendExpr(AR->getStart(), Ty);
1144 return SE->getAddExpr(SE->getSignExtendExpr(AR->getStepRecurrence(*SE), Ty),
1145 SE->getSignExtendExpr(PreStart, Ty));
1148 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1150 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1151 "This is not an extending conversion!");
1152 assert(isSCEVable(Ty) &&
1153 "This is not a conversion to a SCEVable type!");
1154 Ty = getEffectiveSCEVType(Ty);
1156 // Fold if the operand is constant.
1157 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1159 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(), Ty)));
1161 // sext(sext(x)) --> sext(x)
1162 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1163 return getSignExtendExpr(SS->getOperand(), Ty);
1165 // sext(zext(x)) --> zext(x)
1166 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
1167 return getZeroExtendExpr(SZ->getOperand(), Ty);
1169 // Before doing any expensive analysis, check to see if we've already
1170 // computed a SCEV for this Op and Ty.
1171 FoldingSetNodeID ID;
1172 ID.AddInteger(scSignExtend);
1176 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1178 // If the input value is provably positive, build a zext instead.
1179 if (isKnownNonNegative(Op))
1180 return getZeroExtendExpr(Op, Ty);
1182 // sext(trunc(x)) --> sext(x) or x or trunc(x)
1183 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
1184 // It's possible the bits taken off by the truncate were all sign bits. If
1185 // so, we should be able to simplify this further.
1186 const SCEV *X = ST->getOperand();
1187 ConstantRange CR = getSignedRange(X);
1188 unsigned TruncBits = getTypeSizeInBits(ST->getType());
1189 unsigned NewBits = getTypeSizeInBits(Ty);
1190 if (CR.truncate(TruncBits).signExtend(NewBits).contains(
1191 CR.sextOrTrunc(NewBits)))
1192 return getTruncateOrSignExtend(X, Ty);
1195 // If the input value is a chrec scev, and we can prove that the value
1196 // did not overflow the old, smaller, value, we can sign extend all of the
1197 // operands (often constants). This allows analysis of something like
1198 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1199 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1200 if (AR->isAffine()) {
1201 const SCEV *Start = AR->getStart();
1202 const SCEV *Step = AR->getStepRecurrence(*this);
1203 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1204 const Loop *L = AR->getLoop();
1206 // If we have special knowledge that this addrec won't overflow,
1207 // we don't need to do any further analysis.
1208 if (AR->getNoWrapFlags(SCEV::FlagNSW))
1209 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1210 getSignExtendExpr(Step, Ty),
1213 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1214 // Note that this serves two purposes: It filters out loops that are
1215 // simply not analyzable, and it covers the case where this code is
1216 // being called from within backedge-taken count analysis, such that
1217 // attempting to ask for the backedge-taken count would likely result
1218 // in infinite recursion. In the later case, the analysis code will
1219 // cope with a conservative value, and it will take care to purge
1220 // that value once it has finished.
1221 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1222 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1223 // Manually compute the final value for AR, checking for
1226 // Check whether the backedge-taken count can be losslessly casted to
1227 // the addrec's type. The count is always unsigned.
1228 const SCEV *CastedMaxBECount =
1229 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1230 const SCEV *RecastedMaxBECount =
1231 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1232 if (MaxBECount == RecastedMaxBECount) {
1233 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1234 // Check whether Start+Step*MaxBECount has no signed overflow.
1235 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1236 const SCEV *SAdd = getSignExtendExpr(getAddExpr(Start, SMul), WideTy);
1237 const SCEV *WideStart = getSignExtendExpr(Start, WideTy);
1238 const SCEV *WideMaxBECount =
1239 getZeroExtendExpr(CastedMaxBECount, WideTy);
1240 const SCEV *OperandExtendedAdd =
1241 getAddExpr(WideStart,
1242 getMulExpr(WideMaxBECount,
1243 getSignExtendExpr(Step, WideTy)));
1244 if (SAdd == OperandExtendedAdd) {
1245 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1246 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1247 // Return the expression with the addrec on the outside.
1248 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1249 getSignExtendExpr(Step, Ty),
1250 L, AR->getNoWrapFlags());
1252 // Similar to above, only this time treat the step value as unsigned.
1253 // This covers loops that count up with an unsigned step.
1254 OperandExtendedAdd =
1255 getAddExpr(WideStart,
1256 getMulExpr(WideMaxBECount,
1257 getZeroExtendExpr(Step, WideTy)));
1258 if (SAdd == OperandExtendedAdd) {
1259 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1260 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1261 // Return the expression with the addrec on the outside.
1262 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1263 getZeroExtendExpr(Step, Ty),
1264 L, AR->getNoWrapFlags());
1268 // If the backedge is guarded by a comparison with the pre-inc value
1269 // the addrec is safe. Also, if the entry is guarded by a comparison
1270 // with the start value and the backedge is guarded by a comparison
1271 // with the post-inc value, the addrec is safe.
1272 ICmpInst::Predicate Pred;
1273 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, this);
1274 if (OverflowLimit &&
1275 (isLoopBackedgeGuardedByCond(L, Pred, AR, OverflowLimit) ||
1276 (isLoopEntryGuardedByCond(L, Pred, Start, OverflowLimit) &&
1277 isLoopBackedgeGuardedByCond(L, Pred, AR->getPostIncExpr(*this),
1279 // Cache knowledge of AR NSW, then propagate NSW to the wide AddRec.
1280 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1281 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1282 getSignExtendExpr(Step, Ty),
1283 L, AR->getNoWrapFlags());
1288 // The cast wasn't folded; create an explicit cast node.
1289 // Recompute the insert position, as it may have been invalidated.
1290 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1291 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1293 UniqueSCEVs.InsertNode(S, IP);
1297 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1298 /// unspecified bits out to the given type.
1300 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1302 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1303 "This is not an extending conversion!");
1304 assert(isSCEVable(Ty) &&
1305 "This is not a conversion to a SCEVable type!");
1306 Ty = getEffectiveSCEVType(Ty);
1308 // Sign-extend negative constants.
1309 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1310 if (SC->getValue()->getValue().isNegative())
1311 return getSignExtendExpr(Op, Ty);
1313 // Peel off a truncate cast.
1314 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1315 const SCEV *NewOp = T->getOperand();
1316 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1317 return getAnyExtendExpr(NewOp, Ty);
1318 return getTruncateOrNoop(NewOp, Ty);
1321 // Next try a zext cast. If the cast is folded, use it.
1322 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1323 if (!isa<SCEVZeroExtendExpr>(ZExt))
1326 // Next try a sext cast. If the cast is folded, use it.
1327 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1328 if (!isa<SCEVSignExtendExpr>(SExt))
1331 // Force the cast to be folded into the operands of an addrec.
1332 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1333 SmallVector<const SCEV *, 4> Ops;
1334 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1336 Ops.push_back(getAnyExtendExpr(*I, Ty));
1337 return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW);
1340 // If the expression is obviously signed, use the sext cast value.
1341 if (isa<SCEVSMaxExpr>(Op))
1344 // Absent any other information, use the zext cast value.
1348 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1349 /// a list of operands to be added under the given scale, update the given
1350 /// map. This is a helper function for getAddRecExpr. As an example of
1351 /// what it does, given a sequence of operands that would form an add
1352 /// expression like this:
1354 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1356 /// where A and B are constants, update the map with these values:
1358 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1360 /// and add 13 + A*B*29 to AccumulatedConstant.
1361 /// This will allow getAddRecExpr to produce this:
1363 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1365 /// This form often exposes folding opportunities that are hidden in
1366 /// the original operand list.
1368 /// Return true iff it appears that any interesting folding opportunities
1369 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1370 /// the common case where no interesting opportunities are present, and
1371 /// is also used as a check to avoid infinite recursion.
1374 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1375 SmallVector<const SCEV *, 8> &NewOps,
1376 APInt &AccumulatedConstant,
1377 const SCEV *const *Ops, size_t NumOperands,
1379 ScalarEvolution &SE) {
1380 bool Interesting = false;
1382 // Iterate over the add operands. They are sorted, with constants first.
1384 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1386 // Pull a buried constant out to the outside.
1387 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1389 AccumulatedConstant += Scale * C->getValue()->getValue();
1392 // Next comes everything else. We're especially interested in multiplies
1393 // here, but they're in the middle, so just visit the rest with one loop.
1394 for (; i != NumOperands; ++i) {
1395 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1396 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1398 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1399 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1400 // A multiplication of a constant with another add; recurse.
1401 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1403 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1404 Add->op_begin(), Add->getNumOperands(),
1407 // A multiplication of a constant with some other value. Update
1409 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1410 const SCEV *Key = SE.getMulExpr(MulOps);
1411 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1412 M.insert(std::make_pair(Key, NewScale));
1414 NewOps.push_back(Pair.first->first);
1416 Pair.first->second += NewScale;
1417 // The map already had an entry for this value, which may indicate
1418 // a folding opportunity.
1423 // An ordinary operand. Update the map.
1424 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1425 M.insert(std::make_pair(Ops[i], Scale));
1427 NewOps.push_back(Pair.first->first);
1429 Pair.first->second += Scale;
1430 // The map already had an entry for this value, which may indicate
1431 // a folding opportunity.
1441 struct APIntCompare {
1442 bool operator()(const APInt &LHS, const APInt &RHS) const {
1443 return LHS.ult(RHS);
1448 /// getAddExpr - Get a canonical add expression, or something simpler if
1450 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1451 SCEV::NoWrapFlags Flags) {
1452 assert(!(Flags & ~(SCEV::FlagNUW | SCEV::FlagNSW)) &&
1453 "only nuw or nsw allowed");
1454 assert(!Ops.empty() && "Cannot get empty add!");
1455 if (Ops.size() == 1) return Ops[0];
1457 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1458 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1459 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1460 "SCEVAddExpr operand types don't match!");
1463 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1465 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1466 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1467 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1469 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1470 E = Ops.end(); I != E; ++I)
1471 if (!isKnownNonNegative(*I)) {
1475 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1478 // Sort by complexity, this groups all similar expression types together.
1479 GroupByComplexity(Ops, LI);
1481 // If there are any constants, fold them together.
1483 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1485 assert(Idx < Ops.size());
1486 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1487 // We found two constants, fold them together!
1488 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1489 RHSC->getValue()->getValue());
1490 if (Ops.size() == 2) return Ops[0];
1491 Ops.erase(Ops.begin()+1); // Erase the folded element
1492 LHSC = cast<SCEVConstant>(Ops[0]);
1495 // If we are left with a constant zero being added, strip it off.
1496 if (LHSC->getValue()->isZero()) {
1497 Ops.erase(Ops.begin());
1501 if (Ops.size() == 1) return Ops[0];
1504 // Okay, check to see if the same value occurs in the operand list more than
1505 // once. If so, merge them together into an multiply expression. Since we
1506 // sorted the list, these values are required to be adjacent.
1507 Type *Ty = Ops[0]->getType();
1508 bool FoundMatch = false;
1509 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1510 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1511 // Scan ahead to count how many equal operands there are.
1513 while (i+Count != e && Ops[i+Count] == Ops[i])
1515 // Merge the values into a multiply.
1516 const SCEV *Scale = getConstant(Ty, Count);
1517 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1518 if (Ops.size() == Count)
1521 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1522 --i; e -= Count - 1;
1526 return getAddExpr(Ops, Flags);
1528 // Check for truncates. If all the operands are truncated from the same
1529 // type, see if factoring out the truncate would permit the result to be
1530 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1531 // if the contents of the resulting outer trunc fold to something simple.
1532 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1533 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1534 Type *DstType = Trunc->getType();
1535 Type *SrcType = Trunc->getOperand()->getType();
1536 SmallVector<const SCEV *, 8> LargeOps;
1538 // Check all the operands to see if they can be represented in the
1539 // source type of the truncate.
1540 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1541 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1542 if (T->getOperand()->getType() != SrcType) {
1546 LargeOps.push_back(T->getOperand());
1547 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1548 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1549 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1550 SmallVector<const SCEV *, 8> LargeMulOps;
1551 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1552 if (const SCEVTruncateExpr *T =
1553 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1554 if (T->getOperand()->getType() != SrcType) {
1558 LargeMulOps.push_back(T->getOperand());
1559 } else if (const SCEVConstant *C =
1560 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1561 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1568 LargeOps.push_back(getMulExpr(LargeMulOps));
1575 // Evaluate the expression in the larger type.
1576 const SCEV *Fold = getAddExpr(LargeOps, Flags);
1577 // If it folds to something simple, use it. Otherwise, don't.
1578 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1579 return getTruncateExpr(Fold, DstType);
1583 // Skip past any other cast SCEVs.
1584 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1587 // If there are add operands they would be next.
1588 if (Idx < Ops.size()) {
1589 bool DeletedAdd = false;
1590 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1591 // If we have an add, expand the add operands onto the end of the operands
1593 Ops.erase(Ops.begin()+Idx);
1594 Ops.append(Add->op_begin(), Add->op_end());
1598 // If we deleted at least one add, we added operands to the end of the list,
1599 // and they are not necessarily sorted. Recurse to resort and resimplify
1600 // any operands we just acquired.
1602 return getAddExpr(Ops);
1605 // Skip over the add expression until we get to a multiply.
1606 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1609 // Check to see if there are any folding opportunities present with
1610 // operands multiplied by constant values.
1611 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1612 uint64_t BitWidth = getTypeSizeInBits(Ty);
1613 DenseMap<const SCEV *, APInt> M;
1614 SmallVector<const SCEV *, 8> NewOps;
1615 APInt AccumulatedConstant(BitWidth, 0);
1616 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1617 Ops.data(), Ops.size(),
1618 APInt(BitWidth, 1), *this)) {
1619 // Some interesting folding opportunity is present, so its worthwhile to
1620 // re-generate the operands list. Group the operands by constant scale,
1621 // to avoid multiplying by the same constant scale multiple times.
1622 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1623 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1624 E = NewOps.end(); I != E; ++I)
1625 MulOpLists[M.find(*I)->second].push_back(*I);
1626 // Re-generate the operands list.
1628 if (AccumulatedConstant != 0)
1629 Ops.push_back(getConstant(AccumulatedConstant));
1630 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1631 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1633 Ops.push_back(getMulExpr(getConstant(I->first),
1634 getAddExpr(I->second)));
1636 return getConstant(Ty, 0);
1637 if (Ops.size() == 1)
1639 return getAddExpr(Ops);
1643 // If we are adding something to a multiply expression, make sure the
1644 // something is not already an operand of the multiply. If so, merge it into
1646 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1647 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1648 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1649 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1650 if (isa<SCEVConstant>(MulOpSCEV))
1652 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1653 if (MulOpSCEV == Ops[AddOp]) {
1654 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1655 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1656 if (Mul->getNumOperands() != 2) {
1657 // If the multiply has more than two operands, we must get the
1659 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1660 Mul->op_begin()+MulOp);
1661 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1662 InnerMul = getMulExpr(MulOps);
1664 const SCEV *One = getConstant(Ty, 1);
1665 const SCEV *AddOne = getAddExpr(One, InnerMul);
1666 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1667 if (Ops.size() == 2) return OuterMul;
1669 Ops.erase(Ops.begin()+AddOp);
1670 Ops.erase(Ops.begin()+Idx-1);
1672 Ops.erase(Ops.begin()+Idx);
1673 Ops.erase(Ops.begin()+AddOp-1);
1675 Ops.push_back(OuterMul);
1676 return getAddExpr(Ops);
1679 // Check this multiply against other multiplies being added together.
1680 for (unsigned OtherMulIdx = Idx+1;
1681 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1683 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1684 // If MulOp occurs in OtherMul, we can fold the two multiplies
1686 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1687 OMulOp != e; ++OMulOp)
1688 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1689 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1690 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1691 if (Mul->getNumOperands() != 2) {
1692 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1693 Mul->op_begin()+MulOp);
1694 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1695 InnerMul1 = getMulExpr(MulOps);
1697 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1698 if (OtherMul->getNumOperands() != 2) {
1699 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1700 OtherMul->op_begin()+OMulOp);
1701 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1702 InnerMul2 = getMulExpr(MulOps);
1704 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1705 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1706 if (Ops.size() == 2) return OuterMul;
1707 Ops.erase(Ops.begin()+Idx);
1708 Ops.erase(Ops.begin()+OtherMulIdx-1);
1709 Ops.push_back(OuterMul);
1710 return getAddExpr(Ops);
1716 // If there are any add recurrences in the operands list, see if any other
1717 // added values are loop invariant. If so, we can fold them into the
1719 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1722 // Scan over all recurrences, trying to fold loop invariants into them.
1723 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1724 // Scan all of the other operands to this add and add them to the vector if
1725 // they are loop invariant w.r.t. the recurrence.
1726 SmallVector<const SCEV *, 8> LIOps;
1727 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1728 const Loop *AddRecLoop = AddRec->getLoop();
1729 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1730 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1731 LIOps.push_back(Ops[i]);
1732 Ops.erase(Ops.begin()+i);
1736 // If we found some loop invariants, fold them into the recurrence.
1737 if (!LIOps.empty()) {
1738 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1739 LIOps.push_back(AddRec->getStart());
1741 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1743 AddRecOps[0] = getAddExpr(LIOps);
1745 // Build the new addrec. Propagate the NUW and NSW flags if both the
1746 // outer add and the inner addrec are guaranteed to have no overflow.
1747 // Always propagate NW.
1748 Flags = AddRec->getNoWrapFlags(setFlags(Flags, SCEV::FlagNW));
1749 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop, Flags);
1751 // If all of the other operands were loop invariant, we are done.
1752 if (Ops.size() == 1) return NewRec;
1754 // Otherwise, add the folded AddRec by the non-invariant parts.
1755 for (unsigned i = 0;; ++i)
1756 if (Ops[i] == AddRec) {
1760 return getAddExpr(Ops);
1763 // Okay, if there weren't any loop invariants to be folded, check to see if
1764 // there are multiple AddRec's with the same loop induction variable being
1765 // added together. If so, we can fold them.
1766 for (unsigned OtherIdx = Idx+1;
1767 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1769 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1770 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1771 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1773 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1775 if (const SCEVAddRecExpr *OtherAddRec =
1776 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1777 if (OtherAddRec->getLoop() == AddRecLoop) {
1778 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1780 if (i >= AddRecOps.size()) {
1781 AddRecOps.append(OtherAddRec->op_begin()+i,
1782 OtherAddRec->op_end());
1785 AddRecOps[i] = getAddExpr(AddRecOps[i],
1786 OtherAddRec->getOperand(i));
1788 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1790 // Step size has changed, so we cannot guarantee no self-wraparound.
1791 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop, SCEV::FlagAnyWrap);
1792 return getAddExpr(Ops);
1795 // Otherwise couldn't fold anything into this recurrence. Move onto the
1799 // Okay, it looks like we really DO need an add expr. Check to see if we
1800 // already have one, otherwise create a new one.
1801 FoldingSetNodeID ID;
1802 ID.AddInteger(scAddExpr);
1803 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1804 ID.AddPointer(Ops[i]);
1807 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1809 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1810 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1811 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1813 UniqueSCEVs.InsertNode(S, IP);
1815 S->setNoWrapFlags(Flags);
1819 static uint64_t umul_ov(uint64_t i, uint64_t j, bool &Overflow) {
1821 if (j > 1 && k / j != i) Overflow = true;
1825 /// Compute the result of "n choose k", the binomial coefficient. If an
1826 /// intermediate computation overflows, Overflow will be set and the return will
1827 /// be garbage. Overflow is not cleared on absence of overflow.
1828 static uint64_t Choose(uint64_t n, uint64_t k, bool &Overflow) {
1829 // We use the multiplicative formula:
1830 // n(n-1)(n-2)...(n-(k-1)) / k(k-1)(k-2)...1 .
1831 // At each iteration, we take the n-th term of the numeral and divide by the
1832 // (k-n)th term of the denominator. This division will always produce an
1833 // integral result, and helps reduce the chance of overflow in the
1834 // intermediate computations. However, we can still overflow even when the
1835 // final result would fit.
1837 if (n == 0 || n == k) return 1;
1838 if (k > n) return 0;
1844 for (uint64_t i = 1; i <= k; ++i) {
1845 r = umul_ov(r, n-(i-1), Overflow);
1851 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1853 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1854 SCEV::NoWrapFlags Flags) {
1855 assert(Flags == maskFlags(Flags, SCEV::FlagNUW | SCEV::FlagNSW) &&
1856 "only nuw or nsw allowed");
1857 assert(!Ops.empty() && "Cannot get empty mul!");
1858 if (Ops.size() == 1) return Ops[0];
1860 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1861 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1862 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1863 "SCEVMulExpr operand types don't match!");
1866 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1868 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1869 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1870 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1872 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1873 E = Ops.end(); I != E; ++I)
1874 if (!isKnownNonNegative(*I)) {
1878 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1881 // Sort by complexity, this groups all similar expression types together.
1882 GroupByComplexity(Ops, LI);
1884 // If there are any constants, fold them together.
1886 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1888 // C1*(C2+V) -> C1*C2 + C1*V
1889 if (Ops.size() == 2)
1890 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1891 if (Add->getNumOperands() == 2 &&
1892 isa<SCEVConstant>(Add->getOperand(0)))
1893 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1894 getMulExpr(LHSC, Add->getOperand(1)));
1897 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1898 // We found two constants, fold them together!
