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 // As a special case, fold trunc(undef) to undef. We don't want to
882 // know too much about SCEVUnknowns, but this special case is handy
884 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
885 if (isa<UndefValue>(U->getValue()))
886 return getSCEV(UndefValue::get(Ty));
888 // The cast wasn't folded; create an explicit cast node. We can reuse
889 // the existing insert position since if we get here, we won't have
890 // made any changes which would invalidate it.
891 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
893 UniqueSCEVs.InsertNode(S, IP);
897 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
899 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
900 "This is not an extending conversion!");
901 assert(isSCEVable(Ty) &&
902 "This is not a conversion to a SCEVable type!");
903 Ty = getEffectiveSCEVType(Ty);
905 // Fold if the operand is constant.
906 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
908 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(), Ty)));
910 // zext(zext(x)) --> zext(x)
911 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
912 return getZeroExtendExpr(SZ->getOperand(), Ty);
914 // Before doing any expensive analysis, check to see if we've already
915 // computed a SCEV for this Op and Ty.
917 ID.AddInteger(scZeroExtend);
921 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
923 // zext(trunc(x)) --> zext(x) or x or trunc(x)
924 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
925 // It's possible the bits taken off by the truncate were all zero bits. If
926 // so, we should be able to simplify this further.
927 const SCEV *X = ST->getOperand();
928 ConstantRange CR = getUnsignedRange(X);
929 unsigned TruncBits = getTypeSizeInBits(ST->getType());
930 unsigned NewBits = getTypeSizeInBits(Ty);
931 if (CR.truncate(TruncBits).zeroExtend(NewBits).contains(
932 CR.zextOrTrunc(NewBits)))
933 return getTruncateOrZeroExtend(X, Ty);
936 // If the input value is a chrec scev, and we can prove that the value
937 // did not overflow the old, smaller, value, we can zero extend all of the
938 // operands (often constants). This allows analysis of something like
939 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
940 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
941 if (AR->isAffine()) {
942 const SCEV *Start = AR->getStart();
943 const SCEV *Step = AR->getStepRecurrence(*this);
944 unsigned BitWidth = getTypeSizeInBits(AR->getType());
945 const Loop *L = AR->getLoop();
947 // If we have special knowledge that this addrec won't overflow,
948 // we don't need to do any further analysis.
949 if (AR->getNoWrapFlags(SCEV::FlagNUW))
950 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
951 getZeroExtendExpr(Step, Ty),
952 L, AR->getNoWrapFlags());
954 // Check whether the backedge-taken count is SCEVCouldNotCompute.
955 // Note that this serves two purposes: It filters out loops that are
956 // simply not analyzable, and it covers the case where this code is
957 // being called from within backedge-taken count analysis, such that
958 // attempting to ask for the backedge-taken count would likely result
959 // in infinite recursion. In the later case, the analysis code will
960 // cope with a conservative value, and it will take care to purge
961 // that value once it has finished.
962 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
963 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
964 // Manually compute the final value for AR, checking for
967 // Check whether the backedge-taken count can be losslessly casted to
968 // the addrec's type. The count is always unsigned.
969 const SCEV *CastedMaxBECount =
970 getTruncateOrZeroExtend(MaxBECount, Start->getType());
971 const SCEV *RecastedMaxBECount =
972 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
973 if (MaxBECount == RecastedMaxBECount) {
974 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
975 // Check whether Start+Step*MaxBECount has no unsigned overflow.
976 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
977 const SCEV *Add = getAddExpr(Start, ZMul);
978 const SCEV *OperandExtendedAdd =
979 getAddExpr(getZeroExtendExpr(Start, WideTy),
980 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
981 getZeroExtendExpr(Step, WideTy)));
982 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) {
983 // Cache knowledge of AR NUW, which is propagated to this AddRec.
984 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
985 // Return the expression with the addrec on the outside.
986 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
987 getZeroExtendExpr(Step, Ty),
988 L, AR->getNoWrapFlags());
990 // Similar to above, only this time treat the step value as signed.
991 // This covers loops that count down.
992 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
993 Add = getAddExpr(Start, SMul);
995 getAddExpr(getZeroExtendExpr(Start, WideTy),
996 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
997 getSignExtendExpr(Step, WideTy)));
998 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) {
999 // Cache knowledge of AR NW, which is propagated to this AddRec.
1000 // Negative step causes unsigned wrap, but it still can't self-wrap.
1001 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1002 // Return the expression with the addrec on the outside.
1003 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1004 getSignExtendExpr(Step, Ty),
1005 L, AR->getNoWrapFlags());
1009 // If the backedge is guarded by a comparison with the pre-inc value
1010 // the addrec is safe. Also, if the entry is guarded by a comparison
1011 // with the start value and the backedge is guarded by a comparison
1012 // with the post-inc value, the addrec is safe.
1013 if (isKnownPositive(Step)) {
1014 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
1015 getUnsignedRange(Step).getUnsignedMax());
1016 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
1017 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
1018 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
1019 AR->getPostIncExpr(*this), N))) {
1020 // Cache knowledge of AR NUW, which is propagated to this AddRec.
1021 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
1022 // Return the expression with the addrec on the outside.
1023 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1024 getZeroExtendExpr(Step, Ty),
1025 L, AR->getNoWrapFlags());
1027 } else if (isKnownNegative(Step)) {
1028 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
1029 getSignedRange(Step).getSignedMin());
1030 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
1031 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
1032 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
1033 AR->getPostIncExpr(*this), N))) {
1034 // Cache knowledge of AR NW, which is propagated to this AddRec.
1035 // Negative step causes unsigned wrap, but it still can't self-wrap.
1036 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1037 // Return the expression with the addrec on the outside.
1038 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1039 getSignExtendExpr(Step, Ty),
1040 L, AR->getNoWrapFlags());
1046 // The cast wasn't folded; create an explicit cast node.
1047 // Recompute the insert position, as it may have been invalidated.
1048 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1049 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1051 UniqueSCEVs.InsertNode(S, IP);
1055 // Get the limit of a recurrence such that incrementing by Step cannot cause
1056 // signed overflow as long as the value of the recurrence within the loop does
1057 // not exceed this limit before incrementing.
1058 static const SCEV *getOverflowLimitForStep(const SCEV *Step,
1059 ICmpInst::Predicate *Pred,
1060 ScalarEvolution *SE) {
1061 unsigned BitWidth = SE->getTypeSizeInBits(Step->getType());
1062 if (SE->isKnownPositive(Step)) {
1063 *Pred = ICmpInst::ICMP_SLT;
1064 return SE->getConstant(APInt::getSignedMinValue(BitWidth) -
1065 SE->getSignedRange(Step).getSignedMax());
1067 if (SE->isKnownNegative(Step)) {
1068 *Pred = ICmpInst::ICMP_SGT;
1069 return SE->getConstant(APInt::getSignedMaxValue(BitWidth) -
1070 SE->getSignedRange(Step).getSignedMin());
1075 // The recurrence AR has been shown to have no signed wrap. Typically, if we can
1076 // prove NSW for AR, then we can just as easily prove NSW for its preincrement
1077 // or postincrement sibling. This allows normalizing a sign extended AddRec as
1078 // such: {sext(Step + Start),+,Step} => {(Step + sext(Start),+,Step} As a
1079 // result, the expression "Step + sext(PreIncAR)" is congruent with
1080 // "sext(PostIncAR)"
1081 static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR,
1083 ScalarEvolution *SE) {
1084 const Loop *L = AR->getLoop();
1085 const SCEV *Start = AR->getStart();
1086 const SCEV *Step = AR->getStepRecurrence(*SE);
1088 // Check for a simple looking step prior to loop entry.
1089 const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start);
1093 // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV
1094 // subtraction is expensive. For this purpose, perform a quick and dirty
1095 // difference, by checking for Step in the operand list.
1096 SmallVector<const SCEV *, 4> DiffOps;
1097 for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end();
1100 DiffOps.push_back(*I);
1102 if (DiffOps.size() == SA->getNumOperands())
1105 // This is a postinc AR. Check for overflow on the preinc recurrence using the
1106 // same three conditions that getSignExtendedExpr checks.
1108 // 1. NSW flags on the step increment.
1109 const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags());
1110 const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>(
1111 SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap));
1113 if (PreAR && PreAR->getNoWrapFlags(SCEV::FlagNSW))
1116 // 2. Direct overflow check on the step operation's expression.
1117 unsigned BitWidth = SE->getTypeSizeInBits(AR->getType());
1118 Type *WideTy = IntegerType::get(SE->getContext(), BitWidth * 2);
1119 const SCEV *OperandExtendedStart =
1120 SE->getAddExpr(SE->getSignExtendExpr(PreStart, WideTy),
1121 SE->getSignExtendExpr(Step, WideTy));
1122 if (SE->getSignExtendExpr(Start, WideTy) == OperandExtendedStart) {
1123 // Cache knowledge of PreAR NSW.
1125 const_cast<SCEVAddRecExpr *>(PreAR)->setNoWrapFlags(SCEV::FlagNSW);
1126 // FIXME: this optimization needs a unit test
1127 DEBUG(dbgs() << "SCEV: untested prestart overflow check\n");
1131 // 3. Loop precondition.
1132 ICmpInst::Predicate Pred;
1133 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, SE);
1135 if (OverflowLimit &&
1136 SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) {
1142 // Get the normalized sign-extended expression for this AddRec's Start.
1143 static const SCEV *getSignExtendAddRecStart(const SCEVAddRecExpr *AR,
1145 ScalarEvolution *SE) {
1146 const SCEV *PreStart = getPreStartForSignExtend(AR, Ty, SE);
1148 return SE->getSignExtendExpr(AR->getStart(), Ty);
1150 return SE->getAddExpr(SE->getSignExtendExpr(AR->getStepRecurrence(*SE), Ty),
1151 SE->getSignExtendExpr(PreStart, Ty));
1154 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1156 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1157 "This is not an extending conversion!");
1158 assert(isSCEVable(Ty) &&
1159 "This is not a conversion to a SCEVable type!");
1160 Ty = getEffectiveSCEVType(Ty);
1162 // Fold if the operand is constant.
1163 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1165 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(), Ty)));
1167 // sext(sext(x)) --> sext(x)
1168 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1169 return getSignExtendExpr(SS->getOperand(), Ty);
1171 // sext(zext(x)) --> zext(x)
1172 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
1173 return getZeroExtendExpr(SZ->getOperand(), Ty);
1175 // Before doing any expensive analysis, check to see if we've already
1176 // computed a SCEV for this Op and Ty.
1177 FoldingSetNodeID ID;
1178 ID.AddInteger(scSignExtend);
1182 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1184 // If the input value is provably positive, build a zext instead.
1185 if (isKnownNonNegative(Op))
1186 return getZeroExtendExpr(Op, Ty);
1188 // sext(trunc(x)) --> sext(x) or x or trunc(x)
1189 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
1190 // It's possible the bits taken off by the truncate were all sign bits. If
1191 // so, we should be able to simplify this further.
1192 const SCEV *X = ST->getOperand();
1193 ConstantRange CR = getSignedRange(X);
1194 unsigned TruncBits = getTypeSizeInBits(ST->getType());
1195 unsigned NewBits = getTypeSizeInBits(Ty);
1196 if (CR.truncate(TruncBits).signExtend(NewBits).contains(
1197 CR.sextOrTrunc(NewBits)))
1198 return getTruncateOrSignExtend(X, Ty);
1201 // If the input value is a chrec scev, and we can prove that the value
1202 // did not overflow the old, smaller, value, we can sign extend all of the
1203 // operands (often constants). This allows analysis of something like
1204 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1205 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1206 if (AR->isAffine()) {
1207 const SCEV *Start = AR->getStart();
1208 const SCEV *Step = AR->getStepRecurrence(*this);
1209 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1210 const Loop *L = AR->getLoop();
1212 // If we have special knowledge that this addrec won't overflow,
1213 // we don't need to do any further analysis.
1214 if (AR->getNoWrapFlags(SCEV::FlagNSW))
1215 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1216 getSignExtendExpr(Step, Ty),
1219 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1220 // Note that this serves two purposes: It filters out loops that are
1221 // simply not analyzable, and it covers the case where this code is
1222 // being called from within backedge-taken count analysis, such that
1223 // attempting to ask for the backedge-taken count would likely result
1224 // in infinite recursion. In the later case, the analysis code will
1225 // cope with a conservative value, and it will take care to purge
1226 // that value once it has finished.
1227 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1228 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1229 // Manually compute the final value for AR, checking for
1232 // Check whether the backedge-taken count can be losslessly casted to
1233 // the addrec's type. The count is always unsigned.
1234 const SCEV *CastedMaxBECount =
1235 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1236 const SCEV *RecastedMaxBECount =
1237 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1238 if (MaxBECount == RecastedMaxBECount) {
1239 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1240 // Check whether Start+Step*MaxBECount has no signed overflow.
1241 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1242 const SCEV *Add = getAddExpr(Start, SMul);
1243 const SCEV *OperandExtendedAdd =
1244 getAddExpr(getSignExtendExpr(Start, WideTy),
1245 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1246 getSignExtendExpr(Step, WideTy)));
1247 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) {
1248 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1249 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1250 // Return the expression with the addrec on the outside.
1251 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1252 getSignExtendExpr(Step, Ty),
1253 L, AR->getNoWrapFlags());
1255 // Similar to above, only this time treat the step value as unsigned.
1256 // This covers loops that count up with an unsigned step.
1257 const SCEV *UMul = getMulExpr(CastedMaxBECount, Step);
1258 Add = getAddExpr(Start, UMul);
1259 OperandExtendedAdd =
1260 getAddExpr(getSignExtendExpr(Start, WideTy),
1261 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1262 getZeroExtendExpr(Step, WideTy)));
1263 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) {
1264 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1265 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1266 // Return the expression with the addrec on the outside.
1267 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1268 getZeroExtendExpr(Step, Ty),
1269 L, AR->getNoWrapFlags());
1273 // If the backedge is guarded by a comparison with the pre-inc value
1274 // the addrec is safe. Also, if the entry is guarded by a comparison
1275 // with the start value and the backedge is guarded by a comparison
1276 // with the post-inc value, the addrec is safe.
1277 ICmpInst::Predicate Pred;
1278 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, this);
1279 if (OverflowLimit &&
1280 (isLoopBackedgeGuardedByCond(L, Pred, AR, OverflowLimit) ||
1281 (isLoopEntryGuardedByCond(L, Pred, Start, OverflowLimit) &&
1282 isLoopBackedgeGuardedByCond(L, Pred, AR->getPostIncExpr(*this),
1284 // Cache knowledge of AR NSW, then propagate NSW to the wide AddRec.
1285 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1286 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1287 getSignExtendExpr(Step, Ty),
1288 L, AR->getNoWrapFlags());
1293 // The cast wasn't folded; create an explicit cast node.
1294 // Recompute the insert position, as it may have been invalidated.
1295 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1296 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1298 UniqueSCEVs.InsertNode(S, IP);
1302 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1303 /// unspecified bits out to the given type.
1305 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1307 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1308 "This is not an extending conversion!");
1309 assert(isSCEVable(Ty) &&
1310 "This is not a conversion to a SCEVable type!");
1311 Ty = getEffectiveSCEVType(Ty);
1313 // Sign-extend negative constants.
1314 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1315 if (SC->getValue()->getValue().isNegative())
1316 return getSignExtendExpr(Op, Ty);
1318 // Peel off a truncate cast.
1319 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1320 const SCEV *NewOp = T->getOperand();
1321 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1322 return getAnyExtendExpr(NewOp, Ty);
1323 return getTruncateOrNoop(NewOp, Ty);
1326 // Next try a zext cast. If the cast is folded, use it.
1327 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1328 if (!isa<SCEVZeroExtendExpr>(ZExt))
1331 // Next try a sext cast. If the cast is folded, use it.
1332 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1333 if (!isa<SCEVSignExtendExpr>(SExt))
1336 // Force the cast to be folded into the operands of an addrec.
1337 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1338 SmallVector<const SCEV *, 4> Ops;
1339 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1341 Ops.push_back(getAnyExtendExpr(*I, Ty));
1342 return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW);
1345 // As a special case, fold anyext(undef) to undef. We don't want to
1346 // know too much about SCEVUnknowns, but this special case is handy
1348 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
1349 if (isa<UndefValue>(U->getValue()))
1350 return getSCEV(UndefValue::get(Ty));
1352 // If the expression is obviously signed, use the sext cast value.
1353 if (isa<SCEVSMaxExpr>(Op))
1356 // Absent any other information, use the zext cast value.
1360 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1361 /// a list of operands to be added under the given scale, update the given
1362 /// map. This is a helper function for getAddRecExpr. As an example of
1363 /// what it does, given a sequence of operands that would form an add
1364 /// expression like this:
1366 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1368 /// where A and B are constants, update the map with these values:
1370 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1372 /// and add 13 + A*B*29 to AccumulatedConstant.
1373 /// This will allow getAddRecExpr to produce this:
1375 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1377 /// This form often exposes folding opportunities that are hidden in
1378 /// the original operand list.
1380 /// Return true iff it appears that any interesting folding opportunities
1381 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1382 /// the common case where no interesting opportunities are present, and
1383 /// is also used as a check to avoid infinite recursion.
1386 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1387 SmallVector<const SCEV *, 8> &NewOps,
1388 APInt &AccumulatedConstant,
1389 const SCEV *const *Ops, size_t NumOperands,
1391 ScalarEvolution &SE) {
1392 bool Interesting = false;
1394 // Iterate over the add operands. They are sorted, with constants first.
1396 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1398 // Pull a buried constant out to the outside.
1399 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1401 AccumulatedConstant += Scale * C->getValue()->getValue();
1404 // Next comes everything else. We're especially interested in multiplies
1405 // here, but they're in the middle, so just visit the rest with one loop.
1406 for (; i != NumOperands; ++i) {
1407 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1408 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1410 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1411 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1412 // A multiplication of a constant with another add; recurse.
1413 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1415 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1416 Add->op_begin(), Add->getNumOperands(),
1419 // A multiplication of a constant with some other value. Update
1421 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1422 const SCEV *Key = SE.getMulExpr(MulOps);
1423 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1424 M.insert(std::make_pair(Key, NewScale));
1426 NewOps.push_back(Pair.first->first);
1428 Pair.first->second += NewScale;
1429 // The map already had an entry for this value, which may indicate
1430 // a folding opportunity.
1435 // An ordinary operand. Update the map.
1436 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1437 M.insert(std::make_pair(Ops[i], Scale));
1439 NewOps.push_back(Pair.first->first);
1441 Pair.first->second += Scale;
1442 // The map already had an entry for this value, which may indicate
1443 // a folding opportunity.
1453 struct APIntCompare {
1454 bool operator()(const APInt &LHS, const APInt &RHS) const {
1455 return LHS.ult(RHS);
1460 /// getAddExpr - Get a canonical add expression, or something simpler if
1462 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1463 SCEV::NoWrapFlags Flags) {
1464 assert(!(Flags & ~(SCEV::FlagNUW | SCEV::FlagNSW)) &&
1465 "only nuw or nsw allowed");
1466 assert(!Ops.empty() && "Cannot get empty add!");
1467 if (Ops.size() == 1) return Ops[0];
1469 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1470 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1471 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1472 "SCEVAddExpr operand types don't match!");
1475 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1477 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1478 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1479 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1481 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1482 E = Ops.end(); I != E; ++I)
1483 if (!isKnownNonNegative(*I)) {
1487 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1490 // Sort by complexity, this groups all similar expression types together.
1491 GroupByComplexity(Ops, LI);
1493 // If there are any constants, fold them together.
1495 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1497 assert(Idx < Ops.size());
1498 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1499 // We found two constants, fold them together!
1500 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1501 RHSC->getValue()->getValue());
1502 if (Ops.size() == 2) return Ops[0];
1503 Ops.erase(Ops.begin()+1); // Erase the folded element
1504 LHSC = cast<SCEVConstant>(Ops[0]);
1507 // If we are left with a constant zero being added, strip it off.
1508 if (LHSC->getValue()->isZero()) {
1509 Ops.erase(Ops.begin());
1513 if (Ops.size() == 1) return Ops[0];
1516 // Okay, check to see if the same value occurs in the operand list more than
1517 // once. If so, merge them together into an multiply expression. Since we
1518 // sorted the list, these values are required to be adjacent.
1519 Type *Ty = Ops[0]->getType();
1520 bool FoundMatch = false;
1521 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1522 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1523 // Scan ahead to count how many equal operands there are.
1525 while (i+Count != e && Ops[i+Count] == Ops[i])
1527 // Merge the values into a multiply.
1528 const SCEV *Scale = getConstant(Ty, Count);
1529 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1530 if (Ops.size() == Count)
1533 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1534 --i; e -= Count - 1;
1538 return getAddExpr(Ops, Flags);
1540 // Check for truncates. If all the operands are truncated from the same
1541 // type, see if factoring out the truncate would permit the result to be
1542 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1543 // if the contents of the resulting outer trunc fold to something simple.
1544 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1545 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1546 Type *DstType = Trunc->getType();
1547 Type *SrcType = Trunc->getOperand()->getType();
1548 SmallVector<const SCEV *, 8> LargeOps;
1550 // Check all the operands to see if they can be represented in the
1551 // source type of the truncate.
1552 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1553 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1554 if (T->getOperand()->getType() != SrcType) {
1558 LargeOps.push_back(T->getOperand());
1559 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1560 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1561 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1562 SmallVector<const SCEV *, 8> LargeMulOps;
1563 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1564 if (const SCEVTruncateExpr *T =
1565 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1566 if (T->getOperand()->getType() != SrcType) {
1570 LargeMulOps.push_back(T->getOperand());
1571 } else if (const SCEVConstant *C =
1572 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1573 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1580 LargeOps.push_back(getMulExpr(LargeMulOps));
1587 // Evaluate the expression in the larger type.
1588 const SCEV *Fold = getAddExpr(LargeOps, Flags);
1589 // If it folds to something simple, use it. Otherwise, don't.
1590 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1591 return getTruncateExpr(Fold, DstType);
1595 // Skip past any other cast SCEVs.
1596 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1599 // If there are add operands they would be next.
