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/ScalarEvolution.h"
63 #include "llvm/ADT/STLExtras.h"
64 #include "llvm/ADT/SmallPtrSet.h"
65 #include "llvm/ADT/Statistic.h"
66 #include "llvm/Analysis/ConstantFolding.h"
67 #include "llvm/Analysis/InstructionSimplify.h"
68 #include "llvm/Analysis/LoopInfo.h"
69 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
70 #include "llvm/Analysis/ValueTracking.h"
71 #include "llvm/IR/Constants.h"
72 #include "llvm/IR/DataLayout.h"
73 #include "llvm/IR/DerivedTypes.h"
74 #include "llvm/IR/Dominators.h"
75 #include "llvm/IR/GlobalAlias.h"
76 #include "llvm/IR/GlobalVariable.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/LLVMContext.h"
79 #include "llvm/IR/Operator.h"
80 #include "llvm/Support/CommandLine.h"
81 #include "llvm/Support/ConstantRange.h"
82 #include "llvm/Support/Debug.h"
83 #include "llvm/Support/ErrorHandling.h"
84 #include "llvm/Support/GetElementPtrTypeIterator.h"
85 #include "llvm/Support/InstIterator.h"
86 #include "llvm/Support/MathExtras.h"
87 #include "llvm/Support/raw_ostream.h"
88 #include "llvm/Target/TargetLibraryInfo.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 // FIXME: Enable this with XDEBUG when the test suite is clean.
110 VerifySCEV("verify-scev",
111 cl::desc("Verify ScalarEvolution's backedge taken counts (slow)"));
113 INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
114 "Scalar Evolution Analysis", false, true)
115 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
116 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
117 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
118 INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
119 "Scalar Evolution Analysis", false, true)
120 char ScalarEvolution::ID = 0;
122 //===----------------------------------------------------------------------===//
123 // SCEV class definitions
124 //===----------------------------------------------------------------------===//
126 //===----------------------------------------------------------------------===//
127 // Implementation of the SCEV class.
130 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
131 void SCEV::dump() const {
137 void SCEV::print(raw_ostream &OS) const {
138 switch (getSCEVType()) {
140 cast<SCEVConstant>(this)->getValue()->printAsOperand(OS, false);
143 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(this);
144 const SCEV *Op = Trunc->getOperand();
145 OS << "(trunc " << *Op->getType() << " " << *Op << " to "
146 << *Trunc->getType() << ")";
150 const SCEVZeroExtendExpr *ZExt = cast<SCEVZeroExtendExpr>(this);
151 const SCEV *Op = ZExt->getOperand();
152 OS << "(zext " << *Op->getType() << " " << *Op << " to "
153 << *ZExt->getType() << ")";
157 const SCEVSignExtendExpr *SExt = cast<SCEVSignExtendExpr>(this);
158 const SCEV *Op = SExt->getOperand();
159 OS << "(sext " << *Op->getType() << " " << *Op << " to "
160 << *SExt->getType() << ")";
164 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(this);
165 OS << "{" << *AR->getOperand(0);
166 for (unsigned i = 1, e = AR->getNumOperands(); i != e; ++i)
167 OS << ",+," << *AR->getOperand(i);
169 if (AR->getNoWrapFlags(FlagNUW))
171 if (AR->getNoWrapFlags(FlagNSW))
173 if (AR->getNoWrapFlags(FlagNW) &&
174 !AR->getNoWrapFlags((NoWrapFlags)(FlagNUW | FlagNSW)))
176 AR->getLoop()->getHeader()->printAsOperand(OS, /*PrintType=*/false);
184 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(this);
185 const char *OpStr = 0;
186 switch (NAry->getSCEVType()) {
187 case scAddExpr: OpStr = " + "; break;
188 case scMulExpr: OpStr = " * "; break;
189 case scUMaxExpr: OpStr = " umax "; break;
190 case scSMaxExpr: OpStr = " smax "; break;
193 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
196 if (llvm::next(I) != E)
200 switch (NAry->getSCEVType()) {
203 if (NAry->getNoWrapFlags(FlagNUW))
205 if (NAry->getNoWrapFlags(FlagNSW))
211 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(this);
212 OS << "(" << *UDiv->getLHS() << " /u " << *UDiv->getRHS() << ")";
216 const SCEVUnknown *U = cast<SCEVUnknown>(this);
218 if (U->isSizeOf(AllocTy)) {
219 OS << "sizeof(" << *AllocTy << ")";
222 if (U->isAlignOf(AllocTy)) {
223 OS << "alignof(" << *AllocTy << ")";
229 if (U->isOffsetOf(CTy, FieldNo)) {
230 OS << "offsetof(" << *CTy << ", ";
231 FieldNo->printAsOperand(OS, false);
236 // Otherwise just print it normally.
237 U->getValue()->printAsOperand(OS, false);
240 case scCouldNotCompute:
241 OS << "***COULDNOTCOMPUTE***";
245 llvm_unreachable("Unknown SCEV kind!");
248 Type *SCEV::getType() const {
249 switch (getSCEVType()) {
251 return cast<SCEVConstant>(this)->getType();
255 return cast<SCEVCastExpr>(this)->getType();
260 return cast<SCEVNAryExpr>(this)->getType();
262 return cast<SCEVAddExpr>(this)->getType();
264 return cast<SCEVUDivExpr>(this)->getType();
266 return cast<SCEVUnknown>(this)->getType();
267 case scCouldNotCompute:
268 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
270 llvm_unreachable("Unknown SCEV kind!");
274 bool SCEV::isZero() const {
275 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
276 return SC->getValue()->isZero();
280 bool SCEV::isOne() const {
281 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
282 return SC->getValue()->isOne();
286 bool SCEV::isAllOnesValue() const {
287 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
288 return SC->getValue()->isAllOnesValue();
292 /// isNonConstantNegative - Return true if the specified scev is negated, but
294 bool SCEV::isNonConstantNegative() const {
295 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(this);
296 if (!Mul) return false;
298 // If there is a constant factor, it will be first.
299 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
300 if (!SC) return false;
302 // Return true if the value is negative, this matches things like (-42 * V).
303 return SC->getValue()->getValue().isNegative();
306 SCEVCouldNotCompute::SCEVCouldNotCompute() :
307 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
309 bool SCEVCouldNotCompute::classof(const SCEV *S) {
310 return S->getSCEVType() == scCouldNotCompute;
313 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
315 ID.AddInteger(scConstant);
318 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
319 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V);
320 UniqueSCEVs.InsertNode(S, IP);
324 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
325 return getConstant(ConstantInt::get(getContext(), Val));
329 ScalarEvolution::getConstant(Type *Ty, uint64_t V, bool isSigned) {
330 IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
331 return getConstant(ConstantInt::get(ITy, V, isSigned));
334 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
335 unsigned SCEVTy, const SCEV *op, Type *ty)
336 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
338 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
339 const SCEV *op, Type *ty)
340 : SCEVCastExpr(ID, scTruncate, op, ty) {
341 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
342 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
343 "Cannot truncate non-integer value!");
346 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
347 const SCEV *op, Type *ty)
348 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
349 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
350 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
351 "Cannot zero extend non-integer value!");
354 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
355 const SCEV *op, Type *ty)
356 : SCEVCastExpr(ID, scSignExtend, op, ty) {
357 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
358 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
359 "Cannot sign extend non-integer value!");
362 void SCEVUnknown::deleted() {
363 // Clear this SCEVUnknown from various maps.