1899 ConstantInt *Fold = ConstantInt::get(getContext(),
1900 LHSC->getValue()->getValue() *
1901 RHSC->getValue()->getValue());
1902 Ops[0] = getConstant(Fold);
1903 Ops.erase(Ops.begin()+1); // Erase the folded element
1904 if (Ops.size() == 1) return Ops[0];
1905 LHSC = cast<SCEVConstant>(Ops[0]);
1908 // If we are left with a constant one being multiplied, strip it off.
1909 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1910 Ops.erase(Ops.begin());
1912 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1913 // If we have a multiply of zero, it will always be zero.
1915 } else if (Ops[0]->isAllOnesValue()) {
1916 // If we have a mul by -1 of an add, try distributing the -1 among the
1918 if (Ops.size() == 2) {
1919 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1920 SmallVector<const SCEV *, 4> NewOps;
1921 bool AnyFolded = false;
1922 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(),
1923 E = Add->op_end(); I != E; ++I) {
1924 const SCEV *Mul = getMulExpr(Ops[0], *I);
1925 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1926 NewOps.push_back(Mul);
1929 return getAddExpr(NewOps);
1931 else if (const SCEVAddRecExpr *
1932 AddRec = dyn_cast<SCEVAddRecExpr>(Ops[1])) {
1933 // Negation preserves a recurrence's no self-wrap property.
1934 SmallVector<const SCEV *, 4> Operands;
1935 for (SCEVAddRecExpr::op_iterator I = AddRec->op_begin(),
1936 E = AddRec->op_end(); I != E; ++I) {
1937 Operands.push_back(getMulExpr(Ops[0], *I));
1939 return getAddRecExpr(Operands, AddRec->getLoop(),
1940 AddRec->getNoWrapFlags(SCEV::FlagNW));
1945 if (Ops.size() == 1)
1949 // Skip over the add expression until we get to a multiply.
1950 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1953 // If there are mul operands inline them all into this expression.
1954 if (Idx < Ops.size()) {
1955 bool DeletedMul = false;
1956 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1957 // If we have an mul, expand the mul operands onto the end of the operands
1959 Ops.erase(Ops.begin()+Idx);
1960 Ops.append(Mul->op_begin(), Mul->op_end());
1964 // If we deleted at least one mul, we added operands to the end of the list,
1965 // and they are not necessarily sorted. Recurse to resort and resimplify
1966 // any operands we just acquired.
1968 return getMulExpr(Ops);
1971 // If there are any add recurrences in the operands list, see if any other
1972 // added values are loop invariant. If so, we can fold them into the
1974 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1977 // Scan over all recurrences, trying to fold loop invariants into them.
1978 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1979 // Scan all of the other operands to this mul and add them to the vector if
1980 // they are loop invariant w.r.t. the recurrence.
1981 SmallVector<const SCEV *, 8> LIOps;
1982 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1983 const Loop *AddRecLoop = AddRec->getLoop();
1984 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1985 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1986 LIOps.push_back(Ops[i]);
1987 Ops.erase(Ops.begin()+i);
1991 // If we found some loop invariants, fold them into the recurrence.
1992 if (!LIOps.empty()) {
1993 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1994 SmallVector<const SCEV *, 4> NewOps;
1995 NewOps.reserve(AddRec->getNumOperands());
1996 const SCEV *Scale = getMulExpr(LIOps);
1997 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1998 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
2000 // Build the new addrec. Propagate the NUW and NSW flags if both the
2001 // outer mul and the inner addrec are guaranteed to have no overflow.
2003 // No self-wrap cannot be guaranteed after changing the step size, but
2004 // will be inferred if either NUW or NSW is true.
2005 Flags = AddRec->getNoWrapFlags(clearFlags(Flags, SCEV::FlagNW));
2006 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop, Flags);
2008 // If all of the other operands were loop invariant, we are done.
2009 if (Ops.size() == 1) return NewRec;
2011 // Otherwise, multiply the folded AddRec by the non-invariant parts.
2012 for (unsigned i = 0;; ++i)
2013 if (Ops[i] == AddRec) {
2017 return getMulExpr(Ops);
2020 // Okay, if there weren't any loop invariants to be folded, check to see if
2021 // there are multiple AddRec's with the same loop induction variable being
2022 // multiplied together. If so, we can fold them.
2023 for (unsigned OtherIdx = Idx+1;
2024 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2026 if (AddRecLoop != cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop())
2029 // {A1,+,A2,+,...,+,An}<L> * {B1,+,B2,+,...,+,Bn}<L>
2030 // = {x=1 in [ sum y=x..2x [ sum z=max(y-x, y-n)..min(x,n) [
2031 // choose(x, 2x)*choose(2x-y, x-z)*A_{y-z}*B_z
2032 // ]]],+,...up to x=2n}.
2033 // Note that the arguments to choose() are always integers with values
2034 // known at compile time, never SCEV objects.
2036 // The implementation avoids pointless extra computations when the two
2037 // addrec's are of different length (mathematically, it's equivalent to
2038 // an infinite stream of zeros on the right).
2039 bool OpsModified = false;
2040 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2042 const SCEVAddRecExpr *OtherAddRec =
2043 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]);
2044 if (!OtherAddRec || OtherAddRec->getLoop() != AddRecLoop)
2047 bool Overflow = false;
2048 Type *Ty = AddRec->getType();
2049 bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64;
2050 SmallVector<const SCEV*, 7> AddRecOps;
2051 for (int x = 0, xe = AddRec->getNumOperands() +
2052 OtherAddRec->getNumOperands() - 1; x != xe && !Overflow; ++x) {
2053 const SCEV *Term = getConstant(Ty, 0);
2054 for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) {
2055 uint64_t Coeff1 = Choose(x, 2*x - y, Overflow);
2056 for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1),
2057 ze = std::min(x+1, (int)OtherAddRec->getNumOperands());
2058 z < ze && !Overflow; ++z) {
2059 uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow);
2061 if (LargerThan64Bits)
2062 Coeff = umul_ov(Coeff1, Coeff2, Overflow);
2064 Coeff = Coeff1*Coeff2;
2065 const SCEV *CoeffTerm = getConstant(Ty, Coeff);
2066 const SCEV *Term1 = AddRec->getOperand(y-z);
2067 const SCEV *Term2 = OtherAddRec->getOperand(z);
2068 Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2));
2071 AddRecOps.push_back(Term);
2074 const SCEV *NewAddRec = getAddRecExpr(AddRecOps, AddRec->getLoop(),
2076 if (Ops.size() == 2) return NewAddRec;
2077 Ops[Idx] = NewAddRec;
2078 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
2080 AddRec = dyn_cast<SCEVAddRecExpr>(NewAddRec);
2086 return getMulExpr(Ops);
2089 // Otherwise couldn't fold anything into this recurrence. Move onto the
2093 // Okay, it looks like we really DO need an mul expr. Check to see if we
2094 // already have one, otherwise create a new one.
2095 FoldingSetNodeID ID;
2096 ID.AddInteger(scMulExpr);
2097 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2098 ID.AddPointer(Ops[i]);
2101 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2103 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2104 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2105 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
2107 UniqueSCEVs.InsertNode(S, IP);
2109 S->setNoWrapFlags(Flags);
2113 /// getUDivExpr - Get a canonical unsigned division expression, or something
2114 /// simpler if possible.
2115 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
2117 assert(getEffectiveSCEVType(LHS->getType()) ==
2118 getEffectiveSCEVType(RHS->getType()) &&
2119 "SCEVUDivExpr operand types don't match!");
2121 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
2122 if (RHSC->getValue()->equalsInt(1))
2123 return LHS; // X udiv 1 --> x
2124 // If the denominator is zero, the result of the udiv is undefined. Don't
2125 // try to analyze it, because the resolution chosen here may differ from
2126 // the resolution chosen in other parts of the compiler.
2127 if (!RHSC->getValue()->isZero()) {
2128 // Determine if the division can be folded into the operands of
2130 // TODO: Generalize this to non-constants by using known-bits information.
2131 Type *Ty = LHS->getType();
2132 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
2133 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
2134 // For non-power-of-two values, effectively round the value up to the
2135 // nearest power of two.
2136 if (!RHSC->getValue()->getValue().isPowerOf2())
2138 IntegerType *ExtTy =
2139 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
2140 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
2141 if (const SCEVConstant *Step =
2142 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this))) {
2143 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
2144 const APInt &StepInt = Step->getValue()->getValue();
2145 const APInt &DivInt = RHSC->getValue()->getValue();
2146 if (!StepInt.urem(DivInt) &&
2147 getZeroExtendExpr(AR, ExtTy) ==
2148 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2149 getZeroExtendExpr(Step, ExtTy),
2150 AR->getLoop(), SCEV::FlagAnyWrap)) {
2151 SmallVector<const SCEV *, 4> Operands;
2152 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
2153 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
2154 return getAddRecExpr(Operands, AR->getLoop(),
2157 /// Get a canonical UDivExpr for a recurrence.
2158 /// {X,+,N}/C => {Y,+,N}/C where Y=X-(X%N). Safe when C%N=0.
2159 // We can currently only fold X%N if X is constant.
2160 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
2161 if (StartC && !DivInt.urem(StepInt) &&
2162 getZeroExtendExpr(AR, ExtTy) ==
2163 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2164 getZeroExtendExpr(Step, ExtTy),
2165 AR->getLoop(), SCEV::FlagAnyWrap)) {
2166 const APInt &StartInt = StartC->getValue()->getValue();
2167 const APInt &StartRem = StartInt.urem(StepInt);
2169 LHS = getAddRecExpr(getConstant(StartInt - StartRem), Step,
2170 AR->getLoop(), SCEV::FlagNW);
2173 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
2174 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
2175 SmallVector<const SCEV *, 4> Operands;
2176 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
2177 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
2178 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
2179 // Find an operand that's safely divisible.
2180 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
2181 const SCEV *Op = M->getOperand(i);
2182 const SCEV *Div = getUDivExpr(Op, RHSC);
2183 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
2184 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
2187 return getMulExpr(Operands);
2191 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
2192 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(LHS)) {
2193 SmallVector<const SCEV *, 4> Operands;
2194 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
2195 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
2196 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
2198 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2199 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
2200 if (isa<SCEVUDivExpr>(Op) ||
2201 getMulExpr(Op, RHS) != A->getOperand(i))
2203 Operands.push_back(Op);
2205 if (Operands.size() == A->getNumOperands())
2206 return getAddExpr(Operands);
2210 // Fold if both operands are constant.
2211 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
2212 Constant *LHSCV = LHSC->getValue();
2213 Constant *RHSCV = RHSC->getValue();
2214 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
2220 FoldingSetNodeID ID;
2221 ID.AddInteger(scUDivExpr);
2225 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2226 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
2228 UniqueSCEVs.InsertNode(S, IP);
2233 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2234 /// Simplify the expression as much as possible.
2235 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start, const SCEV *Step,
2237 SCEV::NoWrapFlags Flags) {
2238 SmallVector<const SCEV *, 4> Operands;
2239 Operands.push_back(Start);
2240 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
2241 if (StepChrec->getLoop() == L) {
2242 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
2243 return getAddRecExpr(Operands, L, maskFlags(Flags, SCEV::FlagNW));
2246 Operands.push_back(Step);
2247 return getAddRecExpr(Operands, L, Flags);
2250 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2251 /// Simplify the expression as much as possible.
2253 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2254 const Loop *L, SCEV::NoWrapFlags Flags) {
2255 if (Operands.size() == 1) return Operands[0];
2257 Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2258 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2259 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2260 "SCEVAddRecExpr operand types don't match!");
2261 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2262 assert(isLoopInvariant(Operands[i], L) &&
2263 "SCEVAddRecExpr operand is not loop-invariant!");
2266 if (Operands.back()->isZero()) {
2267 Operands.pop_back();
2268 return getAddRecExpr(Operands, L, SCEV::FlagAnyWrap); // {X,+,0} --> X
2271 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2272 // use that information to infer NUW and NSW flags. However, computing a
2273 // BE count requires calling getAddRecExpr, so we may not yet have a
2274 // meaningful BE count at this point (and if we don't, we'd be stuck
2275 // with a SCEVCouldNotCompute as the cached BE count).
2277 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
2279 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
2280 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
2281 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
2283 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2284 E = Operands.end(); I != E; ++I)
2285 if (!isKnownNonNegative(*I)) {
2289 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
2292 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2293 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2294 const Loop *NestedLoop = NestedAR->getLoop();
2295 if (L->contains(NestedLoop) ?
2296 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2297 (!NestedLoop->contains(L) &&
2298 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2299 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2300 NestedAR->op_end());
2301 Operands[0] = NestedAR->getStart();
2302 // AddRecs require their operands be loop-invariant with respect to their
2303 // loops. Don't perform this transformation if it would break this
2305 bool AllInvariant = true;
2306 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2307 if (!isLoopInvariant(Operands[i], L)) {
2308 AllInvariant = false;
2312 // Create a recurrence for the outer loop with the same step size.
2314 // The outer recurrence keeps its NW flag but only keeps NUW/NSW if the
2315 // inner recurrence has the same property.
2316 SCEV::NoWrapFlags OuterFlags =
2317 maskFlags(Flags, SCEV::FlagNW | NestedAR->getNoWrapFlags());
2319 NestedOperands[0] = getAddRecExpr(Operands, L, OuterFlags);
2320 AllInvariant = true;
2321 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2322 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2323 AllInvariant = false;
2327 // Ok, both add recurrences are valid after the transformation.
2329 // The inner recurrence keeps its NW flag but only keeps NUW/NSW if
2330 // the outer recurrence has the same property.
2331 SCEV::NoWrapFlags InnerFlags =
2332 maskFlags(NestedAR->getNoWrapFlags(), SCEV::FlagNW | Flags);
2333 return getAddRecExpr(NestedOperands, NestedLoop, InnerFlags);
2336 // Reset Operands to its original state.
2337 Operands[0] = NestedAR;
2341 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2342 // already have one, otherwise create a new one.
2343 FoldingSetNodeID ID;
2344 ID.AddInteger(scAddRecExpr);
2345 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2346 ID.AddPointer(Operands[i]);
2350 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2352 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2353 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2354 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2355 O, Operands.size(), L);
2356 UniqueSCEVs.InsertNode(S, IP);
2358 S->setNoWrapFlags(Flags);
2362 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2364 SmallVector<const SCEV *, 2> Ops;
2367 return getSMaxExpr(Ops);
2371 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2372 assert(!Ops.empty() && "Cannot get empty smax!");
2373 if (Ops.size() == 1) return Ops[0];
2375 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2376 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2377 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2378 "SCEVSMaxExpr operand types don't match!");
2381 // Sort by complexity, this groups all similar expression types together.
2382 GroupByComplexity(Ops, LI);
2384 // If there are any constants, fold them together.
2386 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2388 assert(Idx < Ops.size());
2389 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2390 // We found two constants, fold them together!
2391 ConstantInt *Fold = ConstantInt::get(getContext(),
2392 APIntOps::smax(LHSC->getValue()->getValue(),
2393 RHSC->getValue()->getValue()));
2394 Ops[0] = getConstant(Fold);
2395 Ops.erase(Ops.begin()+1); // Erase the folded element
2396 if (Ops.size() == 1) return Ops[0];
2397 LHSC = cast<SCEVConstant>(Ops[0]);
2400 // If we are left with a constant minimum-int, strip it off.
2401 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2402 Ops.erase(Ops.begin());
2404 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2405 // If we have an smax with a constant maximum-int, it will always be
2410 if (Ops.size() == 1) return Ops[0];
2413 // Find the first SMax
2414 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2417 // Check to see if one of the operands is an SMax. If so, expand its operands
2418 // onto our operand list, and recurse to simplify.
2419 if (Idx < Ops.size()) {
2420 bool DeletedSMax = false;
2421 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2422 Ops.erase(Ops.begin()+Idx);
2423 Ops.append(SMax->op_begin(), SMax->op_end());
2428 return getSMaxExpr(Ops);
2431 // Okay, check to see if the same value occurs in the operand list twice. If
2432 // so, delete one. Since we sorted the list, these values are required to
2434 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2435 // X smax Y smax Y --> X smax Y
2436 // X smax Y --> X, if X is always greater than Y
2437 if (Ops[i] == Ops[i+1] ||
2438 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2439 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2441 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2442 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2446 if (Ops.size() == 1) return Ops[0];
2448 assert(!Ops.empty() && "Reduced smax down to nothing!");
2450 // Okay, it looks like we really DO need an smax expr. Check to see if we
2451 // already have one, otherwise create a new one.
2452 FoldingSetNodeID ID;
2453 ID.AddInteger(scSMaxExpr);
2454 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2455 ID.AddPointer(Ops[i]);
2457 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2458 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2459 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2460 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2462 UniqueSCEVs.InsertNode(S, IP);
2466 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2468 SmallVector<const SCEV *, 2> Ops;
2471 return getUMaxExpr(Ops);
2475 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2476 assert(!Ops.empty() && "Cannot get empty umax!");
2477 if (Ops.size() == 1) return Ops[0];
2479 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2480 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2481 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2482 "SCEVUMaxExpr operand types don't match!");
2485 // Sort by complexity, this groups all similar expression types together.
2486 GroupByComplexity(Ops, LI);
2488 // If there are any constants, fold them together.
2490 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2492 assert(Idx < Ops.size());
2493 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2494 // We found two constants, fold them together!
2495 ConstantInt *Fold = ConstantInt::get(getContext(),
2496 APIntOps::umax(LHSC->getValue()->getValue(),
2497 RHSC->getValue()->getValue()));
2498 Ops[0] = getConstant(Fold);
2499 Ops.erase(Ops.begin()+1); // Erase the folded element
2500 if (Ops.size() == 1) return Ops[0];
2501 LHSC = cast<SCEVConstant>(Ops[0]);
2504 // If we are left with a constant minimum-int, strip it off.
2505 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2506 Ops.erase(Ops.begin());
2508 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2509 // If we have an umax with a constant maximum-int, it will always be
2514 if (Ops.size() == 1) return Ops[0];
2517 // Find the first UMax
2518 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2521 // Check to see if one of the operands is a UMax. If so, expand its operands
2522 // onto our operand list, and recurse to simplify.
2523 if (Idx < Ops.size()) {
2524 bool DeletedUMax = false;
2525 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2526 Ops.erase(Ops.begin()+Idx);
2527 Ops.append(UMax->op_begin(), UMax->op_end());
2532 return getUMaxExpr(Ops);
2535 // Okay, check to see if the same value occurs in the operand list twice. If
2536 // so, delete one. Since we sorted the list, these values are required to
2538 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2539 // X umax Y umax Y --> X umax Y
2540 // X umax Y --> X, if X is always greater than Y
2541 if (Ops[i] == Ops[i+1] ||
2542 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2543 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2545 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2546 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2550 if (Ops.size() == 1) return Ops[0];
2552 assert(!Ops.empty() && "Reduced umax down to nothing!");
2554 // Okay, it looks like we really DO need a umax expr. Check to see if we
2555 // already have one, otherwise create a new one.
2556 FoldingSetNodeID ID;
2557 ID.AddInteger(scUMaxExpr);
2558 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2559 ID.AddPointer(Ops[i]);
2561 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2562 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2563 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2564 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2566 UniqueSCEVs.InsertNode(S, IP);
2570 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2572 // ~smax(~x, ~y) == smin(x, y).
2573 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2576 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2578 // ~umax(~x, ~y) == umin(x, y)
2579 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2582 const SCEV *ScalarEvolution::getSizeOfExpr(Type *AllocTy) {
2583 // If we have TargetData, we can bypass creating a target-independent
2584 // constant expression and then folding it back into a ConstantInt.
2585 // This is just a compile-time optimization.
2587 return getConstant(TD->getIntPtrType(getContext()),
2588 TD->getTypeAllocSize(AllocTy));
2590 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2591 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2592 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2594 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2595 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2598 const SCEV *ScalarEvolution::getAlignOfExpr(Type *AllocTy) {
2599 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2600 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2601 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2603 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2604 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2607 const SCEV *ScalarEvolution::getOffsetOfExpr(StructType *STy,
2609 // If we have TargetData, we can bypass creating a target-independent
2610 // constant expression and then folding it back into a ConstantInt.
2611 // This is just a compile-time optimization.
2613 return getConstant(TD->getIntPtrType(getContext()),
2614 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2616 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2617 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2618 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2620 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2621 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2624 const SCEV *ScalarEvolution::getOffsetOfExpr(Type *CTy,
2625 Constant *FieldNo) {
2626 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2627 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2628 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2630 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2631 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2634 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2635 // Don't attempt to do anything other than create a SCEVUnknown object
2636 // here. createSCEV only calls getUnknown after checking for all other
2637 // interesting possibilities, and any other code that calls getUnknown
2638 // is doing so in order to hide a value from SCEV canonicalization.
2640 FoldingSetNodeID ID;
2641 ID.AddInteger(scUnknown);
2644 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2645 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2646 "Stale SCEVUnknown in uniquing map!");
2649 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2651 FirstUnknown = cast<SCEVUnknown>(S);
2652 UniqueSCEVs.InsertNode(S, IP);
2656 //===----------------------------------------------------------------------===//
2657 // Basic SCEV Analysis and PHI Idiom Recognition Code
2660 /// isSCEVable - Test if values of the given type are analyzable within
2661 /// the SCEV framework. This primarily includes integer types, and it
2662 /// can optionally include pointer types if the ScalarEvolution class
2663 /// has access to target-specific information.