1600 if (Idx < Ops.size()) {
1601 bool DeletedAdd = false;
1602 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1603 // If we have an add, expand the add operands onto the end of the operands
1605 Ops.erase(Ops.begin()+Idx);
1606 Ops.append(Add->op_begin(), Add->op_end());
1610 // If we deleted at least one add, we added operands to the end of the list,
1611 // and they are not necessarily sorted. Recurse to resort and resimplify
1612 // any operands we just acquired.
1614 return getAddExpr(Ops);
1617 // Skip over the add expression until we get to a multiply.
1618 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1621 // Check to see if there are any folding opportunities present with
1622 // operands multiplied by constant values.
1623 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1624 uint64_t BitWidth = getTypeSizeInBits(Ty);
1625 DenseMap<const SCEV *, APInt> M;
1626 SmallVector<const SCEV *, 8> NewOps;
1627 APInt AccumulatedConstant(BitWidth, 0);
1628 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1629 Ops.data(), Ops.size(),
1630 APInt(BitWidth, 1), *this)) {
1631 // Some interesting folding opportunity is present, so its worthwhile to
1632 // re-generate the operands list. Group the operands by constant scale,
1633 // to avoid multiplying by the same constant scale multiple times.
1634 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1635 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1636 E = NewOps.end(); I != E; ++I)
1637 MulOpLists[M.find(*I)->second].push_back(*I);
1638 // Re-generate the operands list.
1640 if (AccumulatedConstant != 0)
1641 Ops.push_back(getConstant(AccumulatedConstant));
1642 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1643 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1645 Ops.push_back(getMulExpr(getConstant(I->first),
1646 getAddExpr(I->second)));
1648 return getConstant(Ty, 0);
1649 if (Ops.size() == 1)
1651 return getAddExpr(Ops);
1655 // If we are adding something to a multiply expression, make sure the
1656 // something is not already an operand of the multiply. If so, merge it into
1658 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1659 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1660 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1661 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1662 if (isa<SCEVConstant>(MulOpSCEV))
1664 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1665 if (MulOpSCEV == Ops[AddOp]) {
1666 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1667 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1668 if (Mul->getNumOperands() != 2) {
1669 // If the multiply has more than two operands, we must get the
1671 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1672 Mul->op_begin()+MulOp);
1673 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1674 InnerMul = getMulExpr(MulOps);
1676 const SCEV *One = getConstant(Ty, 1);
1677 const SCEV *AddOne = getAddExpr(One, InnerMul);
1678 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1679 if (Ops.size() == 2) return OuterMul;
1681 Ops.erase(Ops.begin()+AddOp);
1682 Ops.erase(Ops.begin()+Idx-1);
1684 Ops.erase(Ops.begin()+Idx);
1685 Ops.erase(Ops.begin()+AddOp-1);
1687 Ops.push_back(OuterMul);
1688 return getAddExpr(Ops);
1691 // Check this multiply against other multiplies being added together.
1692 for (unsigned OtherMulIdx = Idx+1;
1693 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1695 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1696 // If MulOp occurs in OtherMul, we can fold the two multiplies
1698 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1699 OMulOp != e; ++OMulOp)
1700 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1701 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1702 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1703 if (Mul->getNumOperands() != 2) {
1704 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1705 Mul->op_begin()+MulOp);
1706 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1707 InnerMul1 = getMulExpr(MulOps);
1709 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1710 if (OtherMul->getNumOperands() != 2) {
1711 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1712 OtherMul->op_begin()+OMulOp);
1713 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1714 InnerMul2 = getMulExpr(MulOps);
1716 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1717 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1718 if (Ops.size() == 2) return OuterMul;
1719 Ops.erase(Ops.begin()+Idx);
1720 Ops.erase(Ops.begin()+OtherMulIdx-1);
1721 Ops.push_back(OuterMul);
1722 return getAddExpr(Ops);
1728 // If there are any add recurrences in the operands list, see if any other
1729 // added values are loop invariant. If so, we can fold them into the
1731 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1734 // Scan over all recurrences, trying to fold loop invariants into them.
1735 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1736 // Scan all of the other operands to this add and add them to the vector if
1737 // they are loop invariant w.r.t. the recurrence.
1738 SmallVector<const SCEV *, 8> LIOps;
1739 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1740 const Loop *AddRecLoop = AddRec->getLoop();
1741 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1742 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1743 LIOps.push_back(Ops[i]);
1744 Ops.erase(Ops.begin()+i);
1748 // If we found some loop invariants, fold them into the recurrence.
1749 if (!LIOps.empty()) {
1750 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1751 LIOps.push_back(AddRec->getStart());
1753 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1755 AddRecOps[0] = getAddExpr(LIOps);
1757 // Build the new addrec. Propagate the NUW and NSW flags if both the
1758 // outer add and the inner addrec are guaranteed to have no overflow.
1759 // Always propagate NW.
1760 Flags = AddRec->getNoWrapFlags(setFlags(Flags, SCEV::FlagNW));
1761 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop, Flags);
1763 // If all of the other operands were loop invariant, we are done.
1764 if (Ops.size() == 1) return NewRec;
1766 // Otherwise, add the folded AddRec by the non-invariant parts.
1767 for (unsigned i = 0;; ++i)
1768 if (Ops[i] == AddRec) {
1772 return getAddExpr(Ops);
1775 // Okay, if there weren't any loop invariants to be folded, check to see if
1776 // there are multiple AddRec's with the same loop induction variable being
1777 // added together. If so, we can fold them.
1778 for (unsigned OtherIdx = Idx+1;
1779 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1781 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1782 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1783 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1785 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1787 if (const SCEVAddRecExpr *OtherAddRec =
1788 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1789 if (OtherAddRec->getLoop() == AddRecLoop) {
1790 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1792 if (i >= AddRecOps.size()) {
1793 AddRecOps.append(OtherAddRec->op_begin()+i,
1794 OtherAddRec->op_end());
1797 AddRecOps[i] = getAddExpr(AddRecOps[i],
1798 OtherAddRec->getOperand(i));
1800 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1802 // Step size has changed, so we cannot guarantee no self-wraparound.
1803 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop, SCEV::FlagAnyWrap);
1804 return getAddExpr(Ops);
1807 // Otherwise couldn't fold anything into this recurrence. Move onto the
1811 // Okay, it looks like we really DO need an add expr. Check to see if we
1812 // already have one, otherwise create a new one.
1813 FoldingSetNodeID ID;
1814 ID.AddInteger(scAddExpr);
1815 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1816 ID.AddPointer(Ops[i]);
1819 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1821 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1822 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1823 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1825 UniqueSCEVs.InsertNode(S, IP);
1827 S->setNoWrapFlags(Flags);
1831 static uint64_t umul_ov(uint64_t i, uint64_t j, bool &Overflow) {
1833 if (j > 1 && k / j != i) Overflow = true;
1837 /// Compute the result of "n choose k", the binomial coefficient. If an
1838 /// intermediate computation overflows, Overflow will be set and the return will
1839 /// be garbage. Overflow is not cleared on absense of overflow.
1840 static uint64_t Choose(uint64_t n, uint64_t k, bool &Overflow) {
1841 // We use the multiplicative formula:
1842 // n(n-1)(n-2)...(n-(k-1)) / k(k-1)(k-2)...1 .
1843 // At each iteration, we take the n-th term of the numeral and divide by the
1844 // (k-n)th term of the denominator. This division will always produce an
1845 // integral result, and helps reduce the chance of overflow in the
1846 // intermediate computations. However, we can still overflow even when the
1847 // final result would fit.
1849 if (n == 0 || n == k) return 1;
1850 if (k > n) return 0;
1856 for (uint64_t i = 1; i <= k; ++i) {
1857 r = umul_ov(r, n-(i-1), Overflow);
1863 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1865 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1866 SCEV::NoWrapFlags Flags) {
1867 assert(Flags == maskFlags(Flags, SCEV::FlagNUW | SCEV::FlagNSW) &&
1868 "only nuw or nsw allowed");
1869 assert(!Ops.empty() && "Cannot get empty mul!");
1870 if (Ops.size() == 1) return Ops[0];
1872 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1873 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1874 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1875 "SCEVMulExpr operand types don't match!");
1878 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1880 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1881 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1882 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1884 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1885 E = Ops.end(); I != E; ++I)
1886 if (!isKnownNonNegative(*I)) {
1890 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1893 // Sort by complexity, this groups all similar expression types together.
1894 GroupByComplexity(Ops, LI);
1896 // If there are any constants, fold them together.
1898 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1900 // C1*(C2+V) -> C1*C2 + C1*V
1901 if (Ops.size() == 2)
1902 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1903 if (Add->getNumOperands() == 2 &&
1904 isa<SCEVConstant>(Add->getOperand(0)))
1905 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1906 getMulExpr(LHSC, Add->getOperand(1)));
1909 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1910 // We found two constants, fold them together!
1911 ConstantInt *Fold = ConstantInt::get(getContext(),
1912 LHSC->getValue()->getValue() *
1913 RHSC->getValue()->getValue());
1914 Ops[0] = getConstant(Fold);
1915 Ops.erase(Ops.begin()+1); // Erase the folded element
1916 if (Ops.size() == 1) return Ops[0];
1917 LHSC = cast<SCEVConstant>(Ops[0]);
1920 // If we are left with a constant one being multiplied, strip it off.
1921 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1922 Ops.erase(Ops.begin());
1924 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1925 // If we have a multiply of zero, it will always be zero.
1927 } else if (Ops[0]->isAllOnesValue()) {
1928 // If we have a mul by -1 of an add, try distributing the -1 among the
1930 if (Ops.size() == 2) {
1931 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1932 SmallVector<const SCEV *, 4> NewOps;
1933 bool AnyFolded = false;
1934 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(),
1935 E = Add->op_end(); I != E; ++I) {
1936 const SCEV *Mul = getMulExpr(Ops[0], *I);
1937 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1938 NewOps.push_back(Mul);
1941 return getAddExpr(NewOps);
1943 else if (const SCEVAddRecExpr *
1944 AddRec = dyn_cast<SCEVAddRecExpr>(Ops[1])) {
1945 // Negation preserves a recurrence's no self-wrap property.
1946 SmallVector<const SCEV *, 4> Operands;
1947 for (SCEVAddRecExpr::op_iterator I = AddRec->op_begin(),
1948 E = AddRec->op_end(); I != E; ++I) {
1949 Operands.push_back(getMulExpr(Ops[0], *I));
1951 return getAddRecExpr(Operands, AddRec->getLoop(),
1952 AddRec->getNoWrapFlags(SCEV::FlagNW));
1957 if (Ops.size() == 1)
1961 // Skip over the add expression until we get to a multiply.
1962 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1965 // If there are mul operands inline them all into this expression.
1966 if (Idx < Ops.size()) {
1967 bool DeletedMul = false;
1968 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1969 // If we have an mul, expand the mul operands onto the end of the operands
1971 Ops.erase(Ops.begin()+Idx);
1972 Ops.append(Mul->op_begin(), Mul->op_end());
1976 // If we deleted at least one mul, we added operands to the end of the list,
1977 // and they are not necessarily sorted. Recurse to resort and resimplify
1978 // any operands we just acquired.
1980 return getMulExpr(Ops);
1983 // If there are any add recurrences in the operands list, see if any other
1984 // added values are loop invariant. If so, we can fold them into the
1986 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1989 // Scan over all recurrences, trying to fold loop invariants into them.
1990 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1991 // Scan all of the other operands to this mul and add them to the vector if
1992 // they are loop invariant w.r.t. the recurrence.
1993 SmallVector<const SCEV *, 8> LIOps;
1994 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1995 const Loop *AddRecLoop = AddRec->getLoop();
1996 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1997 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1998 LIOps.push_back(Ops[i]);
1999 Ops.erase(Ops.begin()+i);
2003 // If we found some loop invariants, fold them into the recurrence.
2004 if (!LIOps.empty()) {
2005 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
2006 SmallVector<const SCEV *, 4> NewOps;
2007 NewOps.reserve(AddRec->getNumOperands());
2008 const SCEV *Scale = getMulExpr(LIOps);
2009 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
2010 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
2012 // Build the new addrec. Propagate the NUW and NSW flags if both the
2013 // outer mul and the inner addrec are guaranteed to have no overflow.
2015 // No self-wrap cannot be guaranteed after changing the step size, but
2016 // will be inferred if either NUW or NSW is true.
2017 Flags = AddRec->getNoWrapFlags(clearFlags(Flags, SCEV::FlagNW));
2018 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop, Flags);
2020 // If all of the other operands were loop invariant, we are done.
2021 if (Ops.size() == 1) return NewRec;
2023 // Otherwise, multiply the folded AddRec by the non-invariant parts.
2024 for (unsigned i = 0;; ++i)
2025 if (Ops[i] == AddRec) {
2029 return getMulExpr(Ops);
2032 // Okay, if there weren't any loop invariants to be folded, check to see if
2033 // there are multiple AddRec's with the same loop induction variable being
2034 // multiplied together. If so, we can fold them.
2035 for (unsigned OtherIdx = Idx+1;
2036 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2038 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
2039 // {A1,+,A2,+,...,+,An}<L> * {B1,+,B2,+,...,+,Bn}<L>
2040 // = {x=1 in [ sum y=x..2x [ sum z=max(y-x, y-n)..min(x,n) [
2041 // choose(x, 2x)*choose(2x-y, x-z)*A_{y-z}*B_z
2042 // ]]],+,...up to x=2n}.
2043 // Note that the arguments to choose() are always integers with values
2044 // known at compile time, never SCEV objects.
2046 // The implementation avoids pointless extra computations when the two
2047 // addrec's are of different length (mathematically, it's equivalent to
2048 // an infinite stream of zeros on the right).
2049 bool OpsModified = false;
2050 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2052 if (const SCEVAddRecExpr *OtherAddRec =
2053 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
2054 if (OtherAddRec->getLoop() == AddRecLoop) {
2055 bool Overflow = false;
2056 Type *Ty = AddRec->getType();
2057 bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64;
2058 SmallVector<const SCEV*, 7> AddRecOps;
2059 for (int x = 0, xe = AddRec->getNumOperands() +
2060 OtherAddRec->getNumOperands() - 1;
2061 x != xe && !Overflow; ++x) {
2062 const SCEV *Term = getConstant(Ty, 0);
2063 for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) {
2064 uint64_t Coeff1 = Choose(x, 2*x - y, Overflow);
2065 for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1),
2066 ze = std::min(x+1, (int)OtherAddRec->getNumOperands());
2067 z < ze && !Overflow; ++z) {
2068 uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow);
2070 if (LargerThan64Bits)
2071 Coeff = umul_ov(Coeff1, Coeff2, Overflow);
2073 Coeff = Coeff1*Coeff2;
2074 const SCEV *CoeffTerm = getConstant(Ty, Coeff);
2075 const SCEV *Term1 = AddRec->getOperand(y-z);
2076 const SCEV *Term2 = OtherAddRec->getOperand(z);
2077 Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2));
2080 AddRecOps.push_back(Term);
2083 const SCEV *NewAddRec = getAddRecExpr(AddRecOps,
2086 if (Ops.size() == 2) return NewAddRec;
2087 Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
2088 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
2093 return getMulExpr(Ops);
2097 // Otherwise couldn't fold anything into this recurrence. Move onto the
2101 // Okay, it looks like we really DO need an mul expr. Check to see if we
2102 // already have one, otherwise create a new one.
2103 FoldingSetNodeID ID;
2104 ID.AddInteger(scMulExpr);
2105 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2106 ID.AddPointer(Ops[i]);
2109 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2111 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2112 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2113 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
2115 UniqueSCEVs.InsertNode(S, IP);
2117 S->setNoWrapFlags(Flags);
2121 /// getUDivExpr - Get a canonical unsigned division expression, or something
2122 /// simpler if possible.
2123 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
2125 assert(getEffectiveSCEVType(LHS->getType()) ==
2126 getEffectiveSCEVType(RHS->getType()) &&
2127 "SCEVUDivExpr operand types don't match!");
2129 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
2130 if (RHSC->getValue()->equalsInt(1))
2131 return LHS; // X udiv 1 --> x
2132 // If the denominator is zero, the result of the udiv is undefined. Don't
2133 // try to analyze it, because the resolution chosen here may differ from
2134 // the resolution chosen in other parts of the compiler.
2135 if (!RHSC->getValue()->isZero()) {
2136 // Determine if the division can be folded into the operands of
2138 // TODO: Generalize this to non-constants by using known-bits information.
2139 Type *Ty = LHS->getType();
2140 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
2141 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
2142 // For non-power-of-two values, effectively round the value up to the
2143 // nearest power of two.
2144 if (!RHSC->getValue()->getValue().isPowerOf2())
2146 IntegerType *ExtTy =
2147 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
2148 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
2149 if (const SCEVConstant *Step =
2150 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this))) {
2151 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
2152 const APInt &StepInt = Step->getValue()->getValue();
2153 const APInt &DivInt = RHSC->getValue()->getValue();
2154 if (!StepInt.urem(DivInt) &&
2155 getZeroExtendExpr(AR, ExtTy) ==
2156 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2157 getZeroExtendExpr(Step, ExtTy),
2158 AR->getLoop(), SCEV::FlagAnyWrap)) {
2159 SmallVector<const SCEV *, 4> Operands;
2160 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
2161 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
2162 return getAddRecExpr(Operands, AR->getLoop(),
2165 /// Get a canonical UDivExpr for a recurrence.
2166 /// {X,+,N}/C => {Y,+,N}/C where Y=X-(X%N). Safe when C%N=0.
2167 // We can currently only fold X%N if X is constant.
2168 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
2169 if (StartC && !DivInt.urem(StepInt) &&
2170 getZeroExtendExpr(AR, ExtTy) ==
2171 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2172 getZeroExtendExpr(Step, ExtTy),
2173 AR->getLoop(), SCEV::FlagAnyWrap)) {
2174 const APInt &StartInt = StartC->getValue()->getValue();
2175 const APInt &StartRem = StartInt.urem(StepInt);
2177 LHS = getAddRecExpr(getConstant(StartInt - StartRem), Step,
2178 AR->getLoop(), SCEV::FlagNW);
2181 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
2182 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
2183 SmallVector<const SCEV *, 4> Operands;
2184 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
2185 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
2186 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
2187 // Find an operand that's safely divisible.
2188 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
2189 const SCEV *Op = M->getOperand(i);
2190 const SCEV *Div = getUDivExpr(Op, RHSC);
2191 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
2192 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
2195 return getMulExpr(Operands);
2199 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
2200 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(LHS)) {
2201 SmallVector<const SCEV *, 4> Operands;
2202 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
2203 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
2204 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
2206 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2207 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
2208 if (isa<SCEVUDivExpr>(Op) ||
2209 getMulExpr(Op, RHS) != A->getOperand(i))
2211 Operands.push_back(Op);
2213 if (Operands.size() == A->getNumOperands())
2214 return getAddExpr(Operands);
2218 // Fold if both operands are constant.
2219 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
2220 Constant *LHSCV = LHSC->getValue();
2221 Constant *RHSCV = RHSC->getValue();
2222 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
2228 FoldingSetNodeID ID;
2229 ID.AddInteger(scUDivExpr);
2233 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2234 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
2236 UniqueSCEVs.InsertNode(S, IP);
2241 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2242 /// Simplify the expression as much as possible.
2243 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start, const SCEV *Step,
2245 SCEV::NoWrapFlags Flags) {
2246 SmallVector<const SCEV *, 4> Operands;
2247 Operands.push_back(Start);
2248 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
2249 if (StepChrec->getLoop() == L) {
2250 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
2251 return getAddRecExpr(Operands, L, maskFlags(Flags, SCEV::FlagNW));
2254 Operands.push_back(Step);
2255 return getAddRecExpr(Operands, L, Flags);
2258 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2259 /// Simplify the expression as much as possible.
2261 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2262 const Loop *L, SCEV::NoWrapFlags Flags) {
2263 if (Operands.size() == 1) return Operands[0];
2265 Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2266 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2267 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2268 "SCEVAddRecExpr operand types don't match!");
2269 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2270 assert(isLoopInvariant(Operands[i], L) &&
2271 "SCEVAddRecExpr operand is not loop-invariant!");
2274 if (Operands.back()->isZero()) {
2275 Operands.pop_back();
2276 return getAddRecExpr(Operands, L, SCEV::FlagAnyWrap); // {X,+,0} --> X
2279 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2280 // use that information to infer NUW and NSW flags. However, computing a
2281 // BE count requires calling getAddRecExpr, so we may not yet have a
2282 // meaningful BE count at this point (and if we don't, we'd be stuck
2283 // with a SCEVCouldNotCompute as the cached BE count).
2285 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
2287 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
2288 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
2289 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
2291 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2292 E = Operands.end(); I != E; ++I)
2293 if (!isKnownNonNegative(*I)) {
2297 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
2300 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2301 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2302 const Loop *NestedLoop = NestedAR->getLoop();
2303 if (L->contains(NestedLoop) ?
2304 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2305 (!NestedLoop->contains(L) &&
2306 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2307 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2308 NestedAR->op_end());
2309 Operands[0] = NestedAR->getStart();
2310 // AddRecs require their operands be loop-invariant with respect to their
2311 // loops. Don't perform this transformation if it would break this
2313 bool AllInvariant = true;
2314 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2315 if (!isLoopInvariant(Operands[i], L)) {
2316 AllInvariant = false;
2320 // Create a recurrence for the outer loop with the same step size.
2322 // The outer recurrence keeps its NW flag but only keeps NUW/NSW if the
2323 // inner recurrence has the same property.
2324 SCEV::NoWrapFlags OuterFlags =
2325 maskFlags(Flags, SCEV::FlagNW | NestedAR->getNoWrapFlags());
2327 NestedOperands[0] = getAddRecExpr(Operands, L, OuterFlags);
2328 AllInvariant = true;
2329 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2330 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2331 AllInvariant = false;
2335 // Ok, both add recurrences are valid after the transformation.
2337 // The inner recurrence keeps its NW flag but only keeps NUW/NSW if
2338 // the outer recurrence has the same property.
2339 SCEV::NoWrapFlags InnerFlags =
2340 maskFlags(NestedAR->getNoWrapFlags(), SCEV::FlagNW | Flags);
2341 return getAddRecExpr(NestedOperands, NestedLoop, InnerFlags);
2344 // Reset Operands to its original state.