364 SE->forgetMemoizedResults(this);
366 // Remove this SCEVUnknown from the uniquing map.
367 SE->UniqueSCEVs.RemoveNode(this);
369 // Release the value.
373 void SCEVUnknown::allUsesReplacedWith(Value *New) {
374 // Clear this SCEVUnknown from various maps.
375 SE->forgetMemoizedResults(this);
377 // Remove this SCEVUnknown from the uniquing map.
378 SE->UniqueSCEVs.RemoveNode(this);
380 // Update this SCEVUnknown to point to the new value. This is needed
381 // because there may still be outstanding SCEVs which still point to
386 bool SCEVUnknown::isSizeOf(Type *&AllocTy) const {
387 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
388 if (VCE->getOpcode() == Instruction::PtrToInt)
389 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
390 if (CE->getOpcode() == Instruction::GetElementPtr &&
391 CE->getOperand(0)->isNullValue() &&
392 CE->getNumOperands() == 2)
393 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
395 AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
403 bool SCEVUnknown::isAlignOf(Type *&AllocTy) const {
404 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
405 if (VCE->getOpcode() == Instruction::PtrToInt)
406 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
407 if (CE->getOpcode() == Instruction::GetElementPtr &&
408 CE->getOperand(0)->isNullValue()) {
410 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
411 if (StructType *STy = dyn_cast<StructType>(Ty))
412 if (!STy->isPacked() &&
413 CE->getNumOperands() == 3 &&
414 CE->getOperand(1)->isNullValue()) {
415 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
417 STy->getNumElements() == 2 &&
418 STy->getElementType(0)->isIntegerTy(1)) {
419 AllocTy = STy->getElementType(1);
428 bool SCEVUnknown::isOffsetOf(Type *&CTy, Constant *&FieldNo) const {
429 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
430 if (VCE->getOpcode() == Instruction::PtrToInt)
431 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
432 if (CE->getOpcode() == Instruction::GetElementPtr &&
433 CE->getNumOperands() == 3 &&
434 CE->getOperand(0)->isNullValue() &&
435 CE->getOperand(1)->isNullValue()) {
437 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
438 // Ignore vector types here so that ScalarEvolutionExpander doesn't
439 // emit getelementptrs that index into vectors.
440 if (Ty->isStructTy() || Ty->isArrayTy()) {
442 FieldNo = CE->getOperand(2);
450 //===----------------------------------------------------------------------===//
452 //===----------------------------------------------------------------------===//
455 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
456 /// than the complexity of the RHS. This comparator is used to canonicalize
458 class SCEVComplexityCompare {
459 const LoopInfo *const LI;
461 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
463 // Return true or false if LHS is less than, or at least RHS, respectively.
464 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
465 return compare(LHS, RHS) < 0;
468 // Return negative, zero, or positive, if LHS is less than, equal to, or
469 // greater than RHS, respectively. A three-way result allows recursive
470 // comparisons to be more efficient.
471 int compare(const SCEV *LHS, const SCEV *RHS) const {
472 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
476 // Primarily, sort the SCEVs by their getSCEVType().
477 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
479 return (int)LType - (int)RType;
481 // Aside from the getSCEVType() ordering, the particular ordering
482 // isn't very important except that it's beneficial to be consistent,
483 // so that (a + b) and (b + a) don't end up as different expressions.
486 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS);
487 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
489 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
490 // not as complete as it could be.
491 const Value *LV = LU->getValue(), *RV = RU->getValue();
493 // Order pointer values after integer values. This helps SCEVExpander
495 bool LIsPointer = LV->getType()->isPointerTy(),
496 RIsPointer = RV->getType()->isPointerTy();
497 if (LIsPointer != RIsPointer)
498 return (int)LIsPointer - (int)RIsPointer;
500 // Compare getValueID values.
501 unsigned LID = LV->getValueID(),
502 RID = RV->getValueID();
504 return (int)LID - (int)RID;
506 // Sort arguments by their position.
507 if (const Argument *LA = dyn_cast<Argument>(LV)) {
508 const Argument *RA = cast<Argument>(RV);
509 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo();
510 return (int)LArgNo - (int)RArgNo;
513 // For instructions, compare their loop depth, and their operand
514 // count. This is pretty loose.
515 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) {
516 const Instruction *RInst = cast<Instruction>(RV);
518 // Compare loop depths.
519 const BasicBlock *LParent = LInst->getParent(),
520 *RParent = RInst->getParent();
521 if (LParent != RParent) {
522 unsigned LDepth = LI->getLoopDepth(LParent),
523 RDepth = LI->getLoopDepth(RParent);
524 if (LDepth != RDepth)
525 return (int)LDepth - (int)RDepth;
528 // Compare the number of operands.
529 unsigned LNumOps = LInst->getNumOperands(),
530 RNumOps = RInst->getNumOperands();
531 return (int)LNumOps - (int)RNumOps;
538 const SCEVConstant *LC = cast<SCEVConstant>(LHS);
539 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
541 // Compare constant values.
542 const APInt &LA = LC->getValue()->getValue();
543 const APInt &RA = RC->getValue()->getValue();
544 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth();
545 if (LBitWidth != RBitWidth)
546 return (int)LBitWidth - (int)RBitWidth;
547 return LA.ult(RA) ? -1 : 1;
551 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS);
552 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
554 // Compare addrec loop depths.
555 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop();
556 if (LLoop != RLoop) {
557 unsigned LDepth = LLoop->getLoopDepth(),
558 RDepth = RLoop->getLoopDepth();
559 if (LDepth != RDepth)
560 return (int)LDepth - (int)RDepth;
563 // Addrec complexity grows with operand count.
564 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands();
565 if (LNumOps != RNumOps)
566 return (int)LNumOps - (int)RNumOps;
568 // Lexicographically compare.
569 for (unsigned i = 0; i != LNumOps; ++i) {
570 long X = compare(LA->getOperand(i), RA->getOperand(i));
582 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS);
583 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
585 // Lexicographically compare n-ary expressions.
586 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
587 if (LNumOps != RNumOps)
588 return (int)LNumOps - (int)RNumOps;
590 for (unsigned i = 0; i != LNumOps; ++i) {
593 long X = compare(LC->getOperand(i), RC->getOperand(i));
597 return (int)LNumOps - (int)RNumOps;
601 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
602 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
604 // Lexicographically compare udiv expressions.
605 long X = compare(LC->getLHS(), RC->getLHS());
608 return compare(LC->getRHS(), RC->getRHS());
614 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
615 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
617 // Compare cast expressions by operand.