2664 bool ScalarEvolution::isSCEVable(Type *Ty) const {
2665 // Integers and pointers are always SCEVable.
2666 return Ty->isIntegerTy() || Ty->isPointerTy();
2669 /// getTypeSizeInBits - Return the size in bits of the specified type,
2670 /// for which isSCEVable must return true.
2671 uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const {
2672 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2674 // If we have a TargetData, use it!
2676 return TD->getTypeSizeInBits(Ty);
2678 // Integer types have fixed sizes.
2679 if (Ty->isIntegerTy())
2680 return Ty->getPrimitiveSizeInBits();
2682 // The only other support type is pointer. Without TargetData, conservatively
2683 // assume pointers are 64-bit.
2684 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2688 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2689 /// the given type and which represents how SCEV will treat the given
2690 /// type, for which isSCEVable must return true. For pointer types,
2691 /// this is the pointer-sized integer type.
2692 Type *ScalarEvolution::getEffectiveSCEVType(Type *Ty) const {
2693 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2695 if (Ty->isIntegerTy())
2698 // The only other support type is pointer.
2699 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2700 if (TD) return TD->getIntPtrType(getContext());
2702 // Without TargetData, conservatively assume pointers are 64-bit.
2703 return Type::getInt64Ty(getContext());
2706 const SCEV *ScalarEvolution::getCouldNotCompute() {
2707 return &CouldNotCompute;
2710 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2711 /// expression and create a new one.
2712 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2713 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2715 ValueExprMapType::const_iterator I = ValueExprMap.find_as(V);
2716 if (I != ValueExprMap.end()) return I->second;
2717 const SCEV *S = createSCEV(V);
2719 // The process of creating a SCEV for V may have caused other SCEVs
2720 // to have been created, so it's necessary to insert the new entry
2721 // from scratch, rather than trying to remember the insert position
2723 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2727 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2729 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2730 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2732 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2734 Type *Ty = V->getType();
2735 Ty = getEffectiveSCEVType(Ty);
2736 return getMulExpr(V,
2737 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2740 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2741 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2742 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2744 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2746 Type *Ty = V->getType();
2747 Ty = getEffectiveSCEVType(Ty);
2748 const SCEV *AllOnes =
2749 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2750 return getMinusSCEV(AllOnes, V);
2753 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1.
2754 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2755 SCEV::NoWrapFlags Flags) {
2756 assert(!maskFlags(Flags, SCEV::FlagNUW) && "subtraction does not have NUW");
2758 // Fast path: X - X --> 0.
2760 return getConstant(LHS->getType(), 0);
2763 return getAddExpr(LHS, getNegativeSCEV(RHS), Flags);
2766 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2767 /// input value to the specified type. If the type must be extended, it is zero
2770 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, Type *Ty) {
2771 Type *SrcTy = V->getType();
2772 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2773 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2774 "Cannot truncate or zero extend with non-integer arguments!");
2775 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2776 return V; // No conversion
2777 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2778 return getTruncateExpr(V, Ty);
2779 return getZeroExtendExpr(V, Ty);
2782 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2783 /// input value to the specified type. If the type must be extended, it is sign
2786 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2788 Type *SrcTy = V->getType();
2789 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2790 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2791 "Cannot truncate or zero extend with non-integer arguments!");
2792 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2793 return V; // No conversion
2794 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2795 return getTruncateExpr(V, Ty);
2796 return getSignExtendExpr(V, Ty);
2799 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2800 /// input value to the specified type. If the type must be extended, it is zero
2801 /// extended. The conversion must not be narrowing.
2803 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, Type *Ty) {
2804 Type *SrcTy = V->getType();
2805 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2806 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2807 "Cannot noop or zero extend with non-integer arguments!");
2808 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2809 "getNoopOrZeroExtend cannot truncate!");
2810 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2811 return V; // No conversion
2812 return getZeroExtendExpr(V, Ty);
2815 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2816 /// input value to the specified type. If the type must be extended, it is sign
2817 /// extended. The conversion must not be narrowing.
2819 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, Type *Ty) {
2820 Type *SrcTy = V->getType();
2821 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2822 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2823 "Cannot noop or sign extend with non-integer arguments!");
2824 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2825 "getNoopOrSignExtend cannot truncate!");
2826 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2827 return V; // No conversion
2828 return getSignExtendExpr(V, Ty);
2831 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2832 /// the input value to the specified type. If the type must be extended,
2833 /// it is extended with unspecified bits. The conversion must not be
2836 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, Type *Ty) {
2837 Type *SrcTy = V->getType();
2838 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2839 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2840 "Cannot noop or any extend with non-integer arguments!");
2841 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2842 "getNoopOrAnyExtend cannot truncate!");
2843 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2844 return V; // No conversion
2845 return getAnyExtendExpr(V, Ty);
2848 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2849 /// input value to the specified type. The conversion must not be widening.
2851 ScalarEvolution::getTruncateOrNoop(const SCEV *V, Type *Ty) {
2852 Type *SrcTy = V->getType();
2853 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2854 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2855 "Cannot truncate or noop with non-integer arguments!");
2856 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2857 "getTruncateOrNoop cannot extend!");
2858 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2859 return V; // No conversion
2860 return getTruncateExpr(V, Ty);
2863 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2864 /// the types using zero-extension, and then perform a umax operation
2866 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2868 const SCEV *PromotedLHS = LHS;
2869 const SCEV *PromotedRHS = RHS;
2871 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2872 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2874 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2876 return getUMaxExpr(PromotedLHS, PromotedRHS);
2879 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2880 /// the types using zero-extension, and then perform a umin operation
2882 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2884 const SCEV *PromotedLHS = LHS;
2885 const SCEV *PromotedRHS = RHS;
2887 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2888 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2890 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2892 return getUMinExpr(PromotedLHS, PromotedRHS);
2895 /// getPointerBase - Transitively follow the chain of pointer-type operands
2896 /// until reaching a SCEV that does not have a single pointer operand. This
2897 /// returns a SCEVUnknown pointer for well-formed pointer-type expressions,
2898 /// but corner cases do exist.
2899 const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) {
2900 // A pointer operand may evaluate to a nonpointer expression, such as null.
2901 if (!V->getType()->isPointerTy())
2904 if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(V)) {
2905 return getPointerBase(Cast->getOperand());
2907 else if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(V)) {
2908 const SCEV *PtrOp = 0;
2909 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
2911 if ((*I)->getType()->isPointerTy()) {
2912 // Cannot find the base of an expression with multiple pointer operands.
2920 return getPointerBase(PtrOp);
2925 /// PushDefUseChildren - Push users of the given Instruction
2926 /// onto the given Worklist.
2928 PushDefUseChildren(Instruction *I,
2929 SmallVectorImpl<Instruction *> &Worklist) {
2930 // Push the def-use children onto the Worklist stack.
2931 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2933 Worklist.push_back(cast<Instruction>(*UI));
2936 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2937 /// instructions that depend on the given instruction and removes them from
2938 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2941 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2942 SmallVector<Instruction *, 16> Worklist;
2943 PushDefUseChildren(PN, Worklist);
2945 SmallPtrSet<Instruction *, 8> Visited;
2947 while (!Worklist.empty()) {
2948 Instruction *I = Worklist.pop_back_val();
2949 if (!Visited.insert(I)) continue;
2951 ValueExprMapType::iterator It =
2952 ValueExprMap.find_as(static_cast<Value *>(I));
2953 if (It != ValueExprMap.end()) {
2954 const SCEV *Old = It->second;
2956 // Short-circuit the def-use traversal if the symbolic name
2957 // ceases to appear in expressions.
2958 if (Old != SymName && !hasOperand(Old, SymName))
2961 // SCEVUnknown for a PHI either means that it has an unrecognized
2962 // structure, it's a PHI that's in the progress of being computed
2963 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2964 // additional loop trip count information isn't going to change anything.
2965 // In the second case, createNodeForPHI will perform the necessary
2966 // updates on its own when it gets to that point. In the third, we do
2967 // want to forget the SCEVUnknown.
2968 if (!isa<PHINode>(I) ||
2969 !isa<SCEVUnknown>(Old) ||
2970 (I != PN && Old == SymName)) {
2971 forgetMemoizedResults(Old);
2972 ValueExprMap.erase(It);
2976 PushDefUseChildren(I, Worklist);
2980 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2981 /// a loop header, making it a potential recurrence, or it doesn't.
2983 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2984 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2985 if (L->getHeader() == PN->getParent()) {
2986 // The loop may have multiple entrances or multiple exits; we can analyze
2987 // this phi as an addrec if it has a unique entry value and a unique
2989 Value *BEValueV = 0, *StartValueV = 0;
2990 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2991 Value *V = PN->getIncomingValue(i);
2992 if (L->contains(PN->getIncomingBlock(i))) {
2995 } else if (BEValueV != V) {
2999 } else if (!StartValueV) {
3001 } else if (StartValueV != V) {
3006 if (BEValueV && StartValueV) {
3007 // While we are analyzing this PHI node, handle its value symbolically.
3008 const SCEV *SymbolicName = getUnknown(PN);
3009 assert(ValueExprMap.find_as(PN) == ValueExprMap.end() &&
3010 "PHI node already processed?");
3011 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
3013 // Using this symbolic name for the PHI, analyze the value coming around
3015 const SCEV *BEValue = getSCEV(BEValueV);
3017 // NOTE: If BEValue is loop invariant, we know that the PHI node just
3018 // has a special value for the first iteration of the loop.
3020 // If the value coming around the backedge is an add with the symbolic
3021 // value we just inserted, then we found a simple induction variable!
3022 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
3023 // If there is a single occurrence of the symbolic value, replace it
3024 // with a recurrence.
3025 unsigned FoundIndex = Add->getNumOperands();
3026 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3027 if (Add->getOperand(i) == SymbolicName)
3028 if (FoundIndex == e) {
3033 if (FoundIndex != Add->getNumOperands()) {
3034 // Create an add with everything but the specified operand.
3035 SmallVector<const SCEV *, 8> Ops;
3036 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3037 if (i != FoundIndex)
3038 Ops.push_back(Add->getOperand(i));
3039 const SCEV *Accum = getAddExpr(Ops);
3041 // This is not a valid addrec if the step amount is varying each
3042 // loop iteration, but is not itself an addrec in this loop.
3043 if (isLoopInvariant(Accum, L) ||
3044 (isa<SCEVAddRecExpr>(Accum) &&
3045 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
3046 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap;
3048 // If the increment doesn't overflow, then neither the addrec nor
3049 // the post-increment will overflow.
3050 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
3051 if (OBO->hasNoUnsignedWrap())
3052 Flags = setFlags(Flags, SCEV::FlagNUW);
3053 if (OBO->hasNoSignedWrap())
3054 Flags = setFlags(Flags, SCEV::FlagNSW);
3055 } else if (const GEPOperator *GEP =
3056 dyn_cast<GEPOperator>(BEValueV)) {
3057 // If the increment is an inbounds GEP, then we know the address
3058 // space cannot be wrapped around. We cannot make any guarantee
3059 // about signed or unsigned overflow because pointers are
3060 // unsigned but we may have a negative index from the base
3062 if (GEP->isInBounds())
3063 Flags = setFlags(Flags, SCEV::FlagNW);
3066 const SCEV *StartVal = getSCEV(StartValueV);
3067 const SCEV *PHISCEV = getAddRecExpr(StartVal, Accum, L, Flags);
3069 // Since the no-wrap flags are on the increment, they apply to the
3070 // post-incremented value as well.
3071 if (isLoopInvariant(Accum, L))
3072 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
3075 // Okay, for the entire analysis of this edge we assumed the PHI
3076 // to be symbolic. We now need to go back and purge all of the
3077 // entries for the scalars that use the symbolic expression.
3078 ForgetSymbolicName(PN, SymbolicName);
3079 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3083 } else if (const SCEVAddRecExpr *AddRec =
3084 dyn_cast<SCEVAddRecExpr>(BEValue)) {
3085 // Otherwise, this could be a loop like this:
3086 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
3087 // In this case, j = {1,+,1} and BEValue is j.
3088 // Because the other in-value of i (0) fits the evolution of BEValue
3089 // i really is an addrec evolution.
3090 if (AddRec->getLoop() == L && AddRec->isAffine()) {
3091 const SCEV *StartVal = getSCEV(StartValueV);
3093 // If StartVal = j.start - j.stride, we can use StartVal as the
3094 // initial step of the addrec evolution.
3095 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
3096 AddRec->getOperand(1))) {
3097 // FIXME: For constant StartVal, we should be able to infer
3099 const SCEV *PHISCEV =
3100 getAddRecExpr(StartVal, AddRec->getOperand(1), L,
3103 // Okay, for the entire analysis of this edge we assumed the PHI
3104 // to be symbolic. We now need to go back and purge all of the
3105 // entries for the scalars that use the symbolic expression.
3106 ForgetSymbolicName(PN, SymbolicName);
3107 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3115 // If the PHI has a single incoming value, follow that value, unless the
3116 // PHI's incoming blocks are in a different loop, in which case doing so
3117 // risks breaking LCSSA form. Instcombine would normally zap these, but
3118 // it doesn't have DominatorTree information, so it may miss cases.
3119 if (Value *V = SimplifyInstruction(PN, TD, TLI, DT))
3120 if (LI->replacementPreservesLCSSAForm(PN, V))
3123 // If it's not a loop phi, we can't handle it yet.
3124 return getUnknown(PN);
3127 /// createNodeForGEP - Expand GEP instructions into add and multiply
3128 /// operations. This allows them to be analyzed by regular SCEV code.
3130 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
3132 // Don't blindly transfer the inbounds flag from the GEP instruction to the
3133 // Add expression, because the Instruction may be guarded by control flow
3134 // and the no-overflow bits may not be valid for the expression in any
3136 bool isInBounds = GEP->isInBounds();
3138 Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
3139 Value *Base = GEP->getOperand(0);
3140 // Don't attempt to analyze GEPs over unsized objects.
3141 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
3142 return getUnknown(GEP);
3143 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
3144 gep_type_iterator GTI = gep_type_begin(GEP);
3145 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
3149 // Compute the (potentially symbolic) offset in bytes for this index.
3150 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
3151 // For a struct, add the member offset.
3152 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
3153 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
3155 // Add the field offset to the running total offset.
3156 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
3158 // For an array, add the element offset, explicitly scaled.
3159 const SCEV *ElementSize = getSizeOfExpr(*GTI);
3160 const SCEV *IndexS = getSCEV(Index);
3161 // Getelementptr indices are signed.
3162 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
3164 // Multiply the index by the element size to compute the element offset.
3165 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize,
3166 isInBounds ? SCEV::FlagNSW :
3169 // Add the element offset to the running total offset.
3170 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
3174 // Get the SCEV for the GEP base.
3175 const SCEV *BaseS = getSCEV(Base);
3177 // Add the total offset from all the GEP indices to the base.
3178 return getAddExpr(BaseS, TotalOffset,
3179 isInBounds ? SCEV::FlagNSW : SCEV::FlagAnyWrap);
3182 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
3183 /// guaranteed to end in (at every loop iteration). It is, at the same time,
3184 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
3185 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
3187 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
3188 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3189 return C->getValue()->getValue().countTrailingZeros();
3191 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
3192 return std::min(GetMinTrailingZeros(T->getOperand()),
3193 (uint32_t)getTypeSizeInBits(T->getType()));
3195 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
3196 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3197 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3198 getTypeSizeInBits(E->getType()) : OpRes;
3201 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
3202 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3203 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3204 getTypeSizeInBits(E->getType()) : OpRes;
3207 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
3208 // The result is the min of all operands results.
3209 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3210 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3211 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3215 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
3216 // The result is the sum of all operands results.
3217 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
3218 uint32_t BitWidth = getTypeSizeInBits(M->getType());
3219 for (unsigned i = 1, e = M->getNumOperands();
3220 SumOpRes != BitWidth && i != e; ++i)
3221 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
3226 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
3227 // The result is the min of all operands results.
3228 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3229 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3230 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3234 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
3235 // The result is the min of all operands results.
3236 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3237 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3238 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3242 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
3243 // The result is the min of all operands results.
3244 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3245 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3246 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3250 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3251 // For a SCEVUnknown, ask ValueTracking.
3252 unsigned BitWidth = getTypeSizeInBits(U->getType());
3253 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3254 ComputeMaskedBits(U->getValue(), Zeros, Ones);
3255 return Zeros.countTrailingOnes();
3262 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
3265 ScalarEvolution::getUnsignedRange(const SCEV *S) {
3266 // See if we've computed this range already.
3267 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
3268 if (I != UnsignedRanges.end())
3271 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3272 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
3274 unsigned BitWidth = getTypeSizeInBits(S->getType());
3275 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3277 // If the value has known zeros, the maximum unsigned value will have those
3278 // known zeros as well.
3279 uint32_t TZ = GetMinTrailingZeros(S);
3281 ConservativeResult =
3282 ConstantRange(APInt::getMinValue(BitWidth),
3283 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
3285 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3286 ConstantRange X = getUnsignedRange(Add->getOperand(0));
3287 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3288 X = X.add(getUnsignedRange(Add->getOperand(i)));
3289 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
3292 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3293 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
3294 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3295 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
3296 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
3299 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3300 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
3301 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3302 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
3303 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
3306 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3307 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3308 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3309 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3310 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
3313 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3314 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3315 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3316 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3319 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3320 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3321 return setUnsignedRange(ZExt,
3322 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3325 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3326 ConstantRange X = getUnsignedRange(SExt->getOperand());
3327 return setUnsignedRange(SExt,
3328 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3331 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3332 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3333 return setUnsignedRange(Trunc,
3334 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3337 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3338 // If there's no unsigned wrap, the value will never be less than its
3340 if (AddRec->getNoWrapFlags(SCEV::FlagNUW))
3341 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3342 if (!C->getValue()->isZero())
3343 ConservativeResult =
3344 ConservativeResult.intersectWith(
3345 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3347 // TODO: non-affine addrec
3348 if (AddRec->isAffine()) {
3349 Type *Ty = AddRec->getType();
3350 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3351 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3352 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3353 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3355 const SCEV *Start = AddRec->getStart();
3356 const SCEV *Step = AddRec->getStepRecurrence(*this);
3358 ConstantRange StartRange = getUnsignedRange(Start);
3359 ConstantRange StepRange = getSignedRange(Step);
3360 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3361 ConstantRange EndRange =
3362 StartRange.add(MaxBECountRange.multiply(StepRange));
3364 // Check for overflow. This must be done with ConstantRange arithmetic
3365 // because we could be called from within the ScalarEvolution overflow
3367 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3368 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3369 ConstantRange ExtMaxBECountRange =
3370 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3371 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3372 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3374 return setUnsignedRange(AddRec, ConservativeResult);
3376 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3377 EndRange.getUnsignedMin());
3378 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3379 EndRange.getUnsignedMax());
3380 if (Min.isMinValue() && Max.isMaxValue())
3381 return setUnsignedRange(AddRec, ConservativeResult);
3382 return setUnsignedRange(AddRec,
3383 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3387 return setUnsignedRange(AddRec, ConservativeResult);
3390 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3391 // For a SCEVUnknown, ask ValueTracking.
3392 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3393 ComputeMaskedBits(U->getValue(), Zeros, Ones, TD);
3394 if (Ones == ~Zeros + 1)
3395 return setUnsignedRange(U, ConservativeResult);
3396 return setUnsignedRange(U,
3397 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3400 return setUnsignedRange(S, ConservativeResult);
3403 /// getSignedRange - Determine the signed range for a particular SCEV.
3406 ScalarEvolution::getSignedRange(const SCEV *S) {
3407 // See if we've computed this range already.
3408 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3409 if (I != SignedRanges.end())
3412 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3413 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3415 unsigned BitWidth = getTypeSizeInBits(S->getType());
3416 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3418 // If the value has known zeros, the maximum signed value will have those
3419 // known zeros as well.
3420 uint32_t TZ = GetMinTrailingZeros(S);
3422 ConservativeResult =
3423 ConstantRange(APInt::getSignedMinValue(BitWidth),
3424 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3426 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3427 ConstantRange X = getSignedRange(Add->getOperand(0));
3428 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3429 X = X.add(getSignedRange(Add->getOperand(i)));
3430 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3433 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3434 ConstantRange X = getSignedRange(Mul->getOperand(0));
3435 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3436 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3437 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3440 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3441 ConstantRange X = getSignedRange(SMax->getOperand(0));
3442 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3443 X = X.smax(getSignedRange(SMax->getOperand(i)));
3444 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3447 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3448 ConstantRange X = getSignedRange(UMax->getOperand(0));
3449 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3450 X = X.umax(getSignedRange(UMax->getOperand(i)));
3451 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3454 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3455 ConstantRange X = getSignedRange(UDiv->getLHS());
3456 ConstantRange Y = getSignedRange(UDiv->getRHS());
3457 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3460 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3461 ConstantRange X = getSignedRange(ZExt->getOperand());
3462 return setSignedRange(ZExt,
3463 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3466 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3467 ConstantRange X = getSignedRange(SExt->getOperand());
3468 return setSignedRange(SExt,
3469 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3472 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3473 ConstantRange X = getSignedRange(Trunc->getOperand());
3474 return setSignedRange(Trunc,
3475 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3478 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3479 // If there's no signed wrap, and all the operands have the same sign or
3480 // zero, the value won't ever change sign.