2345 Operands[0] = NestedAR;
2349 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2350 // already have one, otherwise create a new one.
2351 FoldingSetNodeID ID;
2352 ID.AddInteger(scAddRecExpr);
2353 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2354 ID.AddPointer(Operands[i]);
2358 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2360 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2361 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2362 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2363 O, Operands.size(), L);
2364 UniqueSCEVs.InsertNode(S, IP);
2366 S->setNoWrapFlags(Flags);
2370 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2372 SmallVector<const SCEV *, 2> Ops;
2375 return getSMaxExpr(Ops);
2379 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2380 assert(!Ops.empty() && "Cannot get empty smax!");
2381 if (Ops.size() == 1) return Ops[0];
2383 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2384 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2385 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2386 "SCEVSMaxExpr operand types don't match!");
2389 // Sort by complexity, this groups all similar expression types together.
2390 GroupByComplexity(Ops, LI);
2392 // If there are any constants, fold them together.
2394 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2396 assert(Idx < Ops.size());
2397 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2398 // We found two constants, fold them together!
2399 ConstantInt *Fold = ConstantInt::get(getContext(),
2400 APIntOps::smax(LHSC->getValue()->getValue(),
2401 RHSC->getValue()->getValue()));
2402 Ops[0] = getConstant(Fold);
2403 Ops.erase(Ops.begin()+1); // Erase the folded element
2404 if (Ops.size() == 1) return Ops[0];
2405 LHSC = cast<SCEVConstant>(Ops[0]);
2408 // If we are left with a constant minimum-int, strip it off.
2409 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2410 Ops.erase(Ops.begin());
2412 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2413 // If we have an smax with a constant maximum-int, it will always be
2418 if (Ops.size() == 1) return Ops[0];
2421 // Find the first SMax
2422 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2425 // Check to see if one of the operands is an SMax. If so, expand its operands
2426 // onto our operand list, and recurse to simplify.
2427 if (Idx < Ops.size()) {
2428 bool DeletedSMax = false;
2429 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2430 Ops.erase(Ops.begin()+Idx);
2431 Ops.append(SMax->op_begin(), SMax->op_end());
2436 return getSMaxExpr(Ops);
2439 // Okay, check to see if the same value occurs in the operand list twice. If
2440 // so, delete one. Since we sorted the list, these values are required to
2442 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2443 // X smax Y smax Y --> X smax Y
2444 // X smax Y --> X, if X is always greater than Y
2445 if (Ops[i] == Ops[i+1] ||
2446 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2447 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2449 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2450 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2454 if (Ops.size() == 1) return Ops[0];
2456 assert(!Ops.empty() && "Reduced smax down to nothing!");
2458 // Okay, it looks like we really DO need an smax expr. Check to see if we
2459 // already have one, otherwise create a new one.
2460 FoldingSetNodeID ID;
2461 ID.AddInteger(scSMaxExpr);
2462 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2463 ID.AddPointer(Ops[i]);
2465 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2466 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2467 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2468 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2470 UniqueSCEVs.InsertNode(S, IP);
2474 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2476 SmallVector<const SCEV *, 2> Ops;
2479 return getUMaxExpr(Ops);
2483 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2484 assert(!Ops.empty() && "Cannot get empty umax!");
2485 if (Ops.size() == 1) return Ops[0];
2487 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2488 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2489 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2490 "SCEVUMaxExpr operand types don't match!");
2493 // Sort by complexity, this groups all similar expression types together.
2494 GroupByComplexity(Ops, LI);
2496 // If there are any constants, fold them together.
2498 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2500 assert(Idx < Ops.size());
2501 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2502 // We found two constants, fold them together!
2503 ConstantInt *Fold = ConstantInt::get(getContext(),
2504 APIntOps::umax(LHSC->getValue()->getValue(),
2505 RHSC->getValue()->getValue()));
2506 Ops[0] = getConstant(Fold);
2507 Ops.erase(Ops.begin()+1); // Erase the folded element
2508 if (Ops.size() == 1) return Ops[0];
2509 LHSC = cast<SCEVConstant>(Ops[0]);
2512 // If we are left with a constant minimum-int, strip it off.
2513 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2514 Ops.erase(Ops.begin());
2516 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2517 // If we have an umax with a constant maximum-int, it will always be
2522 if (Ops.size() == 1) return Ops[0];
2525 // Find the first UMax
2526 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2529 // Check to see if one of the operands is a UMax. If so, expand its operands
2530 // onto our operand list, and recurse to simplify.
2531 if (Idx < Ops.size()) {
2532 bool DeletedUMax = false;
2533 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2534 Ops.erase(Ops.begin()+Idx);
2535 Ops.append(UMax->op_begin(), UMax->op_end());
2540 return getUMaxExpr(Ops);
2543 // Okay, check to see if the same value occurs in the operand list twice. If
2544 // so, delete one. Since we sorted the list, these values are required to
2546 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2547 // X umax Y umax Y --> X umax Y
2548 // X umax Y --> X, if X is always greater than Y
2549 if (Ops[i] == Ops[i+1] ||
2550 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2551 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2553 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2554 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2558 if (Ops.size() == 1) return Ops[0];
2560 assert(!Ops.empty() && "Reduced umax down to nothing!");
2562 // Okay, it looks like we really DO need a umax expr. Check to see if we
2563 // already have one, otherwise create a new one.
2564 FoldingSetNodeID ID;
2565 ID.AddInteger(scUMaxExpr);
2566 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2567 ID.AddPointer(Ops[i]);
2569 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2570 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2571 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2572 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2574 UniqueSCEVs.InsertNode(S, IP);
2578 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2580 // ~smax(~x, ~y) == smin(x, y).
2581 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2584 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2586 // ~umax(~x, ~y) == umin(x, y)
2587 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2590 const SCEV *ScalarEvolution::getSizeOfExpr(Type *AllocTy) {
2591 // If we have TargetData, we can bypass creating a target-independent
2592 // constant expression and then folding it back into a ConstantInt.
2593 // This is just a compile-time optimization.
2595 return getConstant(TD->getIntPtrType(getContext()),
2596 TD->getTypeAllocSize(AllocTy));
2598 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2599 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2600 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2602 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2603 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2606 const SCEV *ScalarEvolution::getAlignOfExpr(Type *AllocTy) {
2607 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2608 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2609 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2611 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2612 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2615 const SCEV *ScalarEvolution::getOffsetOfExpr(StructType *STy,
2617 // If we have TargetData, we can bypass creating a target-independent
2618 // constant expression and then folding it back into a ConstantInt.
2619 // This is just a compile-time optimization.
2621 return getConstant(TD->getIntPtrType(getContext()),
2622 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2624 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2625 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2626 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2628 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2629 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2632 const SCEV *ScalarEvolution::getOffsetOfExpr(Type *CTy,
2633 Constant *FieldNo) {
2634 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2635 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2636 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2638 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2639 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2642 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2643 // Don't attempt to do anything other than create a SCEVUnknown object
2644 // here. createSCEV only calls getUnknown after checking for all other
2645 // interesting possibilities, and any other code that calls getUnknown
2646 // is doing so in order to hide a value from SCEV canonicalization.
2648 FoldingSetNodeID ID;
2649 ID.AddInteger(scUnknown);
2652 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2653 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2654 "Stale SCEVUnknown in uniquing map!");
2657 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2659 FirstUnknown = cast<SCEVUnknown>(S);
2660 UniqueSCEVs.InsertNode(S, IP);
2664 //===----------------------------------------------------------------------===//
2665 // Basic SCEV Analysis and PHI Idiom Recognition Code
2668 /// isSCEVable - Test if values of the given type are analyzable within
2669 /// the SCEV framework. This primarily includes integer types, and it
2670 /// can optionally include pointer types if the ScalarEvolution class
2671 /// has access to target-specific information.
2672 bool ScalarEvolution::isSCEVable(Type *Ty) const {
2673 // Integers and pointers are always SCEVable.
2674 return Ty->isIntegerTy() || Ty->isPointerTy();
2677 /// getTypeSizeInBits - Return the size in bits of the specified type,
2678 /// for which isSCEVable must return true.
2679 uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const {
2680 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2682 // If we have a TargetData, use it!
2684 return TD->getTypeSizeInBits(Ty);
2686 // Integer types have fixed sizes.
2687 if (Ty->isIntegerTy())
2688 return Ty->getPrimitiveSizeInBits();
2690 // The only other support type is pointer. Without TargetData, conservatively
2691 // assume pointers are 64-bit.
2692 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2696 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2697 /// the given type and which represents how SCEV will treat the given
2698 /// type, for which isSCEVable must return true. For pointer types,
2699 /// this is the pointer-sized integer type.
2700 Type *ScalarEvolution::getEffectiveSCEVType(Type *Ty) const {
2701 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2703 if (Ty->isIntegerTy())
2706 // The only other support type is pointer.
2707 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2708 if (TD) return TD->getIntPtrType(getContext());
2710 // Without TargetData, conservatively assume pointers are 64-bit.
2711 return Type::getInt64Ty(getContext());
2714 const SCEV *ScalarEvolution::getCouldNotCompute() {
2715 return &CouldNotCompute;
2718 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2719 /// expression and create a new one.
2720 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2721 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2723 ValueExprMapType::const_iterator I = ValueExprMap.find(V);
2724 if (I != ValueExprMap.end()) return I->second;
2725 const SCEV *S = createSCEV(V);
2727 // The process of creating a SCEV for V may have caused other SCEVs
2728 // to have been created, so it's necessary to insert the new entry
2729 // from scratch, rather than trying to remember the insert position
2731 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2735 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2737 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2738 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2740 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2742 Type *Ty = V->getType();
2743 Ty = getEffectiveSCEVType(Ty);
2744 return getMulExpr(V,
2745 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2748 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2749 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2750 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2752 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2754 Type *Ty = V->getType();
2755 Ty = getEffectiveSCEVType(Ty);
2756 const SCEV *AllOnes =
2757 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2758 return getMinusSCEV(AllOnes, V);
2761 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1.
2762 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2763 SCEV::NoWrapFlags Flags) {
2764 assert(!maskFlags(Flags, SCEV::FlagNUW) && "subtraction does not have NUW");
2766 // Fast path: X - X --> 0.
2768 return getConstant(LHS->getType(), 0);
2771 return getAddExpr(LHS, getNegativeSCEV(RHS), Flags);
2774 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2775 /// input value to the specified type. If the type must be extended, it is zero
2778 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, Type *Ty) {
2779 Type *SrcTy = V->getType();
2780 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2781 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2782 "Cannot truncate or zero extend with non-integer arguments!");
2783 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2784 return V; // No conversion
2785 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2786 return getTruncateExpr(V, Ty);
2787 return getZeroExtendExpr(V, Ty);
2790 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2791 /// input value to the specified type. If the type must be extended, it is sign
2794 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2796 Type *SrcTy = V->getType();
2797 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2798 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2799 "Cannot truncate or zero extend with non-integer arguments!");
2800 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2801 return V; // No conversion
2802 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2803 return getTruncateExpr(V, Ty);
2804 return getSignExtendExpr(V, Ty);
2807 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2808 /// input value to the specified type. If the type must be extended, it is zero
2809 /// extended. The conversion must not be narrowing.
2811 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, Type *Ty) {
2812 Type *SrcTy = V->getType();
2813 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2814 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2815 "Cannot noop or zero extend with non-integer arguments!");
2816 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2817 "getNoopOrZeroExtend cannot truncate!");
2818 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2819 return V; // No conversion
2820 return getZeroExtendExpr(V, Ty);
2823 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2824 /// input value to the specified type. If the type must be extended, it is sign
2825 /// extended. The conversion must not be narrowing.
2827 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, Type *Ty) {
2828 Type *SrcTy = V->getType();
2829 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2830 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2831 "Cannot noop or sign extend with non-integer arguments!");
2832 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2833 "getNoopOrSignExtend cannot truncate!");
2834 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2835 return V; // No conversion
2836 return getSignExtendExpr(V, Ty);
2839 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2840 /// the input value to the specified type. If the type must be extended,
2841 /// it is extended with unspecified bits. The conversion must not be
2844 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, Type *Ty) {
2845 Type *SrcTy = V->getType();
2846 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2847 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2848 "Cannot noop or any extend with non-integer arguments!");
2849 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2850 "getNoopOrAnyExtend cannot truncate!");
2851 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2852 return V; // No conversion
2853 return getAnyExtendExpr(V, Ty);
2856 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2857 /// input value to the specified type. The conversion must not be widening.
2859 ScalarEvolution::getTruncateOrNoop(const SCEV *V, Type *Ty) {
2860 Type *SrcTy = V->getType();
2861 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2862 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2863 "Cannot truncate or noop with non-integer arguments!");
2864 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2865 "getTruncateOrNoop cannot extend!");
2866 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2867 return V; // No conversion
2868 return getTruncateExpr(V, Ty);
2871 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2872 /// the types using zero-extension, and then perform a umax operation
2874 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2876 const SCEV *PromotedLHS = LHS;
2877 const SCEV *PromotedRHS = RHS;
2879 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2880 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2882 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2884 return getUMaxExpr(PromotedLHS, PromotedRHS);
2887 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2888 /// the types using zero-extension, and then perform a umin operation
2890 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2892 const SCEV *PromotedLHS = LHS;
2893 const SCEV *PromotedRHS = RHS;
2895 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2896 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2898 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2900 return getUMinExpr(PromotedLHS, PromotedRHS);
2903 /// getPointerBase - Transitively follow the chain of pointer-type operands
2904 /// until reaching a SCEV that does not have a single pointer operand. This
2905 /// returns a SCEVUnknown pointer for well-formed pointer-type expressions,
2906 /// but corner cases do exist.
2907 const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) {
2908 // A pointer operand may evaluate to a nonpointer expression, such as null.
2909 if (!V->getType()->isPointerTy())
2912 if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(V)) {
2913 return getPointerBase(Cast->getOperand());
2915 else if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(V)) {
2916 const SCEV *PtrOp = 0;
2917 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
2919 if ((*I)->getType()->isPointerTy()) {
2920 // Cannot find the base of an expression with multiple pointer operands.
2928 return getPointerBase(PtrOp);
2933 /// PushDefUseChildren - Push users of the given Instruction
2934 /// onto the given Worklist.
2936 PushDefUseChildren(Instruction *I,
2937 SmallVectorImpl<Instruction *> &Worklist) {
2938 // Push the def-use children onto the Worklist stack.
2939 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2941 Worklist.push_back(cast<Instruction>(*UI));
2944 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2945 /// instructions that depend on the given instruction and removes them from
2946 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2949 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2950 SmallVector<Instruction *, 16> Worklist;
2951 PushDefUseChildren(PN, Worklist);
2953 SmallPtrSet<Instruction *, 8> Visited;
2955 while (!Worklist.empty()) {
2956 Instruction *I = Worklist.pop_back_val();
2957 if (!Visited.insert(I)) continue;
2959 ValueExprMapType::iterator It =
2960 ValueExprMap.find(static_cast<Value *>(I));
2961 if (It != ValueExprMap.end()) {
2962 const SCEV *Old = It->second;
2964 // Short-circuit the def-use traversal if the symbolic name
2965 // ceases to appear in expressions.
2966 if (Old != SymName && !hasOperand(Old, SymName))
2969 // SCEVUnknown for a PHI either means that it has an unrecognized
2970 // structure, it's a PHI that's in the progress of being computed
2971 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2972 // additional loop trip count information isn't going to change anything.
2973 // In the second case, createNodeForPHI will perform the necessary
2974 // updates on its own when it gets to that point. In the third, we do
2975 // want to forget the SCEVUnknown.
2976 if (!isa<PHINode>(I) ||
2977 !isa<SCEVUnknown>(Old) ||
2978 (I != PN && Old == SymName)) {
2979 forgetMemoizedResults(Old);
2980 ValueExprMap.erase(It);
2984 PushDefUseChildren(I, Worklist);
2988 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2989 /// a loop header, making it a potential recurrence, or it doesn't.
2991 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2992 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2993 if (L->getHeader() == PN->getParent()) {
2994 // The loop may have multiple entrances or multiple exits; we can analyze
2995 // this phi as an addrec if it has a unique entry value and a unique
2997 Value *BEValueV = 0, *StartValueV = 0;
2998 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2999 Value *V = PN->getIncomingValue(i);
3000 if (L->contains(PN->getIncomingBlock(i))) {
3003 } else if (BEValueV != V) {
3007 } else if (!StartValueV) {
3009 } else if (StartValueV != V) {
3014 if (BEValueV && StartValueV) {
3015 // While we are analyzing this PHI node, handle its value symbolically.
3016 const SCEV *SymbolicName = getUnknown(PN);
3017 assert(ValueExprMap.find(PN) == ValueExprMap.end() &&
3018 "PHI node already processed?");
3019 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
3021 // Using this symbolic name for the PHI, analyze the value coming around
3023 const SCEV *BEValue = getSCEV(BEValueV);
3025 // NOTE: If BEValue is loop invariant, we know that the PHI node just
3026 // has a special value for the first iteration of the loop.
3028 // If the value coming around the backedge is an add with the symbolic
3029 // value we just inserted, then we found a simple induction variable!
3030 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
3031 // If there is a single occurrence of the symbolic value, replace it
3032 // with a recurrence.
3033 unsigned FoundIndex = Add->getNumOperands();
3034 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3035 if (Add->getOperand(i) == SymbolicName)
3036 if (FoundIndex == e) {
3041 if (FoundIndex != Add->getNumOperands()) {
3042 // Create an add with everything but the specified operand.
3043 SmallVector<const SCEV *, 8> Ops;
3044 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3045 if (i != FoundIndex)
3046 Ops.push_back(Add->getOperand(i));
3047 const SCEV *Accum = getAddExpr(Ops);
3049 // This is not a valid addrec if the step amount is varying each
3050 // loop iteration, but is not itself an addrec in this loop.
3051 if (isLoopInvariant(Accum, L) ||
3052 (isa<SCEVAddRecExpr>(Accum) &&
3053 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
3054 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap;
3056 // If the increment doesn't overflow, then neither the addrec nor
3057 // the post-increment will overflow.
3058 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
3059 if (OBO->hasNoUnsignedWrap())
3060 Flags = setFlags(Flags, SCEV::FlagNUW);
3061 if (OBO->hasNoSignedWrap())
3062 Flags = setFlags(Flags, SCEV::FlagNSW);
3063 } else if (const GEPOperator *GEP =
3064 dyn_cast<GEPOperator>(BEValueV)) {
3065 // If the increment is an inbounds GEP, then we know the address
3066 // space cannot be wrapped around. We cannot make any guarantee
3067 // about signed or unsigned overflow because pointers are
3068 // unsigned but we may have a negative index from the base
3070 if (GEP->isInBounds())
3071 Flags = setFlags(Flags, SCEV::FlagNW);
3074 const SCEV *StartVal = getSCEV(StartValueV);
3075 const SCEV *PHISCEV = getAddRecExpr(StartVal, Accum, L, Flags);
3077 // Since the no-wrap flags are on the increment, they apply to the
3078 // post-incremented value as well.
3079 if (isLoopInvariant(Accum, L))
3080 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
3083 // Okay, for the entire analysis of this edge we assumed the PHI
3084 // to be symbolic. We now need to go back and purge all of the
3085 // entries for the scalars that use the symbolic expression.
3086 ForgetSymbolicName(PN, SymbolicName);
3087 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3091 } else if (const SCEVAddRecExpr *AddRec =
3092 dyn_cast<SCEVAddRecExpr>(BEValue)) {
3093 // Otherwise, this could be a loop like this:
3094 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
3095 // In this case, j = {1,+,1} and BEValue is j.
3096 // Because the other in-value of i (0) fits the evolution of BEValue
3097 // i really is an addrec evolution.
3098 if (AddRec->getLoop() == L && AddRec->isAffine()) {
3099 const SCEV *StartVal = getSCEV(StartValueV);
3101 // If StartVal = j.start - j.stride, we can use StartVal as the
3102 // initial step of the addrec evolution.
3103 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
3104 AddRec->getOperand(1))) {
3105 // FIXME: For constant StartVal, we should be able to infer
3107 const SCEV *PHISCEV =
3108 getAddRecExpr(StartVal, AddRec->getOperand(1), L,
3111 // Okay, for the entire analysis of this edge we assumed the PHI
3112 // to be symbolic. We now need to go back and purge all of the
3113 // entries for the scalars that use the symbolic expression.
3114 ForgetSymbolicName(PN, SymbolicName);
3115 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3123 // If the PHI has a single incoming value, follow that value, unless the
3124 // PHI's incoming blocks are in a different loop, in which case doing so
3125 // risks breaking LCSSA form. Instcombine would normally zap these, but
3126 // it doesn't have DominatorTree information, so it may miss cases.
3127 if (Value *V = SimplifyInstruction(PN, TD, TLI, DT))
3128 if (LI->replacementPreservesLCSSAForm(PN, V))
3131 // If it's not a loop phi, we can't handle it yet.
3132 return getUnknown(PN);
3135 /// createNodeForGEP - Expand GEP instructions into add and multiply
3136 /// operations. This allows them to be analyzed by regular SCEV code.
3138 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
3140 // Don't blindly transfer the inbounds flag from the GEP instruction to the
3141 // Add expression, because the Instruction may be guarded by control flow
3142 // and the no-overflow bits may not be valid for the expression in any
3144 bool isInBounds = GEP->isInBounds();
3146 Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
3147 Value *Base = GEP->getOperand(0);
3148 // Don't attempt to analyze GEPs over unsized objects.
3149 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
3150 return getUnknown(GEP);
3151 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
3152 gep_type_iterator GTI = gep_type_begin(GEP);
3153 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
3157 // Compute the (potentially symbolic) offset in bytes for this index.
3158 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
3159 // For a struct, add the member offset.
3160 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
3161 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
3163 // Add the field offset to the running total offset.
3164 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
3166 // For an array, add the element offset, explicitly scaled.
3167 const SCEV *ElementSize = getSizeOfExpr(*GTI);
3168 const SCEV *IndexS = getSCEV(Index);
3169 // Getelementptr indices are signed.
3170 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
3172 // Multiply the index by the element size to compute the element offset.
3173 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize,
3174 isInBounds ? SCEV::FlagNSW :
3177 // Add the element offset to the running total offset.
3178 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
3182 // Get the SCEV for the GEP base.