618 return compare(LC->getOperand(), RC->getOperand());
622 llvm_unreachable("Unknown SCEV kind!");
628 /// GroupByComplexity - Given a list of SCEV objects, order them by their
629 /// complexity, and group objects of the same complexity together by value.
630 /// When this routine is finished, we know that any duplicates in the vector are
631 /// consecutive and that complexity is monotonically increasing.
633 /// Note that we go take special precautions to ensure that we get deterministic
634 /// results from this routine. In other words, we don't want the results of
635 /// this to depend on where the addresses of various SCEV objects happened to
638 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
640 if (Ops.size() < 2) return; // Noop
641 if (Ops.size() == 2) {
642 // This is the common case, which also happens to be trivially simple.
644 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
645 if (SCEVComplexityCompare(LI)(RHS, LHS))
650 // Do the rough sort by complexity.
651 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
653 // Now that we are sorted by complexity, group elements of the same
654 // complexity. Note that this is, at worst, N^2, but the vector is likely to
655 // be extremely short in practice. Note that we take this approach because we
656 // do not want to depend on the addresses of the objects we are grouping.
657 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
658 const SCEV *S = Ops[i];
659 unsigned Complexity = S->getSCEVType();
661 // If there are any objects of the same complexity and same value as this
663 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
664 if (Ops[j] == S) { // Found a duplicate.
665 // Move it to immediately after i'th element.
666 std::swap(Ops[i+1], Ops[j]);
667 ++i; // no need to rescan it.
668 if (i == e-2) return; // Done!
676 //===----------------------------------------------------------------------===//
677 // Simple SCEV method implementations
678 //===----------------------------------------------------------------------===//
680 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
682 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
685 // Handle the simplest case efficiently.
687 return SE.getTruncateOrZeroExtend(It, ResultTy);
689 // We are using the following formula for BC(It, K):
691 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
693 // Suppose, W is the bitwidth of the return value. We must be prepared for
694 // overflow. Hence, we must assure that the result of our computation is
695 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
696 // safe in modular arithmetic.
698 // However, this code doesn't use exactly that formula; the formula it uses
699 // is something like the following, where T is the number of factors of 2 in
700 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
703 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
705 // This formula is trivially equivalent to the previous formula. However,
706 // this formula can be implemented much more efficiently. The trick is that
707 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
708 // arithmetic. To do exact division in modular arithmetic, all we have
709 // to do is multiply by the inverse. Therefore, this step can be done at
712 // The next issue is how to safely do the division by 2^T. The way this
713 // is done is by doing the multiplication step at a width of at least W + T
714 // bits. This way, the bottom W+T bits of the product are accurate. Then,
715 // when we perform the division by 2^T (which is equivalent to a right shift
716 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
717 // truncated out after the division by 2^T.
719 // In comparison to just directly using the first formula, this technique
720 // is much more efficient; using the first formula requires W * K bits,
721 // but this formula less than W + K bits. Also, the first formula requires
722 // a division step, whereas this formula only requires multiplies and shifts.
724 // It doesn't matter whether the subtraction step is done in the calculation
725 // width or the input iteration count's width; if the subtraction overflows,
726 // the result must be zero anyway. We prefer here to do it in the width of
727 // the induction variable because it helps a lot for certain cases; CodeGen
728 // isn't smart enough to ignore the overflow, which leads to much less
729 // efficient code if the width of the subtraction is wider than the native
732 // (It's possible to not widen at all by pulling out factors of 2 before
733 // the multiplication; for example, K=2 can be calculated as
734 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
735 // extra arithmetic, so it's not an obvious win, and it gets
736 // much more complicated for K > 3.)
738 // Protection from insane SCEVs; this bound is conservative,
739 // but it probably doesn't matter.
741 return SE.getCouldNotCompute();
743 unsigned W = SE.getTypeSizeInBits(ResultTy);
745 // Calculate K! / 2^T and T; we divide out the factors of two before
746 // multiplying for calculating K! / 2^T to avoid overflow.
747 // Other overflow doesn't matter because we only care about the bottom
748 // W bits of the result.
749 APInt OddFactorial(W, 1);
751 for (unsigned i = 3; i <= K; ++i) {
753 unsigned TwoFactors = Mult.countTrailingZeros();
755 Mult = Mult.lshr(TwoFactors);
756 OddFactorial *= Mult;
759 // We need at least W + T bits for the multiplication step
760 unsigned CalculationBits = W + T;
762 // Calculate 2^T, at width T+W.
763 APInt DivFactor = APInt::getOneBitSet(CalculationBits, T);
765 // Calculate the multiplicative inverse of K! / 2^T;
766 // this multiplication factor will perform the exact division by
768 APInt Mod = APInt::getSignedMinValue(W+1);
769 APInt MultiplyFactor = OddFactorial.zext(W+1);
770 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
771 MultiplyFactor = MultiplyFactor.trunc(W);
773 // Calculate the product, at width T+W
774 IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
776 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
777 for (unsigned i = 1; i != K; ++i) {
778 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
779 Dividend = SE.getMulExpr(Dividend,
780 SE.getTruncateOrZeroExtend(S, CalculationTy));
784 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
786 // Truncate the result, and divide by K! / 2^T.
788 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
789 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
792 /// evaluateAtIteration - Return the value of this chain of recurrences at
793 /// the specified iteration number. We can evaluate this recurrence by
794 /// multiplying each element in the chain by the binomial coefficient
795 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
797 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
799 /// where BC(It, k) stands for binomial coefficient.
801 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
802 ScalarEvolution &SE) const {
803 const SCEV *Result = getStart();
804 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
805 // The computation is correct in the face of overflow provided that the
806 // multiplication is performed _after_ the evaluation of the binomial
808 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
809 if (isa<SCEVCouldNotCompute>(Coeff))
812 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
817 //===----------------------------------------------------------------------===//
818 // SCEV Expression folder implementations
819 //===----------------------------------------------------------------------===//
821 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
823 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
824 "This is not a truncating conversion!");
825 assert(isSCEVable(Ty) &&
826 "This is not a conversion to a SCEVable type!");
827 Ty = getEffectiveSCEVType(Ty);
830 ID.AddInteger(scTruncate);
834 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
836 // Fold if the operand is constant.
837 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
839 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
841 // trunc(trunc(x)) --> trunc(x)
842 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
843 return getTruncateExpr(ST->getOperand(), Ty);
845 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
846 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
847 return getTruncateOrSignExtend(SS->getOperand(), Ty);
849 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
850 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
851 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
853 // trunc(x1+x2+...+xN) --> trunc(x1)+trunc(x2)+...+trunc(xN) if we can
854 // eliminate all the truncates.
855 if (const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Op)) {
856 SmallVector<const SCEV *, 4> Operands;
857 bool hasTrunc = false;
858 for (unsigned i = 0, e = SA->getNumOperands(); i != e && !hasTrunc; ++i) {
859 const SCEV *S = getTruncateExpr(SA->getOperand(i), Ty);
860 hasTrunc = isa<SCEVTruncateExpr>(S);
861 Operands.push_back(S);
864 return getAddExpr(Operands);
865 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
868 // trunc(x1*x2*...*xN) --> trunc(x1)*trunc(x2)*...*trunc(xN) if we can
869 // eliminate all the truncates.