3481 if (AddRec->getNoWrapFlags(SCEV::FlagNSW)) {
3482 bool AllNonNeg = true;
3483 bool AllNonPos = true;
3484 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3485 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3486 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3489 ConservativeResult = ConservativeResult.intersectWith(
3490 ConstantRange(APInt(BitWidth, 0),
3491 APInt::getSignedMinValue(BitWidth)));
3493 ConservativeResult = ConservativeResult.intersectWith(
3494 ConstantRange(APInt::getSignedMinValue(BitWidth),
3495 APInt(BitWidth, 1)));
3498 // TODO: non-affine addrec
3499 if (AddRec->isAffine()) {
3500 Type *Ty = AddRec->getType();
3501 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3502 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3503 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3504 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3506 const SCEV *Start = AddRec->getStart();
3507 const SCEV *Step = AddRec->getStepRecurrence(*this);
3509 ConstantRange StartRange = getSignedRange(Start);
3510 ConstantRange StepRange = getSignedRange(Step);
3511 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3512 ConstantRange EndRange =
3513 StartRange.add(MaxBECountRange.multiply(StepRange));
3515 // Check for overflow. This must be done with ConstantRange arithmetic
3516 // because we could be called from within the ScalarEvolution overflow
3518 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3519 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3520 ConstantRange ExtMaxBECountRange =
3521 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3522 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3523 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3525 return setSignedRange(AddRec, ConservativeResult);
3527 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3528 EndRange.getSignedMin());
3529 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3530 EndRange.getSignedMax());
3531 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3532 return setSignedRange(AddRec, ConservativeResult);
3533 return setSignedRange(AddRec,
3534 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3538 return setSignedRange(AddRec, ConservativeResult);
3541 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3542 // For a SCEVUnknown, ask ValueTracking.
3543 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3544 return setSignedRange(U, ConservativeResult);
3545 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3547 return setSignedRange(U, ConservativeResult);
3548 return setSignedRange(U, ConservativeResult.intersectWith(
3549 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3550 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3553 return setSignedRange(S, ConservativeResult);
3556 /// createSCEV - We know that there is no SCEV for the specified value.
3557 /// Analyze the expression.
3559 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3560 if (!isSCEVable(V->getType()))
3561 return getUnknown(V);
3563 unsigned Opcode = Instruction::UserOp1;
3564 if (Instruction *I = dyn_cast<Instruction>(V)) {
3565 Opcode = I->getOpcode();
3567 // Don't attempt to analyze instructions in blocks that aren't
3568 // reachable. Such instructions don't matter, and they aren't required
3569 // to obey basic rules for definitions dominating uses which this
3570 // analysis depends on.
3571 if (!DT->isReachableFromEntry(I->getParent()))
3572 return getUnknown(V);
3573 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3574 Opcode = CE->getOpcode();
3575 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3576 return getConstant(CI);
3577 else if (isa<ConstantPointerNull>(V))
3578 return getConstant(V->getType(), 0);
3579 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3580 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3582 return getUnknown(V);
3584 Operator *U = cast<Operator>(V);
3586 case Instruction::Add: {
3587 // The simple thing to do would be to just call getSCEV on both operands
3588 // and call getAddExpr with the result. However if we're looking at a
3589 // bunch of things all added together, this can be quite inefficient,
3590 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3591 // Instead, gather up all the operands and make a single getAddExpr call.
3592 // LLVM IR canonical form means we need only traverse the left operands.
3594 // Don't apply this instruction's NSW or NUW flags to the new
3595 // expression. The instruction may be guarded by control flow that the
3596 // no-wrap behavior depends on. Non-control-equivalent instructions can be
3597 // mapped to the same SCEV expression, and it would be incorrect to transfer
3598 // NSW/NUW semantics to those operations.
3599 SmallVector<const SCEV *, 4> AddOps;
3600 AddOps.push_back(getSCEV(U->getOperand(1)));
3601 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3602 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3603 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3605 U = cast<Operator>(Op);
3606 const SCEV *Op1 = getSCEV(U->getOperand(1));
3607 if (Opcode == Instruction::Sub)
3608 AddOps.push_back(getNegativeSCEV(Op1));
3610 AddOps.push_back(Op1);
3612 AddOps.push_back(getSCEV(U->getOperand(0)));
3613 return getAddExpr(AddOps);
3615 case Instruction::Mul: {
3616 // Don't transfer NSW/NUW for the same reason as AddExpr.
3617 SmallVector<const SCEV *, 4> MulOps;
3618 MulOps.push_back(getSCEV(U->getOperand(1)));
3619 for (Value *Op = U->getOperand(0);
3620 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3621 Op = U->getOperand(0)) {
3622 U = cast<Operator>(Op);
3623 MulOps.push_back(getSCEV(U->getOperand(1)));
3625 MulOps.push_back(getSCEV(U->getOperand(0)));
3626 return getMulExpr(MulOps);
3628 case Instruction::UDiv:
3629 return getUDivExpr(getSCEV(U->getOperand(0)),
3630 getSCEV(U->getOperand(1)));
3631 case Instruction::Sub:
3632 return getMinusSCEV(getSCEV(U->getOperand(0)),
3633 getSCEV(U->getOperand(1)));
3634 case Instruction::And:
3635 // For an expression like x&255 that merely masks off the high bits,
3636 // use zext(trunc(x)) as the SCEV expression.
3637 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3638 if (CI->isNullValue())
3639 return getSCEV(U->getOperand(1));
3640 if (CI->isAllOnesValue())
3641 return getSCEV(U->getOperand(0));
3642 const APInt &A = CI->getValue();
3644 // Instcombine's ShrinkDemandedConstant may strip bits out of
3645 // constants, obscuring what would otherwise be a low-bits mask.
3646 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3647 // knew about to reconstruct a low-bits mask value.
3648 unsigned LZ = A.countLeadingZeros();
3649 unsigned BitWidth = A.getBitWidth();
3650 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3651 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD);
3653 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3655 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3657 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3658 IntegerType::get(getContext(), BitWidth - LZ)),
3663 case Instruction::Or:
3664 // If the RHS of the Or is a constant, we may have something like:
3665 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3666 // optimizations will transparently handle this case.
3668 // In order for this transformation to be safe, the LHS must be of the
3669 // form X*(2^n) and the Or constant must be less than 2^n.
3670 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3671 const SCEV *LHS = getSCEV(U->getOperand(0));
3672 const APInt &CIVal = CI->getValue();
3673 if (GetMinTrailingZeros(LHS) >=
3674 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3675 // Build a plain add SCEV.
3676 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3677 // If the LHS of the add was an addrec and it has no-wrap flags,
3678 // transfer the no-wrap flags, since an or won't introduce a wrap.
3679 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3680 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3681 const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags(
3682 OldAR->getNoWrapFlags());
3688 case Instruction::Xor:
3689 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3690 // If the RHS of the xor is a signbit, then this is just an add.
3691 // Instcombine turns add of signbit into xor as a strength reduction step.
3692 if (CI->getValue().isSignBit())
3693 return getAddExpr(getSCEV(U->getOperand(0)),
3694 getSCEV(U->getOperand(1)));
3696 // If the RHS of xor is -1, then this is a not operation.
3697 if (CI->isAllOnesValue())
3698 return getNotSCEV(getSCEV(U->getOperand(0)));
3700 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3701 // This is a variant of the check for xor with -1, and it handles
3702 // the case where instcombine has trimmed non-demanded bits out
3703 // of an xor with -1.
3704 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3705 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3706 if (BO->getOpcode() == Instruction::And &&
3707 LCI->getValue() == CI->getValue())
3708 if (const SCEVZeroExtendExpr *Z =
3709 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3710 Type *UTy = U->getType();
3711 const SCEV *Z0 = Z->getOperand();
3712 Type *Z0Ty = Z0->getType();
3713 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3715 // If C is a low-bits mask, the zero extend is serving to
3716 // mask off the high bits. Complement the operand and
3717 // re-apply the zext.
3718 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3719 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3721 // If C is a single bit, it may be in the sign-bit position
3722 // before the zero-extend. In this case, represent the xor
3723 // using an add, which is equivalent, and re-apply the zext.
3724 APInt Trunc = CI->getValue().trunc(Z0TySize);
3725 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3727 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3733 case Instruction::Shl:
3734 // Turn shift left of a constant amount into a multiply.
3735 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3736 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3738 // If the shift count is not less than the bitwidth, the result of
3739 // the shift is undefined. Don't try to analyze it, because the
3740 // resolution chosen here may differ from the resolution chosen in
3741 // other parts of the compiler.
3742 if (SA->getValue().uge(BitWidth))
3745 Constant *X = ConstantInt::get(getContext(),
3746 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3747 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3751 case Instruction::LShr:
3752 // Turn logical shift right of a constant into a unsigned divide.
3753 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3754 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3756 // If the shift count is not less than the bitwidth, the result of
3757 // the shift is undefined. Don't try to analyze it, because the
3758 // resolution chosen here may differ from the resolution chosen in
3759 // other parts of the compiler.
3760 if (SA->getValue().uge(BitWidth))
3763 Constant *X = ConstantInt::get(getContext(),
3764 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3765 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3769 case Instruction::AShr:
3770 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3771 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3772 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3773 if (L->getOpcode() == Instruction::Shl &&
3774 L->getOperand(1) == U->getOperand(1)) {
3775 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3777 // If the shift count is not less than the bitwidth, the result of
3778 // the shift is undefined. Don't try to analyze it, because the
3779 // resolution chosen here may differ from the resolution chosen in
3780 // other parts of the compiler.
3781 if (CI->getValue().uge(BitWidth))
3784 uint64_t Amt = BitWidth - CI->getZExtValue();
3785 if (Amt == BitWidth)
3786 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3788 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3789 IntegerType::get(getContext(),
3795 case Instruction::Trunc:
3796 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3798 case Instruction::ZExt:
3799 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3801 case Instruction::SExt:
3802 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3804 case Instruction::BitCast:
3805 // BitCasts are no-op casts so we just eliminate the cast.
3806 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3807 return getSCEV(U->getOperand(0));
3810 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3811 // lead to pointer expressions which cannot safely be expanded to GEPs,
3812 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3813 // simplifying integer expressions.
3815 case Instruction::GetElementPtr:
3816 return createNodeForGEP(cast<GEPOperator>(U));
3818 case Instruction::PHI:
3819 return createNodeForPHI(cast<PHINode>(U));
3821 case Instruction::Select:
3822 // This could be a smax or umax that was lowered earlier.
3823 // Try to recover it.
3824 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3825 Value *LHS = ICI->getOperand(0);
3826 Value *RHS = ICI->getOperand(1);
3827 switch (ICI->getPredicate()) {
3828 case ICmpInst::ICMP_SLT:
3829 case ICmpInst::ICMP_SLE:
3830 std::swap(LHS, RHS);
3832 case ICmpInst::ICMP_SGT:
3833 case ICmpInst::ICMP_SGE:
3834 // a >s b ? a+x : b+x -> smax(a, b)+x
3835 // a >s b ? b+x : a+x -> smin(a, b)+x
3836 if (LHS->getType() == U->getType()) {
3837 const SCEV *LS = getSCEV(LHS);
3838 const SCEV *RS = getSCEV(RHS);
3839 const SCEV *LA = getSCEV(U->getOperand(1));
3840 const SCEV *RA = getSCEV(U->getOperand(2));
3841 const SCEV *LDiff = getMinusSCEV(LA, LS);
3842 const SCEV *RDiff = getMinusSCEV(RA, RS);
3844 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3845 LDiff = getMinusSCEV(LA, RS);
3846 RDiff = getMinusSCEV(RA, LS);
3848 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3851 case ICmpInst::ICMP_ULT:
3852 case ICmpInst::ICMP_ULE:
3853 std::swap(LHS, RHS);
3855 case ICmpInst::ICMP_UGT:
3856 case ICmpInst::ICMP_UGE:
3857 // a >u b ? a+x : b+x -> umax(a, b)+x
3858 // a >u b ? b+x : a+x -> umin(a, b)+x
3859 if (LHS->getType() == U->getType()) {
3860 const SCEV *LS = getSCEV(LHS);
3861 const SCEV *RS = getSCEV(RHS);
3862 const SCEV *LA = getSCEV(U->getOperand(1));
3863 const SCEV *RA = getSCEV(U->getOperand(2));
3864 const SCEV *LDiff = getMinusSCEV(LA, LS);
3865 const SCEV *RDiff = getMinusSCEV(RA, RS);
3867 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3868 LDiff = getMinusSCEV(LA, RS);
3869 RDiff = getMinusSCEV(RA, LS);
3871 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3874 case ICmpInst::ICMP_NE:
3875 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3876 if (LHS->getType() == U->getType() &&
3877 isa<ConstantInt>(RHS) &&
3878 cast<ConstantInt>(RHS)->isZero()) {
3879 const SCEV *One = getConstant(LHS->getType(), 1);
3880 const SCEV *LS = getSCEV(LHS);
3881 const SCEV *LA = getSCEV(U->getOperand(1));
3882 const SCEV *RA = getSCEV(U->getOperand(2));
3883 const SCEV *LDiff = getMinusSCEV(LA, LS);
3884 const SCEV *RDiff = getMinusSCEV(RA, One);
3886 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3889 case ICmpInst::ICMP_EQ:
3890 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3891 if (LHS->getType() == U->getType() &&
3892 isa<ConstantInt>(RHS) &&
3893 cast<ConstantInt>(RHS)->isZero()) {
3894 const SCEV *One = getConstant(LHS->getType(), 1);
3895 const SCEV *LS = getSCEV(LHS);
3896 const SCEV *LA = getSCEV(U->getOperand(1));
3897 const SCEV *RA = getSCEV(U->getOperand(2));
3898 const SCEV *LDiff = getMinusSCEV(LA, One);
3899 const SCEV *RDiff = getMinusSCEV(RA, LS);
3901 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3909 default: // We cannot analyze this expression.
3913 return getUnknown(V);
3918 //===----------------------------------------------------------------------===//
3919 // Iteration Count Computation Code
3922 /// getSmallConstantTripCount - Returns the maximum trip count of this loop as a
3923 /// normal unsigned value. Returns 0 if the trip count is unknown or not
3924 /// constant. Will also return 0 if the maximum trip count is very large (>=
3927 /// This "trip count" assumes that control exits via ExitingBlock. More
3928 /// precisely, it is the number of times that control may reach ExitingBlock
3929 /// before taking the branch. For loops with multiple exits, it may not be the
3930 /// number times that the loop header executes because the loop may exit
3931 /// prematurely via another branch.
3932 unsigned ScalarEvolution::
3933 getSmallConstantTripCount(Loop *L, BasicBlock *ExitingBlock) {
3934 const SCEVConstant *ExitCount =
3935 dyn_cast<SCEVConstant>(getExitCount(L, ExitingBlock));
3939 ConstantInt *ExitConst = ExitCount->getValue();
3941 // Guard against huge trip counts.
3942 if (ExitConst->getValue().getActiveBits() > 32)
3945 // In case of integer overflow, this returns 0, which is correct.
3946 return ((unsigned)ExitConst->getZExtValue()) + 1;
3949 /// getSmallConstantTripMultiple - Returns the largest constant divisor of the
3950 /// trip count of this loop as a normal unsigned value, if possible. This
3951 /// means that the actual trip count is always a multiple of the returned
3952 /// value (don't forget the trip count could very well be zero as well!).
3954 /// Returns 1 if the trip count is unknown or not guaranteed to be the
3955 /// multiple of a constant (which is also the case if the trip count is simply
3956 /// constant, use getSmallConstantTripCount for that case), Will also return 1
3957 /// if the trip count is very large (>= 2^32).
3959 /// As explained in the comments for getSmallConstantTripCount, this assumes
3960 /// that control exits the loop via ExitingBlock.
3961 unsigned ScalarEvolution::
3962 getSmallConstantTripMultiple(Loop *L, BasicBlock *ExitingBlock) {
3963 const SCEV *ExitCount = getExitCount(L, ExitingBlock);
3964 if (ExitCount == getCouldNotCompute())
3967 // Get the trip count from the BE count by adding 1.
3968 const SCEV *TCMul = getAddExpr(ExitCount,
3969 getConstant(ExitCount->getType(), 1));
3970 // FIXME: SCEV distributes multiplication as V1*C1 + V2*C1. We could attempt
3971 // to factor simple cases.
3972 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(TCMul))
3973 TCMul = Mul->getOperand(0);
3975 const SCEVConstant *MulC = dyn_cast<SCEVConstant>(TCMul);
3979 ConstantInt *Result = MulC->getValue();
3981 // Guard against huge trip counts.
3982 if (!Result || Result->getValue().getActiveBits() > 32)
3985 return (unsigned)Result->getZExtValue();
3988 // getExitCount - Get the expression for the number of loop iterations for which
3989 // this loop is guaranteed not to exit via ExitintBlock. Otherwise return
3990 // SCEVCouldNotCompute.
3991 const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) {
3992 return getBackedgeTakenInfo(L).getExact(ExitingBlock, this);
3995 /// getBackedgeTakenCount - If the specified loop has a predictable
3996 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3997 /// object. The backedge-taken count is the number of times the loop header
3998 /// will be branched to from within the loop. This is one less than the
3999 /// trip count of the loop, since it doesn't count the first iteration,
4000 /// when the header is branched to from outside the loop.
4002 /// Note that it is not valid to call this method on a loop without a
4003 /// loop-invariant backedge-taken count (see
4004 /// hasLoopInvariantBackedgeTakenCount).
4006 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
4007 return getBackedgeTakenInfo(L).getExact(this);
4010 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
4011 /// return the least SCEV value that is known never to be less than the
4012 /// actual backedge taken count.
4013 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
4014 return getBackedgeTakenInfo(L).getMax(this);
4017 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
4018 /// onto the given Worklist.
4020 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
4021 BasicBlock *Header = L->getHeader();
4023 // Push all Loop-header PHIs onto the Worklist stack.
4024 for (BasicBlock::iterator I = Header->begin();
4025 PHINode *PN = dyn_cast<PHINode>(I); ++I)
4026 Worklist.push_back(PN);
4029 const ScalarEvolution::BackedgeTakenInfo &
4030 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
4031 // Initially insert an invalid entry for this loop. If the insertion
4032 // succeeds, proceed to actually compute a backedge-taken count and
4033 // update the value. The temporary CouldNotCompute value tells SCEV
4034 // code elsewhere that it shouldn't attempt to request a new
4035 // backedge-taken count, which could result in infinite recursion.
4036 std::pair<DenseMap<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
4037 BackedgeTakenCounts.insert(std::make_pair(L, BackedgeTakenInfo()));
4039 return Pair.first->second;
4041 // ComputeBackedgeTakenCount may allocate memory for its result. Inserting it
4042 // into the BackedgeTakenCounts map transfers ownership. Otherwise, the result
4043 // must be cleared in this scope.
4044 BackedgeTakenInfo Result = ComputeBackedgeTakenCount(L);
4046 if (Result.getExact(this) != getCouldNotCompute()) {
4047 assert(isLoopInvariant(Result.getExact(this), L) &&
4048 isLoopInvariant(Result.getMax(this), L) &&
4049 "Computed backedge-taken count isn't loop invariant for loop!");
4050 ++NumTripCountsComputed;
4052 else if (Result.getMax(this) == getCouldNotCompute() &&
4053 isa<PHINode>(L->getHeader()->begin())) {
4054 // Only count loops that have phi nodes as not being computable.
4055 ++NumTripCountsNotComputed;
4058 // Now that we know more about the trip count for this loop, forget any
4059 // existing SCEV values for PHI nodes in this loop since they are only
4060 // conservative estimates made without the benefit of trip count
4061 // information. This is similar to the code in forgetLoop, except that
4062 // it handles SCEVUnknown PHI nodes specially.
4063 if (Result.hasAnyInfo()) {
4064 SmallVector<Instruction *, 16> Worklist;
4065 PushLoopPHIs(L, Worklist);
4067 SmallPtrSet<Instruction *, 8> Visited;
4068 while (!Worklist.empty()) {
4069 Instruction *I = Worklist.pop_back_val();
4070 if (!Visited.insert(I)) continue;
4072 ValueExprMapType::iterator It =
4073 ValueExprMap.find_as(static_cast<Value *>(I));
4074 if (It != ValueExprMap.end()) {
4075 const SCEV *Old = It->second;
4077 // SCEVUnknown for a PHI either means that it has an unrecognized
4078 // structure, or it's a PHI that's in the progress of being computed
4079 // by createNodeForPHI. In the former case, additional loop trip
4080 // count information isn't going to change anything. In the later
4081 // case, createNodeForPHI will perform the necessary updates on its
4082 // own when it gets to that point.