3183 const SCEV *BaseS = getSCEV(Base);
3185 // Add the total offset from all the GEP indices to the base.
3186 return getAddExpr(BaseS, TotalOffset,
3187 isInBounds ? SCEV::FlagNSW : SCEV::FlagAnyWrap);
3190 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
3191 /// guaranteed to end in (at every loop iteration). It is, at the same time,
3192 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
3193 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
3195 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
3196 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3197 return C->getValue()->getValue().countTrailingZeros();
3199 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
3200 return std::min(GetMinTrailingZeros(T->getOperand()),
3201 (uint32_t)getTypeSizeInBits(T->getType()));
3203 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
3204 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3205 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3206 getTypeSizeInBits(E->getType()) : OpRes;
3209 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
3210 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3211 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3212 getTypeSizeInBits(E->getType()) : OpRes;
3215 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
3216 // The result is the min of all operands results.
3217 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3218 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3219 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3223 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
3224 // The result is the sum of all operands results.
3225 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
3226 uint32_t BitWidth = getTypeSizeInBits(M->getType());
3227 for (unsigned i = 1, e = M->getNumOperands();
3228 SumOpRes != BitWidth && i != e; ++i)
3229 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
3234 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
3235 // The result is the min of all operands results.
3236 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3237 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3238 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3242 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(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 SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
3251 // The result is the min of all operands results.
3252 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3253 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3254 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3258 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3259 // For a SCEVUnknown, ask ValueTracking.
3260 unsigned BitWidth = getTypeSizeInBits(U->getType());
3261 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3262 ComputeMaskedBits(U->getValue(), Zeros, Ones);
3263 return Zeros.countTrailingOnes();
3270 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
3273 ScalarEvolution::getUnsignedRange(const SCEV *S) {
3274 // See if we've computed this range already.
3275 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
3276 if (I != UnsignedRanges.end())
3279 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3280 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
3282 unsigned BitWidth = getTypeSizeInBits(S->getType());
3283 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3285 // If the value has known zeros, the maximum unsigned value will have those
3286 // known zeros as well.
3287 uint32_t TZ = GetMinTrailingZeros(S);
3289 ConservativeResult =
3290 ConstantRange(APInt::getMinValue(BitWidth),
3291 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
3293 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3294 ConstantRange X = getUnsignedRange(Add->getOperand(0));
3295 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3296 X = X.add(getUnsignedRange(Add->getOperand(i)));
3297 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
3300 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3301 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
3302 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3303 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
3304 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
3307 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3308 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
3309 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3310 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
3311 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
3314 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3315 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3316 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3317 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3318 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
3321 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3322 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3323 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3324 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3327 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3328 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3329 return setUnsignedRange(ZExt,
3330 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3333 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3334 ConstantRange X = getUnsignedRange(SExt->getOperand());
3335 return setUnsignedRange(SExt,
3336 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3339 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3340 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3341 return setUnsignedRange(Trunc,
3342 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3345 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3346 // If there's no unsigned wrap, the value will never be less than its
3348 if (AddRec->getNoWrapFlags(SCEV::FlagNUW))
3349 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3350 if (!C->getValue()->isZero())
3351 ConservativeResult =
3352 ConservativeResult.intersectWith(
3353 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3355 // TODO: non-affine addrec
3356 if (AddRec->isAffine()) {
3357 Type *Ty = AddRec->getType();
3358 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3359 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3360 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3361 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3363 const SCEV *Start = AddRec->getStart();
3364 const SCEV *Step = AddRec->getStepRecurrence(*this);
3366 ConstantRange StartRange = getUnsignedRange(Start);
3367 ConstantRange StepRange = getSignedRange(Step);
3368 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3369 ConstantRange EndRange =
3370 StartRange.add(MaxBECountRange.multiply(StepRange));
3372 // Check for overflow. This must be done with ConstantRange arithmetic
3373 // because we could be called from within the ScalarEvolution overflow
3375 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3376 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3377 ConstantRange ExtMaxBECountRange =
3378 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3379 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3380 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3382 return setUnsignedRange(AddRec, ConservativeResult);
3384 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3385 EndRange.getUnsignedMin());
3386 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3387 EndRange.getUnsignedMax());
3388 if (Min.isMinValue() && Max.isMaxValue())
3389 return setUnsignedRange(AddRec, ConservativeResult);
3390 return setUnsignedRange(AddRec,
3391 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3395 return setUnsignedRange(AddRec, ConservativeResult);
3398 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3399 // For a SCEVUnknown, ask ValueTracking.
3400 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3401 ComputeMaskedBits(U->getValue(), Zeros, Ones, TD);
3402 if (Ones == ~Zeros + 1)
3403 return setUnsignedRange(U, ConservativeResult);
3404 return setUnsignedRange(U,
3405 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3408 return setUnsignedRange(S, ConservativeResult);
3411 /// getSignedRange - Determine the signed range for a particular SCEV.
3414 ScalarEvolution::getSignedRange(const SCEV *S) {
3415 // See if we've computed this range already.
3416 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3417 if (I != SignedRanges.end())
3420 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3421 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3423 unsigned BitWidth = getTypeSizeInBits(S->getType());
3424 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3426 // If the value has known zeros, the maximum signed value will have those
3427 // known zeros as well.
3428 uint32_t TZ = GetMinTrailingZeros(S);
3430 ConservativeResult =
3431 ConstantRange(APInt::getSignedMinValue(BitWidth),
3432 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3434 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3435 ConstantRange X = getSignedRange(Add->getOperand(0));
3436 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3437 X = X.add(getSignedRange(Add->getOperand(i)));
3438 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3441 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3442 ConstantRange X = getSignedRange(Mul->getOperand(0));
3443 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3444 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3445 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3448 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3449 ConstantRange X = getSignedRange(SMax->getOperand(0));
3450 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3451 X = X.smax(getSignedRange(SMax->getOperand(i)));
3452 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3455 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3456 ConstantRange X = getSignedRange(UMax->getOperand(0));
3457 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3458 X = X.umax(getSignedRange(UMax->getOperand(i)));
3459 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3462 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3463 ConstantRange X = getSignedRange(UDiv->getLHS());
3464 ConstantRange Y = getSignedRange(UDiv->getRHS());
3465 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3468 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3469 ConstantRange X = getSignedRange(ZExt->getOperand());
3470 return setSignedRange(ZExt,
3471 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3474 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3475 ConstantRange X = getSignedRange(SExt->getOperand());
3476 return setSignedRange(SExt,
3477 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3480 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3481 ConstantRange X = getSignedRange(Trunc->getOperand());
3482 return setSignedRange(Trunc,
3483 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3486 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3487 // If there's no signed wrap, and all the operands have the same sign or
3488 // zero, the value won't ever change sign.
3489 if (AddRec->getNoWrapFlags(SCEV::FlagNSW)) {
3490 bool AllNonNeg = true;
3491 bool AllNonPos = true;
3492 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3493 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3494 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3497 ConservativeResult = ConservativeResult.intersectWith(
3498 ConstantRange(APInt(BitWidth, 0),
3499 APInt::getSignedMinValue(BitWidth)));
3501 ConservativeResult = ConservativeResult.intersectWith(
3502 ConstantRange(APInt::getSignedMinValue(BitWidth),
3503 APInt(BitWidth, 1)));
3506 // TODO: non-affine addrec
3507 if (AddRec->isAffine()) {
3508 Type *Ty = AddRec->getType();
3509 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3510 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3511 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3512 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3514 const SCEV *Start = AddRec->getStart();
3515 const SCEV *Step = AddRec->getStepRecurrence(*this);
3517 ConstantRange StartRange = getSignedRange(Start);
3518 ConstantRange StepRange = getSignedRange(Step);
3519 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3520 ConstantRange EndRange =
3521 StartRange.add(MaxBECountRange.multiply(StepRange));
3523 // Check for overflow. This must be done with ConstantRange arithmetic
3524 // because we could be called from within the ScalarEvolution overflow
3526 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3527 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3528 ConstantRange ExtMaxBECountRange =
3529 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3530 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3531 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3533 return setSignedRange(AddRec, ConservativeResult);
3535 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3536 EndRange.getSignedMin());
3537 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3538 EndRange.getSignedMax());
3539 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3540 return setSignedRange(AddRec, ConservativeResult);
3541 return setSignedRange(AddRec,
3542 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3546 return setSignedRange(AddRec, ConservativeResult);
3549 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3550 // For a SCEVUnknown, ask ValueTracking.
3551 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3552 return setSignedRange(U, ConservativeResult);
3553 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3555 return setSignedRange(U, ConservativeResult);
3556 return setSignedRange(U, ConservativeResult.intersectWith(
3557 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3558 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3561 return setSignedRange(S, ConservativeResult);
3564 /// createSCEV - We know that there is no SCEV for the specified value.
3565 /// Analyze the expression.
3567 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3568 if (!isSCEVable(V->getType()))
3569 return getUnknown(V);
3571 unsigned Opcode = Instruction::UserOp1;
3572 if (Instruction *I = dyn_cast<Instruction>(V)) {
3573 Opcode = I->getOpcode();
3575 // Don't attempt to analyze instructions in blocks that aren't
3576 // reachable. Such instructions don't matter, and they aren't required
3577 // to obey basic rules for definitions dominating uses which this
3578 // analysis depends on.
3579 if (!DT->isReachableFromEntry(I->getParent()))
3580 return getUnknown(V);
3581 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3582 Opcode = CE->getOpcode();
3583 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3584 return getConstant(CI);
3585 else if (isa<ConstantPointerNull>(V))
3586 return getConstant(V->getType(), 0);
3587 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3588 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3590 return getUnknown(V);
3592 Operator *U = cast<Operator>(V);
3594 case Instruction::Add: {
3595 // The simple thing to do would be to just call getSCEV on both operands
3596 // and call getAddExpr with the result. However if we're looking at a
3597 // bunch of things all added together, this can be quite inefficient,
3598 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3599 // Instead, gather up all the operands and make a single getAddExpr call.
3600 // LLVM IR canonical form means we need only traverse the left operands.
3602 // Don't apply this instruction's NSW or NUW flags to the new
3603 // expression. The instruction may be guarded by control flow that the
3604 // no-wrap behavior depends on. Non-control-equivalent instructions can be
3605 // mapped to the same SCEV expression, and it would be incorrect to transfer
3606 // NSW/NUW semantics to those operations.
3607 SmallVector<const SCEV *, 4> AddOps;
3608 AddOps.push_back(getSCEV(U->getOperand(1)));
3609 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3610 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3611 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3613 U = cast<Operator>(Op);
3614 const SCEV *Op1 = getSCEV(U->getOperand(1));
3615 if (Opcode == Instruction::Sub)
3616 AddOps.push_back(getNegativeSCEV(Op1));
3618 AddOps.push_back(Op1);
3620 AddOps.push_back(getSCEV(U->getOperand(0)));
3621 return getAddExpr(AddOps);
3623 case Instruction::Mul: {
3624 // Don't transfer NSW/NUW for the same reason as AddExpr.
3625 SmallVector<const SCEV *, 4> MulOps;
3626 MulOps.push_back(getSCEV(U->getOperand(1)));
3627 for (Value *Op = U->getOperand(0);
3628 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3629 Op = U->getOperand(0)) {
3630 U = cast<Operator>(Op);
3631 MulOps.push_back(getSCEV(U->getOperand(1)));
3633 MulOps.push_back(getSCEV(U->getOperand(0)));
3634 return getMulExpr(MulOps);
3636 case Instruction::UDiv:
3637 return getUDivExpr(getSCEV(U->getOperand(0)),
3638 getSCEV(U->getOperand(1)));
3639 case Instruction::Sub:
3640 return getMinusSCEV(getSCEV(U->getOperand(0)),
3641 getSCEV(U->getOperand(1)));
3642 case Instruction::And:
3643 // For an expression like x&255 that merely masks off the high bits,
3644 // use zext(trunc(x)) as the SCEV expression.
3645 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3646 if (CI->isNullValue())
3647 return getSCEV(U->getOperand(1));
3648 if (CI->isAllOnesValue())
3649 return getSCEV(U->getOperand(0));
3650 const APInt &A = CI->getValue();
3652 // Instcombine's ShrinkDemandedConstant may strip bits out of
3653 // constants, obscuring what would otherwise be a low-bits mask.
3654 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3655 // knew about to reconstruct a low-bits mask value.
3656 unsigned LZ = A.countLeadingZeros();
3657 unsigned BitWidth = A.getBitWidth();
3658 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3659 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD);
3661 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3663 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3665 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3666 IntegerType::get(getContext(), BitWidth - LZ)),
3671 case Instruction::Or:
3672 // If the RHS of the Or is a constant, we may have something like:
3673 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3674 // optimizations will transparently handle this case.
3676 // In order for this transformation to be safe, the LHS must be of the
3677 // form X*(2^n) and the Or constant must be less than 2^n.
3678 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3679 const SCEV *LHS = getSCEV(U->getOperand(0));
3680 const APInt &CIVal = CI->getValue();
3681 if (GetMinTrailingZeros(LHS) >=
3682 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3683 // Build a plain add SCEV.
3684 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3685 // If the LHS of the add was an addrec and it has no-wrap flags,
3686 // transfer the no-wrap flags, since an or won't introduce a wrap.
3687 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3688 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3689 const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags(
3690 OldAR->getNoWrapFlags());
3696 case Instruction::Xor:
3697 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3698 // If the RHS of the xor is a signbit, then this is just an add.
3699 // Instcombine turns add of signbit into xor as a strength reduction step.
3700 if (CI->getValue().isSignBit())
3701 return getAddExpr(getSCEV(U->getOperand(0)),
3702 getSCEV(U->getOperand(1)));
3704 // If the RHS of xor is -1, then this is a not operation.
3705 if (CI->isAllOnesValue())
3706 return getNotSCEV(getSCEV(U->getOperand(0)));
3708 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3709 // This is a variant of the check for xor with -1, and it handles
3710 // the case where instcombine has trimmed non-demanded bits out
3711 // of an xor with -1.
3712 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3713 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3714 if (BO->getOpcode() == Instruction::And &&
3715 LCI->getValue() == CI->getValue())
3716 if (const SCEVZeroExtendExpr *Z =
3717 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3718 Type *UTy = U->getType();
3719 const SCEV *Z0 = Z->getOperand();
3720 Type *Z0Ty = Z0->getType();
3721 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3723 // If C is a low-bits mask, the zero extend is serving to
3724 // mask off the high bits. Complement the operand and
3725 // re-apply the zext.
3726 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3727 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3729 // If C is a single bit, it may be in the sign-bit position
3730 // before the zero-extend. In this case, represent the xor
3731 // using an add, which is equivalent, and re-apply the zext.
3732 APInt Trunc = CI->getValue().trunc(Z0TySize);
3733 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3735 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3741 case Instruction::Shl:
3742 // Turn shift left of a constant amount into a multiply.
3743 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3744 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3746 // If the shift count is not less than the bitwidth, the result of
3747 // the shift is undefined. Don't try to analyze it, because the
3748 // resolution chosen here may differ from the resolution chosen in
3749 // other parts of the compiler.
3750 if (SA->getValue().uge(BitWidth))
3753 Constant *X = ConstantInt::get(getContext(),
3754 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3755 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3759 case Instruction::LShr:
3760 // Turn logical shift right of a constant into a unsigned divide.
3761 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3762 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3764 // If the shift count is not less than the bitwidth, the result of
3765 // the shift is undefined. Don't try to analyze it, because the
3766 // resolution chosen here may differ from the resolution chosen in
3767 // other parts of the compiler.
3768 if (SA->getValue().uge(BitWidth))
3771 Constant *X = ConstantInt::get(getContext(),
3772 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3773 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3777 case Instruction::AShr:
3778 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3779 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3780 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3781 if (L->getOpcode() == Instruction::Shl &&
3782 L->getOperand(1) == U->getOperand(1)) {
3783 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3785 // If the shift count is not less than the bitwidth, the result of
3786 // the shift is undefined. Don't try to analyze it, because the
3787 // resolution chosen here may differ from the resolution chosen in
3788 // other parts of the compiler.
3789 if (CI->getValue().uge(BitWidth))
3792 uint64_t Amt = BitWidth - CI->getZExtValue();
3793 if (Amt == BitWidth)
3794 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3796 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3797 IntegerType::get(getContext(),
3803 case Instruction::Trunc:
3804 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3806 case Instruction::ZExt:
3807 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3809 case Instruction::SExt:
3810 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3812 case Instruction::BitCast:
3813 // BitCasts are no-op casts so we just eliminate the cast.
3814 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3815 return getSCEV(U->getOperand(0));
3818 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3819 // lead to pointer expressions which cannot safely be expanded to GEPs,
3820 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3821 // simplifying integer expressions.
3823 case Instruction::GetElementPtr:
3824 return createNodeForGEP(cast<GEPOperator>(U));
3826 case Instruction::PHI:
3827 return createNodeForPHI(cast<PHINode>(U));
3829 case Instruction::Select:
3830 // This could be a smax or umax that was lowered earlier.
3831 // Try to recover it.
3832 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3833 Value *LHS = ICI->getOperand(0);
3834 Value *RHS = ICI->getOperand(1);
3835 switch (ICI->getPredicate()) {
3836 case ICmpInst::ICMP_SLT:
3837 case ICmpInst::ICMP_SLE:
3838 std::swap(LHS, RHS);
3840 case ICmpInst::ICMP_SGT:
3841 case ICmpInst::ICMP_SGE:
3842 // a >s b ? a+x : b+x -> smax(a, b)+x
3843 // a >s b ? b+x : a+x -> smin(a, b)+x
3844 if (LHS->getType() == U->getType()) {
3845 const SCEV *LS = getSCEV(LHS);
3846 const SCEV *RS = getSCEV(RHS);
3847 const SCEV *LA = getSCEV(U->getOperand(1));
3848 const SCEV *RA = getSCEV(U->getOperand(2));
3849 const SCEV *LDiff = getMinusSCEV(LA, LS);
3850 const SCEV *RDiff = getMinusSCEV(RA, RS);
3852 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3853 LDiff = getMinusSCEV(LA, RS);
3854 RDiff = getMinusSCEV(RA, LS);
3856 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3859 case ICmpInst::ICMP_ULT:
3860 case ICmpInst::ICMP_ULE:
3861 std::swap(LHS, RHS);
3863 case ICmpInst::ICMP_UGT:
3864 case ICmpInst::ICMP_UGE:
3865 // a >u b ? a+x : b+x -> umax(a, b)+x
3866 // a >u b ? b+x : a+x -> umin(a, b)+x
3867 if (LHS->getType() == U->getType()) {
3868 const SCEV *LS = getSCEV(LHS);
3869 const SCEV *RS = getSCEV(RHS);
3870 const SCEV *LA = getSCEV(U->getOperand(1));
3871 const SCEV *RA = getSCEV(U->getOperand(2));
3872 const SCEV *LDiff = getMinusSCEV(LA, LS);
3873 const SCEV *RDiff = getMinusSCEV(RA, RS);
3875 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3876 LDiff = getMinusSCEV(LA, RS);
3877 RDiff = getMinusSCEV(RA, LS);
3879 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3882 case ICmpInst::ICMP_NE:
3883 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3884 if (LHS->getType() == U->getType() &&
3885 isa<ConstantInt>(RHS) &&
3886 cast<ConstantInt>(RHS)->isZero()) {
3887 const SCEV *One = getConstant(LHS->getType(), 1);
3888 const SCEV *LS = getSCEV(LHS);
3889 const SCEV *LA = getSCEV(U->getOperand(1));
3890 const SCEV *RA = getSCEV(U->getOperand(2));
3891 const SCEV *LDiff = getMinusSCEV(LA, LS);
3892 const SCEV *RDiff = getMinusSCEV(RA, One);
3894 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3897 case ICmpInst::ICMP_EQ:
3898 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3899 if (LHS->getType() == U->getType() &&
3900 isa<ConstantInt>(RHS) &&
3901 cast<ConstantInt>(RHS)->isZero()) {
3902 const SCEV *One = getConstant(LHS->getType(), 1);
3903 const SCEV *LS = getSCEV(LHS);
3904 const SCEV *LA = getSCEV(U->getOperand(1));
3905 const SCEV *RA = getSCEV(U->getOperand(2));
3906 const SCEV *LDiff = getMinusSCEV(LA, One);
3907 const SCEV *RDiff = getMinusSCEV(RA, LS);
3909 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3917 default: // We cannot analyze this expression.
3921 return getUnknown(V);
3926 //===----------------------------------------------------------------------===//
3927 // Iteration Count Computation Code
3930 /// getSmallConstantTripCount - Returns the maximum trip count of this loop as a
3931 /// normal unsigned value. Returns 0 if the trip count is unknown or not
3932 /// constant. Will also return 0 if the maximum trip count is very large (>=
3935 /// This "trip count" assumes that control exits via ExitingBlock. More
3936 /// precisely, it is the number of times that control may reach ExitingBlock
3937 /// before taking the branch. For loops with multiple exits, it may not be the
3938 /// number times that the loop header executes because the loop may exit
3939 /// prematurely via another branch.
3940 unsigned ScalarEvolution::
3941 getSmallConstantTripCount(Loop *L, BasicBlock *ExitingBlock) {
3942 const SCEVConstant *ExitCount =
3943 dyn_cast<SCEVConstant>(getExitCount(L, ExitingBlock));
3947 ConstantInt *ExitConst = ExitCount->getValue();
3949 // Guard against huge trip counts.
3950 if (ExitConst->getValue().getActiveBits() > 32)
3953 // In case of integer overflow, this returns 0, which is correct.
3954 return ((unsigned)ExitConst->getZExtValue()) + 1;
3957 /// getSmallConstantTripMultiple - Returns the largest constant divisor of the
3958 /// trip count of this loop as a normal unsigned value, if possible. This
3959 /// means that the actual trip count is always a multiple of the returned
3960 /// value (don't forget the trip count could very well be zero as well!).
3962 /// Returns 1 if the trip count is unknown or not guaranteed to be the
3963 /// multiple of a constant (which is also the case if the trip count is simply
3964 /// constant, use getSmallConstantTripCount for that case), Will also return 1
3965 /// if the trip count is very large (>= 2^32).