870 if (const SCEVMulExpr *SM = dyn_cast<SCEVMulExpr>(Op)) {
871 SmallVector<const SCEV *, 4> Operands;
872 bool hasTrunc = false;
873 for (unsigned i = 0, e = SM->getNumOperands(); i != e && !hasTrunc; ++i) {
874 const SCEV *S = getTruncateExpr(SM->getOperand(i), Ty);
875 hasTrunc = isa<SCEVTruncateExpr>(S);
876 Operands.push_back(S);
879 return getMulExpr(Operands);
880 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
883 // If the input value is a chrec scev, truncate the chrec's operands.
884 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
885 SmallVector<const SCEV *, 4> Operands;
886 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
887 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
888 return getAddRecExpr(Operands, AddRec->getLoop(), SCEV::FlagAnyWrap);
891 // The cast wasn't folded; create an explicit cast node. We can reuse
892 // the existing insert position since if we get here, we won't have
893 // made any changes which would invalidate it.
894 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
896 UniqueSCEVs.InsertNode(S, IP);
900 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
902 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
903 "This is not an extending conversion!");
904 assert(isSCEVable(Ty) &&
905 "This is not a conversion to a SCEVable type!");
906 Ty = getEffectiveSCEVType(Ty);
908 // Fold if the operand is constant.
909 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
911 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(), Ty)));
913 // zext(zext(x)) --> zext(x)
914 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
915 return getZeroExtendExpr(SZ->getOperand(), Ty);
917 // Before doing any expensive analysis, check to see if we've already
918 // computed a SCEV for this Op and Ty.
920 ID.AddInteger(scZeroExtend);
924 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
926 // zext(trunc(x)) --> zext(x) or x or trunc(x)
927 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
928 // It's possible the bits taken off by the truncate were all zero bits. If
929 // so, we should be able to simplify this further.
930 const SCEV *X = ST->getOperand();
931 ConstantRange CR = getUnsignedRange(X);
932 unsigned TruncBits = getTypeSizeInBits(ST->getType());
933 unsigned NewBits = getTypeSizeInBits(Ty);
934 if (CR.truncate(TruncBits).zeroExtend(NewBits).contains(
935 CR.zextOrTrunc(NewBits)))
936 return getTruncateOrZeroExtend(X, Ty);
939 // If the input value is a chrec scev, and we can prove that the value
940 // did not overflow the old, smaller, value, we can zero extend all of the
941 // operands (often constants). This allows analysis of something like
942 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
943 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
944 if (AR->isAffine()) {
945 const SCEV *Start = AR->getStart();
946 const SCEV *Step = AR->getStepRecurrence(*this);
947 unsigned BitWidth = getTypeSizeInBits(AR->getType());
948 const Loop *L = AR->getLoop();
950 // If we have special knowledge that this addrec won't overflow,
951 // we don't need to do any further analysis.
952 if (AR->getNoWrapFlags(SCEV::FlagNUW))
953 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
954 getZeroExtendExpr(Step, Ty),
955 L, AR->getNoWrapFlags());
957 // Check whether the backedge-taken count is SCEVCouldNotCompute.
958 // Note that this serves two purposes: It filters out loops that are
959 // simply not analyzable, and it covers the case where this code is
960 // being called from within backedge-taken count analysis, such that
961 // attempting to ask for the backedge-taken count would likely result
962 // in infinite recursion. In the later case, the analysis code will
963 // cope with a conservative value, and it will take care to purge
964 // that value once it has finished.
965 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
966 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
967 // Manually compute the final value for AR, checking for
970 // Check whether the backedge-taken count can be losslessly casted to
971 // the addrec's type. The count is always unsigned.
972 const SCEV *CastedMaxBECount =
973 getTruncateOrZeroExtend(MaxBECount, Start->getType());
974 const SCEV *RecastedMaxBECount =
975 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
976 if (MaxBECount == RecastedMaxBECount) {
977 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
978 // Check whether Start+Step*MaxBECount has no unsigned overflow.
979 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
980 const SCEV *ZAdd = getZeroExtendExpr(getAddExpr(Start, ZMul), WideTy);
981 const SCEV *WideStart = getZeroExtendExpr(Start, WideTy);
982 const SCEV *WideMaxBECount =
983 getZeroExtendExpr(CastedMaxBECount, WideTy);
984 const SCEV *OperandExtendedAdd =
985 getAddExpr(WideStart,
986 getMulExpr(WideMaxBECount,
987 getZeroExtendExpr(Step, WideTy)));
988 if (ZAdd == OperandExtendedAdd) {
989 // Cache knowledge of AR NUW, which is propagated to this AddRec.
990 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
991 // Return the expression with the addrec on the outside.
992 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
993 getZeroExtendExpr(Step, Ty),
994 L, AR->getNoWrapFlags());
996 // Similar to above, only this time treat the step value as signed.
997 // This covers loops that count down.
999 getAddExpr(WideStart,
1000 getMulExpr(WideMaxBECount,
1001 getSignExtendExpr(Step, WideTy)));
1002 if (ZAdd == OperandExtendedAdd) {
1003 // Cache knowledge of AR NW, which is propagated to this AddRec.
1004 // Negative step causes unsigned wrap, but it still can't self-wrap.
1005 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1006 // Return the expression with the addrec on the outside.
1007 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1008 getSignExtendExpr(Step, Ty),
1009 L, AR->getNoWrapFlags());
1013 // If the backedge is guarded by a comparison with the pre-inc value
1014 // the addrec is safe. Also, if the entry is guarded by a comparison
1015 // with the start value and the backedge is guarded by a comparison
1016 // with the post-inc value, the addrec is safe.
1017 if (isKnownPositive(Step)) {
1018 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
1019 getUnsignedRange(Step).getUnsignedMax());
1020 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
1021 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
1022 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
1023 AR->getPostIncExpr(*this), N))) {
1024 // Cache knowledge of AR NUW, which is propagated to this AddRec.
1025 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
1026 // Return the expression with the addrec on the outside.
1027 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1028 getZeroExtendExpr(Step, Ty),
1029 L, AR->getNoWrapFlags());
1031 } else if (isKnownNegative(Step)) {
1032 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
1033 getSignedRange(Step).getSignedMin());
1034 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
1035 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
1036 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
1037 AR->getPostIncExpr(*this), N))) {
1038 // Cache knowledge of AR NW, which is propagated to this AddRec.
1039 // Negative step causes unsigned wrap, but it still can't self-wrap.
1040 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1041 // Return the expression with the addrec on the outside.
1042 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1043 getSignExtendExpr(Step, Ty),
1044 L, AR->getNoWrapFlags());
1050 // The cast wasn't folded; create an explicit cast node.
1051 // Recompute the insert position, as it may have been invalidated.