4083 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
4084 forgetMemoizedResults(Old);
4085 ValueExprMap.erase(It);
4087 if (PHINode *PN = dyn_cast<PHINode>(I))
4088 ConstantEvolutionLoopExitValue.erase(PN);
4091 PushDefUseChildren(I, Worklist);
4095 // Re-lookup the insert position, since the call to
4096 // ComputeBackedgeTakenCount above could result in a
4097 // recusive call to getBackedgeTakenInfo (on a different
4098 // loop), which would invalidate the iterator computed
4100 return BackedgeTakenCounts.find(L)->second = Result;
4103 /// forgetLoop - This method should be called by the client when it has
4104 /// changed a loop in a way that may effect ScalarEvolution's ability to
4105 /// compute a trip count, or if the loop is deleted.
4106 void ScalarEvolution::forgetLoop(const Loop *L) {
4107 // Drop any stored trip count value.
4108 DenseMap<const Loop*, BackedgeTakenInfo>::iterator BTCPos =
4109 BackedgeTakenCounts.find(L);
4110 if (BTCPos != BackedgeTakenCounts.end()) {
4111 BTCPos->second.clear();
4112 BackedgeTakenCounts.erase(BTCPos);
4115 // Drop information about expressions based on loop-header PHIs.
4116 SmallVector<Instruction *, 16> Worklist;
4117 PushLoopPHIs(L, Worklist);
4119 SmallPtrSet<Instruction *, 8> Visited;
4120 while (!Worklist.empty()) {
4121 Instruction *I = Worklist.pop_back_val();
4122 if (!Visited.insert(I)) continue;
4124 ValueExprMapType::iterator It =
4125 ValueExprMap.find_as(static_cast<Value *>(I));
4126 if (It != ValueExprMap.end()) {
4127 forgetMemoizedResults(It->second);
4128 ValueExprMap.erase(It);
4129 if (PHINode *PN = dyn_cast<PHINode>(I))
4130 ConstantEvolutionLoopExitValue.erase(PN);
4133 PushDefUseChildren(I, Worklist);
4136 // Forget all contained loops too, to avoid dangling entries in the
4137 // ValuesAtScopes map.
4138 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
4142 /// forgetValue - This method should be called by the client when it has
4143 /// changed a value in a way that may effect its value, or which may
4144 /// disconnect it from a def-use chain linking it to a loop.
4145 void ScalarEvolution::forgetValue(Value *V) {
4146 Instruction *I = dyn_cast<Instruction>(V);
4149 // Drop information about expressions based on loop-header PHIs.
4150 SmallVector<Instruction *, 16> Worklist;
4151 Worklist.push_back(I);
4153 SmallPtrSet<Instruction *, 8> Visited;
4154 while (!Worklist.empty()) {
4155 I = Worklist.pop_back_val();
4156 if (!Visited.insert(I)) continue;
4158 ValueExprMapType::iterator It =
4159 ValueExprMap.find_as(static_cast<Value *>(I));
4160 if (It != ValueExprMap.end()) {
4161 forgetMemoizedResults(It->second);
4162 ValueExprMap.erase(It);
4163 if (PHINode *PN = dyn_cast<PHINode>(I))
4164 ConstantEvolutionLoopExitValue.erase(PN);
4167 PushDefUseChildren(I, Worklist);
4171 /// getExact - Get the exact loop backedge taken count considering all loop
4172 /// exits. A computable result can only be return for loops with a single exit.
4173 /// Returning the minimum taken count among all exits is incorrect because one
4174 /// of the loop's exit limit's may have been skipped. HowFarToZero assumes that
4175 /// the limit of each loop test is never skipped. This is a valid assumption as
4176 /// long as the loop exits via that test. For precise results, it is the
4177 /// caller's responsibility to specify the relevant loop exit using
4178 /// getExact(ExitingBlock, SE).
4180 ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
4181 // If any exits were not computable, the loop is not computable.
4182 if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute();
4184 // We need exactly one computable exit.
4185 if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
4186 assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
4188 const SCEV *BECount = 0;
4189 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4190 ENT != 0; ENT = ENT->getNextExit()) {
4192 assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV");
4195 BECount = ENT->ExactNotTaken;
4196 else if (BECount != ENT->ExactNotTaken)
4197 return SE->getCouldNotCompute();
4199 assert(BECount && "Invalid not taken count for loop exit");
4203 /// getExact - Get the exact not taken count for this loop exit.
4205 ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock,
4206 ScalarEvolution *SE) const {
4207 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4208 ENT != 0; ENT = ENT->getNextExit()) {
4210 if (ENT->ExitingBlock == ExitingBlock)
4211 return ENT->ExactNotTaken;
4213 return SE->getCouldNotCompute();
4216 /// getMax - Get the max backedge taken count for the loop.
4218 ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const {
4219 return Max ? Max : SE->getCouldNotCompute();
4222 /// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each
4223 /// computable exit into a persistent ExitNotTakenInfo array.
4224 ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo(
4225 SmallVectorImpl< std::pair<BasicBlock *, const SCEV *> > &ExitCounts,
4226 bool Complete, const SCEV *MaxCount) : Max(MaxCount) {
4229 ExitNotTaken.setIncomplete();
4231 unsigned NumExits = ExitCounts.size();
4232 if (NumExits == 0) return;
4234 ExitNotTaken.ExitingBlock = ExitCounts[0].first;
4235 ExitNotTaken.ExactNotTaken = ExitCounts[0].second;
4236 if (NumExits == 1) return;
4238 // Handle the rare case of multiple computable exits.
4239 ExitNotTakenInfo *ENT = new ExitNotTakenInfo[NumExits-1];
4241 ExitNotTakenInfo *PrevENT = &ExitNotTaken;
4242 for (unsigned i = 1; i < NumExits; ++i, PrevENT = ENT, ++ENT) {
4243 PrevENT->setNextExit(ENT);
4244 ENT->ExitingBlock = ExitCounts[i].first;
4245 ENT->ExactNotTaken = ExitCounts[i].second;
4249 /// clear - Invalidate this result and free the ExitNotTakenInfo array.
4250 void ScalarEvolution::BackedgeTakenInfo::clear() {
4251 ExitNotTaken.ExitingBlock = 0;
4252 ExitNotTaken.ExactNotTaken = 0;
4253 delete[] ExitNotTaken.getNextExit();
4256 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
4257 /// of the specified loop will execute.
4258 ScalarEvolution::BackedgeTakenInfo
4259 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
4260 SmallVector<BasicBlock *, 8> ExitingBlocks;
4261 L->getExitingBlocks(ExitingBlocks);
4263 // Examine all exits and pick the most conservative values.
4264 const SCEV *MaxBECount = getCouldNotCompute();
4265 bool CouldComputeBECount = true;
4266 SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts;
4267 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
4268 ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]);
4269 if (EL.Exact == getCouldNotCompute())
4270 // We couldn't compute an exact value for this exit, so
4271 // we won't be able to compute an exact value for the loop.
4272 CouldComputeBECount = false;
4274 ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact));
4276 if (MaxBECount == getCouldNotCompute())
4277 MaxBECount = EL.Max;
4278 else if (EL.Max != getCouldNotCompute()) {
4279 // We cannot take the "min" MaxBECount, because non-unit stride loops may
4280 // skip some loop tests. Taking the max over the exits is sufficiently
4281 // conservative. TODO: We could do better taking into consideration
4282 // that (1) the loop has unit stride (2) the last loop test is
4283 // less-than/greater-than (3) any loop test is less-than/greater-than AND
4284 // falls-through some constant times less then the other tests.
4285 MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
4289 return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
4292 /// ComputeExitLimit - Compute the number of times the backedge of the specified
4293 /// loop will execute if it exits via the specified block.
4294 ScalarEvolution::ExitLimit
4295 ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) {
4297 // Okay, we've chosen an exiting block. See what condition causes us to
4298 // exit at this block.
4300 // FIXME: we should be able to handle switch instructions (with a single exit)
4301 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
4302 if (ExitBr == 0) return getCouldNotCompute();
4303 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
4305 // At this point, we know we have a conditional branch that determines whether
4306 // the loop is exited. However, we don't know if the branch is executed each
4307 // time through the loop. If not, then the execution count of the branch will
4308 // not be equal to the trip count of the loop.
4310 // Currently we check for this by checking to see if the Exit branch goes to
4311 // the loop header. If so, we know it will always execute the same number of
4312 // times as the loop. We also handle the case where the exit block *is* the
4313 // loop header. This is common for un-rotated loops.
4315 // If both of those tests fail, walk up the unique predecessor chain to the
4316 // header, stopping if there is an edge that doesn't exit the loop. If the
4317 // header is reached, the execution count of the branch will be equal to the
4318 // trip count of the loop.
4320 // More extensive analysis could be done to handle more cases here.
4322 if (ExitBr->getSuccessor(0) != L->getHeader() &&
4323 ExitBr->getSuccessor(1) != L->getHeader() &&
4324 ExitBr->getParent() != L->getHeader()) {
4325 // The simple checks failed, try climbing the unique predecessor chain
4326 // up to the header.
4328 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
4329 BasicBlock *Pred = BB->getUniquePredecessor();
4331 return getCouldNotCompute();
4332 TerminatorInst *PredTerm = Pred->getTerminator();
4333 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
4334 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
4337 // If the predecessor has a successor that isn't BB and isn't
4338 // outside the loop, assume the worst.
4339 if (L->contains(PredSucc))
4340 return getCouldNotCompute();
4342 if (Pred == L->getHeader()) {
4349 return getCouldNotCompute();
4352 // Proceed to the next level to examine the exit condition expression.
4353 return ComputeExitLimitFromCond(L, ExitBr->getCondition(),
4354 ExitBr->getSuccessor(0),
4355 ExitBr->getSuccessor(1));
4358 /// ComputeExitLimitFromCond - Compute the number of times the
4359 /// backedge of the specified loop will execute if its exit condition
4360 /// were a conditional branch of ExitCond, TBB, and FBB.
4361 ScalarEvolution::ExitLimit
4362 ScalarEvolution::ComputeExitLimitFromCond(const Loop *L,
4366 // Check if the controlling expression for this loop is an And or Or.
4367 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
4368 if (BO->getOpcode() == Instruction::And) {
4369 // Recurse on the operands of the and.
4370 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB);
4371 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB);
4372 const SCEV *BECount = getCouldNotCompute();
4373 const SCEV *MaxBECount = getCouldNotCompute();
4374 if (L->contains(TBB)) {
4375 // Both conditions must be true for the loop to continue executing.
4376 // Choose the less conservative count.
4377 if (EL0.Exact == getCouldNotCompute() ||
4378 EL1.Exact == getCouldNotCompute())
4379 BECount = getCouldNotCompute();
4381 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4382 if (EL0.Max == getCouldNotCompute())
4383 MaxBECount = EL1.Max;
4384 else if (EL1.Max == getCouldNotCompute())
4385 MaxBECount = EL0.Max;
4387 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4389 // Both conditions must be true at the same time for the loop to exit.
4390 // For now, be conservative.
4391 assert(L->contains(FBB) && "Loop block has no successor in loop!");
4392 if (EL0.Max == EL1.Max)
4393 MaxBECount = EL0.Max;
4394 if (EL0.Exact == EL1.Exact)
4395 BECount = EL0.Exact;
4398 return ExitLimit(BECount, MaxBECount);
4400 if (BO->getOpcode() == Instruction::Or) {
4401 // Recurse on the operands of the or.
4402 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB);
4403 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB);
4404 const SCEV *BECount = getCouldNotCompute();
4405 const SCEV *MaxBECount = getCouldNotCompute();
4406 if (L->contains(FBB)) {
4407 // Both conditions must be false for the loop to continue executing.
4408 // Choose the less conservative count.
4409 if (EL0.Exact == getCouldNotCompute() ||
4410 EL1.Exact == getCouldNotCompute())
4411 BECount = getCouldNotCompute();
4413 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4414 if (EL0.Max == getCouldNotCompute())
4415 MaxBECount = EL1.Max;
4416 else if (EL1.Max == getCouldNotCompute())
4417 MaxBECount = EL0.Max;
4419 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4421 // Both conditions must be false at the same time for the loop to exit.
4422 // For now, be conservative.
4423 assert(L->contains(TBB) && "Loop block has no successor in loop!");
4424 if (EL0.Max == EL1.Max)
4425 MaxBECount = EL0.Max;
4426 if (EL0.Exact == EL1.Exact)
4427 BECount = EL0.Exact;
4430 return ExitLimit(BECount, MaxBECount);
4434 // With an icmp, it may be feasible to compute an exact backedge-taken count.
4435 // Proceed to the next level to examine the icmp.
4436 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
4437 return ComputeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB);
4439 // Check for a constant condition. These are normally stripped out by
4440 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
4441 // preserve the CFG and is temporarily leaving constant conditions
4443 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
4444 if (L->contains(FBB) == !CI->getZExtValue())
4445 // The backedge is always taken.
4446 return getCouldNotCompute();
4448 // The backedge is never taken.
4449 return getConstant(CI->getType(), 0);
4452 // If it's not an integer or pointer comparison then compute it the hard way.
4453 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4456 /// ComputeExitLimitFromICmp - Compute the number of times the
4457 /// backedge of the specified loop will execute if its exit condition
4458 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4459 ScalarEvolution::ExitLimit
4460 ScalarEvolution::ComputeExitLimitFromICmp(const Loop *L,
4465 // If the condition was exit on true, convert the condition to exit on false
4466 ICmpInst::Predicate Cond;
4467 if (!L->contains(FBB))
4468 Cond = ExitCond->getPredicate();
4470 Cond = ExitCond->getInversePredicate();
4472 // Handle common loops like: for (X = "string"; *X; ++X)
4473 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4474 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4476 ComputeLoadConstantCompareExitLimit(LI, RHS, L, Cond);
4477 if (ItCnt.hasAnyInfo())
4481 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4482 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4484 // Try to evaluate any dependencies out of the loop.
4485 LHS = getSCEVAtScope(LHS, L);
4486 RHS = getSCEVAtScope(RHS, L);
4488 // At this point, we would like to compute how many iterations of the
4489 // loop the predicate will return true for these inputs.
4490 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4491 // If there is a loop-invariant, force it into the RHS.
4492 std::swap(LHS, RHS);
4493 Cond = ICmpInst::getSwappedPredicate(Cond);
4496 // Simplify the operands before analyzing them.
4497 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4499 // If we have a comparison of a chrec against a constant, try to use value
4500 // ranges to answer this query.
4501 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4502 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4503 if (AddRec->getLoop() == L) {
4504 // Form the constant range.
4505 ConstantRange CompRange(
4506 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4508 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4509 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4513 case ICmpInst::ICMP_NE: { // while (X != Y)
4514 // Convert to: while (X-Y != 0)
4515 ExitLimit EL = HowFarToZero(getMinusSCEV(LHS, RHS), L);
4516 if (EL.hasAnyInfo()) return EL;
4519 case ICmpInst::ICMP_EQ: { // while (X == Y)
4520 // Convert to: while (X-Y == 0)
4521 ExitLimit EL = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4522 if (EL.hasAnyInfo()) return EL;
4525 case ICmpInst::ICMP_SLT: {
4526 ExitLimit EL = HowManyLessThans(LHS, RHS, L, true);
4527 if (EL.hasAnyInfo()) return EL;
4530 case ICmpInst::ICMP_SGT: {
4531 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4532 getNotSCEV(RHS), L, true);
4533 if (EL.hasAnyInfo()) return EL;
4536 case ICmpInst::ICMP_ULT: {
4537 ExitLimit EL = HowManyLessThans(LHS, RHS, L, false);
4538 if (EL.hasAnyInfo()) return EL;
4541 case ICmpInst::ICMP_UGT: {
4542 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4543 getNotSCEV(RHS), L, false);
4544 if (EL.hasAnyInfo()) return EL;
4549 dbgs() << "ComputeBackedgeTakenCount ";
4550 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4551 dbgs() << "[unsigned] ";
4552 dbgs() << *LHS << " "
4553 << Instruction::getOpcodeName(Instruction::ICmp)
4554 << " " << *RHS << "\n";
4558 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4561 static ConstantInt *
4562 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4563 ScalarEvolution &SE) {
4564 const SCEV *InVal = SE.getConstant(C);
4565 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4566 assert(isa<SCEVConstant>(Val) &&
4567 "Evaluation of SCEV at constant didn't fold correctly?");
4568 return cast<SCEVConstant>(Val)->getValue();
4571 /// ComputeLoadConstantCompareExitLimit - Given an exit condition of
4572 /// 'icmp op load X, cst', try to see if we can compute the backedge
4573 /// execution count.
4574 ScalarEvolution::ExitLimit
4575 ScalarEvolution::ComputeLoadConstantCompareExitLimit(
4579 ICmpInst::Predicate predicate) {
4581 if (LI->isVolatile()) return getCouldNotCompute();
4583 // Check to see if the loaded pointer is a getelementptr of a global.
4584 // TODO: Use SCEV instead of manually grubbing with GEPs.
4585 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4586 if (!GEP) return getCouldNotCompute();
4588 // Make sure that it is really a constant global we are gepping, with an
4589 // initializer, and make sure the first IDX is really 0.
4590 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4591 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4592 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4593 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4594 return getCouldNotCompute();
4596 // Okay, we allow one non-constant index into the GEP instruction.
4598 std::vector<Constant*> Indexes;
4599 unsigned VarIdxNum = 0;
4600 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4601 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4602 Indexes.push_back(CI);
4603 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4604 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4605 VarIdx = GEP->getOperand(i);
4607 Indexes.push_back(0);
4610 // Loop-invariant loads may be a byproduct of loop optimization. Skip them.
4612 return getCouldNotCompute();
4614 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4615 // Check to see if X is a loop variant variable value now.
4616 const SCEV *Idx = getSCEV(VarIdx);
4617 Idx = getSCEVAtScope(Idx, L);
4619 // We can only recognize very limited forms of loop index expressions, in
4620 // particular, only affine AddRec's like {C1,+,C2}.
4621 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4622 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4623 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4624 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4625 return getCouldNotCompute();
4627 unsigned MaxSteps = MaxBruteForceIterations;
4628 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4629 ConstantInt *ItCst = ConstantInt::get(
4630 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4631 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4633 // Form the GEP offset.
4634 Indexes[VarIdxNum] = Val;
4636 Constant *Result = ConstantFoldLoadThroughGEPIndices(GV->getInitializer(),
4638 if (Result == 0) break; // Cannot compute!
4640 // Evaluate the condition for this iteration.
4641 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4642 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4643 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4645 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4646 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4649 ++NumArrayLenItCounts;
4650 return getConstant(ItCst); // Found terminating iteration!
4653 return getCouldNotCompute();
4657 /// CanConstantFold - Return true if we can constant fold an instruction of the
4658 /// specified type, assuming that all operands were constants.
4659 static bool CanConstantFold(const Instruction *I) {
4660 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4661 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) ||
4665 if (const CallInst *CI = dyn_cast<CallInst>(I))
4666 if (const Function *F = CI->getCalledFunction())
4667 return canConstantFoldCallTo(F);
4671 /// Determine whether this instruction can constant evolve within this loop
4672 /// assuming its operands can all constant evolve.
4673 static bool canConstantEvolve(Instruction *I, const Loop *L) {
4674 // An instruction outside of the loop can't be derived from a loop PHI.
4675 if (!L->contains(I)) return false;
4677 if (isa<PHINode>(I)) {
4678 if (L->getHeader() == I->getParent())
4681 // We don't currently keep track of the control flow needed to evaluate
4682 // PHIs, so we cannot handle PHIs inside of loops.
4686 // If we won't be able to constant fold this expression even if the operands
4687 // are constants, bail early.
4688 return CanConstantFold(I);
4691 /// getConstantEvolvingPHIOperands - Implement getConstantEvolvingPHI by
4692 /// recursing through each instruction operand until reaching a loop header phi.
4694 getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L,
4695 DenseMap<Instruction *, PHINode *> &PHIMap) {
4697 // Otherwise, we can evaluate this instruction if all of its operands are
4698 // constant or derived from a PHI node themselves.
4700 for (Instruction::op_iterator OpI = UseInst->op_begin(),
4701 OpE = UseInst->op_end(); OpI != OpE; ++OpI) {
4703 if (isa<Constant>(*OpI)) continue;
4705 Instruction *OpInst = dyn_cast<Instruction>(*OpI);
4706 if (!OpInst || !canConstantEvolve(OpInst, L)) return 0;
4708 PHINode *P = dyn_cast<PHINode>(OpInst);
4710 // If this operand is already visited, reuse the prior result.
4711 // We may have P != PHI if this is the deepest point at which the
4712 // inconsistent paths meet.
4713 P = PHIMap.lookup(OpInst);
4715 // Recurse and memoize the results, whether a phi is found or not.
4716 // This recursive call invalidates pointers into PHIMap.
4717 P = getConstantEvolvingPHIOperands(OpInst, L, PHIMap);
4720 if (P == 0) return 0; // Not evolving from PHI
4721 if (PHI && PHI != P) return 0; // Evolving from multiple different PHIs.
4724 // This is a expression evolving from a constant PHI!
4728 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4729 /// in the loop that V is derived from. We allow arbitrary operations along the
4730 /// way, but the operands of an operation must either be constants or a value
4731 /// derived from a constant PHI. If this expression does not fit with these
4732 /// constraints, return null.
4733 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4734 Instruction *I = dyn_cast<Instruction>(V);
4735 if (I == 0 || !canConstantEvolve(I, L)) return 0;
4737 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4741 // Record non-constant instructions contained by the loop.