3967 /// As explained in the comments for getSmallConstantTripCount, this assumes
3968 /// that control exits the loop via ExitingBlock.
3969 unsigned ScalarEvolution::
3970 getSmallConstantTripMultiple(Loop *L, BasicBlock *ExitingBlock) {
3971 const SCEV *ExitCount = getExitCount(L, ExitingBlock);
3972 if (ExitCount == getCouldNotCompute())
3975 // Get the trip count from the BE count by adding 1.
3976 const SCEV *TCMul = getAddExpr(ExitCount,
3977 getConstant(ExitCount->getType(), 1));
3978 // FIXME: SCEV distributes multiplication as V1*C1 + V2*C1. We could attempt
3979 // to factor simple cases.
3980 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(TCMul))
3981 TCMul = Mul->getOperand(0);
3983 const SCEVConstant *MulC = dyn_cast<SCEVConstant>(TCMul);
3987 ConstantInt *Result = MulC->getValue();
3989 // Guard against huge trip counts.
3990 if (!Result || Result->getValue().getActiveBits() > 32)
3993 return (unsigned)Result->getZExtValue();
3996 // getExitCount - Get the expression for the number of loop iterations for which
3997 // this loop is guaranteed not to exit via ExitintBlock. Otherwise return
3998 // SCEVCouldNotCompute.
3999 const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) {
4000 return getBackedgeTakenInfo(L).getExact(ExitingBlock, this);
4003 /// getBackedgeTakenCount - If the specified loop has a predictable
4004 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
4005 /// object. The backedge-taken count is the number of times the loop header
4006 /// will be branched to from within the loop. This is one less than the
4007 /// trip count of the loop, since it doesn't count the first iteration,
4008 /// when the header is branched to from outside the loop.
4010 /// Note that it is not valid to call this method on a loop without a
4011 /// loop-invariant backedge-taken count (see
4012 /// hasLoopInvariantBackedgeTakenCount).
4014 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
4015 return getBackedgeTakenInfo(L).getExact(this);
4018 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
4019 /// return the least SCEV value that is known never to be less than the
4020 /// actual backedge taken count.
4021 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
4022 return getBackedgeTakenInfo(L).getMax(this);
4025 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
4026 /// onto the given Worklist.
4028 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
4029 BasicBlock *Header = L->getHeader();
4031 // Push all Loop-header PHIs onto the Worklist stack.
4032 for (BasicBlock::iterator I = Header->begin();
4033 PHINode *PN = dyn_cast<PHINode>(I); ++I)
4034 Worklist.push_back(PN);
4037 const ScalarEvolution::BackedgeTakenInfo &
4038 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
4039 // Initially insert an invalid entry for this loop. If the insertion
4040 // succeeds, proceed to actually compute a backedge-taken count and
4041 // update the value. The temporary CouldNotCompute value tells SCEV
4042 // code elsewhere that it shouldn't attempt to request a new
4043 // backedge-taken count, which could result in infinite recursion.
4044 std::pair<DenseMap<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
4045 BackedgeTakenCounts.insert(std::make_pair(L, BackedgeTakenInfo()));
4047 return Pair.first->second;
4049 // ComputeBackedgeTakenCount may allocate memory for its result. Inserting it
4050 // into the BackedgeTakenCounts map transfers ownership. Otherwise, the result
4051 // must be cleared in this scope.
4052 BackedgeTakenInfo Result = ComputeBackedgeTakenCount(L);
4054 if (Result.getExact(this) != getCouldNotCompute()) {
4055 assert(isLoopInvariant(Result.getExact(this), L) &&
4056 isLoopInvariant(Result.getMax(this), L) &&
4057 "Computed backedge-taken count isn't loop invariant for loop!");
4058 ++NumTripCountsComputed;
4060 else if (Result.getMax(this) == getCouldNotCompute() &&
4061 isa<PHINode>(L->getHeader()->begin())) {
4062 // Only count loops that have phi nodes as not being computable.
4063 ++NumTripCountsNotComputed;
4066 // Now that we know more about the trip count for this loop, forget any
4067 // existing SCEV values for PHI nodes in this loop since they are only
4068 // conservative estimates made without the benefit of trip count
4069 // information. This is similar to the code in forgetLoop, except that
4070 // it handles SCEVUnknown PHI nodes specially.
4071 if (Result.hasAnyInfo()) {
4072 SmallVector<Instruction *, 16> Worklist;
4073 PushLoopPHIs(L, Worklist);
4075 SmallPtrSet<Instruction *, 8> Visited;
4076 while (!Worklist.empty()) {
4077 Instruction *I = Worklist.pop_back_val();
4078 if (!Visited.insert(I)) continue;
4080 ValueExprMapType::iterator It =
4081 ValueExprMap.find(static_cast<Value *>(I));
4082 if (It != ValueExprMap.end()) {
4083 const SCEV *Old = It->second;
4085 // SCEVUnknown for a PHI either means that it has an unrecognized
4086 // structure, or it's a PHI that's in the progress of being computed
4087 // by createNodeForPHI. In the former case, additional loop trip
4088 // count information isn't going to change anything. In the later
4089 // case, createNodeForPHI will perform the necessary updates on its
4090 // own when it gets to that point.
4091 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
4092 forgetMemoizedResults(Old);
4093 ValueExprMap.erase(It);
4095 if (PHINode *PN = dyn_cast<PHINode>(I))
4096 ConstantEvolutionLoopExitValue.erase(PN);
4099 PushDefUseChildren(I, Worklist);
4103 // Re-lookup the insert position, since the call to
4104 // ComputeBackedgeTakenCount above could result in a
4105 // recusive call to getBackedgeTakenInfo (on a different
4106 // loop), which would invalidate the iterator computed
4108 return BackedgeTakenCounts.find(L)->second = Result;
4111 /// forgetLoop - This method should be called by the client when it has
4112 /// changed a loop in a way that may effect ScalarEvolution's ability to
4113 /// compute a trip count, or if the loop is deleted.
4114 void ScalarEvolution::forgetLoop(const Loop *L) {
4115 // Drop any stored trip count value.
4116 DenseMap<const Loop*, BackedgeTakenInfo>::iterator BTCPos =
4117 BackedgeTakenCounts.find(L);
4118 if (BTCPos != BackedgeTakenCounts.end()) {
4119 BTCPos->second.clear();
4120 BackedgeTakenCounts.erase(BTCPos);
4123 // Drop information about expressions based on loop-header PHIs.
4124 SmallVector<Instruction *, 16> Worklist;
4125 PushLoopPHIs(L, Worklist);
4127 SmallPtrSet<Instruction *, 8> Visited;
4128 while (!Worklist.empty()) {
4129 Instruction *I = Worklist.pop_back_val();
4130 if (!Visited.insert(I)) continue;
4132 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
4133 if (It != ValueExprMap.end()) {
4134 forgetMemoizedResults(It->second);
4135 ValueExprMap.erase(It);
4136 if (PHINode *PN = dyn_cast<PHINode>(I))
4137 ConstantEvolutionLoopExitValue.erase(PN);
4140 PushDefUseChildren(I, Worklist);
4143 // Forget all contained loops too, to avoid dangling entries in the
4144 // ValuesAtScopes map.
4145 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
4149 /// forgetValue - This method should be called by the client when it has
4150 /// changed a value in a way that may effect its value, or which may
4151 /// disconnect it from a def-use chain linking it to a loop.
4152 void ScalarEvolution::forgetValue(Value *V) {
4153 Instruction *I = dyn_cast<Instruction>(V);
4156 // Drop information about expressions based on loop-header PHIs.
4157 SmallVector<Instruction *, 16> Worklist;
4158 Worklist.push_back(I);
4160 SmallPtrSet<Instruction *, 8> Visited;
4161 while (!Worklist.empty()) {
4162 I = Worklist.pop_back_val();
4163 if (!Visited.insert(I)) continue;
4165 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
4166 if (It != ValueExprMap.end()) {
4167 forgetMemoizedResults(It->second);
4168 ValueExprMap.erase(It);
4169 if (PHINode *PN = dyn_cast<PHINode>(I))
4170 ConstantEvolutionLoopExitValue.erase(PN);
4173 PushDefUseChildren(I, Worklist);
4177 /// getExact - Get the exact loop backedge taken count considering all loop
4178 /// exits. A computable result can only be return for loops with a single exit.
4179 /// Returning the minimum taken count among all exits is incorrect because one
4180 /// of the loop's exit limit's may have been skipped. HowFarToZero assumes that
4181 /// the limit of each loop test is never skipped. This is a valid assumption as
4182 /// long as the loop exits via that test. For precise results, it is the
4183 /// caller's responsibility to specify the relevant loop exit using
4184 /// getExact(ExitingBlock, SE).
4186 ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
4187 // If any exits were not computable, the loop is not computable.
4188 if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute();
4190 // We need exactly one computable exit.
4191 if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
4192 assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
4194 const SCEV *BECount = 0;
4195 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4196 ENT != 0; ENT = ENT->getNextExit()) {
4198 assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV");
4201 BECount = ENT->ExactNotTaken;
4202 else if (BECount != ENT->ExactNotTaken)
4203 return SE->getCouldNotCompute();
4205 assert(BECount && "Invalid not taken count for loop exit");
4209 /// getExact - Get the exact not taken count for this loop exit.
4211 ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock,
4212 ScalarEvolution *SE) const {
4213 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4214 ENT != 0; ENT = ENT->getNextExit()) {
4216 if (ENT->ExitingBlock == ExitingBlock)
4217 return ENT->ExactNotTaken;
4219 return SE->getCouldNotCompute();
4222 /// getMax - Get the max backedge taken count for the loop.
4224 ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const {
4225 return Max ? Max : SE->getCouldNotCompute();
4228 /// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each
4229 /// computable exit into a persistent ExitNotTakenInfo array.
4230 ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo(
4231 SmallVectorImpl< std::pair<BasicBlock *, const SCEV *> > &ExitCounts,
4232 bool Complete, const SCEV *MaxCount) : Max(MaxCount) {
4235 ExitNotTaken.setIncomplete();
4237 unsigned NumExits = ExitCounts.size();
4238 if (NumExits == 0) return;
4240 ExitNotTaken.ExitingBlock = ExitCounts[0].first;
4241 ExitNotTaken.ExactNotTaken = ExitCounts[0].second;
4242 if (NumExits == 1) return;
4244 // Handle the rare case of multiple computable exits.
4245 ExitNotTakenInfo *ENT = new ExitNotTakenInfo[NumExits-1];
4247 ExitNotTakenInfo *PrevENT = &ExitNotTaken;
4248 for (unsigned i = 1; i < NumExits; ++i, PrevENT = ENT, ++ENT) {
4249 PrevENT->setNextExit(ENT);
4250 ENT->ExitingBlock = ExitCounts[i].first;
4251 ENT->ExactNotTaken = ExitCounts[i].second;
4255 /// clear - Invalidate this result and free the ExitNotTakenInfo array.
4256 void ScalarEvolution::BackedgeTakenInfo::clear() {
4257 ExitNotTaken.ExitingBlock = 0;
4258 ExitNotTaken.ExactNotTaken = 0;
4259 delete[] ExitNotTaken.getNextExit();
4262 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
4263 /// of the specified loop will execute.
4264 ScalarEvolution::BackedgeTakenInfo
4265 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
4266 SmallVector<BasicBlock *, 8> ExitingBlocks;
4267 L->getExitingBlocks(ExitingBlocks);
4269 // Examine all exits and pick the most conservative values.
4270 const SCEV *MaxBECount = getCouldNotCompute();
4271 bool CouldComputeBECount = true;
4272 SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts;
4273 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
4274 ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]);
4275 if (EL.Exact == getCouldNotCompute())
4276 // We couldn't compute an exact value for this exit, so
4277 // we won't be able to compute an exact value for the loop.
4278 CouldComputeBECount = false;
4280 ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact));
4282 if (MaxBECount == getCouldNotCompute())
4283 MaxBECount = EL.Max;
4284 else if (EL.Max != getCouldNotCompute()) {
4285 // We cannot take the "min" MaxBECount, because non-unit stride loops may
4286 // skip some loop tests. Taking the max over the exits is sufficiently
4287 // conservative. TODO: We could do better taking into consideration
4288 // that (1) the loop has unit stride (2) the last loop test is
4289 // less-than/greater-than (3) any loop test is less-than/greater-than AND
4290 // falls-through some constant times less then the other tests.
4291 MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
4295 return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
4298 /// ComputeExitLimit - Compute the number of times the backedge of the specified
4299 /// loop will execute if it exits via the specified block.
4300 ScalarEvolution::ExitLimit
4301 ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) {
4303 // Okay, we've chosen an exiting block. See what condition causes us to
4304 // exit at this block.
4306 // FIXME: we should be able to handle switch instructions (with a single exit)
4307 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
4308 if (ExitBr == 0) return getCouldNotCompute();
4309 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
4311 // At this point, we know we have a conditional branch that determines whether
4312 // the loop is exited. However, we don't know if the branch is executed each
4313 // time through the loop. If not, then the execution count of the branch will
4314 // not be equal to the trip count of the loop.
4316 // Currently we check for this by checking to see if the Exit branch goes to
4317 // the loop header. If so, we know it will always execute the same number of
4318 // times as the loop. We also handle the case where the exit block *is* the
4319 // loop header. This is common for un-rotated loops.
4321 // If both of those tests fail, walk up the unique predecessor chain to the
4322 // header, stopping if there is an edge that doesn't exit the loop. If the
4323 // header is reached, the execution count of the branch will be equal to the
4324 // trip count of the loop.
4326 // More extensive analysis could be done to handle more cases here.
4328 if (ExitBr->getSuccessor(0) != L->getHeader() &&
4329 ExitBr->getSuccessor(1) != L->getHeader() &&
4330 ExitBr->getParent() != L->getHeader()) {
4331 // The simple checks failed, try climbing the unique predecessor chain
4332 // up to the header.
4334 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
4335 BasicBlock *Pred = BB->getUniquePredecessor();
4337 return getCouldNotCompute();
4338 TerminatorInst *PredTerm = Pred->getTerminator();
4339 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
4340 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
4343 // If the predecessor has a successor that isn't BB and isn't
4344 // outside the loop, assume the worst.
4345 if (L->contains(PredSucc))
4346 return getCouldNotCompute();
4348 if (Pred == L->getHeader()) {
4355 return getCouldNotCompute();
4358 // Proceed to the next level to examine the exit condition expression.
4359 return ComputeExitLimitFromCond(L, ExitBr->getCondition(),
4360 ExitBr->getSuccessor(0),
4361 ExitBr->getSuccessor(1));
4364 /// ComputeExitLimitFromCond - Compute the number of times the
4365 /// backedge of the specified loop will execute if its exit condition
4366 /// were a conditional branch of ExitCond, TBB, and FBB.
4367 ScalarEvolution::ExitLimit
4368 ScalarEvolution::ComputeExitLimitFromCond(const Loop *L,
4372 // Check if the controlling expression for this loop is an And or Or.
4373 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
4374 if (BO->getOpcode() == Instruction::And) {
4375 // Recurse on the operands of the and.
4376 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB);
4377 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB);
4378 const SCEV *BECount = getCouldNotCompute();
4379 const SCEV *MaxBECount = getCouldNotCompute();
4380 if (L->contains(TBB)) {
4381 // Both conditions must be true for the loop to continue executing.
4382 // Choose the less conservative count.
4383 if (EL0.Exact == getCouldNotCompute() ||
4384 EL1.Exact == getCouldNotCompute())
4385 BECount = getCouldNotCompute();
4387 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4388 if (EL0.Max == getCouldNotCompute())
4389 MaxBECount = EL1.Max;
4390 else if (EL1.Max == getCouldNotCompute())
4391 MaxBECount = EL0.Max;
4393 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4395 // Both conditions must be true at the same time for the loop to exit.
4396 // For now, be conservative.
4397 assert(L->contains(FBB) && "Loop block has no successor in loop!");
4398 if (EL0.Max == EL1.Max)
4399 MaxBECount = EL0.Max;
4400 if (EL0.Exact == EL1.Exact)
4401 BECount = EL0.Exact;
4404 return ExitLimit(BECount, MaxBECount);
4406 if (BO->getOpcode() == Instruction::Or) {
4407 // Recurse on the operands of the or.
4408 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB);
4409 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB);
4410 const SCEV *BECount = getCouldNotCompute();
4411 const SCEV *MaxBECount = getCouldNotCompute();
4412 if (L->contains(FBB)) {
4413 // Both conditions must be false for the loop to continue executing.
4414 // Choose the less conservative count.
4415 if (EL0.Exact == getCouldNotCompute() ||
4416 EL1.Exact == getCouldNotCompute())
4417 BECount = getCouldNotCompute();
4419 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4420 if (EL0.Max == getCouldNotCompute())
4421 MaxBECount = EL1.Max;
4422 else if (EL1.Max == getCouldNotCompute())
4423 MaxBECount = EL0.Max;
4425 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4427 // Both conditions must be false at the same time for the loop to exit.
4428 // For now, be conservative.
4429 assert(L->contains(TBB) && "Loop block has no successor in loop!");
4430 if (EL0.Max == EL1.Max)
4431 MaxBECount = EL0.Max;
4432 if (EL0.Exact == EL1.Exact)
4433 BECount = EL0.Exact;
4436 return ExitLimit(BECount, MaxBECount);
4440 // With an icmp, it may be feasible to compute an exact backedge-taken count.
4441 // Proceed to the next level to examine the icmp.
4442 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
4443 return ComputeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB);
4445 // Check for a constant condition. These are normally stripped out by
4446 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
4447 // preserve the CFG and is temporarily leaving constant conditions
4449 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
4450 if (L->contains(FBB) == !CI->getZExtValue())
4451 // The backedge is always taken.
4452 return getCouldNotCompute();
4454 // The backedge is never taken.
4455 return getConstant(CI->getType(), 0);
4458 // If it's not an integer or pointer comparison then compute it the hard way.
4459 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4462 /// ComputeExitLimitFromICmp - Compute the number of times the
4463 /// backedge of the specified loop will execute if its exit condition
4464 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4465 ScalarEvolution::ExitLimit
4466 ScalarEvolution::ComputeExitLimitFromICmp(const Loop *L,
4471 // If the condition was exit on true, convert the condition to exit on false
4472 ICmpInst::Predicate Cond;
4473 if (!L->contains(FBB))
4474 Cond = ExitCond->getPredicate();
4476 Cond = ExitCond->getInversePredicate();
4478 // Handle common loops like: for (X = "string"; *X; ++X)
4479 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4480 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4482 ComputeLoadConstantCompareExitLimit(LI, RHS, L, Cond);
4483 if (ItCnt.hasAnyInfo())
4487 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4488 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4490 // Try to evaluate any dependencies out of the loop.
4491 LHS = getSCEVAtScope(LHS, L);
4492 RHS = getSCEVAtScope(RHS, L);
4494 // At this point, we would like to compute how many iterations of the
4495 // loop the predicate will return true for these inputs.
4496 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4497 // If there is a loop-invariant, force it into the RHS.
4498 std::swap(LHS, RHS);
4499 Cond = ICmpInst::getSwappedPredicate(Cond);
4502 // Simplify the operands before analyzing them.
4503 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4505 // If we have a comparison of a chrec against a constant, try to use value
4506 // ranges to answer this query.
4507 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4508 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4509 if (AddRec->getLoop() == L) {
4510 // Form the constant range.
4511 ConstantRange CompRange(
4512 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4514 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4515 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4519 case ICmpInst::ICMP_NE: { // while (X != Y)
4520 // Convert to: while (X-Y != 0)
4521 ExitLimit EL = HowFarToZero(getMinusSCEV(LHS, RHS), L);
4522 if (EL.hasAnyInfo()) return EL;
4525 case ICmpInst::ICMP_EQ: { // while (X == Y)
4526 // Convert to: while (X-Y == 0)
4527 ExitLimit EL = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4528 if (EL.hasAnyInfo()) return EL;
4531 case ICmpInst::ICMP_SLT: {
4532 ExitLimit EL = HowManyLessThans(LHS, RHS, L, true);
4533 if (EL.hasAnyInfo()) return EL;
4536 case ICmpInst::ICMP_SGT: {
4537 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4538 getNotSCEV(RHS), L, true);
4539 if (EL.hasAnyInfo()) return EL;
4542 case ICmpInst::ICMP_ULT: {
4543 ExitLimit EL = HowManyLessThans(LHS, RHS, L, false);
4544 if (EL.hasAnyInfo()) return EL;
4547 case ICmpInst::ICMP_UGT: {
4548 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4549 getNotSCEV(RHS), L, false);
4550 if (EL.hasAnyInfo()) return EL;
4555 dbgs() << "ComputeBackedgeTakenCount ";
4556 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4557 dbgs() << "[unsigned] ";
4558 dbgs() << *LHS << " "
4559 << Instruction::getOpcodeName(Instruction::ICmp)
4560 << " " << *RHS << "\n";
4564 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4567 static ConstantInt *
4568 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4569 ScalarEvolution &SE) {
4570 const SCEV *InVal = SE.getConstant(C);
4571 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4572 assert(isa<SCEVConstant>(Val) &&
4573 "Evaluation of SCEV at constant didn't fold correctly?");
4574 return cast<SCEVConstant>(Val)->getValue();
4577 /// ComputeLoadConstantCompareExitLimit - Given an exit condition of
4578 /// 'icmp op load X, cst', try to see if we can compute the backedge
4579 /// execution count.
4580 ScalarEvolution::ExitLimit
4581 ScalarEvolution::ComputeLoadConstantCompareExitLimit(
4585 ICmpInst::Predicate predicate) {
4587 if (LI->isVolatile()) return getCouldNotCompute();
4589 // Check to see if the loaded pointer is a getelementptr of a global.
4590 // TODO: Use SCEV instead of manually grubbing with GEPs.
4591 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4592 if (!GEP) return getCouldNotCompute();
4594 // Make sure that it is really a constant global we are gepping, with an
4595 // initializer, and make sure the first IDX is really 0.
4596 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4597 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4598 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4599 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4600 return getCouldNotCompute();
4602 // Okay, we allow one non-constant index into the GEP instruction.
4604 std::vector<Constant*> Indexes;
4605 unsigned VarIdxNum = 0;
4606 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4607 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4608 Indexes.push_back(CI);
4609 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4610 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4611 VarIdx = GEP->getOperand(i);
4613 Indexes.push_back(0);
4616 // Loop-invariant loads may be a byproduct of loop optimization. Skip them.