1052 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1053 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1055 UniqueSCEVs.InsertNode(S, IP);
1059 // Get the limit of a recurrence such that incrementing by Step cannot cause
1060 // signed overflow as long as the value of the recurrence within the loop does
1061 // not exceed this limit before incrementing.
1062 static const SCEV *getOverflowLimitForStep(const SCEV *Step,
1063 ICmpInst::Predicate *Pred,
1064 ScalarEvolution *SE) {
1065 unsigned BitWidth = SE->getTypeSizeInBits(Step->getType());
1066 if (SE->isKnownPositive(Step)) {
1067 *Pred = ICmpInst::ICMP_SLT;
1068 return SE->getConstant(APInt::getSignedMinValue(BitWidth) -
1069 SE->getSignedRange(Step).getSignedMax());
1071 if (SE->isKnownNegative(Step)) {
1072 *Pred = ICmpInst::ICMP_SGT;
1073 return SE->getConstant(APInt::getSignedMaxValue(BitWidth) -
1074 SE->getSignedRange(Step).getSignedMin());
1079 // The recurrence AR has been shown to have no signed wrap. Typically, if we can
1080 // prove NSW for AR, then we can just as easily prove NSW for its preincrement
1081 // or postincrement sibling. This allows normalizing a sign extended AddRec as
1082 // such: {sext(Step + Start),+,Step} => {(Step + sext(Start),+,Step} As a
1083 // result, the expression "Step + sext(PreIncAR)" is congruent with
1084 // "sext(PostIncAR)"
1085 static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR,
1087 ScalarEvolution *SE) {
1088 const Loop *L = AR->getLoop();
1089 const SCEV *Start = AR->getStart();
1090 const SCEV *Step = AR->getStepRecurrence(*SE);
1092 // Check for a simple looking step prior to loop entry.
1093 const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start);
1097 // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV
1098 // subtraction is expensive. For this purpose, perform a quick and dirty
1099 // difference, by checking for Step in the operand list.
1100 SmallVector<const SCEV *, 4> DiffOps;
1101 for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end();
1104 DiffOps.push_back(*I);
1106 if (DiffOps.size() == SA->getNumOperands())
1109 // This is a postinc AR. Check for overflow on the preinc recurrence using the
1110 // same three conditions that getSignExtendedExpr checks.
1112 // 1. NSW flags on the step increment.
1113 const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags());
1114 const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>(
1115 SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap));
1117 if (PreAR && PreAR->getNoWrapFlags(SCEV::FlagNSW))
1120 // 2. Direct overflow check on the step operation's expression.
1121 unsigned BitWidth = SE->getTypeSizeInBits(AR->getType());
1122 Type *WideTy = IntegerType::get(SE->getContext(), BitWidth * 2);
1123 const SCEV *OperandExtendedStart =
1124 SE->getAddExpr(SE->getSignExtendExpr(PreStart, WideTy),
1125 SE->getSignExtendExpr(Step, WideTy));
1126 if (SE->getSignExtendExpr(Start, WideTy) == OperandExtendedStart) {
1127 // Cache knowledge of PreAR NSW.
1129 const_cast<SCEVAddRecExpr *>(PreAR)->setNoWrapFlags(SCEV::FlagNSW);
1130 // FIXME: this optimization needs a unit test
1131 DEBUG(dbgs() << "SCEV: untested prestart overflow check\n");
1135 // 3. Loop precondition.
1136 ICmpInst::Predicate Pred;
1137 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, SE);
1139 if (OverflowLimit &&
1140 SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) {
1146 // Get the normalized sign-extended expression for this AddRec's Start.
1147 static const SCEV *getSignExtendAddRecStart(const SCEVAddRecExpr *AR,
1149 ScalarEvolution *SE) {
1150 const SCEV *PreStart = getPreStartForSignExtend(AR, Ty, SE);
1152 return SE->getSignExtendExpr(AR->getStart(), Ty);
1154 return SE->getAddExpr(SE->getSignExtendExpr(AR->getStepRecurrence(*SE), Ty),
1155 SE->getSignExtendExpr(PreStart, Ty));
1158 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1160 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1161 "This is not an extending conversion!");
1162 assert(isSCEVable(Ty) &&
1163 "This is not a conversion to a SCEVable type!");
1164 Ty = getEffectiveSCEVType(Ty);
1166 // Fold if the operand is constant.
1167 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1169 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(), Ty)));
1171 // sext(sext(x)) --> sext(x)
1172 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1173 return getSignExtendExpr(SS->getOperand(), Ty);
1175 // sext(zext(x)) --> zext(x)
1176 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
1177 return getZeroExtendExpr(SZ->getOperand(), Ty);
1179 // Before doing any expensive analysis, check to see if we've already
1180 // computed a SCEV for this Op and Ty.
1181 FoldingSetNodeID ID;
1182 ID.AddInteger(scSignExtend);
1186 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1188 // If the input value is provably positive, build a zext instead.
1189 if (isKnownNonNegative(Op))
1190 return getZeroExtendExpr(Op, Ty);
1192 // sext(trunc(x)) --> sext(x) or x or trunc(x)
1193 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
1194 // It's possible the bits taken off by the truncate were all sign bits. If
1195 // so, we should be able to simplify this further.
1196 const SCEV *X = ST->getOperand();
1197 ConstantRange CR = getSignedRange(X);
1198 unsigned TruncBits = getTypeSizeInBits(ST->getType());
1199 unsigned NewBits = getTypeSizeInBits(Ty);
1200 if (CR.truncate(TruncBits).signExtend(NewBits).contains(
1201 CR.sextOrTrunc(NewBits)))
1202 return getTruncateOrSignExtend(X, Ty);
1205 // If the input value is a chrec scev, and we can prove that the value
1206 // did not overflow the old, smaller, value, we can sign extend all of the
1207 // operands (often constants). This allows analysis of something like
1208 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1209 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1210 if (AR->isAffine()) {
1211 const SCEV *Start = AR->getStart();
1212 const SCEV *Step = AR->getStepRecurrence(*this);
1213 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1214 const Loop *L = AR->getLoop();
1216 // If we have special knowledge that this addrec won't overflow,
1217 // we don't need to do any further analysis.
1218 if (AR->getNoWrapFlags(SCEV::FlagNSW))
1219 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1220 getSignExtendExpr(Step, Ty),
1223 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1224 // Note that this serves two purposes: It filters out loops that are
1225 // simply not analyzable, and it covers the case where this code is
1226 // being called from within backedge-taken count analysis, such that
1227 // attempting to ask for the backedge-taken count would likely result
1228 // in infinite recursion. In the later case, the analysis code will
1229 // cope with a conservative value, and it will take care to purge
1230 // that value once it has finished.
1231 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1232 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1233 // Manually compute the final value for AR, checking for
1236 // Check whether the backedge-taken count can be losslessly casted to
1237 // the addrec's type. The count is always unsigned.