4742 DenseMap<Instruction *, PHINode *> PHIMap;
4743 return getConstantEvolvingPHIOperands(I, L, PHIMap);
4746 /// EvaluateExpression - Given an expression that passes the
4747 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4748 /// in the loop has the value PHIVal. If we can't fold this expression for some
4749 /// reason, return null.
4750 static Constant *EvaluateExpression(Value *V, const Loop *L,
4751 DenseMap<Instruction *, Constant *> &Vals,
4752 const TargetData *TD,
4753 const TargetLibraryInfo *TLI) {
4754 // Convenient constant check, but redundant for recursive calls.
4755 if (Constant *C = dyn_cast<Constant>(V)) return C;
4756 Instruction *I = dyn_cast<Instruction>(V);
4759 if (Constant *C = Vals.lookup(I)) return C;
4761 // An instruction inside the loop depends on a value outside the loop that we
4762 // weren't given a mapping for, or a value such as a call inside the loop.
4763 if (!canConstantEvolve(I, L)) return 0;
4765 // An unmapped PHI can be due to a branch or another loop inside this loop,
4766 // or due to this not being the initial iteration through a loop where we
4767 // couldn't compute the evolution of this particular PHI last time.
4768 if (isa<PHINode>(I)) return 0;
4770 std::vector<Constant*> Operands(I->getNumOperands());
4772 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4773 Instruction *Operand = dyn_cast<Instruction>(I->getOperand(i));
4775 Operands[i] = dyn_cast<Constant>(I->getOperand(i));
4776 if (!Operands[i]) return 0;
4779 Constant *C = EvaluateExpression(Operand, L, Vals, TD, TLI);
4785 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4786 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4787 Operands[1], TD, TLI);
4788 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4789 if (!LI->isVolatile())
4790 return ConstantFoldLoadFromConstPtr(Operands[0], TD);
4792 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Operands, TD,
4796 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4797 /// in the header of its containing loop, we know the loop executes a
4798 /// constant number of times, and the PHI node is just a recurrence
4799 /// involving constants, fold it.
4801 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4804 DenseMap<PHINode*, Constant*>::const_iterator I =
4805 ConstantEvolutionLoopExitValue.find(PN);
4806 if (I != ConstantEvolutionLoopExitValue.end())
4809 if (BEs.ugt(MaxBruteForceIterations))
4810 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4812 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4814 DenseMap<Instruction *, Constant *> CurrentIterVals;
4815 BasicBlock *Header = L->getHeader();
4816 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
4818 // Since the loop is canonicalized, the PHI node must have two entries. One
4819 // entry must be a constant (coming in from outside of the loop), and the
4820 // second must be derived from the same PHI.
4821 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4823 for (BasicBlock::iterator I = Header->begin();
4824 (PHI = dyn_cast<PHINode>(I)); ++I) {
4825 Constant *StartCST =
4826 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
4827 if (StartCST == 0) continue;
4828 CurrentIterVals[PHI] = StartCST;
4830 if (!CurrentIterVals.count(PN))
4833 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4835 // Execute the loop symbolically to determine the exit value.
4836 if (BEs.getActiveBits() >= 32)
4837 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4839 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4840 unsigned IterationNum = 0;
4841 for (; ; ++IterationNum) {
4842 if (IterationNum == NumIterations)
4843 return RetVal = CurrentIterVals[PN]; // Got exit value!
4845 // Compute the value of the PHIs for the next iteration.
4846 // EvaluateExpression adds non-phi values to the CurrentIterVals map.
4847 DenseMap<Instruction *, Constant *> NextIterVals;
4848 Constant *NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD,
4851 return 0; // Couldn't evaluate!
4852 NextIterVals[PN] = NextPHI;
4854 bool StoppedEvolving = NextPHI == CurrentIterVals[PN];
4856 // Also evaluate the other PHI nodes. However, we don't get to stop if we
4857 // cease to be able to evaluate one of them or if they stop evolving,
4858 // because that doesn't necessarily prevent us from computing PN.
4859 SmallVector<std::pair<PHINode *, Constant *>, 8> PHIsToCompute;
4860 for (DenseMap<Instruction *, Constant *>::const_iterator
4861 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
4862 PHINode *PHI = dyn_cast<PHINode>(I->first);
4863 if (!PHI || PHI == PN || PHI->getParent() != Header) continue;
4864 PHIsToCompute.push_back(std::make_pair(PHI, I->second));
4866 // We use two distinct loops because EvaluateExpression may invalidate any
4867 // iterators into CurrentIterVals.
4868 for (SmallVectorImpl<std::pair<PHINode *, Constant*> >::const_iterator
4869 I = PHIsToCompute.begin(), E = PHIsToCompute.end(); I != E; ++I) {
4870 PHINode *PHI = I->first;
4871 Constant *&NextPHI = NextIterVals[PHI];
4872 if (!NextPHI) { // Not already computed.
4873 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
4874 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
4876 if (NextPHI != I->second)
4877 StoppedEvolving = false;
4880 // If all entries in CurrentIterVals == NextIterVals then we can stop
4881 // iterating, the loop can't continue to change.
4882 if (StoppedEvolving)
4883 return RetVal = CurrentIterVals[PN];
4885 CurrentIterVals.swap(NextIterVals);
4889 /// ComputeExitCountExhaustively - If the loop is known to execute a
4890 /// constant number of times (the condition evolves only from constants),
4891 /// try to evaluate a few iterations of the loop until we get the exit
4892 /// condition gets a value of ExitWhen (true or false). If we cannot
4893 /// evaluate the trip count of the loop, return getCouldNotCompute().
4894 const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L,
4897 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4898 if (PN == 0) return getCouldNotCompute();
4900 // If the loop is canonicalized, the PHI will have exactly two entries.
4901 // That's the only form we support here.
4902 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4904 DenseMap<Instruction *, Constant *> CurrentIterVals;
4905 BasicBlock *Header = L->getHeader();
4906 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
4908 // One entry must be a constant (coming in from outside of the loop), and the
4909 // second must be derived from the same PHI.
4910 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4912 for (BasicBlock::iterator I = Header->begin();
4913 (PHI = dyn_cast<PHINode>(I)); ++I) {
4914 Constant *StartCST =
4915 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
4916 if (StartCST == 0) continue;
4917 CurrentIterVals[PHI] = StartCST;
4919 if (!CurrentIterVals.count(PN))
4920 return getCouldNotCompute();
4922 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4923 // the loop symbolically to determine when the condition gets a value of
4926 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4927 for (unsigned IterationNum = 0; IterationNum != MaxIterations;++IterationNum){
4928 ConstantInt *CondVal =
4929 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, L, CurrentIterVals,
4932 // Couldn't symbolically evaluate.
4933 if (!CondVal) return getCouldNotCompute();
4935 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4936 ++NumBruteForceTripCountsComputed;
4937 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4940 // Update all the PHI nodes for the next iteration.
4941 DenseMap<Instruction *, Constant *> NextIterVals;
4943 // Create a list of which PHIs we need to compute. We want to do this before
4944 // calling EvaluateExpression on them because that may invalidate iterators
4945 // into CurrentIterVals.
4946 SmallVector<PHINode *, 8> PHIsToCompute;
4947 for (DenseMap<Instruction *, Constant *>::const_iterator
4948 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
4949 PHINode *PHI = dyn_cast<PHINode>(I->first);
4950 if (!PHI || PHI->getParent() != Header) continue;
4951 PHIsToCompute.push_back(PHI);
4953 for (SmallVectorImpl<PHINode *>::const_iterator I = PHIsToCompute.begin(),
4954 E = PHIsToCompute.end(); I != E; ++I) {
4956 Constant *&NextPHI = NextIterVals[PHI];
4957 if (NextPHI) continue; // Already computed!
4959 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
4960 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
4962 CurrentIterVals.swap(NextIterVals);
4965 // Too many iterations were needed to evaluate.
4966 return getCouldNotCompute();
4969 /// getSCEVAtScope - Return a SCEV expression for the specified value
4970 /// at the specified scope in the program. The L value specifies a loop
4971 /// nest to evaluate the expression at, where null is the top-level or a
4972 /// specified loop is immediately inside of the loop.
4974 /// This method can be used to compute the exit value for a variable defined
4975 /// in a loop by querying what the value will hold in the parent loop.
4977 /// In the case that a relevant loop exit value cannot be computed, the
4978 /// original value V is returned.
4979 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4980 // Check to see if we've folded this expression at this loop before.
4981 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4982 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4983 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4985 return Pair.first->second ? Pair.first->second : V;
4987 // Otherwise compute it.
4988 const SCEV *C = computeSCEVAtScope(V, L);
4989 ValuesAtScopes[V][L] = C;
4993 /// This builds up a Constant using the ConstantExpr interface. That way, we
4994 /// will return Constants for objects which aren't represented by a
4995 /// SCEVConstant, because SCEVConstant is restricted to ConstantInt.
4996 /// Returns NULL if the SCEV isn't representable as a Constant.
4997 static Constant *BuildConstantFromSCEV(const SCEV *V) {
4998 switch (V->getSCEVType()) {
4999 default: // TODO: smax, umax.
5000 case scCouldNotCompute:
5004 return cast<SCEVConstant>(V)->getValue();
5006 return dyn_cast<Constant>(cast<SCEVUnknown>(V)->getValue());
5007 case scSignExtend: {
5008 const SCEVSignExtendExpr *SS = cast<SCEVSignExtendExpr>(V);
5009 if (Constant *CastOp = BuildConstantFromSCEV(SS->getOperand()))
5010 return ConstantExpr::getSExt(CastOp, SS->getType());
5013 case scZeroExtend: {
5014 const SCEVZeroExtendExpr *SZ = cast<SCEVZeroExtendExpr>(V);
5015 if (Constant *CastOp = BuildConstantFromSCEV(SZ->getOperand()))
5016 return ConstantExpr::getZExt(CastOp, SZ->getType());
5020 const SCEVTruncateExpr *ST = cast<SCEVTruncateExpr>(V);
5021 if (Constant *CastOp = BuildConstantFromSCEV(ST->getOperand()))
5022 return ConstantExpr::getTrunc(CastOp, ST->getType());
5026 const SCEVAddExpr *SA = cast<SCEVAddExpr>(V);
5027 if (Constant *C = BuildConstantFromSCEV(SA->getOperand(0))) {
5028 if (C->getType()->isPointerTy())
5029 C = ConstantExpr::getBitCast(C, Type::getInt8PtrTy(C->getContext()));
5030 for (unsigned i = 1, e = SA->getNumOperands(); i != e; ++i) {
5031 Constant *C2 = BuildConstantFromSCEV(SA->getOperand(i));
5035 if (!C->getType()->isPointerTy() && C2->getType()->isPointerTy()) {
5037 // The offsets have been converted to bytes. We can add bytes to an
5038 // i8* by GEP with the byte count in the first index.
5039 C = ConstantExpr::getBitCast(C,Type::getInt8PtrTy(C->getContext()));
5042 // Don't bother trying to sum two pointers. We probably can't
5043 // statically compute a load that results from it anyway.
5044 if (C2->getType()->isPointerTy())
5047 if (C->getType()->isPointerTy()) {
5048 if (cast<PointerType>(C->getType())->getElementType()->isStructTy())
5049 C2 = ConstantExpr::getIntegerCast(
5050 C2, Type::getInt32Ty(C->getContext()), true);
5051 C = ConstantExpr::getGetElementPtr(C, C2);
5053 C = ConstantExpr::getAdd(C, C2);
5060 const SCEVMulExpr *SM = cast<SCEVMulExpr>(V);
5061 if (Constant *C = BuildConstantFromSCEV(SM->getOperand(0))) {
5062 // Don't bother with pointers at all.
5063 if (C->getType()->isPointerTy()) return 0;
5064 for (unsigned i = 1, e = SM->getNumOperands(); i != e; ++i) {
5065 Constant *C2 = BuildConstantFromSCEV(SM->getOperand(i));
5066 if (!C2 || C2->getType()->isPointerTy()) return 0;
5067 C = ConstantExpr::getMul(C, C2);
5074 const SCEVUDivExpr *SU = cast<SCEVUDivExpr>(V);
5075 if (Constant *LHS = BuildConstantFromSCEV(SU->getLHS()))
5076 if (Constant *RHS = BuildConstantFromSCEV(SU->getRHS()))
5077 if (LHS->getType() == RHS->getType())
5078 return ConstantExpr::getUDiv(LHS, RHS);
5085 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
5086 if (isa<SCEVConstant>(V)) return V;
5088 // If this instruction is evolved from a constant-evolving PHI, compute the
5089 // exit value from the loop without using SCEVs.
5090 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
5091 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
5092 const Loop *LI = (*this->LI)[I->getParent()];
5093 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
5094 if (PHINode *PN = dyn_cast<PHINode>(I))
5095 if (PN->getParent() == LI->getHeader()) {
5096 // Okay, there is no closed form solution for the PHI node. Check
5097 // to see if the loop that contains it has a known backedge-taken
5098 // count. If so, we may be able to force computation of the exit
5100 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
5101 if (const SCEVConstant *BTCC =
5102 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
5103 // Okay, we know how many times the containing loop executes. If
5104 // this is a constant evolving PHI node, get the final value at
5105 // the specified iteration number.
5106 Constant *RV = getConstantEvolutionLoopExitValue(PN,
5107 BTCC->getValue()->getValue(),
5109 if (RV) return getSCEV(RV);
5113 // Okay, this is an expression that we cannot symbolically evaluate
5114 // into a SCEV. Check to see if it's possible to symbolically evaluate
5115 // the arguments into constants, and if so, try to constant propagate the
5116 // result. This is particularly useful for computing loop exit values.
5117 if (CanConstantFold(I)) {
5118 SmallVector<Constant *, 4> Operands;
5119 bool MadeImprovement = false;
5120 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
5121 Value *Op = I->getOperand(i);
5122 if (Constant *C = dyn_cast<Constant>(Op)) {
5123 Operands.push_back(C);
5127 // If any of the operands is non-constant and if they are
5128 // non-integer and non-pointer, don't even try to analyze them
5129 // with scev techniques.
5130 if (!isSCEVable(Op->getType()))
5133 const SCEV *OrigV = getSCEV(Op);
5134 const SCEV *OpV = getSCEVAtScope(OrigV, L);
5135 MadeImprovement |= OrigV != OpV;
5137 Constant *C = BuildConstantFromSCEV(OpV);
5139 if (C->getType() != Op->getType())
5140 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5144 Operands.push_back(C);
5147 // Check to see if getSCEVAtScope actually made an improvement.
5148 if (MadeImprovement) {
5150 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
5151 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
5152 Operands[0], Operands[1], TD,
5154 else if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
5155 if (!LI->isVolatile())
5156 C = ConstantFoldLoadFromConstPtr(Operands[0], TD);
5158 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
5166 // This is some other type of SCEVUnknown, just return it.
5170 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
5171 // Avoid performing the look-up in the common case where the specified
5172 // expression has no loop-variant portions.
5173 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
5174 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5175 if (OpAtScope != Comm->getOperand(i)) {
5176 // Okay, at least one of these operands is loop variant but might be
5177 // foldable. Build a new instance of the folded commutative expression.
5178 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
5179 Comm->op_begin()+i);
5180 NewOps.push_back(OpAtScope);
5182 for (++i; i != e; ++i) {
5183 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5184 NewOps.push_back(OpAtScope);
5186 if (isa<SCEVAddExpr>(Comm))
5187 return getAddExpr(NewOps);
5188 if (isa<SCEVMulExpr>(Comm))
5189 return getMulExpr(NewOps);
5190 if (isa<SCEVSMaxExpr>(Comm))
5191 return getSMaxExpr(NewOps);
5192 if (isa<SCEVUMaxExpr>(Comm))
5193 return getUMaxExpr(NewOps);
5194 llvm_unreachable("Unknown commutative SCEV type!");
5197 // If we got here, all operands are loop invariant.
5201 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
5202 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
5203 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
5204 if (LHS == Div->getLHS() && RHS == Div->getRHS())
5205 return Div; // must be loop invariant
5206 return getUDivExpr(LHS, RHS);
5209 // If this is a loop recurrence for a loop that does not contain L, then we
5210 // are dealing with the final value computed by the loop.
5211 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
5212 // First, attempt to evaluate each operand.
5213 // Avoid performing the look-up in the common case where the specified
5214 // expression has no loop-variant portions.
5215 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
5216 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
5217 if (OpAtScope == AddRec->getOperand(i))
5220 // Okay, at least one of these operands is loop variant but might be
5221 // foldable. Build a new instance of the folded commutative expression.
5222 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
5223 AddRec->op_begin()+i);
5224 NewOps.push_back(OpAtScope);
5225 for (++i; i != e; ++i)
5226 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
5228 const SCEV *FoldedRec =
5229 getAddRecExpr(NewOps, AddRec->getLoop(),
5230 AddRec->getNoWrapFlags(SCEV::FlagNW));
5231 AddRec = dyn_cast<SCEVAddRecExpr>(FoldedRec);
5232 // The addrec may be folded to a nonrecurrence, for example, if the
5233 // induction variable is multiplied by zero after constant folding. Go
5234 // ahead and return the folded value.
5240 // If the scope is outside the addrec's loop, evaluate it by using the
5241 // loop exit value of the addrec.
5242 if (!AddRec->getLoop()->contains(L)) {
5243 // To evaluate this recurrence, we need to know how many times the AddRec
5244 // loop iterates. Compute this now.
5245 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
5246 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
5248 // Then, evaluate the AddRec.
5249 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
5255 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
5256 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5257 if (Op == Cast->getOperand())
5258 return Cast; // must be loop invariant
5259 return getZeroExtendExpr(Op, Cast->getType());
5262 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
5263 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5264 if (Op == Cast->getOperand())
5265 return Cast; // must be loop invariant
5266 return getSignExtendExpr(Op, Cast->getType());
5269 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
5270 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5271 if (Op == Cast->getOperand())
5272 return Cast; // must be loop invariant
5273 return getTruncateExpr(Op, Cast->getType());
5276 llvm_unreachable("Unknown SCEV type!");
5279 /// getSCEVAtScope - This is a convenience function which does
5280 /// getSCEVAtScope(getSCEV(V), L).
5281 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
5282 return getSCEVAtScope(getSCEV(V), L);
5285 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
5286 /// following equation:
5288 /// A * X = B (mod N)
5290 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
5291 /// A and B isn't important.
5293 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
5294 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
5295 ScalarEvolution &SE) {
5296 uint32_t BW = A.getBitWidth();
5297 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
5298 assert(A != 0 && "A must be non-zero.");
5302 // The gcd of A and N may have only one prime factor: 2. The number of
5303 // trailing zeros in A is its multiplicity
5304 uint32_t Mult2 = A.countTrailingZeros();
5307 // 2. Check if B is divisible by D.
5309 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
5310 // is not less than multiplicity of this prime factor for D.
5311 if (B.countTrailingZeros() < Mult2)
5312 return SE.getCouldNotCompute();
5314 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
5317 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
5318 // bit width during computations.
5319 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
5320 APInt Mod(BW + 1, 0);
5321 Mod.setBit(BW - Mult2); // Mod = N / D
5322 APInt I = AD.multiplicativeInverse(Mod);
5324 // 4. Compute the minimum unsigned root of the equation:
5325 // I * (B / D) mod (N / D)
5326 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
5328 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
5330 return SE.getConstant(Result.trunc(BW));
5333 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
5334 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
5335 /// might be the same) or two SCEVCouldNotCompute objects.
5337 static std::pair<const SCEV *,const SCEV *>
5338 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
5339 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
5340 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
5341 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
5342 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
5344 // We currently can only solve this if the coefficients are constants.
5345 if (!LC || !MC || !NC) {
5346 const SCEV *CNC = SE.getCouldNotCompute();
5347 return std::make_pair(CNC, CNC);
5350 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
5351 const APInt &L = LC->getValue()->getValue();
5352 const APInt &M = MC->getValue()->getValue();
5353 const APInt &N = NC->getValue()->getValue();
5354 APInt Two(BitWidth, 2);
5355 APInt Four(BitWidth, 4);
5358 using namespace APIntOps;
5360 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
5361 // The B coefficient is M-N/2
5365 // The A coefficient is N/2
5366 APInt A(N.sdiv(Two));
5368 // Compute the B^2-4ac term.
5371 SqrtTerm -= Four * (A * C);
5373 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
5374 // integer value or else APInt::sqrt() will assert.
5375 APInt SqrtVal(SqrtTerm.sqrt());
5377 // Compute the two solutions for the quadratic formula.
5378 // The divisions must be performed as signed divisions.
5381 if (TwoA.isMinValue()) {
5382 const SCEV *CNC = SE.getCouldNotCompute();
5383 return std::make_pair(CNC, CNC);
5386 LLVMContext &Context = SE.getContext();
5388 ConstantInt *Solution1 =
5389 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
5390 ConstantInt *Solution2 =
5391 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
5393 return std::make_pair(SE.getConstant(Solution1),
5394 SE.getConstant(Solution2));
5395 } // end APIntOps namespace
5398 /// HowFarToZero - Return the number of times a backedge comparing the specified
5399 /// value to zero will execute. If not computable, return CouldNotCompute.