4618 return getCouldNotCompute();
4620 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4621 // Check to see if X is a loop variant variable value now.
4622 const SCEV *Idx = getSCEV(VarIdx);
4623 Idx = getSCEVAtScope(Idx, L);
4625 // We can only recognize very limited forms of loop index expressions, in
4626 // particular, only affine AddRec's like {C1,+,C2}.
4627 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4628 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4629 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4630 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4631 return getCouldNotCompute();
4633 unsigned MaxSteps = MaxBruteForceIterations;
4634 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4635 ConstantInt *ItCst = ConstantInt::get(
4636 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4637 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4639 // Form the GEP offset.
4640 Indexes[VarIdxNum] = Val;
4642 Constant *Result = ConstantFoldLoadThroughGEPIndices(GV->getInitializer(),
4644 if (Result == 0) break; // Cannot compute!
4646 // Evaluate the condition for this iteration.
4647 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4648 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4649 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4651 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4652 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4655 ++NumArrayLenItCounts;
4656 return getConstant(ItCst); // Found terminating iteration!
4659 return getCouldNotCompute();
4663 /// CanConstantFold - Return true if we can constant fold an instruction of the
4664 /// specified type, assuming that all operands were constants.
4665 static bool CanConstantFold(const Instruction *I) {
4666 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4667 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) ||
4671 if (const CallInst *CI = dyn_cast<CallInst>(I))
4672 if (const Function *F = CI->getCalledFunction())
4673 return canConstantFoldCallTo(F);
4677 /// Determine whether this instruction can constant evolve within this loop
4678 /// assuming its operands can all constant evolve.
4679 static bool canConstantEvolve(Instruction *I, const Loop *L) {
4680 // An instruction outside of the loop can't be derived from a loop PHI.
4681 if (!L->contains(I)) return false;
4683 if (isa<PHINode>(I)) {
4684 if (L->getHeader() == I->getParent())
4687 // We don't currently keep track of the control flow needed to evaluate
4688 // PHIs, so we cannot handle PHIs inside of loops.
4692 // If we won't be able to constant fold this expression even if the operands
4693 // are constants, bail early.
4694 return CanConstantFold(I);
4697 /// getConstantEvolvingPHIOperands - Implement getConstantEvolvingPHI by
4698 /// recursing through each instruction operand until reaching a loop header phi.
4700 getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L,
4701 DenseMap<Instruction *, PHINode *> &PHIMap) {
4703 // Otherwise, we can evaluate this instruction if all of its operands are
4704 // constant or derived from a PHI node themselves.
4706 for (Instruction::op_iterator OpI = UseInst->op_begin(),
4707 OpE = UseInst->op_end(); OpI != OpE; ++OpI) {
4709 if (isa<Constant>(*OpI)) continue;
4711 Instruction *OpInst = dyn_cast<Instruction>(*OpI);
4712 if (!OpInst || !canConstantEvolve(OpInst, L)) return 0;
4714 PHINode *P = dyn_cast<PHINode>(OpInst);
4716 // If this operand is already visited, reuse the prior result.
4717 // We may have P != PHI if this is the deepest point at which the
4718 // inconsistent paths meet.
4719 P = PHIMap.lookup(OpInst);
4721 // Recurse and memoize the results, whether a phi is found or not.
4722 // This recursive call invalidates pointers into PHIMap.
4723 P = getConstantEvolvingPHIOperands(OpInst, L, PHIMap);
4726 if (P == 0) return 0; // Not evolving from PHI
4727 if (PHI && PHI != P) return 0; // Evolving from multiple different PHIs.
4730 // This is a expression evolving from a constant PHI!
4734 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4735 /// in the loop that V is derived from. We allow arbitrary operations along the
4736 /// way, but the operands of an operation must either be constants or a value
4737 /// derived from a constant PHI. If this expression does not fit with these
4738 /// constraints, return null.
4739 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4740 Instruction *I = dyn_cast<Instruction>(V);
4741 if (I == 0 || !canConstantEvolve(I, L)) return 0;
4743 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4747 // Record non-constant instructions contained by the loop.
4748 DenseMap<Instruction *, PHINode *> PHIMap;
4749 return getConstantEvolvingPHIOperands(I, L, PHIMap);
4752 /// EvaluateExpression - Given an expression that passes the
4753 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4754 /// in the loop has the value PHIVal. If we can't fold this expression for some
4755 /// reason, return null.
4756 static Constant *EvaluateExpression(Value *V, const Loop *L,
4757 DenseMap<Instruction *, Constant *> &Vals,
4758 const TargetData *TD,
4759 const TargetLibraryInfo *TLI) {
4760 // Convenient constant check, but redundant for recursive calls.
4761 if (Constant *C = dyn_cast<Constant>(V)) return C;
4762 Instruction *I = dyn_cast<Instruction>(V);
4765 if (Constant *C = Vals.lookup(I)) return C;
4767 // An instruction inside the loop depends on a value outside the loop that we
4768 // weren't given a mapping for, or a value such as a call inside the loop.
4769 if (!canConstantEvolve(I, L)) return 0;
4771 // An unmapped PHI can be due to a branch or another loop inside this loop,
4772 // or due to this not being the initial iteration through a loop where we
4773 // couldn't compute the evolution of this particular PHI last time.
4774 if (isa<PHINode>(I)) return 0;
4776 std::vector<Constant*> Operands(I->getNumOperands());
4778 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4779 Instruction *Operand = dyn_cast<Instruction>(I->getOperand(i));
4781 Operands[i] = dyn_cast<Constant>(I->getOperand(i));
4782 if (!Operands[i]) return 0;
4785 Constant *C = EvaluateExpression(Operand, L, Vals, TD, TLI);
4791 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4792 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4793 Operands[1], TD, TLI);
4794 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4795 if (!LI->isVolatile())
4796 return ConstantFoldLoadFromConstPtr(Operands[0], TD);
4798 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Operands, TD,
4802 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4803 /// in the header of its containing loop, we know the loop executes a
4804 /// constant number of times, and the PHI node is just a recurrence
4805 /// involving constants, fold it.
4807 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4810 DenseMap<PHINode*, Constant*>::const_iterator I =
4811 ConstantEvolutionLoopExitValue.find(PN);
4812 if (I != ConstantEvolutionLoopExitValue.end())
4815 if (BEs.ugt(MaxBruteForceIterations))
4816 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4818 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4820 DenseMap<Instruction *, Constant *> CurrentIterVals;
4821 BasicBlock *Header = L->getHeader();
4822 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
4824 // Since the loop is canonicalized, the PHI node must have two entries. One
4825 // entry must be a constant (coming in from outside of the loop), and the
4826 // second must be derived from the same PHI.
4827 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4829 for (BasicBlock::iterator I = Header->begin();
4830 (PHI = dyn_cast<PHINode>(I)); ++I) {
4831 Constant *StartCST =
4832 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
4833 if (StartCST == 0) continue;
4834 CurrentIterVals[PHI] = StartCST;
4836 if (!CurrentIterVals.count(PN))
4839 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4841 // Execute the loop symbolically to determine the exit value.
4842 if (BEs.getActiveBits() >= 32)
4843 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4845 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4846 unsigned IterationNum = 0;
4847 for (; ; ++IterationNum) {
4848 if (IterationNum == NumIterations)
4849 return RetVal = CurrentIterVals[PN]; // Got exit value!
4851 // Compute the value of the PHIs for the next iteration.
4852 // EvaluateExpression adds non-phi values to the CurrentIterVals map.
4853 DenseMap<Instruction *, Constant *> NextIterVals;
4854 Constant *NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD,
4857 return 0; // Couldn't evaluate!
4858 NextIterVals[PN] = NextPHI;
4860 bool StoppedEvolving = NextPHI == CurrentIterVals[PN];
4862 // Also evaluate the other PHI nodes. However, we don't get to stop if we
4863 // cease to be able to evaluate one of them or if they stop evolving,
4864 // because that doesn't necessarily prevent us from computing PN.
4865 SmallVector<std::pair<PHINode *, Constant *>, 8> PHIsToCompute;
4866 for (DenseMap<Instruction *, Constant *>::const_iterator
4867 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
4868 PHINode *PHI = dyn_cast<PHINode>(I->first);
4869 if (!PHI || PHI == PN || PHI->getParent() != Header) continue;
4870 PHIsToCompute.push_back(std::make_pair(PHI, I->second));
4872 // We use two distinct loops because EvaluateExpression may invalidate any
4873 // iterators into CurrentIterVals.
4874 for (SmallVectorImpl<std::pair<PHINode *, Constant*> >::const_iterator
4875 I = PHIsToCompute.begin(), E = PHIsToCompute.end(); I != E; ++I) {
4876 PHINode *PHI = I->first;
4877 Constant *&NextPHI = NextIterVals[PHI];
4878 if (!NextPHI) { // Not already computed.
4879 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
4880 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
4882 if (NextPHI != I->second)
4883 StoppedEvolving = false;
4886 // If all entries in CurrentIterVals == NextIterVals then we can stop
4887 // iterating, the loop can't continue to change.
4888 if (StoppedEvolving)
4889 return RetVal = CurrentIterVals[PN];
4891 CurrentIterVals.swap(NextIterVals);
4895 /// ComputeExitCountExhaustively - If the loop is known to execute a
4896 /// constant number of times (the condition evolves only from constants),
4897 /// try to evaluate a few iterations of the loop until we get the exit
4898 /// condition gets a value of ExitWhen (true or false). If we cannot
4899 /// evaluate the trip count of the loop, return getCouldNotCompute().
4900 const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L,
4903 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4904 if (PN == 0) return getCouldNotCompute();
4906 // If the loop is canonicalized, the PHI will have exactly two entries.
4907 // That's the only form we support here.
4908 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4910 DenseMap<Instruction *, Constant *> CurrentIterVals;
4911 BasicBlock *Header = L->getHeader();
4912 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
4914 // One entry must be a constant (coming in from outside of the loop), and the
4915 // second must be derived from the same PHI.
4916 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4918 for (BasicBlock::iterator I = Header->begin();
4919 (PHI = dyn_cast<PHINode>(I)); ++I) {
4920 Constant *StartCST =
4921 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
4922 if (StartCST == 0) continue;
4923 CurrentIterVals[PHI] = StartCST;
4925 if (!CurrentIterVals.count(PN))
4926 return getCouldNotCompute();
4928 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4929 // the loop symbolically to determine when the condition gets a value of
4932 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4933 for (unsigned IterationNum = 0; IterationNum != MaxIterations;++IterationNum){
4934 ConstantInt *CondVal =
4935 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, L, CurrentIterVals,
4938 // Couldn't symbolically evaluate.
4939 if (!CondVal) return getCouldNotCompute();
4941 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4942 ++NumBruteForceTripCountsComputed;
4943 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4946 // Update all the PHI nodes for the next iteration.
4947 DenseMap<Instruction *, Constant *> NextIterVals;
4949 // Create a list of which PHIs we need to compute. We want to do this before
4950 // calling EvaluateExpression on them because that may invalidate iterators
4951 // into CurrentIterVals.
4952 SmallVector<PHINode *, 8> PHIsToCompute;
4953 for (DenseMap<Instruction *, Constant *>::const_iterator
4954 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
4955 PHINode *PHI = dyn_cast<PHINode>(I->first);
4956 if (!PHI || PHI->getParent() != Header) continue;
4957 PHIsToCompute.push_back(PHI);
4959 for (SmallVectorImpl<PHINode *>::const_iterator I = PHIsToCompute.begin(),
4960 E = PHIsToCompute.end(); I != E; ++I) {
4962 Constant *&NextPHI = NextIterVals[PHI];
4963 if (NextPHI) continue; // Already computed!
4965 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
4966 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
4968 CurrentIterVals.swap(NextIterVals);
4971 // Too many iterations were needed to evaluate.
4972 return getCouldNotCompute();
4975 /// getSCEVAtScope - Return a SCEV expression for the specified value
4976 /// at the specified scope in the program. The L value specifies a loop
4977 /// nest to evaluate the expression at, where null is the top-level or a
4978 /// specified loop is immediately inside of the loop.
4980 /// This method can be used to compute the exit value for a variable defined
4981 /// in a loop by querying what the value will hold in the parent loop.
4983 /// In the case that a relevant loop exit value cannot be computed, the
4984 /// original value V is returned.
4985 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4986 // Check to see if we've folded this expression at this loop before.
4987 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4988 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4989 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4991 return Pair.first->second ? Pair.first->second : V;
4993 // Otherwise compute it.
4994 const SCEV *C = computeSCEVAtScope(V, L);
4995 ValuesAtScopes[V][L] = C;
4999 /// This builds up a Constant using the ConstantExpr interface. That way, we
5000 /// will return Constants for objects which aren't represented by a
5001 /// SCEVConstant, because SCEVConstant is restricted to ConstantInt.
5002 /// Returns NULL if the SCEV isn't representable as a Constant.
5003 static Constant *BuildConstantFromSCEV(const SCEV *V) {
5004 switch (V->getSCEVType()) {
5005 default: // TODO: smax, umax.
5006 case scCouldNotCompute:
5010 return cast<SCEVConstant>(V)->getValue();
5012 return dyn_cast<Constant>(cast<SCEVUnknown>(V)->getValue());
5013 case scSignExtend: {
5014 const SCEVSignExtendExpr *SS = cast<SCEVSignExtendExpr>(V);
5015 if (Constant *CastOp = BuildConstantFromSCEV(SS->getOperand()))
5016 return ConstantExpr::getSExt(CastOp, SS->getType());
5019 case scZeroExtend: {
5020 const SCEVZeroExtendExpr *SZ = cast<SCEVZeroExtendExpr>(V);
5021 if (Constant *CastOp = BuildConstantFromSCEV(SZ->getOperand()))
5022 return ConstantExpr::getZExt(CastOp, SZ->getType());
5026 const SCEVTruncateExpr *ST = cast<SCEVTruncateExpr>(V);
5027 if (Constant *CastOp = BuildConstantFromSCEV(ST->getOperand()))
5028 return ConstantExpr::getTrunc(CastOp, ST->getType());
5032 const SCEVAddExpr *SA = cast<SCEVAddExpr>(V);
5033 if (Constant *C = BuildConstantFromSCEV(SA->getOperand(0))) {
5034 if (C->getType()->isPointerTy())
5035 C = ConstantExpr::getBitCast(C, Type::getInt8PtrTy(C->getContext()));
5036 for (unsigned i = 1, e = SA->getNumOperands(); i != e; ++i) {
5037 Constant *C2 = BuildConstantFromSCEV(SA->getOperand(i));
5041 if (!C->getType()->isPointerTy() && C2->getType()->isPointerTy()) {
5043 // The offsets have been converted to bytes. We can add bytes to an
5044 // i8* by GEP with the byte count in the first index.
5045 C = ConstantExpr::getBitCast(C,Type::getInt8PtrTy(C->getContext()));
5048 // Don't bother trying to sum two pointers. We probably can't
5049 // statically compute a load that results from it anyway.
5050 if (C2->getType()->isPointerTy())
5053 if (C->getType()->isPointerTy()) {
5054 if (cast<PointerType>(C->getType())->getElementType()->isStructTy())
5055 C2 = ConstantExpr::getIntegerCast(
5056 C2, Type::getInt32Ty(C->getContext()), true);
5057 C = ConstantExpr::getGetElementPtr(C, C2);
5059 C = ConstantExpr::getAdd(C, C2);
5066 const SCEVMulExpr *SM = cast<SCEVMulExpr>(V);
5067 if (Constant *C = BuildConstantFromSCEV(SM->getOperand(0))) {
5068 // Don't bother with pointers at all.
5069 if (C->getType()->isPointerTy()) return 0;
5070 for (unsigned i = 1, e = SM->getNumOperands(); i != e; ++i) {
5071 Constant *C2 = BuildConstantFromSCEV(SM->getOperand(i));
5072 if (!C2 || C2->getType()->isPointerTy()) return 0;
5073 C = ConstantExpr::getMul(C, C2);
5080 const SCEVUDivExpr *SU = cast<SCEVUDivExpr>(V);
5081 if (Constant *LHS = BuildConstantFromSCEV(SU->getLHS()))
5082 if (Constant *RHS = BuildConstantFromSCEV(SU->getRHS()))
5083 if (LHS->getType() == RHS->getType())
5084 return ConstantExpr::getUDiv(LHS, RHS);
5091 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
5092 if (isa<SCEVConstant>(V)) return V;
5094 // If this instruction is evolved from a constant-evolving PHI, compute the
5095 // exit value from the loop without using SCEVs.
5096 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
5097 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
5098 const Loop *LI = (*this->LI)[I->getParent()];
5099 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
5100 if (PHINode *PN = dyn_cast<PHINode>(I))
5101 if (PN->getParent() == LI->getHeader()) {
5102 // Okay, there is no closed form solution for the PHI node. Check
5103 // to see if the loop that contains it has a known backedge-taken
5104 // count. If so, we may be able to force computation of the exit
5106 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
5107 if (const SCEVConstant *BTCC =
5108 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
5109 // Okay, we know how many times the containing loop executes. If
5110 // this is a constant evolving PHI node, get the final value at
5111 // the specified iteration number.
5112 Constant *RV = getConstantEvolutionLoopExitValue(PN,
5113 BTCC->getValue()->getValue(),
5115 if (RV) return getSCEV(RV);
5119 // Okay, this is an expression that we cannot symbolically evaluate
5120 // into a SCEV. Check to see if it's possible to symbolically evaluate
5121 // the arguments into constants, and if so, try to constant propagate the
5122 // result. This is particularly useful for computing loop exit values.
5123 if (CanConstantFold(I)) {
5124 SmallVector<Constant *, 4> Operands;
5125 bool MadeImprovement = false;
5126 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
5127 Value *Op = I->getOperand(i);
5128 if (Constant *C = dyn_cast<Constant>(Op)) {
5129 Operands.push_back(C);
5133 // If any of the operands is non-constant and if they are
5134 // non-integer and non-pointer, don't even try to analyze them
5135 // with scev techniques.
5136 if (!isSCEVable(Op->getType()))
5139 const SCEV *OrigV = getSCEV(Op);
5140 const SCEV *OpV = getSCEVAtScope(OrigV, L);
5141 MadeImprovement |= OrigV != OpV;
5143 Constant *C = BuildConstantFromSCEV(OpV);
5145 if (C->getType() != Op->getType())
5146 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5150 Operands.push_back(C);
5153 // Check to see if getSCEVAtScope actually made an improvement.
5154 if (MadeImprovement) {
5156 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
5157 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
5158 Operands[0], Operands[1], TD,
5160 else if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
5161 if (!LI->isVolatile())
5162 C = ConstantFoldLoadFromConstPtr(Operands[0], TD);
5164 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
5172 // This is some other type of SCEVUnknown, just return it.
5176 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
5177 // Avoid performing the look-up in the common case where the specified
5178 // expression has no loop-variant portions.
5179 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
5180 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5181 if (OpAtScope != Comm->getOperand(i)) {
5182 // Okay, at least one of these operands is loop variant but might be
5183 // foldable. Build a new instance of the folded commutative expression.
5184 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
5185 Comm->op_begin()+i);
5186 NewOps.push_back(OpAtScope);
5188 for (++i; i != e; ++i) {
5189 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5190 NewOps.push_back(OpAtScope);
5192 if (isa<SCEVAddExpr>(Comm))
5193 return getAddExpr(NewOps);
5194 if (isa<SCEVMulExpr>(Comm))
5195 return getMulExpr(NewOps);
5196 if (isa<SCEVSMaxExpr>(Comm))
5197 return getSMaxExpr(NewOps);
5198 if (isa<SCEVUMaxExpr>(Comm))
5199 return getUMaxExpr(NewOps);
5200 llvm_unreachable("Unknown commutative SCEV type!");
5203 // If we got here, all operands are loop invariant.
5207 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
5208 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
5209 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
5210 if (LHS == Div->getLHS() && RHS == Div->getRHS())
5211 return Div; // must be loop invariant
5212 return getUDivExpr(LHS, RHS);
5215 // If this is a loop recurrence for a loop that does not contain L, then we
5216 // are dealing with the final value computed by the loop.
5217 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
5218 // First, attempt to evaluate each operand.
5219 // Avoid performing the look-up in the common case where the specified
5220 // expression has no loop-variant portions.
5221 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
5222 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
5223 if (OpAtScope == AddRec->getOperand(i))
5226 // Okay, at least one of these operands is loop variant but might be
5227 // foldable. Build a new instance of the folded commutative expression.
5228 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
5229 AddRec->op_begin()+i);
5230 NewOps.push_back(OpAtScope);
5231 for (++i; i != e; ++i)
5232 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
5234 const SCEV *FoldedRec =
5235 getAddRecExpr(NewOps, AddRec->getLoop(),
5236 AddRec->getNoWrapFlags(SCEV::FlagNW));
5237 AddRec = dyn_cast<SCEVAddRecExpr>(FoldedRec);
5238 // The addrec may be folded to a nonrecurrence, for example, if the
5239 // induction variable is multiplied by zero after constant folding. Go
5240 // ahead and return the folded value.
5246 // If the scope is outside the addrec's loop, evaluate it by using the
5247 // loop exit value of the addrec.
5248 if (!AddRec->getLoop()->contains(L)) {
5249 // To evaluate this recurrence, we need to know how many times the AddRec
5250 // loop iterates. Compute this now.
5251 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
5252 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
5254 // Then, evaluate the AddRec.
5255 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
5261 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
5262 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5263 if (Op == Cast->getOperand())
5264 return Cast; // must be loop invariant
5265 return getZeroExtendExpr(Op, Cast->getType());
5268 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
5269 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5270 if (Op == Cast->getOperand())
5271 return Cast; // must be loop invariant
5272 return getSignExtendExpr(Op, Cast->getType());
5275 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
5276 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5277 if (Op == Cast->getOperand())
5278 return Cast; // must be loop invariant
5279 return getTruncateExpr(Op, Cast->getType());
5282 llvm_unreachable("Unknown SCEV type!");
5285 /// getSCEVAtScope - This is a convenience function which does
5286 /// getSCEVAtScope(getSCEV(V), L).