1238 const SCEV *CastedMaxBECount =
1239 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1240 const SCEV *RecastedMaxBECount =
1241 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1242 if (MaxBECount == RecastedMaxBECount) {
1243 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1244 // Check whether Start+Step*MaxBECount has no signed overflow.
1245 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1246 const SCEV *SAdd = getSignExtendExpr(getAddExpr(Start, SMul), WideTy);
1247 const SCEV *WideStart = getSignExtendExpr(Start, WideTy);
1248 const SCEV *WideMaxBECount =
1249 getZeroExtendExpr(CastedMaxBECount, WideTy);
1250 const SCEV *OperandExtendedAdd =
1251 getAddExpr(WideStart,
1252 getMulExpr(WideMaxBECount,
1253 getSignExtendExpr(Step, WideTy)));
1254 if (SAdd == OperandExtendedAdd) {
1255 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1256 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1257 // Return the expression with the addrec on the outside.
1258 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1259 getSignExtendExpr(Step, Ty),
1260 L, AR->getNoWrapFlags());
1262 // Similar to above, only this time treat the step value as unsigned.
1263 // This covers loops that count up with an unsigned step.
1264 OperandExtendedAdd =
1265 getAddExpr(WideStart,
1266 getMulExpr(WideMaxBECount,
1267 getZeroExtendExpr(Step, WideTy)));
1268 if (SAdd == OperandExtendedAdd) {
1269 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1270 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1271 // Return the expression with the addrec on the outside.
1272 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1273 getZeroExtendExpr(Step, Ty),
1274 L, AR->getNoWrapFlags());
1278 // If the backedge is guarded by a comparison with the pre-inc value
1279 // the addrec is safe. Also, if the entry is guarded by a comparison
1280 // with the start value and the backedge is guarded by a comparison
1281 // with the post-inc value, the addrec is safe.
1282 ICmpInst::Predicate Pred;
1283 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, this);
1284 if (OverflowLimit &&
1285 (isLoopBackedgeGuardedByCond(L, Pred, AR, OverflowLimit) ||
1286 (isLoopEntryGuardedByCond(L, Pred, Start, OverflowLimit) &&
1287 isLoopBackedgeGuardedByCond(L, Pred, AR->getPostIncExpr(*this),
1289 // Cache knowledge of AR NSW, then propagate NSW to the wide AddRec.
1290 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1291 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1292 getSignExtendExpr(Step, Ty),
1293 L, AR->getNoWrapFlags());
1298 // The cast wasn't folded; create an explicit cast node.
1299 // Recompute the insert position, as it may have been invalidated.
1300 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1301 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1303 UniqueSCEVs.InsertNode(S, IP);
1307 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1308 /// unspecified bits out to the given type.
1310 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1312 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1313 "This is not an extending conversion!");
1314 assert(isSCEVable(Ty) &&
1315 "This is not a conversion to a SCEVable type!");
1316 Ty = getEffectiveSCEVType(Ty);
1318 // Sign-extend negative constants.
1319 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1320 if (SC->getValue()->getValue().isNegative())
1321 return getSignExtendExpr(Op, Ty);
1323 // Peel off a truncate cast.
1324 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1325 const SCEV *NewOp = T->getOperand();
1326 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1327 return getAnyExtendExpr(NewOp, Ty);
1328 return getTruncateOrNoop(NewOp, Ty);
1331 // Next try a zext cast. If the cast is folded, use it.
1332 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1333 if (!isa<SCEVZeroExtendExpr>(ZExt))
1336 // Next try a sext cast. If the cast is folded, use it.
1337 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1338 if (!isa<SCEVSignExtendExpr>(SExt))
1341 // Force the cast to be folded into the operands of an addrec.
1342 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1343 SmallVector<const SCEV *, 4> Ops;
1344 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1346 Ops.push_back(getAnyExtendExpr(*I, Ty));
1347 return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW);
1350 // If the expression is obviously signed, use the sext cast value.
1351 if (isa<SCEVSMaxExpr>(Op))
1354 // Absent any other information, use the zext cast value.
1358 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1359 /// a list of operands to be added under the given scale, update the given
1360 /// map. This is a helper function for getAddRecExpr. As an example of
1361 /// what it does, given a sequence of operands that would form an add
1362 /// expression like this:
1364 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1366 /// where A and B are constants, update the map with these values:
1368 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1370 /// and add 13 + A*B*29 to AccumulatedConstant.
1371 /// This will allow getAddRecExpr to produce this:
1373 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1375 /// This form often exposes folding opportunities that are hidden in
1376 /// the original operand list.
1378 /// Return true iff it appears that any interesting folding opportunities
1379 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1380 /// the common case where no interesting opportunities are present, and
1381 /// is also used as a check to avoid infinite recursion.
1384 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1385 SmallVectorImpl<const SCEV *> &NewOps,
1386 APInt &AccumulatedConstant,
1387 const SCEV *const *Ops, size_t NumOperands,
1389 ScalarEvolution &SE) {
1390 bool Interesting = false;
1392 // Iterate over the add operands. They are sorted, with constants first.
1394 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1396 // Pull a buried constant out to the outside.
1397 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1399 AccumulatedConstant += Scale * C->getValue()->getValue();
1402 // Next comes everything else. We're especially interested in multiplies
1403 // here, but they're in the middle, so just visit the rest with one loop.
1404 for (; i != NumOperands; ++i) {
1405 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1406 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1408 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1409 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1410 // A multiplication of a constant with another add; recurse.
1411 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1413 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1414 Add->op_begin(), Add->getNumOperands(),
1417 // A multiplication of a constant with some other value. Update
1419 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1420 const SCEV *Key = SE.getMulExpr(MulOps);
1421 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1422 M.insert(std::make_pair(Key, NewScale));
1424 NewOps.push_back(Pair.first->first);
1426 Pair.first->second += NewScale;
1427 // The map already had an entry for this value, which may indicate
1428 // a folding opportunity.
1433 // An ordinary operand. Update the map.
1434 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1435 M.insert(std::make_pair(Ops[i], Scale));
1437 NewOps.push_back(Pair.first->first);
1439 Pair.first->second += Scale;
1440 // The map already had an entry for this value, which may indicate
1441 // a folding opportunity.
1451 struct APIntCompare {
1452 bool operator()(const APInt &LHS, const APInt &RHS) const {
1453 return LHS.ult(RHS);
1458 /// getAddExpr - Get a canonical add expression, or something simpler if
1460 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1461 SCEV::NoWrapFlags Flags) {
1462 assert(!(Flags & ~(SCEV::FlagNUW | SCEV::FlagNSW)) &&
1463 "only nuw or nsw allowed");
1464 assert(!Ops.empty() && "Cannot get empty add!");
1465 if (Ops.size() == 1) return Ops[0];
1467 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1468 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1469 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1470 "SCEVAddExpr operand types don't match!");
1473 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1475 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1476 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1477 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1479 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1480 E = Ops.end(); I != E; ++I)
1481 if (!isKnownNonNegative(*I)) {
1485 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1488 // Sort by complexity, this groups all similar expression types together.