5401 /// This is only used for loops with a "x != y" exit test. The exit condition is
5402 /// now expressed as a single expression, V = x-y. So the exit test is
5403 /// effectively V != 0. We know and take advantage of the fact that this
5404 /// expression only being used in a comparison by zero context.
5405 ScalarEvolution::ExitLimit
5406 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
5407 // If the value is a constant
5408 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5409 // If the value is already zero, the branch will execute zero times.
5410 if (C->getValue()->isZero()) return C;
5411 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5414 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
5415 if (!AddRec || AddRec->getLoop() != L)
5416 return getCouldNotCompute();
5418 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
5419 // the quadratic equation to solve it.
5420 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
5421 std::pair<const SCEV *,const SCEV *> Roots =
5422 SolveQuadraticEquation(AddRec, *this);
5423 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5424 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5427 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
5428 << " sol#2: " << *R2 << "\n";
5430 // Pick the smallest positive root value.
5431 if (ConstantInt *CB =
5432 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
5435 if (CB->getZExtValue() == false)
5436 std::swap(R1, R2); // R1 is the minimum root now.
5438 // We can only use this value if the chrec ends up with an exact zero
5439 // value at this index. When solving for "X*X != 5", for example, we
5440 // should not accept a root of 2.
5441 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
5443 return R1; // We found a quadratic root!
5446 return getCouldNotCompute();
5449 // Otherwise we can only handle this if it is affine.
5450 if (!AddRec->isAffine())
5451 return getCouldNotCompute();
5453 // If this is an affine expression, the execution count of this branch is
5454 // the minimum unsigned root of the following equation:
5456 // Start + Step*N = 0 (mod 2^BW)
5460 // Step*N = -Start (mod 2^BW)
5462 // where BW is the common bit width of Start and Step.
5464 // Get the initial value for the loop.
5465 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
5466 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
5468 // For now we handle only constant steps.
5470 // TODO: Handle a nonconstant Step given AddRec<NUW>. If the
5471 // AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap
5472 // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step.
5473 // We have not yet seen any such cases.
5474 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
5475 if (StepC == 0 || StepC->getValue()->equalsInt(0))
5476 return getCouldNotCompute();
5478 // For positive steps (counting up until unsigned overflow):
5479 // N = -Start/Step (as unsigned)
5480 // For negative steps (counting down to zero):
5482 // First compute the unsigned distance from zero in the direction of Step.
5483 bool CountDown = StepC->getValue()->getValue().isNegative();
5484 const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
5486 // Handle unitary steps, which cannot wraparound.
5487 // 1*N = -Start; -1*N = Start (mod 2^BW), so:
5488 // N = Distance (as unsigned)
5489 if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) {
5490 ConstantRange CR = getUnsignedRange(Start);
5491 const SCEV *MaxBECount;
5492 if (!CountDown && CR.getUnsignedMin().isMinValue())
5493 // When counting up, the worst starting value is 1, not 0.
5494 MaxBECount = CR.getUnsignedMax().isMinValue()
5495 ? getConstant(APInt::getMinValue(CR.getBitWidth()))
5496 : getConstant(APInt::getMaxValue(CR.getBitWidth()));
5498 MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
5499 : -CR.getUnsignedMin());
5500 return ExitLimit(Distance, MaxBECount);
5503 // If the recurrence is known not to wraparound, unsigned divide computes the
5504 // back edge count. We know that the value will either become zero (and thus
5505 // the loop terminates), that the loop will terminate through some other exit
5506 // condition first, or that the loop has undefined behavior. This means
5507 // we can't "miss" the exit value, even with nonunit stride.
5509 // FIXME: Prove that loops always exhibits *acceptable* undefined
5510 // behavior. Loops must exhibit defined behavior until a wrapped value is
5511 // actually used. So the trip count computed by udiv could be smaller than the
5512 // number of well-defined iterations.
5513 if (AddRec->getNoWrapFlags(SCEV::FlagNW)) {
5514 // FIXME: We really want an "isexact" bit for udiv.
5515 return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
5517 // Then, try to solve the above equation provided that Start is constant.
5518 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
5519 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
5520 -StartC->getValue()->getValue(),
5522 return getCouldNotCompute();
5525 /// HowFarToNonZero - Return the number of times a backedge checking the
5526 /// specified value for nonzero will execute. If not computable, return
5528 ScalarEvolution::ExitLimit
5529 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
5530 // Loops that look like: while (X == 0) are very strange indeed. We don't
5531 // handle them yet except for the trivial case. This could be expanded in the
5532 // future as needed.
5534 // If the value is a constant, check to see if it is known to be non-zero
5535 // already. If so, the backedge will execute zero times.
5536 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5537 if (!C->getValue()->isNullValue())
5538 return getConstant(C->getType(), 0);
5539 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5542 // We could implement others, but I really doubt anyone writes loops like
5543 // this, and if they did, they would already be constant folded.
5544 return getCouldNotCompute();
5547 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
5548 /// (which may not be an immediate predecessor) which has exactly one
5549 /// successor from which BB is reachable, or null if no such block is
5552 std::pair<BasicBlock *, BasicBlock *>
5553 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
5554 // If the block has a unique predecessor, then there is no path from the
5555 // predecessor to the block that does not go through the direct edge
5556 // from the predecessor to the block.
5557 if (BasicBlock *Pred = BB->getSinglePredecessor())
5558 return std::make_pair(Pred, BB);
5560 // A loop's header is defined to be a block that dominates the loop.
5561 // If the header has a unique predecessor outside the loop, it must be
5562 // a block that has exactly one successor that can reach the loop.
5563 if (Loop *L = LI->getLoopFor(BB))
5564 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
5566 return std::pair<BasicBlock *, BasicBlock *>();
5569 /// HasSameValue - SCEV structural equivalence is usually sufficient for
5570 /// testing whether two expressions are equal, however for the purposes of
5571 /// looking for a condition guarding a loop, it can be useful to be a little
5572 /// more general, since a front-end may have replicated the controlling
5575 static bool HasSameValue(const SCEV *A, const SCEV *B) {
5576 // Quick check to see if they are the same SCEV.
5577 if (A == B) return true;
5579 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
5580 // two different instructions with the same value. Check for this case.
5581 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
5582 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
5583 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
5584 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
5585 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
5588 // Otherwise assume they may have a different value.
5592 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
5593 /// predicate Pred. Return true iff any changes were made.
5595 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
5596 const SCEV *&LHS, const SCEV *&RHS,
5598 bool Changed = false;
5600 // If we hit the max recursion limit bail out.
5604 // Canonicalize a constant to the right side.
5605 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
5606 // Check for both operands constant.
5607 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
5608 if (ConstantExpr::getICmp(Pred,
5610 RHSC->getValue())->isNullValue())
5611 goto trivially_false;
5613 goto trivially_true;
5615 // Otherwise swap the operands to put the constant on the right.
5616 std::swap(LHS, RHS);
5617 Pred = ICmpInst::getSwappedPredicate(Pred);
5621 // If we're comparing an addrec with a value which is loop-invariant in the
5622 // addrec's loop, put the addrec on the left. Also make a dominance check,
5623 // as both operands could be addrecs loop-invariant in each other's loop.
5624 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
5625 const Loop *L = AR->getLoop();
5626 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
5627 std::swap(LHS, RHS);
5628 Pred = ICmpInst::getSwappedPredicate(Pred);
5633 // If there's a constant operand, canonicalize comparisons with boundary
5634 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
5635 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
5636 const APInt &RA = RC->getValue()->getValue();
5638 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5639 case ICmpInst::ICMP_EQ:
5640 case ICmpInst::ICMP_NE:
5641 // Fold ((-1) * %a) + %b == 0 (equivalent to %b-%a == 0) into %a == %b.
5643 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(LHS))
5644 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(AE->getOperand(0)))
5645 if (AE->getNumOperands() == 2 && ME->getNumOperands() == 2 &&
5646 ME->getOperand(0)->isAllOnesValue()) {
5647 RHS = AE->getOperand(1);
5648 LHS = ME->getOperand(1);
5652 case ICmpInst::ICMP_UGE:
5653 if ((RA - 1).isMinValue()) {
5654 Pred = ICmpInst::ICMP_NE;
5655 RHS = getConstant(RA - 1);
5659 if (RA.isMaxValue()) {
5660 Pred = ICmpInst::ICMP_EQ;
5664 if (RA.isMinValue()) goto trivially_true;
5666 Pred = ICmpInst::ICMP_UGT;
5667 RHS = getConstant(RA - 1);
5670 case ICmpInst::ICMP_ULE:
5671 if ((RA + 1).isMaxValue()) {
5672 Pred = ICmpInst::ICMP_NE;
5673 RHS = getConstant(RA + 1);
5677 if (RA.isMinValue()) {
5678 Pred = ICmpInst::ICMP_EQ;
5682 if (RA.isMaxValue()) goto trivially_true;
5684 Pred = ICmpInst::ICMP_ULT;
5685 RHS = getConstant(RA + 1);
5688 case ICmpInst::ICMP_SGE:
5689 if ((RA - 1).isMinSignedValue()) {
5690 Pred = ICmpInst::ICMP_NE;
5691 RHS = getConstant(RA - 1);
5695 if (RA.isMaxSignedValue()) {
5696 Pred = ICmpInst::ICMP_EQ;
5700 if (RA.isMinSignedValue()) goto trivially_true;
5702 Pred = ICmpInst::ICMP_SGT;
5703 RHS = getConstant(RA - 1);
5706 case ICmpInst::ICMP_SLE:
5707 if ((RA + 1).isMaxSignedValue()) {
5708 Pred = ICmpInst::ICMP_NE;
5709 RHS = getConstant(RA + 1);
5713 if (RA.isMinSignedValue()) {
5714 Pred = ICmpInst::ICMP_EQ;
5718 if (RA.isMaxSignedValue()) goto trivially_true;
5720 Pred = ICmpInst::ICMP_SLT;
5721 RHS = getConstant(RA + 1);
5724 case ICmpInst::ICMP_UGT:
5725 if (RA.isMinValue()) {
5726 Pred = ICmpInst::ICMP_NE;
5730 if ((RA + 1).isMaxValue()) {
5731 Pred = ICmpInst::ICMP_EQ;
5732 RHS = getConstant(RA + 1);
5736 if (RA.isMaxValue()) goto trivially_false;
5738 case ICmpInst::ICMP_ULT:
5739 if (RA.isMaxValue()) {
5740 Pred = ICmpInst::ICMP_NE;
5744 if ((RA - 1).isMinValue()) {
5745 Pred = ICmpInst::ICMP_EQ;
5746 RHS = getConstant(RA - 1);
5750 if (RA.isMinValue()) goto trivially_false;
5752 case ICmpInst::ICMP_SGT:
5753 if (RA.isMinSignedValue()) {
5754 Pred = ICmpInst::ICMP_NE;
5758 if ((RA + 1).isMaxSignedValue()) {
5759 Pred = ICmpInst::ICMP_EQ;
5760 RHS = getConstant(RA + 1);
5764 if (RA.isMaxSignedValue()) goto trivially_false;
5766 case ICmpInst::ICMP_SLT:
5767 if (RA.isMaxSignedValue()) {
5768 Pred = ICmpInst::ICMP_NE;
5772 if ((RA - 1).isMinSignedValue()) {
5773 Pred = ICmpInst::ICMP_EQ;
5774 RHS = getConstant(RA - 1);
5778 if (RA.isMinSignedValue()) goto trivially_false;
5783 // Check for obvious equality.
5784 if (HasSameValue(LHS, RHS)) {
5785 if (ICmpInst::isTrueWhenEqual(Pred))
5786 goto trivially_true;
5787 if (ICmpInst::isFalseWhenEqual(Pred))
5788 goto trivially_false;
5791 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5792 // adding or subtracting 1 from one of the operands.
5794 case ICmpInst::ICMP_SLE:
5795 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5796 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5798 Pred = ICmpInst::ICMP_SLT;
5800 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5801 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5803 Pred = ICmpInst::ICMP_SLT;
5807 case ICmpInst::ICMP_SGE:
5808 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5809 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5811 Pred = ICmpInst::ICMP_SGT;
5813 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5814 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5816 Pred = ICmpInst::ICMP_SGT;
5820 case ICmpInst::ICMP_ULE:
5821 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5822 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5824 Pred = ICmpInst::ICMP_ULT;
5826 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5827 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5829 Pred = ICmpInst::ICMP_ULT;
5833 case ICmpInst::ICMP_UGE:
5834 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5835 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5837 Pred = ICmpInst::ICMP_UGT;
5839 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5840 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5842 Pred = ICmpInst::ICMP_UGT;
5850 // TODO: More simplifications are possible here.
5852 // Recursively simplify until we either hit a recursion limit or nothing
5855 return SimplifyICmpOperands(Pred, LHS, RHS, Depth+1);
5861 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5862 Pred = ICmpInst::ICMP_EQ;
5867 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5868 Pred = ICmpInst::ICMP_NE;
5872 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5873 return getSignedRange(S).getSignedMax().isNegative();
5876 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5877 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5880 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5881 return !getSignedRange(S).getSignedMin().isNegative();
5884 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5885 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5888 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5889 return isKnownNegative(S) || isKnownPositive(S);
5892 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5893 const SCEV *LHS, const SCEV *RHS) {
5894 // Canonicalize the inputs first.
5895 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5897 // If LHS or RHS is an addrec, check to see if the condition is true in
5898 // every iteration of the loop.
5899 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5900 if (isLoopEntryGuardedByCond(
5901 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5902 isLoopBackedgeGuardedByCond(
5903 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5905 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5906 if (isLoopEntryGuardedByCond(
5907 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5908 isLoopBackedgeGuardedByCond(
5909 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5912 // Otherwise see what can be done with known constant ranges.
5913 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5917 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5918 const SCEV *LHS, const SCEV *RHS) {
5919 if (HasSameValue(LHS, RHS))
5920 return ICmpInst::isTrueWhenEqual(Pred);
5922 // This code is split out from isKnownPredicate because it is called from
5923 // within isLoopEntryGuardedByCond.
5926 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5927 case ICmpInst::ICMP_SGT:
5928 Pred = ICmpInst::ICMP_SLT;
5929 std::swap(LHS, RHS);
5930 case ICmpInst::ICMP_SLT: {
5931 ConstantRange LHSRange = getSignedRange(LHS);
5932 ConstantRange RHSRange = getSignedRange(RHS);
5933 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5935 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5939 case ICmpInst::ICMP_SGE:
5940 Pred = ICmpInst::ICMP_SLE;
5941 std::swap(LHS, RHS);
5942 case ICmpInst::ICMP_SLE: {
5943 ConstantRange LHSRange = getSignedRange(LHS);
5944 ConstantRange RHSRange = getSignedRange(RHS);
5945 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5947 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
5951 case ICmpInst::ICMP_UGT:
5952 Pred = ICmpInst::ICMP_ULT;
5953 std::swap(LHS, RHS);
5954 case ICmpInst::ICMP_ULT: {
5955 ConstantRange LHSRange = getUnsignedRange(LHS);
5956 ConstantRange RHSRange = getUnsignedRange(RHS);
5957 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
5959 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
5963 case ICmpInst::ICMP_UGE:
5964 Pred = ICmpInst::ICMP_ULE;
5965 std::swap(LHS, RHS);
5966 case ICmpInst::ICMP_ULE: {
5967 ConstantRange LHSRange = getUnsignedRange(LHS);
5968 ConstantRange RHSRange = getUnsignedRange(RHS);
5969 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
5971 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
5975 case ICmpInst::ICMP_NE: {
5976 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
5978 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
5981 const SCEV *Diff = getMinusSCEV(LHS, RHS);
5982 if (isKnownNonZero(Diff))
5986 case ICmpInst::ICMP_EQ:
5987 // The check at the top of the function catches the case where
5988 // the values are known to be equal.
5994 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
5995 /// protected by a conditional between LHS and RHS. This is used to
5996 /// to eliminate casts.
5998 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
5999 ICmpInst::Predicate Pred,
6000 const SCEV *LHS, const SCEV *RHS) {
6001 // Interpret a null as meaning no loop, where there is obviously no guard
6002 // (interprocedural conditions notwithstanding).
6003 if (!L) return true;
6005 BasicBlock *Latch = L->getLoopLatch();
6009 BranchInst *LoopContinuePredicate =
6010 dyn_cast<BranchInst>(Latch->getTerminator());
6011 if (!LoopContinuePredicate ||
6012 LoopContinuePredicate->isUnconditional())
6015 return isImpliedCond(Pred, LHS, RHS,
6016 LoopContinuePredicate->getCondition(),
6017 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
6020 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
6021 /// by a conditional between LHS and RHS. This is used to help avoid max
6022 /// expressions in loop trip counts, and to eliminate casts.
6024 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
6025 ICmpInst::Predicate Pred,
6026 const SCEV *LHS, const SCEV *RHS) {
6027 // Interpret a null as meaning no loop, where there is obviously no guard
6028 // (interprocedural conditions notwithstanding).
6029 if (!L) return false;
6031 // Starting at the loop predecessor, climb up the predecessor chain, as long
6032 // as there are predecessors that can be found that have unique successors
6033 // leading to the original header.
6034 for (std::pair<BasicBlock *, BasicBlock *>
6035 Pair(L->getLoopPredecessor(), L->getHeader());
6037 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
6039 BranchInst *LoopEntryPredicate =
6040 dyn_cast<BranchInst>(Pair.first->getTerminator());
6041 if (!LoopEntryPredicate ||
6042 LoopEntryPredicate->isUnconditional())
6045 if (isImpliedCond(Pred, LHS, RHS,
6046 LoopEntryPredicate->getCondition(),
6047 LoopEntryPredicate->getSuccessor(0) != Pair.second))
6054 /// RAII wrapper to prevent recursive application of isImpliedCond.
6055 /// ScalarEvolution's PendingLoopPredicates set must be empty unless we are
6056 /// currently evaluating isImpliedCond.
6057 struct MarkPendingLoopPredicate {
6059 DenseSet<Value*> &LoopPreds;
6062 MarkPendingLoopPredicate(Value *C, DenseSet<Value*> &LP)
6063 : Cond(C), LoopPreds(LP) {
6064 Pending = !LoopPreds.insert(Cond).second;
6066 ~MarkPendingLoopPredicate() {
6068 LoopPreds.erase(Cond);
6072 /// isImpliedCond - Test whether the condition described by Pred, LHS,
6073 /// and RHS is true whenever the given Cond value evaluates to true.
6074 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
6075 const SCEV *LHS, const SCEV *RHS,
6076 Value *FoundCondValue,
6078 MarkPendingLoopPredicate Mark(FoundCondValue, PendingLoopPredicates);
6082 // Recursively handle And and Or conditions.
6083 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
6084 if (BO->getOpcode() == Instruction::And) {
6086 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6087 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6088 } else if (BO->getOpcode() == Instruction::Or) {
6090 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6091 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6095 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
6096 if (!ICI) return false;
6098 // Bail if the ICmp's operands' types are wider than the needed type
6099 // before attempting to call getSCEV on them. This avoids infinite
6100 // recursion, since the analysis of widening casts can require loop
6101 // exit condition information for overflow checking, which would
6103 if (getTypeSizeInBits(LHS->getType()) <
6104 getTypeSizeInBits(ICI->getOperand(0)->getType()))
6107 // Now that we found a conditional branch that dominates the loop, check to
6108 // see if it is the comparison we are looking for.
6109 ICmpInst::Predicate FoundPred;
6111 FoundPred = ICI->getInversePredicate();
6113 FoundPred = ICI->getPredicate();
6115 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
6116 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
6118 // Balance the types. The case where FoundLHS' type is wider than
6119 // LHS' type is checked for above.
6120 if (getTypeSizeInBits(LHS->getType()) >
6121 getTypeSizeInBits(FoundLHS->getType())) {
6122 if (CmpInst::isSigned(Pred)) {
6123 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
6124 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
6126 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
6127 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
6131 // Canonicalize the query to match the way instcombine will have
6132 // canonicalized the comparison.
6133 if (SimplifyICmpOperands(Pred, LHS, RHS))
6135 return CmpInst::isTrueWhenEqual(Pred);
6136 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
6137 if (FoundLHS == FoundRHS)
6138 return CmpInst::isFalseWhenEqual(Pred);
6140 // Check to see if we can make the LHS or RHS match.
6141 if (LHS == FoundRHS || RHS == FoundLHS) {
6142 if (isa<SCEVConstant>(RHS)) {
6143 std::swap(FoundLHS, FoundRHS);
6144 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
6146 std::swap(LHS, RHS);
6147 Pred = ICmpInst::getSwappedPredicate(Pred);
6151 // Check whether the found predicate is the same as the desired predicate.
6152 if (FoundPred == Pred)
6153 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
6155 // Check whether swapping the found predicate makes it the same as the
6156 // desired predicate.
6157 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
6158 if (isa<SCEVConstant>(RHS))
6159 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
6161 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
6162 RHS, LHS, FoundLHS, FoundRHS);
6165 // Check whether the actual condition is beyond sufficient.
6166 if (FoundPred == ICmpInst::ICMP_EQ)
6167 if (ICmpInst::isTrueWhenEqual(Pred))
6168 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
6170 if (Pred == ICmpInst::ICMP_NE)
6171 if (!ICmpInst::isTrueWhenEqual(FoundPred))
6172 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
6175 // Otherwise assume the worst.