5287 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
5288 return getSCEVAtScope(getSCEV(V), L);
5291 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
5292 /// following equation:
5294 /// A * X = B (mod N)
5296 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
5297 /// A and B isn't important.
5299 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
5300 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
5301 ScalarEvolution &SE) {
5302 uint32_t BW = A.getBitWidth();
5303 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
5304 assert(A != 0 && "A must be non-zero.");
5308 // The gcd of A and N may have only one prime factor: 2. The number of
5309 // trailing zeros in A is its multiplicity
5310 uint32_t Mult2 = A.countTrailingZeros();
5313 // 2. Check if B is divisible by D.
5315 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
5316 // is not less than multiplicity of this prime factor for D.
5317 if (B.countTrailingZeros() < Mult2)
5318 return SE.getCouldNotCompute();
5320 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
5323 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
5324 // bit width during computations.
5325 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
5326 APInt Mod(BW + 1, 0);
5327 Mod.setBit(BW - Mult2); // Mod = N / D
5328 APInt I = AD.multiplicativeInverse(Mod);
5330 // 4. Compute the minimum unsigned root of the equation:
5331 // I * (B / D) mod (N / D)
5332 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
5334 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
5336 return SE.getConstant(Result.trunc(BW));
5339 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
5340 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
5341 /// might be the same) or two SCEVCouldNotCompute objects.
5343 static std::pair<const SCEV *,const SCEV *>
5344 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
5345 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
5346 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
5347 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
5348 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
5350 // We currently can only solve this if the coefficients are constants.
5351 if (!LC || !MC || !NC) {
5352 const SCEV *CNC = SE.getCouldNotCompute();
5353 return std::make_pair(CNC, CNC);
5356 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
5357 const APInt &L = LC->getValue()->getValue();
5358 const APInt &M = MC->getValue()->getValue();
5359 const APInt &N = NC->getValue()->getValue();
5360 APInt Two(BitWidth, 2);
5361 APInt Four(BitWidth, 4);
5364 using namespace APIntOps;
5366 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
5367 // The B coefficient is M-N/2
5371 // The A coefficient is N/2
5372 APInt A(N.sdiv(Two));
5374 // Compute the B^2-4ac term.
5377 SqrtTerm -= Four * (A * C);
5379 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
5380 // integer value or else APInt::sqrt() will assert.
5381 APInt SqrtVal(SqrtTerm.sqrt());
5383 // Compute the two solutions for the quadratic formula.
5384 // The divisions must be performed as signed divisions.
5387 if (TwoA.isMinValue()) {
5388 const SCEV *CNC = SE.getCouldNotCompute();
5389 return std::make_pair(CNC, CNC);
5392 LLVMContext &Context = SE.getContext();
5394 ConstantInt *Solution1 =
5395 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
5396 ConstantInt *Solution2 =
5397 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
5399 return std::make_pair(SE.getConstant(Solution1),
5400 SE.getConstant(Solution2));
5401 } // end APIntOps namespace
5404 /// HowFarToZero - Return the number of times a backedge comparing the specified
5405 /// value to zero will execute. If not computable, return CouldNotCompute.
5407 /// This is only used for loops with a "x != y" exit test. The exit condition is
5408 /// now expressed as a single expression, V = x-y. So the exit test is
5409 /// effectively V != 0. We know and take advantage of the fact that this
5410 /// expression only being used in a comparison by zero context.
5411 ScalarEvolution::ExitLimit
5412 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
5413 // If the value is a constant
5414 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5415 // If the value is already zero, the branch will execute zero times.
5416 if (C->getValue()->isZero()) return C;
5417 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5420 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
5421 if (!AddRec || AddRec->getLoop() != L)
5422 return getCouldNotCompute();
5424 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
5425 // the quadratic equation to solve it.
5426 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
5427 std::pair<const SCEV *,const SCEV *> Roots =
5428 SolveQuadraticEquation(AddRec, *this);
5429 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5430 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5433 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
5434 << " sol#2: " << *R2 << "\n";
5436 // Pick the smallest positive root value.
5437 if (ConstantInt *CB =
5438 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
5441 if (CB->getZExtValue() == false)
5442 std::swap(R1, R2); // R1 is the minimum root now.
5444 // We can only use this value if the chrec ends up with an exact zero
5445 // value at this index. When solving for "X*X != 5", for example, we
5446 // should not accept a root of 2.
5447 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
5449 return R1; // We found a quadratic root!
5452 return getCouldNotCompute();
5455 // Otherwise we can only handle this if it is affine.
5456 if (!AddRec->isAffine())
5457 return getCouldNotCompute();
5459 // If this is an affine expression, the execution count of this branch is
5460 // the minimum unsigned root of the following equation:
5462 // Start + Step*N = 0 (mod 2^BW)
5466 // Step*N = -Start (mod 2^BW)
5468 // where BW is the common bit width of Start and Step.
5470 // Get the initial value for the loop.
5471 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
5472 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
5474 // For now we handle only constant steps.
5476 // TODO: Handle a nonconstant Step given AddRec<NUW>. If the
5477 // AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap
5478 // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step.
5479 // We have not yet seen any such cases.
5480 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
5482 return getCouldNotCompute();
5484 // For positive steps (counting up until unsigned overflow):
5485 // N = -Start/Step (as unsigned)
5486 // For negative steps (counting down to zero):
5488 // First compute the unsigned distance from zero in the direction of Step.
5489 bool CountDown = StepC->getValue()->getValue().isNegative();
5490 const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
5492 // Handle unitary steps, which cannot wraparound.
5493 // 1*N = -Start; -1*N = Start (mod 2^BW), so:
5494 // N = Distance (as unsigned)
5495 if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) {
5496 ConstantRange CR = getUnsignedRange(Start);
5497 const SCEV *MaxBECount;
5498 if (!CountDown && CR.getUnsignedMin().isMinValue())
5499 // When counting up, the worst starting value is 1, not 0.
5500 MaxBECount = CR.getUnsignedMax().isMinValue()
5501 ? getConstant(APInt::getMinValue(CR.getBitWidth()))
5502 : getConstant(APInt::getMaxValue(CR.getBitWidth()));
5504 MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
5505 : -CR.getUnsignedMin());
5506 return ExitLimit(Distance, MaxBECount);
5509 // If the recurrence is known not to wraparound, unsigned divide computes the
5510 // back edge count. We know that the value will either become zero (and thus
5511 // the loop terminates), that the loop will terminate through some other exit
5512 // condition first, or that the loop has undefined behavior. This means
5513 // we can't "miss" the exit value, even with nonunit stride.
5515 // FIXME: Prove that loops always exhibits *acceptable* undefined
5516 // behavior. Loops must exhibit defined behavior until a wrapped value is
5517 // actually used. So the trip count computed by udiv could be smaller than the
5518 // number of well-defined iterations.
5519 if (AddRec->getNoWrapFlags(SCEV::FlagNW)) {
5520 // FIXME: We really want an "isexact" bit for udiv.
5521 return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
5523 // Then, try to solve the above equation provided that Start is constant.
5524 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
5525 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
5526 -StartC->getValue()->getValue(),
5528 return getCouldNotCompute();
5531 /// HowFarToNonZero - Return the number of times a backedge checking the
5532 /// specified value for nonzero will execute. If not computable, return
5534 ScalarEvolution::ExitLimit
5535 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
5536 // Loops that look like: while (X == 0) are very strange indeed. We don't
5537 // handle them yet except for the trivial case. This could be expanded in the
5538 // future as needed.
5540 // If the value is a constant, check to see if it is known to be non-zero
5541 // already. If so, the backedge will execute zero times.
5542 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5543 if (!C->getValue()->isNullValue())
5544 return getConstant(C->getType(), 0);
5545 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5548 // We could implement others, but I really doubt anyone writes loops like
5549 // this, and if they did, they would already be constant folded.
5550 return getCouldNotCompute();
5553 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
5554 /// (which may not be an immediate predecessor) which has exactly one
5555 /// successor from which BB is reachable, or null if no such block is
5558 std::pair<BasicBlock *, BasicBlock *>
5559 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
5560 // If the block has a unique predecessor, then there is no path from the
5561 // predecessor to the block that does not go through the direct edge
5562 // from the predecessor to the block.
5563 if (BasicBlock *Pred = BB->getSinglePredecessor())
5564 return std::make_pair(Pred, BB);
5566 // A loop's header is defined to be a block that dominates the loop.
5567 // If the header has a unique predecessor outside the loop, it must be
5568 // a block that has exactly one successor that can reach the loop.
5569 if (Loop *L = LI->getLoopFor(BB))
5570 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
5572 return std::pair<BasicBlock *, BasicBlock *>();
5575 /// HasSameValue - SCEV structural equivalence is usually sufficient for
5576 /// testing whether two expressions are equal, however for the purposes of
5577 /// looking for a condition guarding a loop, it can be useful to be a little
5578 /// more general, since a front-end may have replicated the controlling
5581 static bool HasSameValue(const SCEV *A, const SCEV *B) {
5582 // Quick check to see if they are the same SCEV.
5583 if (A == B) return true;
5585 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
5586 // two different instructions with the same value. Check for this case.
5587 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
5588 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
5589 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
5590 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
5591 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
5594 // Otherwise assume they may have a different value.
5598 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
5599 /// predicate Pred. Return true iff any changes were made.
5601 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
5602 const SCEV *&LHS, const SCEV *&RHS) {
5603 bool Changed = false;
5605 // Canonicalize a constant to the right side.
5606 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
5607 // Check for both operands constant.
5608 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
5609 if (ConstantExpr::getICmp(Pred,
5611 RHSC->getValue())->isNullValue())
5612 goto trivially_false;
5614 goto trivially_true;
5616 // Otherwise swap the operands to put the constant on the right.
5617 std::swap(LHS, RHS);
5618 Pred = ICmpInst::getSwappedPredicate(Pred);
5622 // If we're comparing an addrec with a value which is loop-invariant in the
5623 // addrec's loop, put the addrec on the left. Also make a dominance check,
5624 // as both operands could be addrecs loop-invariant in each other's loop.
5625 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
5626 const Loop *L = AR->getLoop();
5627 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
5628 std::swap(LHS, RHS);
5629 Pred = ICmpInst::getSwappedPredicate(Pred);
5634 // If there's a constant operand, canonicalize comparisons with boundary
5635 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
5636 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
5637 const APInt &RA = RC->getValue()->getValue();
5639 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5640 case ICmpInst::ICMP_EQ:
5641 case ICmpInst::ICMP_NE:
5643 case ICmpInst::ICMP_UGE:
5644 if ((RA - 1).isMinValue()) {
5645 Pred = ICmpInst::ICMP_NE;
5646 RHS = getConstant(RA - 1);
5650 if (RA.isMaxValue()) {
5651 Pred = ICmpInst::ICMP_EQ;
5655 if (RA.isMinValue()) goto trivially_true;
5657 Pred = ICmpInst::ICMP_UGT;
5658 RHS = getConstant(RA - 1);
5661 case ICmpInst::ICMP_ULE:
5662 if ((RA + 1).isMaxValue()) {
5663 Pred = ICmpInst::ICMP_NE;
5664 RHS = getConstant(RA + 1);
5668 if (RA.isMinValue()) {
5669 Pred = ICmpInst::ICMP_EQ;
5673 if (RA.isMaxValue()) goto trivially_true;
5675 Pred = ICmpInst::ICMP_ULT;
5676 RHS = getConstant(RA + 1);
5679 case ICmpInst::ICMP_SGE:
5680 if ((RA - 1).isMinSignedValue()) {
5681 Pred = ICmpInst::ICMP_NE;
5682 RHS = getConstant(RA - 1);
5686 if (RA.isMaxSignedValue()) {
5687 Pred = ICmpInst::ICMP_EQ;
5691 if (RA.isMinSignedValue()) goto trivially_true;
5693 Pred = ICmpInst::ICMP_SGT;
5694 RHS = getConstant(RA - 1);
5697 case ICmpInst::ICMP_SLE:
5698 if ((RA + 1).isMaxSignedValue()) {
5699 Pred = ICmpInst::ICMP_NE;
5700 RHS = getConstant(RA + 1);
5704 if (RA.isMinSignedValue()) {
5705 Pred = ICmpInst::ICMP_EQ;
5709 if (RA.isMaxSignedValue()) goto trivially_true;
5711 Pred = ICmpInst::ICMP_SLT;
5712 RHS = getConstant(RA + 1);
5715 case ICmpInst::ICMP_UGT:
5716 if (RA.isMinValue()) {
5717 Pred = ICmpInst::ICMP_NE;
5721 if ((RA + 1).isMaxValue()) {
5722 Pred = ICmpInst::ICMP_EQ;
5723 RHS = getConstant(RA + 1);
5727 if (RA.isMaxValue()) goto trivially_false;
5729 case ICmpInst::ICMP_ULT:
5730 if (RA.isMaxValue()) {
5731 Pred = ICmpInst::ICMP_NE;
5735 if ((RA - 1).isMinValue()) {
5736 Pred = ICmpInst::ICMP_EQ;
5737 RHS = getConstant(RA - 1);
5741 if (RA.isMinValue()) goto trivially_false;
5743 case ICmpInst::ICMP_SGT:
5744 if (RA.isMinSignedValue()) {
5745 Pred = ICmpInst::ICMP_NE;
5749 if ((RA + 1).isMaxSignedValue()) {
5750 Pred = ICmpInst::ICMP_EQ;
5751 RHS = getConstant(RA + 1);
5755 if (RA.isMaxSignedValue()) goto trivially_false;
5757 case ICmpInst::ICMP_SLT:
5758 if (RA.isMaxSignedValue()) {
5759 Pred = ICmpInst::ICMP_NE;
5763 if ((RA - 1).isMinSignedValue()) {
5764 Pred = ICmpInst::ICMP_EQ;
5765 RHS = getConstant(RA - 1);
5769 if (RA.isMinSignedValue()) goto trivially_false;
5774 // Check for obvious equality.
5775 if (HasSameValue(LHS, RHS)) {
5776 if (ICmpInst::isTrueWhenEqual(Pred))
5777 goto trivially_true;
5778 if (ICmpInst::isFalseWhenEqual(Pred))
5779 goto trivially_false;
5782 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5783 // adding or subtracting 1 from one of the operands.
5785 case ICmpInst::ICMP_SLE:
5786 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5787 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5789 Pred = ICmpInst::ICMP_SLT;
5791 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5792 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5794 Pred = ICmpInst::ICMP_SLT;
5798 case ICmpInst::ICMP_SGE:
5799 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5800 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5802 Pred = ICmpInst::ICMP_SGT;
5804 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5805 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5807 Pred = ICmpInst::ICMP_SGT;
5811 case ICmpInst::ICMP_ULE:
5812 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5813 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5815 Pred = ICmpInst::ICMP_ULT;
5817 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5818 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5820 Pred = ICmpInst::ICMP_ULT;
5824 case ICmpInst::ICMP_UGE:
5825 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5826 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5828 Pred = ICmpInst::ICMP_UGT;
5830 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5831 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5833 Pred = ICmpInst::ICMP_UGT;
5841 // TODO: More simplifications are possible here.
5847 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5848 Pred = ICmpInst::ICMP_EQ;
5853 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5854 Pred = ICmpInst::ICMP_NE;
5858 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5859 return getSignedRange(S).getSignedMax().isNegative();
5862 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5863 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5866 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5867 return !getSignedRange(S).getSignedMin().isNegative();
5870 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5871 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5874 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5875 return isKnownNegative(S) || isKnownPositive(S);
5878 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5879 const SCEV *LHS, const SCEV *RHS) {
5880 // Canonicalize the inputs first.
5881 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5883 // If LHS or RHS is an addrec, check to see if the condition is true in
5884 // every iteration of the loop.
5885 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5886 if (isLoopEntryGuardedByCond(
5887 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5888 isLoopBackedgeGuardedByCond(
5889 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5891 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5892 if (isLoopEntryGuardedByCond(
5893 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5894 isLoopBackedgeGuardedByCond(
5895 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5898 // Otherwise see what can be done with known constant ranges.
5899 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5903 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5904 const SCEV *LHS, const SCEV *RHS) {
5905 if (HasSameValue(LHS, RHS))
5906 return ICmpInst::isTrueWhenEqual(Pred);
5908 // This code is split out from isKnownPredicate because it is called from
5909 // within isLoopEntryGuardedByCond.
5912 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5913 case ICmpInst::ICMP_SGT:
5914 Pred = ICmpInst::ICMP_SLT;
5915 std::swap(LHS, RHS);
5916 case ICmpInst::ICMP_SLT: {
5917 ConstantRange LHSRange = getSignedRange(LHS);
5918 ConstantRange RHSRange = getSignedRange(RHS);
5919 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5921 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5925 case ICmpInst::ICMP_SGE:
5926 Pred = ICmpInst::ICMP_SLE;
5927 std::swap(LHS, RHS);
5928 case ICmpInst::ICMP_SLE: {
5929 ConstantRange LHSRange = getSignedRange(LHS);
5930 ConstantRange RHSRange = getSignedRange(RHS);
5931 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5933 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
5937 case ICmpInst::ICMP_UGT:
5938 Pred = ICmpInst::ICMP_ULT;
5939 std::swap(LHS, RHS);
5940 case ICmpInst::ICMP_ULT: {
5941 ConstantRange LHSRange = getUnsignedRange(LHS);
5942 ConstantRange RHSRange = getUnsignedRange(RHS);
5943 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
5945 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
5949 case ICmpInst::ICMP_UGE:
5950 Pred = ICmpInst::ICMP_ULE;
5951 std::swap(LHS, RHS);
5952 case ICmpInst::ICMP_ULE: {
5953 ConstantRange LHSRange = getUnsignedRange(LHS);
5954 ConstantRange RHSRange = getUnsignedRange(RHS);
5955 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
5957 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
5961 case ICmpInst::ICMP_NE: {
5962 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
5964 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
5967 const SCEV *Diff = getMinusSCEV(LHS, RHS);
5968 if (isKnownNonZero(Diff))
5972 case ICmpInst::ICMP_EQ:
5973 // The check at the top of the function catches the case where
5974 // the values are known to be equal.
5980 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
5981 /// protected by a conditional between LHS and RHS. This is used to
5982 /// to eliminate casts.
5984 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
5985 ICmpInst::Predicate Pred,
5986 const SCEV *LHS, const SCEV *RHS) {
5987 // Interpret a null as meaning no loop, where there is obviously no guard
5988 // (interprocedural conditions notwithstanding).
5989 if (!L) return true;
5991 BasicBlock *Latch = L->getLoopLatch();
5995 BranchInst *LoopContinuePredicate =
5996 dyn_cast<BranchInst>(Latch->getTerminator());
5997 if (!LoopContinuePredicate ||
5998 LoopContinuePredicate->isUnconditional())
6001 return isImpliedCond(Pred, LHS, RHS,
6002 LoopContinuePredicate->getCondition(),
6003 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
6006 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
6007 /// by a conditional between LHS and RHS. This is used to help avoid max
6008 /// expressions in loop trip counts, and to eliminate casts.
6010 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
6011 ICmpInst::Predicate Pred,
6012 const SCEV *LHS, const SCEV *RHS) {
6013 // Interpret a null as meaning no loop, where there is obviously no guard
6014 // (interprocedural conditions notwithstanding).
6015 if (!L) return false;
6017 // Starting at the loop predecessor, climb up the predecessor chain, as long
6018 // as there are predecessors that can be found that have unique successors
6019 // leading to the original header.
6020 for (std::pair<BasicBlock *, BasicBlock *>
6021 Pair(L->getLoopPredecessor(), L->getHeader());
6023 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
6025 BranchInst *LoopEntryPredicate =
6026 dyn_cast<BranchInst>(Pair.first->getTerminator());
6027 if (!LoopEntryPredicate ||
6028 LoopEntryPredicate->isUnconditional())
6031 if (isImpliedCond(Pred, LHS, RHS,
6032 LoopEntryPredicate->getCondition(),
6033 LoopEntryPredicate->getSuccessor(0) != Pair.second))
6040 /// isImpliedCond - Test whether the condition described by Pred, LHS,
6041 /// and RHS is true whenever the given Cond value evaluates to true.
6042 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
6043 const SCEV *LHS, const SCEV *RHS,
6044 Value *FoundCondValue,
6046 // Recursively handle And and Or conditions.
6047 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
6048 if (BO->getOpcode() == Instruction::And) {
6050 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6051 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6052 } else if (BO->getOpcode() == Instruction::Or) {
6054 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6055 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6059 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
6060 if (!ICI) return false;
6062 // Bail if the ICmp's operands' types are wider than the needed type
6063 // before attempting to call getSCEV on them. This avoids infinite
6064 // recursion, since the analysis of widening casts can require loop
6065 // exit condition information for overflow checking, which would
6067 if (getTypeSizeInBits(LHS->getType()) <
6068 getTypeSizeInBits(ICI->getOperand(0)->getType()))
6071 // Now that we found a conditional branch that dominates the loop, check to
6072 // see if it is the comparison we are looking for.
6073 ICmpInst::Predicate FoundPred;
6075 FoundPred = ICI->getInversePredicate();
6077 FoundPred = ICI->getPredicate();
6079 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
6080 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
6082 // Balance the types. The case where FoundLHS' type is wider than
6083 // LHS' type is checked for above.
6084 if (getTypeSizeInBits(LHS->getType()) >
6085 getTypeSizeInBits(FoundLHS->getType())) {
6086 if (CmpInst::isSigned(Pred)) {
6087 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
6088 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
6090 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
6091 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
6095 // Canonicalize the query to match the way instcombine will have
6096 // canonicalized the comparison.
6097 if (SimplifyICmpOperands(Pred, LHS, RHS))
6099 return CmpInst::isTrueWhenEqual(Pred);
6100 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
6101 if (FoundLHS == FoundRHS)
6102 return CmpInst::isFalseWhenEqual(Pred);
6104 // Check to see if we can make the LHS or RHS match.
6105 if (LHS == FoundRHS || RHS == FoundLHS) {
6106 if (isa<SCEVConstant>(RHS)) {
6107 std::swap(FoundLHS, FoundRHS);
6108 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
6110 std::swap(LHS, RHS);
6111 Pred = ICmpInst::getSwappedPredicate(Pred);
6115 // Check whether the found predicate is the same as the desired predicate.