1489 GroupByComplexity(Ops, LI);
1491 // If there are any constants, fold them together.
1493 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1495 assert(Idx < Ops.size());
1496 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1497 // We found two constants, fold them together!
1498 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1499 RHSC->getValue()->getValue());
1500 if (Ops.size() == 2) return Ops[0];
1501 Ops.erase(Ops.begin()+1); // Erase the folded element
1502 LHSC = cast<SCEVConstant>(Ops[0]);
1505 // If we are left with a constant zero being added, strip it off.
1506 if (LHSC->getValue()->isZero()) {
1507 Ops.erase(Ops.begin());
1511 if (Ops.size() == 1) return Ops[0];
1514 // Okay, check to see if the same value occurs in the operand list more than
1515 // once. If so, merge them together into an multiply expression. Since we
1516 // sorted the list, these values are required to be adjacent.
1517 Type *Ty = Ops[0]->getType();
1518 bool FoundMatch = false;
1519 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1520 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1521 // Scan ahead to count how many equal operands there are.
1523 while (i+Count != e && Ops[i+Count] == Ops[i])
1525 // Merge the values into a multiply.
1526 const SCEV *Scale = getConstant(Ty, Count);
1527 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1528 if (Ops.size() == Count)
1531 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1532 --i; e -= Count - 1;
1536 return getAddExpr(Ops, Flags);
1538 // Check for truncates. If all the operands are truncated from the same
1539 // type, see if factoring out the truncate would permit the result to be
1540 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1541 // if the contents of the resulting outer trunc fold to something simple.
1542 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1543 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1544 Type *DstType = Trunc->getType();
1545 Type *SrcType = Trunc->getOperand()->getType();
1546 SmallVector<const SCEV *, 8> LargeOps;
1548 // Check all the operands to see if they can be represented in the
1549 // source type of the truncate.
1550 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1551 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1552 if (T->getOperand()->getType() != SrcType) {
1556 LargeOps.push_back(T->getOperand());
1557 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1558 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1559 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1560 SmallVector<const SCEV *, 8> LargeMulOps;
1561 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1562 if (const SCEVTruncateExpr *T =
1563 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1564 if (T->getOperand()->getType() != SrcType) {
1568 LargeMulOps.push_back(T->getOperand());
1569 } else if (const SCEVConstant *C =
1570 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1571 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1578 LargeOps.push_back(getMulExpr(LargeMulOps));
1585 // Evaluate the expression in the larger type.
1586 const SCEV *Fold = getAddExpr(LargeOps, Flags);
1587 // If it folds to something simple, use it. Otherwise, don't.
1588 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1589 return getTruncateExpr(Fold, DstType);
1593 // Skip past any other cast SCEVs.
1594 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1597 // If there are add operands they would be next.
1598 if (Idx < Ops.size()) {
1599 bool DeletedAdd = false;
1600 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1601 // If we have an add, expand the add operands onto the end of the operands
1603 Ops.erase(Ops.begin()+Idx);
1604 Ops.append(Add->op_begin(), Add->op_end());
1608 // If we deleted at least one add, we added operands to the end of the list,
1609 // and they are not necessarily sorted. Recurse to resort and resimplify
1610 // any operands we just acquired.
1612 return getAddExpr(Ops);
1615 // Skip over the add expression until we get to a multiply.
1616 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1619 // Check to see if there are any folding opportunities present with
1620 // operands multiplied by constant values.
1621 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1622 uint64_t BitWidth = getTypeSizeInBits(Ty);
1623 DenseMap<const SCEV *, APInt> M;
1624 SmallVector<const SCEV *, 8> NewOps;
1625 APInt AccumulatedConstant(BitWidth, 0);
1626 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1627 Ops.data(), Ops.size(),
1628 APInt(BitWidth, 1), *this)) {
1629 // Some interesting folding opportunity is present, so its worthwhile to
1630 // re-generate the operands list. Group the operands by constant scale,
1631 // to avoid multiplying by the same constant scale multiple times.
1632 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1633 for (SmallVectorImpl<const SCEV *>::const_iterator I = NewOps.begin(),
1634 E = NewOps.end(); I != E; ++I)
1635 MulOpLists[M.find(*I)->second].push_back(*I);
1636 // Re-generate the operands list.
1638 if (AccumulatedConstant != 0)
1639 Ops.push_back(getConstant(AccumulatedConstant));
1640 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1641 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1643 Ops.push_back(getMulExpr(getConstant(I->first),
1644 getAddExpr(I->second)));
1646 return getConstant(Ty, 0);
1647 if (Ops.size() == 1)
1649 return getAddExpr(Ops);
1653 // If we are adding something to a multiply expression, make sure the
1654 // something is not already an operand of the multiply. If so, merge it into
1656 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1657 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1658 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1659 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1660 if (isa<SCEVConstant>(MulOpSCEV))
1662 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1663 if (MulOpSCEV == Ops[AddOp]) {
1664 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1665 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1666 if (Mul->getNumOperands() != 2) {
1667 // If the multiply has more than two operands, we must get the
1669 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1670 Mul->op_begin()+MulOp);
1671 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1672 InnerMul = getMulExpr(MulOps);
1674 const SCEV *One = getConstant(Ty, 1);
1675 const SCEV *AddOne = getAddExpr(One, InnerMul);
1676 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1677 if (Ops.size() == 2) return OuterMul;
1679 Ops.erase(Ops.begin()+AddOp);
1680 Ops.erase(Ops.begin()+Idx-1);
1682 Ops.erase(Ops.begin()+Idx);
1683 Ops.erase(Ops.begin()+AddOp-1);
1685 Ops.push_back(OuterMul);
1686 return getAddExpr(Ops);
1689 // Check this multiply against other multiplies being added together.
1690 for (unsigned OtherMulIdx = Idx+1;
1691 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1693 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1694 // If MulOp occurs in OtherMul, we can fold the two multiplies
1696 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1697 OMulOp != e; ++OMulOp)
1698 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1699 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1700 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1701 if (Mul->getNumOperands() != 2) {
1702 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1703 Mul->op_begin()+MulOp);
1704 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1705 InnerMul1 = getMulExpr(MulOps);
1707 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1708 if (OtherMul->getNumOperands() != 2) {
1709 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1710 OtherMul->op_begin()+OMulOp);
1711 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1712 InnerMul2 = getMulExpr(MulOps);
1714 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1715 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1716 if (Ops.size() == 2) return OuterMul;
1717 Ops.erase(Ops.begin()+Idx);
1718 Ops.erase(Ops.begin()+OtherMulIdx-1);
1719 Ops.push_back(OuterMul);
1720 return getAddExpr(Ops);
1726 // If there are any add recurrences in the operands list, see if any other
1727 // added values are loop invariant. If so, we can fold them into the
1729 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1732 // Scan over all recurrences, trying to fold loop invariants into them.
1733 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1734 // Scan all of the other operands to this add and add them to the vector if
1735 // they are loop invariant w.r.t. the recurrence.