6179 /// isImpliedCondOperands - Test whether the condition described by Pred,
6180 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
6181 /// and FoundRHS is true.
6182 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
6183 const SCEV *LHS, const SCEV *RHS,
6184 const SCEV *FoundLHS,
6185 const SCEV *FoundRHS) {
6186 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
6187 FoundLHS, FoundRHS) ||
6188 // ~x < ~y --> x > y
6189 isImpliedCondOperandsHelper(Pred, LHS, RHS,
6190 getNotSCEV(FoundRHS),
6191 getNotSCEV(FoundLHS));
6194 /// isImpliedCondOperandsHelper - Test whether the condition described by
6195 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
6196 /// FoundLHS, and FoundRHS is true.
6198 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
6199 const SCEV *LHS, const SCEV *RHS,
6200 const SCEV *FoundLHS,
6201 const SCEV *FoundRHS) {
6203 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
6204 case ICmpInst::ICMP_EQ:
6205 case ICmpInst::ICMP_NE:
6206 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
6209 case ICmpInst::ICMP_SLT:
6210 case ICmpInst::ICMP_SLE:
6211 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
6212 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
6215 case ICmpInst::ICMP_SGT:
6216 case ICmpInst::ICMP_SGE:
6217 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
6218 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
6221 case ICmpInst::ICMP_ULT:
6222 case ICmpInst::ICMP_ULE:
6223 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
6224 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
6227 case ICmpInst::ICMP_UGT:
6228 case ICmpInst::ICMP_UGE:
6229 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
6230 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
6238 /// getBECount - Subtract the end and start values and divide by the step,
6239 /// rounding up, to get the number of times the backedge is executed. Return
6240 /// CouldNotCompute if an intermediate computation overflows.
6241 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
6245 assert(!isKnownNegative(Step) &&
6246 "This code doesn't handle negative strides yet!");
6248 Type *Ty = Start->getType();
6250 // When Start == End, we have an exact BECount == 0. Short-circuit this case
6251 // here because SCEV may not be able to determine that the unsigned division
6252 // after rounding is zero.
6254 return getConstant(Ty, 0);
6256 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
6257 const SCEV *Diff = getMinusSCEV(End, Start);
6258 const SCEV *RoundUp = getAddExpr(Step, NegOne);
6260 // Add an adjustment to the difference between End and Start so that
6261 // the division will effectively round up.
6262 const SCEV *Add = getAddExpr(Diff, RoundUp);
6265 // Check Add for unsigned overflow.
6266 // TODO: More sophisticated things could be done here.
6267 Type *WideTy = IntegerType::get(getContext(),
6268 getTypeSizeInBits(Ty) + 1);
6269 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
6270 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
6271 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
6272 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
6273 return getCouldNotCompute();
6276 return getUDivExpr(Add, Step);
6279 /// HowManyLessThans - Return the number of times a backedge containing the
6280 /// specified less-than comparison will execute. If not computable, return
6281 /// CouldNotCompute.
6282 ScalarEvolution::ExitLimit
6283 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
6284 const Loop *L, bool isSigned) {
6285 // Only handle: "ADDREC < LoopInvariant".
6286 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute();
6288 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
6289 if (!AddRec || AddRec->getLoop() != L)
6290 return getCouldNotCompute();
6292 // Check to see if we have a flag which makes analysis easy.
6293 bool NoWrap = isSigned ?
6294 AddRec->getNoWrapFlags((SCEV::NoWrapFlags)(SCEV::FlagNSW | SCEV::FlagNW)) :
6295 AddRec->getNoWrapFlags((SCEV::NoWrapFlags)(SCEV::FlagNUW | SCEV::FlagNW));
6297 if (AddRec->isAffine()) {
6298 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
6299 const SCEV *Step = AddRec->getStepRecurrence(*this);
6302 return getCouldNotCompute();
6303 if (Step->isOne()) {
6304 // With unit stride, the iteration never steps past the limit value.
6305 } else if (isKnownPositive(Step)) {
6306 // Test whether a positive iteration can step past the limit
6307 // value and past the maximum value for its type in a single step.
6308 // Note that it's not sufficient to check NoWrap here, because even
6309 // though the value after a wrap is undefined, it's not undefined
6310 // behavior, so if wrap does occur, the loop could either terminate or
6311 // loop infinitely, but in either case, the loop is guaranteed to
6312 // iterate at least until the iteration where the wrapping occurs.
6313 const SCEV *One = getConstant(Step->getType(), 1);
6315 APInt Max = APInt::getSignedMaxValue(BitWidth);
6316 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
6317 .slt(getSignedRange(RHS).getSignedMax()))
6318 return getCouldNotCompute();
6320 APInt Max = APInt::getMaxValue(BitWidth);
6321 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
6322 .ult(getUnsignedRange(RHS).getUnsignedMax()))
6323 return getCouldNotCompute();
6326 // TODO: Handle negative strides here and below.
6327 return getCouldNotCompute();
6329 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
6330 // m. So, we count the number of iterations in which {n,+,s} < m is true.
6331 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
6332 // treat m-n as signed nor unsigned due to overflow possibility.
6334 // First, we get the value of the LHS in the first iteration: n
6335 const SCEV *Start = AddRec->getOperand(0);
6337 // Determine the minimum constant start value.
6338 const SCEV *MinStart = getConstant(isSigned ?
6339 getSignedRange(Start).getSignedMin() :
6340 getUnsignedRange(Start).getUnsignedMin());
6342 // If we know that the condition is true in order to enter the loop,
6343 // then we know that it will run exactly (m-n)/s times. Otherwise, we
6344 // only know that it will execute (max(m,n)-n)/s times. In both cases,
6345 // the division must round up.
6346 const SCEV *End = RHS;
6347 if (!isLoopEntryGuardedByCond(L,
6348 isSigned ? ICmpInst::ICMP_SLT :
6350 getMinusSCEV(Start, Step), RHS))
6351 End = isSigned ? getSMaxExpr(RHS, Start)
6352 : getUMaxExpr(RHS, Start);
6354 // Determine the maximum constant end value.
6355 const SCEV *MaxEnd = getConstant(isSigned ?
6356 getSignedRange(End).getSignedMax() :
6357 getUnsignedRange(End).getUnsignedMax());
6359 // If MaxEnd is within a step of the maximum integer value in its type,
6360 // adjust it down to the minimum value which would produce the same effect.
6361 // This allows the subsequent ceiling division of (N+(step-1))/step to
6362 // compute the correct value.
6363 const SCEV *StepMinusOne = getMinusSCEV(Step,
6364 getConstant(Step->getType(), 1));
6367 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
6370 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
6373 // Finally, we subtract these two values and divide, rounding up, to get
6374 // the number of times the backedge is executed.
6375 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
6377 // The maximum backedge count is similar, except using the minimum start
6378 // value and the maximum end value.
6379 // If we already have an exact constant BECount, use it instead.
6380 const SCEV *MaxBECount = isa<SCEVConstant>(BECount) ? BECount
6381 : getBECount(MinStart, MaxEnd, Step, NoWrap);
6383 // If the stride is nonconstant, and NoWrap == true, then
6384 // getBECount(MinStart, MaxEnd) may not compute. This would result in an
6385 // exact BECount and invalid MaxBECount, which should be avoided to catch
6386 // more optimization opportunities.
6387 if (isa<SCEVCouldNotCompute>(MaxBECount))
6388 MaxBECount = BECount;
6390 return ExitLimit(BECount, MaxBECount);
6393 return getCouldNotCompute();
6396 /// getNumIterationsInRange - Return the number of iterations of this loop that
6397 /// produce values in the specified constant range. Another way of looking at
6398 /// this is that it returns the first iteration number where the value is not in
6399 /// the condition, thus computing the exit count. If the iteration count can't
6400 /// be computed, an instance of SCEVCouldNotCompute is returned.
6401 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
6402 ScalarEvolution &SE) const {
6403 if (Range.isFullSet()) // Infinite loop.
6404 return SE.getCouldNotCompute();
6406 // If the start is a non-zero constant, shift the range to simplify things.
6407 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
6408 if (!SC->getValue()->isZero()) {
6409 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
6410 Operands[0] = SE.getConstant(SC->getType(), 0);
6411 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop(),
6412 getNoWrapFlags(FlagNW));
6413 if (const SCEVAddRecExpr *ShiftedAddRec =
6414 dyn_cast<SCEVAddRecExpr>(Shifted))
6415 return ShiftedAddRec->getNumIterationsInRange(
6416 Range.subtract(SC->getValue()->getValue()), SE);
6417 // This is strange and shouldn't happen.
6418 return SE.getCouldNotCompute();
6421 // The only time we can solve this is when we have all constant indices.
6422 // Otherwise, we cannot determine the overflow conditions.
6423 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
6424 if (!isa<SCEVConstant>(getOperand(i)))
6425 return SE.getCouldNotCompute();
6428 // Okay at this point we know that all elements of the chrec are constants and
6429 // that the start element is zero.
6431 // First check to see if the range contains zero. If not, the first
6433 unsigned BitWidth = SE.getTypeSizeInBits(getType());
6434 if (!Range.contains(APInt(BitWidth, 0)))
6435 return SE.getConstant(getType(), 0);
6438 // If this is an affine expression then we have this situation:
6439 // Solve {0,+,A} in Range === Ax in Range
6441 // We know that zero is in the range. If A is positive then we know that
6442 // the upper value of the range must be the first possible exit value.
6443 // If A is negative then the lower of the range is the last possible loop
6444 // value. Also note that we already checked for a full range.
6445 APInt One(BitWidth,1);
6446 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
6447 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
6449 // The exit value should be (End+A)/A.
6450 APInt ExitVal = (End + A).udiv(A);
6451 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
6453 // Evaluate at the exit value. If we really did fall out of the valid
6454 // range, then we computed our trip count, otherwise wrap around or other
6455 // things must have happened.
6456 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
6457 if (Range.contains(Val->getValue()))
6458 return SE.getCouldNotCompute(); // Something strange happened
6460 // Ensure that the previous value is in the range. This is a sanity check.
6461 assert(Range.contains(
6462 EvaluateConstantChrecAtConstant(this,
6463 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
6464 "Linear scev computation is off in a bad way!");
6465 return SE.getConstant(ExitValue);
6466 } else if (isQuadratic()) {
6467 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
6468 // quadratic equation to solve it. To do this, we must frame our problem in
6469 // terms of figuring out when zero is crossed, instead of when
6470 // Range.getUpper() is crossed.
6471 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
6472 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
6473 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop(),
6474 // getNoWrapFlags(FlagNW)
6477 // Next, solve the constructed addrec
6478 std::pair<const SCEV *,const SCEV *> Roots =
6479 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
6480 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
6481 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
6483 // Pick the smallest positive root value.
6484 if (ConstantInt *CB =
6485 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
6486 R1->getValue(), R2->getValue()))) {
6487 if (CB->getZExtValue() == false)
6488 std::swap(R1, R2); // R1 is the minimum root now.
6490 // Make sure the root is not off by one. The returned iteration should
6491 // not be in the range, but the previous one should be. When solving
6492 // for "X*X < 5", for example, we should not return a root of 2.
6493 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
6496 if (Range.contains(R1Val->getValue())) {
6497 // The next iteration must be out of the range...
6498 ConstantInt *NextVal =
6499 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
6501 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6502 if (!Range.contains(R1Val->getValue()))
6503 return SE.getConstant(NextVal);
6504 return SE.getCouldNotCompute(); // Something strange happened
6507 // If R1 was not in the range, then it is a good return value. Make
6508 // sure that R1-1 WAS in the range though, just in case.
6509 ConstantInt *NextVal =
6510 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
6511 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6512 if (Range.contains(R1Val->getValue()))
6514 return SE.getCouldNotCompute(); // Something strange happened
6519 return SE.getCouldNotCompute();
6524 //===----------------------------------------------------------------------===//
6525 // SCEVCallbackVH Class Implementation
6526 //===----------------------------------------------------------------------===//
6528 void ScalarEvolution::SCEVCallbackVH::deleted() {
6529 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6530 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
6531 SE->ConstantEvolutionLoopExitValue.erase(PN);
6532 SE->ValueExprMap.erase(getValPtr());
6533 // this now dangles!
6536 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
6537 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6539 // Forget all the expressions associated with users of the old value,
6540 // so that future queries will recompute the expressions using the new
6542 Value *Old = getValPtr();
6543 SmallVector<User *, 16> Worklist;
6544 SmallPtrSet<User *, 8> Visited;
6545 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
6547 Worklist.push_back(*UI);
6548 while (!Worklist.empty()) {
6549 User *U = Worklist.pop_back_val();
6550 // Deleting the Old value will cause this to dangle. Postpone
6551 // that until everything else is done.
6554 if (!Visited.insert(U))
6556 if (PHINode *PN = dyn_cast<PHINode>(U))
6557 SE->ConstantEvolutionLoopExitValue.erase(PN);
6558 SE->ValueExprMap.erase(U);
6559 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
6561 Worklist.push_back(*UI);
6563 // Delete the Old value.
6564 if (PHINode *PN = dyn_cast<PHINode>(Old))
6565 SE->ConstantEvolutionLoopExitValue.erase(PN);
6566 SE->ValueExprMap.erase(Old);
6567 // this now dangles!
6570 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
6571 : CallbackVH(V), SE(se) {}
6573 //===----------------------------------------------------------------------===//
6574 // ScalarEvolution Class Implementation
6575 //===----------------------------------------------------------------------===//
6577 ScalarEvolution::ScalarEvolution()
6578 : FunctionPass(ID), FirstUnknown(0) {
6579 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
6582 bool ScalarEvolution::runOnFunction(Function &F) {
6584 LI = &getAnalysis<LoopInfo>();
6585 TD = getAnalysisIfAvailable<TargetData>();
6586 TLI = &getAnalysis<TargetLibraryInfo>();
6587 DT = &getAnalysis<DominatorTree>();
6591 void ScalarEvolution::releaseMemory() {
6592 // Iterate through all the SCEVUnknown instances and call their
6593 // destructors, so that they release their references to their values.
6594 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
6598 ValueExprMap.clear();
6600 // Free any extra memory created for ExitNotTakenInfo in the unlikely event
6601 // that a loop had multiple computable exits.
6602 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
6603 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end();
6608 assert(PendingLoopPredicates.empty() && "isImpliedCond garbage");
6610 BackedgeTakenCounts.clear();
6611 ConstantEvolutionLoopExitValue.clear();
6612 ValuesAtScopes.clear();
6613 LoopDispositions.clear();
6614 BlockDispositions.clear();
6615 UnsignedRanges.clear();
6616 SignedRanges.clear();
6617 UniqueSCEVs.clear();
6618 SCEVAllocator.Reset();
6621 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
6622 AU.setPreservesAll();
6623 AU.addRequiredTransitive<LoopInfo>();
6624 AU.addRequiredTransitive<DominatorTree>();
6625 AU.addRequired<TargetLibraryInfo>();
6628 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
6629 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
6632 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
6634 // Print all inner loops first
6635 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
6636 PrintLoopInfo(OS, SE, *I);
6639 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6642 SmallVector<BasicBlock *, 8> ExitBlocks;
6643 L->getExitBlocks(ExitBlocks);
6644 if (ExitBlocks.size() != 1)
6645 OS << "<multiple exits> ";
6647 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
6648 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
6650 OS << "Unpredictable backedge-taken count. ";
6655 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6658 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
6659 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
6661 OS << "Unpredictable max backedge-taken count. ";
6667 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
6668 // ScalarEvolution's implementation of the print method is to print
6669 // out SCEV values of all instructions that are interesting. Doing
6670 // this potentially causes it to create new SCEV objects though,
6671 // which technically conflicts with the const qualifier. This isn't
6672 // observable from outside the class though, so casting away the
6673 // const isn't dangerous.
6674 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
6676 OS << "Classifying expressions for: ";
6677 WriteAsOperand(OS, F, /*PrintType=*/false);
6679 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
6680 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
6683 const SCEV *SV = SE.getSCEV(&*I);
6686 const Loop *L = LI->getLoopFor((*I).getParent());
6688 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
6695 OS << "\t\t" "Exits: ";
6696 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
6697 if (!SE.isLoopInvariant(ExitValue, L)) {
6698 OS << "<<Unknown>>";
6707 OS << "Determining loop execution counts for: ";
6708 WriteAsOperand(OS, F, /*PrintType=*/false);
6710 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6711 PrintLoopInfo(OS, &SE, *I);
6714 ScalarEvolution::LoopDisposition
6715 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
6716 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S];
6717 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair =
6718 Values.insert(std::make_pair(L, LoopVariant));
6720 return Pair.first->second;
6722 LoopDisposition D = computeLoopDisposition(S, L);
6723 return LoopDispositions[S][L] = D;
6726 ScalarEvolution::LoopDisposition
6727 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
6728 switch (S->getSCEVType()) {
6730 return LoopInvariant;
6734 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
6735 case scAddRecExpr: {
6736 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6738 // If L is the addrec's loop, it's computable.
6739 if (AR->getLoop() == L)
6740 return LoopComputable;
6742 // Add recurrences are never invariant in the function-body (null loop).
6746 // This recurrence is variant w.r.t. L if L contains AR's loop.
6747 if (L->contains(AR->getLoop()))
6750 // This recurrence is invariant w.r.t. L if AR's loop contains L.
6751 if (AR->getLoop()->contains(L))
6752 return LoopInvariant;
6754 // This recurrence is variant w.r.t. L if any of its operands
6756 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
6758 if (!isLoopInvariant(*I, L))
6761 // Otherwise it's loop-invariant.
6762 return LoopInvariant;
6768 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6769 bool HasVarying = false;
6770 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6772 LoopDisposition D = getLoopDisposition(*I, L);
6773 if (D == LoopVariant)
6775 if (D == LoopComputable)
6778 return HasVarying ? LoopComputable : LoopInvariant;
6781 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6782 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
6783 if (LD == LoopVariant)
6785 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
6786 if (RD == LoopVariant)
6788 return (LD == LoopInvariant && RD == LoopInvariant) ?
6789 LoopInvariant : LoopComputable;
6792 // All non-instruction values are loop invariant. All instructions are loop
6793 // invariant if they are not contained in the specified loop.
6794 // Instructions are never considered invariant in the function body
6795 // (null loop) because they are defined within the "loop".
6796 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
6797 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
6798 return LoopInvariant;
6799 case scCouldNotCompute:
6800 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6801 default: llvm_unreachable("Unknown SCEV kind!");
6805 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
6806 return getLoopDisposition(S, L) == LoopInvariant;
6809 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
6810 return getLoopDisposition(S, L) == LoopComputable;
6813 ScalarEvolution::BlockDisposition
6814 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6815 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S];
6816 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool>
6817 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock));
6819 return Pair.first->second;
6821 BlockDisposition D = computeBlockDisposition(S, BB);
6822 return BlockDispositions[S][BB] = D;
6825 ScalarEvolution::BlockDisposition
6826 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6827 switch (S->getSCEVType()) {
6829 return ProperlyDominatesBlock;
6833 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
6834 case scAddRecExpr: {
6835 // This uses a "dominates" query instead of "properly dominates" query
6836 // to test for proper dominance too, because the instruction which
6837 // produces the addrec's value is a PHI, and a PHI effectively properly
6838 // dominates its entire containing block.
6839 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6840 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
6841 return DoesNotDominateBlock;
6843 // FALL THROUGH into SCEVNAryExpr handling.
6848 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6850 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6852 BlockDisposition D = getBlockDisposition(*I, BB);
6853 if (D == DoesNotDominateBlock)
6854 return DoesNotDominateBlock;
6855 if (D == DominatesBlock)
6858 return Proper ? ProperlyDominatesBlock : DominatesBlock;
6861 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6862 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6863 BlockDisposition LD = getBlockDisposition(LHS, BB);
6864 if (LD == DoesNotDominateBlock)
6865 return DoesNotDominateBlock;
6866 BlockDisposition RD = getBlockDisposition(RHS, BB);
6867 if (RD == DoesNotDominateBlock)
6868 return DoesNotDominateBlock;
6869 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
6870 ProperlyDominatesBlock : DominatesBlock;
6873 if (Instruction *I =
6874 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
6875 if (I->getParent() == BB)
6876 return DominatesBlock;
6877 if (DT->properlyDominates(I->getParent(), BB))
6878 return ProperlyDominatesBlock;
6879 return DoesNotDominateBlock;
6881 return ProperlyDominatesBlock;
6882 case scCouldNotCompute:
6883 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6885 llvm_unreachable("Unknown SCEV kind!");
6889 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
6890 return getBlockDisposition(S, BB) >= DominatesBlock;
6893 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
6894 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
6898 // Search for a SCEV expression node within an expression tree.
6899 // Implements SCEVTraversal::Visitor.
6904 SCEVSearch(const SCEV *N): Node(N), IsFound(false) {}
6906 bool follow(const SCEV *S) {
6907 IsFound |= (S == Node);
6910 bool isDone() const { return IsFound; }
6914 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
6915 SCEVSearch Search(Op);
6916 visitAll(S, Search);
6917 return Search.IsFound;
6920 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
6921 ValuesAtScopes.erase(S);
6922 LoopDispositions.erase(S);
6923 BlockDispositions.erase(S);
6924 UnsignedRanges.erase(S);
6925 SignedRanges.erase(S);