6116 if (FoundPred == Pred)
6117 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
6119 // Check whether swapping the found predicate makes it the same as the
6120 // desired predicate.
6121 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
6122 if (isa<SCEVConstant>(RHS))
6123 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
6125 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
6126 RHS, LHS, FoundLHS, FoundRHS);
6129 // Check whether the actual condition is beyond sufficient.
6130 if (FoundPred == ICmpInst::ICMP_EQ)
6131 if (ICmpInst::isTrueWhenEqual(Pred))
6132 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
6134 if (Pred == ICmpInst::ICMP_NE)
6135 if (!ICmpInst::isTrueWhenEqual(FoundPred))
6136 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
6139 // Otherwise assume the worst.
6143 /// isImpliedCondOperands - Test whether the condition described by Pred,
6144 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
6145 /// and FoundRHS is true.
6146 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
6147 const SCEV *LHS, const SCEV *RHS,
6148 const SCEV *FoundLHS,
6149 const SCEV *FoundRHS) {
6150 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
6151 FoundLHS, FoundRHS) ||
6152 // ~x < ~y --> x > y
6153 isImpliedCondOperandsHelper(Pred, LHS, RHS,
6154 getNotSCEV(FoundRHS),
6155 getNotSCEV(FoundLHS));
6158 /// isImpliedCondOperandsHelper - Test whether the condition described by
6159 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
6160 /// FoundLHS, and FoundRHS is true.
6162 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
6163 const SCEV *LHS, const SCEV *RHS,
6164 const SCEV *FoundLHS,
6165 const SCEV *FoundRHS) {
6167 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
6168 case ICmpInst::ICMP_EQ:
6169 case ICmpInst::ICMP_NE:
6170 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
6173 case ICmpInst::ICMP_SLT:
6174 case ICmpInst::ICMP_SLE:
6175 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
6176 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
6179 case ICmpInst::ICMP_SGT:
6180 case ICmpInst::ICMP_SGE:
6181 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
6182 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
6185 case ICmpInst::ICMP_ULT:
6186 case ICmpInst::ICMP_ULE:
6187 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
6188 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
6191 case ICmpInst::ICMP_UGT:
6192 case ICmpInst::ICMP_UGE:
6193 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
6194 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
6202 /// getBECount - Subtract the end and start values and divide by the step,
6203 /// rounding up, to get the number of times the backedge is executed. Return
6204 /// CouldNotCompute if an intermediate computation overflows.
6205 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
6209 assert(!isKnownNegative(Step) &&
6210 "This code doesn't handle negative strides yet!");
6212 Type *Ty = Start->getType();
6214 // When Start == End, we have an exact BECount == 0. Short-circuit this case
6215 // here because SCEV may not be able to determine that the unsigned division
6216 // after rounding is zero.
6218 return getConstant(Ty, 0);
6220 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
6221 const SCEV *Diff = getMinusSCEV(End, Start);
6222 const SCEV *RoundUp = getAddExpr(Step, NegOne);
6224 // Add an adjustment to the difference between End and Start so that
6225 // the division will effectively round up.
6226 const SCEV *Add = getAddExpr(Diff, RoundUp);
6229 // Check Add for unsigned overflow.
6230 // TODO: More sophisticated things could be done here.
6231 Type *WideTy = IntegerType::get(getContext(),
6232 getTypeSizeInBits(Ty) + 1);
6233 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
6234 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
6235 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
6236 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
6237 return getCouldNotCompute();
6240 return getUDivExpr(Add, Step);
6243 /// HowManyLessThans - Return the number of times a backedge containing the
6244 /// specified less-than comparison will execute. If not computable, return
6245 /// CouldNotCompute.
6246 ScalarEvolution::ExitLimit
6247 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
6248 const Loop *L, bool isSigned) {
6249 // Only handle: "ADDREC < LoopInvariant".
6250 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute();
6252 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
6253 if (!AddRec || AddRec->getLoop() != L)
6254 return getCouldNotCompute();
6256 // Check to see if we have a flag which makes analysis easy.
6257 bool NoWrap = isSigned ?
6258 AddRec->getNoWrapFlags((SCEV::NoWrapFlags)(SCEV::FlagNSW | SCEV::FlagNW)) :
6259 AddRec->getNoWrapFlags((SCEV::NoWrapFlags)(SCEV::FlagNUW | SCEV::FlagNW));
6261 if (AddRec->isAffine()) {
6262 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
6263 const SCEV *Step = AddRec->getStepRecurrence(*this);
6266 return getCouldNotCompute();
6267 if (Step->isOne()) {
6268 // With unit stride, the iteration never steps past the limit value.
6269 } else if (isKnownPositive(Step)) {
6270 // Test whether a positive iteration can step past the limit
6271 // value and past the maximum value for its type in a single step.
6272 // Note that it's not sufficient to check NoWrap here, because even
6273 // though the value after a wrap is undefined, it's not undefined
6274 // behavior, so if wrap does occur, the loop could either terminate or
6275 // loop infinitely, but in either case, the loop is guaranteed to
6276 // iterate at least until the iteration where the wrapping occurs.
6277 const SCEV *One = getConstant(Step->getType(), 1);
6279 APInt Max = APInt::getSignedMaxValue(BitWidth);
6280 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
6281 .slt(getSignedRange(RHS).getSignedMax()))
6282 return getCouldNotCompute();
6284 APInt Max = APInt::getMaxValue(BitWidth);
6285 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
6286 .ult(getUnsignedRange(RHS).getUnsignedMax()))
6287 return getCouldNotCompute();
6290 // TODO: Handle negative strides here and below.
6291 return getCouldNotCompute();
6293 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
6294 // m. So, we count the number of iterations in which {n,+,s} < m is true.
6295 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
6296 // treat m-n as signed nor unsigned due to overflow possibility.
6298 // First, we get the value of the LHS in the first iteration: n
6299 const SCEV *Start = AddRec->getOperand(0);
6301 // Determine the minimum constant start value.
6302 const SCEV *MinStart = getConstant(isSigned ?
6303 getSignedRange(Start).getSignedMin() :
6304 getUnsignedRange(Start).getUnsignedMin());
6306 // If we know that the condition is true in order to enter the loop,
6307 // then we know that it will run exactly (m-n)/s times. Otherwise, we
6308 // only know that it will execute (max(m,n)-n)/s times. In both cases,
6309 // the division must round up.
6310 const SCEV *End = RHS;
6311 if (!isLoopEntryGuardedByCond(L,
6312 isSigned ? ICmpInst::ICMP_SLT :
6314 getMinusSCEV(Start, Step), RHS))
6315 End = isSigned ? getSMaxExpr(RHS, Start)
6316 : getUMaxExpr(RHS, Start);
6318 // Determine the maximum constant end value.
6319 const SCEV *MaxEnd = getConstant(isSigned ?
6320 getSignedRange(End).getSignedMax() :
6321 getUnsignedRange(End).getUnsignedMax());
6323 // If MaxEnd is within a step of the maximum integer value in its type,
6324 // adjust it down to the minimum value which would produce the same effect.
6325 // This allows the subsequent ceiling division of (N+(step-1))/step to
6326 // compute the correct value.
6327 const SCEV *StepMinusOne = getMinusSCEV(Step,
6328 getConstant(Step->getType(), 1));
6331 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
6334 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
6337 // Finally, we subtract these two values and divide, rounding up, to get
6338 // the number of times the backedge is executed.
6339 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
6341 // The maximum backedge count is similar, except using the minimum start
6342 // value and the maximum end value.
6343 // If we already have an exact constant BECount, use it instead.
6344 const SCEV *MaxBECount = isa<SCEVConstant>(BECount) ? BECount
6345 : getBECount(MinStart, MaxEnd, Step, NoWrap);
6347 // If the stride is nonconstant, and NoWrap == true, then
6348 // getBECount(MinStart, MaxEnd) may not compute. This would result in an
6349 // exact BECount and invalid MaxBECount, which should be avoided to catch
6350 // more optimization opportunities.
6351 if (isa<SCEVCouldNotCompute>(MaxBECount))
6352 MaxBECount = BECount;
6354 return ExitLimit(BECount, MaxBECount);
6357 return getCouldNotCompute();
6360 /// getNumIterationsInRange - Return the number of iterations of this loop that
6361 /// produce values in the specified constant range. Another way of looking at
6362 /// this is that it returns the first iteration number where the value is not in
6363 /// the condition, thus computing the exit count. If the iteration count can't
6364 /// be computed, an instance of SCEVCouldNotCompute is returned.
6365 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
6366 ScalarEvolution &SE) const {
6367 if (Range.isFullSet()) // Infinite loop.
6368 return SE.getCouldNotCompute();
6370 // If the start is a non-zero constant, shift the range to simplify things.
6371 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
6372 if (!SC->getValue()->isZero()) {
6373 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
6374 Operands[0] = SE.getConstant(SC->getType(), 0);
6375 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop(),
6376 getNoWrapFlags(FlagNW));
6377 if (const SCEVAddRecExpr *ShiftedAddRec =
6378 dyn_cast<SCEVAddRecExpr>(Shifted))
6379 return ShiftedAddRec->getNumIterationsInRange(
6380 Range.subtract(SC->getValue()->getValue()), SE);
6381 // This is strange and shouldn't happen.
6382 return SE.getCouldNotCompute();
6385 // The only time we can solve this is when we have all constant indices.
6386 // Otherwise, we cannot determine the overflow conditions.
6387 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
6388 if (!isa<SCEVConstant>(getOperand(i)))
6389 return SE.getCouldNotCompute();
6392 // Okay at this point we know that all elements of the chrec are constants and
6393 // that the start element is zero.
6395 // First check to see if the range contains zero. If not, the first
6397 unsigned BitWidth = SE.getTypeSizeInBits(getType());
6398 if (!Range.contains(APInt(BitWidth, 0)))
6399 return SE.getConstant(getType(), 0);
6402 // If this is an affine expression then we have this situation:
6403 // Solve {0,+,A} in Range === Ax in Range
6405 // We know that zero is in the range. If A is positive then we know that
6406 // the upper value of the range must be the first possible exit value.
6407 // If A is negative then the lower of the range is the last possible loop
6408 // value. Also note that we already checked for a full range.
6409 APInt One(BitWidth,1);
6410 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
6411 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
6413 // The exit value should be (End+A)/A.
6414 APInt ExitVal = (End + A).udiv(A);
6415 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
6417 // Evaluate at the exit value. If we really did fall out of the valid
6418 // range, then we computed our trip count, otherwise wrap around or other
6419 // things must have happened.
6420 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
6421 if (Range.contains(Val->getValue()))
6422 return SE.getCouldNotCompute(); // Something strange happened
6424 // Ensure that the previous value is in the range. This is a sanity check.
6425 assert(Range.contains(
6426 EvaluateConstantChrecAtConstant(this,
6427 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
6428 "Linear scev computation is off in a bad way!");
6429 return SE.getConstant(ExitValue);
6430 } else if (isQuadratic()) {
6431 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
6432 // quadratic equation to solve it. To do this, we must frame our problem in
6433 // terms of figuring out when zero is crossed, instead of when
6434 // Range.getUpper() is crossed.
6435 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
6436 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
6437 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop(),
6438 // getNoWrapFlags(FlagNW)
6441 // Next, solve the constructed addrec
6442 std::pair<const SCEV *,const SCEV *> Roots =
6443 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
6444 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
6445 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
6447 // Pick the smallest positive root value.
6448 if (ConstantInt *CB =
6449 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
6450 R1->getValue(), R2->getValue()))) {
6451 if (CB->getZExtValue() == false)
6452 std::swap(R1, R2); // R1 is the minimum root now.
6454 // Make sure the root is not off by one. The returned iteration should
6455 // not be in the range, but the previous one should be. When solving
6456 // for "X*X < 5", for example, we should not return a root of 2.
6457 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
6460 if (Range.contains(R1Val->getValue())) {
6461 // The next iteration must be out of the range...
6462 ConstantInt *NextVal =
6463 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
6465 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6466 if (!Range.contains(R1Val->getValue()))
6467 return SE.getConstant(NextVal);
6468 return SE.getCouldNotCompute(); // Something strange happened
6471 // If R1 was not in the range, then it is a good return value. Make
6472 // sure that R1-1 WAS in the range though, just in case.
6473 ConstantInt *NextVal =
6474 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
6475 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6476 if (Range.contains(R1Val->getValue()))
6478 return SE.getCouldNotCompute(); // Something strange happened
6483 return SE.getCouldNotCompute();
6488 //===----------------------------------------------------------------------===//
6489 // SCEVCallbackVH Class Implementation
6490 //===----------------------------------------------------------------------===//
6492 void ScalarEvolution::SCEVCallbackVH::deleted() {
6493 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6494 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
6495 SE->ConstantEvolutionLoopExitValue.erase(PN);
6496 SE->ValueExprMap.erase(getValPtr());
6497 // this now dangles!
6500 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
6501 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6503 // Forget all the expressions associated with users of the old value,
6504 // so that future queries will recompute the expressions using the new
6506 Value *Old = getValPtr();
6507 SmallVector<User *, 16> Worklist;
6508 SmallPtrSet<User *, 8> Visited;
6509 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
6511 Worklist.push_back(*UI);
6512 while (!Worklist.empty()) {
6513 User *U = Worklist.pop_back_val();
6514 // Deleting the Old value will cause this to dangle. Postpone
6515 // that until everything else is done.
6518 if (!Visited.insert(U))
6520 if (PHINode *PN = dyn_cast<PHINode>(U))
6521 SE->ConstantEvolutionLoopExitValue.erase(PN);
6522 SE->ValueExprMap.erase(U);
6523 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
6525 Worklist.push_back(*UI);
6527 // Delete the Old value.
6528 if (PHINode *PN = dyn_cast<PHINode>(Old))
6529 SE->ConstantEvolutionLoopExitValue.erase(PN);
6530 SE->ValueExprMap.erase(Old);
6531 // this now dangles!
6534 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
6535 : CallbackVH(V), SE(se) {}
6537 //===----------------------------------------------------------------------===//
6538 // ScalarEvolution Class Implementation
6539 //===----------------------------------------------------------------------===//
6541 ScalarEvolution::ScalarEvolution()
6542 : FunctionPass(ID), FirstUnknown(0) {
6543 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
6546 bool ScalarEvolution::runOnFunction(Function &F) {
6548 LI = &getAnalysis<LoopInfo>();
6549 TD = getAnalysisIfAvailable<TargetData>();
6550 TLI = &getAnalysis<TargetLibraryInfo>();
6551 DT = &getAnalysis<DominatorTree>();
6555 void ScalarEvolution::releaseMemory() {
6556 // Iterate through all the SCEVUnknown instances and call their
6557 // destructors, so that they release their references to their values.
6558 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
6562 ValueExprMap.clear();
6564 // Free any extra memory created for ExitNotTakenInfo in the unlikely event
6565 // that a loop had multiple computable exits.
6566 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
6567 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end();
6572 BackedgeTakenCounts.clear();
6573 ConstantEvolutionLoopExitValue.clear();
6574 ValuesAtScopes.clear();
6575 LoopDispositions.clear();
6576 BlockDispositions.clear();
6577 UnsignedRanges.clear();
6578 SignedRanges.clear();
6579 UniqueSCEVs.clear();
6580 SCEVAllocator.Reset();
6583 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
6584 AU.setPreservesAll();
6585 AU.addRequiredTransitive<LoopInfo>();
6586 AU.addRequiredTransitive<DominatorTree>();
6587 AU.addRequired<TargetLibraryInfo>();
6590 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
6591 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
6594 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
6596 // Print all inner loops first
6597 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
6598 PrintLoopInfo(OS, SE, *I);
6601 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6604 SmallVector<BasicBlock *, 8> ExitBlocks;
6605 L->getExitBlocks(ExitBlocks);
6606 if (ExitBlocks.size() != 1)
6607 OS << "<multiple exits> ";
6609 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
6610 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
6612 OS << "Unpredictable backedge-taken count. ";
6617 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6620 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
6621 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
6623 OS << "Unpredictable max backedge-taken count. ";
6629 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
6630 // ScalarEvolution's implementation of the print method is to print
6631 // out SCEV values of all instructions that are interesting. Doing
6632 // this potentially causes it to create new SCEV objects though,
6633 // which technically conflicts with the const qualifier. This isn't
6634 // observable from outside the class though, so casting away the
6635 // const isn't dangerous.
6636 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
6638 OS << "Classifying expressions for: ";
6639 WriteAsOperand(OS, F, /*PrintType=*/false);
6641 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
6642 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
6645 const SCEV *SV = SE.getSCEV(&*I);
6648 const Loop *L = LI->getLoopFor((*I).getParent());
6650 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
6657 OS << "\t\t" "Exits: ";
6658 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
6659 if (!SE.isLoopInvariant(ExitValue, L)) {
6660 OS << "<<Unknown>>";
6669 OS << "Determining loop execution counts for: ";
6670 WriteAsOperand(OS, F, /*PrintType=*/false);
6672 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6673 PrintLoopInfo(OS, &SE, *I);
6676 ScalarEvolution::LoopDisposition
6677 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
6678 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S];
6679 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair =
6680 Values.insert(std::make_pair(L, LoopVariant));
6682 return Pair.first->second;
6684 LoopDisposition D = computeLoopDisposition(S, L);
6685 return LoopDispositions[S][L] = D;
6688 ScalarEvolution::LoopDisposition
6689 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
6690 switch (S->getSCEVType()) {
6692 return LoopInvariant;
6696 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
6697 case scAddRecExpr: {
6698 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6700 // If L is the addrec's loop, it's computable.
6701 if (AR->getLoop() == L)
6702 return LoopComputable;
6704 // Add recurrences are never invariant in the function-body (null loop).
6708 // This recurrence is variant w.r.t. L if L contains AR's loop.
6709 if (L->contains(AR->getLoop()))
6712 // This recurrence is invariant w.r.t. L if AR's loop contains L.
6713 if (AR->getLoop()->contains(L))
6714 return LoopInvariant;
6716 // This recurrence is variant w.r.t. L if any of its operands
6718 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
6720 if (!isLoopInvariant(*I, L))
6723 // Otherwise it's loop-invariant.
6724 return LoopInvariant;
6730 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6731 bool HasVarying = false;
6732 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6734 LoopDisposition D = getLoopDisposition(*I, L);
6735 if (D == LoopVariant)
6737 if (D == LoopComputable)
6740 return HasVarying ? LoopComputable : LoopInvariant;
6743 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6744 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
6745 if (LD == LoopVariant)
6747 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
6748 if (RD == LoopVariant)
6750 return (LD == LoopInvariant && RD == LoopInvariant) ?
6751 LoopInvariant : LoopComputable;
6754 // All non-instruction values are loop invariant. All instructions are loop
6755 // invariant if they are not contained in the specified loop.
6756 // Instructions are never considered invariant in the function body
6757 // (null loop) because they are defined within the "loop".
6758 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
6759 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
6760 return LoopInvariant;
6761 case scCouldNotCompute:
6762 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6763 default: llvm_unreachable("Unknown SCEV kind!");
6767 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
6768 return getLoopDisposition(S, L) == LoopInvariant;
6771 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
6772 return getLoopDisposition(S, L) == LoopComputable;
6775 ScalarEvolution::BlockDisposition
6776 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6777 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S];
6778 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool>
6779 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock));
6781 return Pair.first->second;
6783 BlockDisposition D = computeBlockDisposition(S, BB);
6784 return BlockDispositions[S][BB] = D;
6787 ScalarEvolution::BlockDisposition
6788 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6789 switch (S->getSCEVType()) {
6791 return ProperlyDominatesBlock;
6795 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
6796 case scAddRecExpr: {
6797 // This uses a "dominates" query instead of "properly dominates" query
6798 // to test for proper dominance too, because the instruction which
6799 // produces the addrec's value is a PHI, and a PHI effectively properly
6800 // dominates its entire containing block.
6801 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6802 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
6803 return DoesNotDominateBlock;
6805 // FALL THROUGH into SCEVNAryExpr handling.
6810 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6812 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6814 BlockDisposition D = getBlockDisposition(*I, BB);
6815 if (D == DoesNotDominateBlock)
6816 return DoesNotDominateBlock;
6817 if (D == DominatesBlock)
6820 return Proper ? ProperlyDominatesBlock : DominatesBlock;
6823 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6824 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6825 BlockDisposition LD = getBlockDisposition(LHS, BB);
6826 if (LD == DoesNotDominateBlock)
6827 return DoesNotDominateBlock;
6828 BlockDisposition RD = getBlockDisposition(RHS, BB);
6829 if (RD == DoesNotDominateBlock)
6830 return DoesNotDominateBlock;
6831 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
6832 ProperlyDominatesBlock : DominatesBlock;
6835 if (Instruction *I =
6836 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
6837 if (I->getParent() == BB)
6838 return DominatesBlock;
6839 if (DT->properlyDominates(I->getParent(), BB))
6840 return ProperlyDominatesBlock;
6841 return DoesNotDominateBlock;
6843 return ProperlyDominatesBlock;
6844 case scCouldNotCompute:
6845 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6847 llvm_unreachable("Unknown SCEV kind!");
6851 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
6852 return getBlockDisposition(S, BB) >= DominatesBlock;
6855 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
6856 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
6859 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
6860 SmallVector<const SCEV *, 8> Worklist;
6861 Worklist.push_back(S);
6863 S = Worklist.pop_back_val();
6865 switch (S->getSCEVType()) {
6870 case scSignExtend: {
6871 const SCEVCastExpr *Cast = cast<SCEVCastExpr>(S);
6872 const SCEV *CastOp = Cast->getOperand();
6875 Worklist.push_back(CastOp);
6883 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6884 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6886 const SCEV *NAryOp = *I;
6889 Worklist.push_back(NAryOp);
6894 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6895 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6896 if (LHS == Op || RHS == Op)
6898 Worklist.push_back(LHS);
6899 Worklist.push_back(RHS);
6904 case scCouldNotCompute:
6905 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6907 llvm_unreachable("Unknown SCEV kind!");
6909 } while (!Worklist.empty());
6914 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
6915 ValuesAtScopes.erase(S);
6916 LoopDispositions.erase(S);
6917 BlockDispositions.erase(S);
6918 UnsignedRanges.erase(S);
6919 SignedRanges.erase(S);