1736 SmallVector<const SCEV *, 8> LIOps;
1737 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1738 const Loop *AddRecLoop = AddRec->getLoop();
1739 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1740 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1741 LIOps.push_back(Ops[i]);
1742 Ops.erase(Ops.begin()+i);
1746 // If we found some loop invariants, fold them into the recurrence.
1747 if (!LIOps.empty()) {
1748 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1749 LIOps.push_back(AddRec->getStart());
1751 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1753 AddRecOps[0] = getAddExpr(LIOps);
1755 // Build the new addrec. Propagate the NUW and NSW flags if both the
1756 // outer add and the inner addrec are guaranteed to have no overflow.
1757 // Always propagate NW.
1758 Flags = AddRec->getNoWrapFlags(setFlags(Flags, SCEV::FlagNW));
1759 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop, Flags);
1761 // If all of the other operands were loop invariant, we are done.
1762 if (Ops.size() == 1) return NewRec;
1764 // Otherwise, add the folded AddRec by the non-invariant parts.
1765 for (unsigned i = 0;; ++i)
1766 if (Ops[i] == AddRec) {
1770 return getAddExpr(Ops);
1773 // Okay, if there weren't any loop invariants to be folded, check to see if
1774 // there are multiple AddRec's with the same loop induction variable being
1775 // added together. If so, we can fold them.
1776 for (unsigned OtherIdx = Idx+1;
1777 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1779 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1780 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1781 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1783 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1785 if (const SCEVAddRecExpr *OtherAddRec =
1786 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1787 if (OtherAddRec->getLoop() == AddRecLoop) {
1788 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1790 if (i >= AddRecOps.size()) {
1791 AddRecOps.append(OtherAddRec->op_begin()+i,
1792 OtherAddRec->op_end());
1795 AddRecOps[i] = getAddExpr(AddRecOps[i],
1796 OtherAddRec->getOperand(i));
1798 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1800 // Step size has changed, so we cannot guarantee no self-wraparound.
1801 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop, SCEV::FlagAnyWrap);
1802 return getAddExpr(Ops);
1805 // Otherwise couldn't fold anything into this recurrence. Move onto the
1809 // Okay, it looks like we really DO need an add expr. Check to see if we
1810 // already have one, otherwise create a new one.
1811 FoldingSetNodeID ID;
1812 ID.AddInteger(scAddExpr);
1813 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1814 ID.AddPointer(Ops[i]);
1817 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1819 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1820 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1821 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1823 UniqueSCEVs.InsertNode(S, IP);
1825 S->setNoWrapFlags(Flags);
1829 static uint64_t umul_ov(uint64_t i, uint64_t j, bool &Overflow) {
1831 if (j > 1 && k / j != i) Overflow = true;
1835 /// Compute the result of "n choose k", the binomial coefficient. If an
1836 /// intermediate computation overflows, Overflow will be set and the return will
1837 /// be garbage. Overflow is not cleared on absence of overflow.
1838 static uint64_t Choose(uint64_t n, uint64_t k, bool &Overflow) {
1839 // We use the multiplicative formula:
1840 // n(n-1)(n-2)...(n-(k-1)) / k(k-1)(k-2)...1 .
1841 // At each iteration, we take the n-th term of the numeral and divide by the
1842 // (k-n)th term of the denominator. This division will always produce an
1843 // integral result, and helps reduce the chance of overflow in the
1844 // intermediate computations. However, we can still overflow even when the
1845 // final result would fit.
1847 if (n == 0 || n == k) return 1;
1848 if (k > n) return 0;
1854 for (uint64_t i = 1; i <= k; ++i) {
1855 r = umul_ov(r, n-(i-1), Overflow);
1861 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1863 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1864 SCEV::NoWrapFlags Flags) {
1865 assert(Flags == maskFlags(Flags, SCEV::FlagNUW | SCEV::FlagNSW) &&
1866 "only nuw or nsw allowed");
1867 assert(!Ops.empty() && "Cannot get empty mul!");
1868 if (Ops.size() == 1) return Ops[0];
1870 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1871 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1872 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1873 "SCEVMulExpr operand types don't match!");
1876 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1878 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1879 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1880 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1882 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1883 E = Ops.end(); I != E; ++I)
1884 if (!isKnownNonNegative(*I)) {
1888 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1891 // Sort by complexity, this groups all similar expression types together.
1892 GroupByComplexity(Ops, LI);
1894 // If there are any constants, fold them together.
1896 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1898 // C1*(C2+V) -> C1*C2 + C1*V
1899 if (Ops.size() == 2)
1900 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1901 if (Add->getNumOperands() == 2 &&
1902 isa<SCEVConstant>(Add->getOperand(0)))
1903 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1904 getMulExpr(LHSC, Add->getOperand(1)));
1907 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1908 // We found two constants, fold them together!
1909 ConstantInt *Fold = ConstantInt::get(getContext(),
1910 LHSC->getValue()->getValue() *
1911 RHSC->getValue()->getValue());
1912 Ops[0] = getConstant(Fold);
1913 Ops.erase(Ops.begin()+1); // Erase the folded element
1914 if (Ops.size() == 1) return Ops[0];
1915 LHSC = cast<SCEVConstant>(Ops[0]);
1918 // If we are left with a constant one being multiplied, strip it off.
1919 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1920 Ops.erase(Ops.begin());
1922 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1923 // If we have a multiply of zero, it will always be zero.
1925 } else if (Ops[0]->isAllOnesValue()) {
1926 // If we have a mul by -1 of an add, try distributing the -1 among the
1928 if (Ops.size() == 2) {
1929 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1930 SmallVector<const SCEV *, 4> NewOps;
1931 bool AnyFolded = false;
1932 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(),
1933 E = Add->op_end(); I != E; ++I) {
1934 const SCEV *Mul = getMulExpr(Ops[0], *I);
1935 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1936 NewOps.push_back(Mul);
1939 return getAddExpr(NewOps);
1941 else if (const SCEVAddRecExpr *
1942 AddRec = dyn_cast<SCEVAddRecExpr>(Ops[1])) {
1943 // Negation preserves a recurrence's no self-wrap property.
1944 SmallVector<const SCEV *, 4> Operands;
1945 for (SCEVAddRecExpr::op_iterator I = AddRec->op_begin(),
1946 E = AddRec->op_end(); I != E; ++I) {
1947 Operands.push_back(getMulExpr(Ops[0], *I));
1949 return getAddRecExpr(Operands, AddRec->getLoop(),
1950 AddRec->getNoWrapFlags(SCEV::FlagNW));
1955 if (Ops.size() == 1)
1959 // Skip over the add expression until we get to a multiply.
1960 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1963 // If there are mul operands inline them all into this expression.
1964 if (Idx < Ops.size()) {
1965 bool DeletedMul = false;
1966 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1967 // If we have an mul, expand the mul operands onto the end of the operands
1969 Ops.erase(Ops.begin()+Idx);
1970 Ops.append(Mul->op_begin(), Mul->op_end());
1974 // If we deleted at least one mul, we added operands to the end of the list,
1975 // and they are not necessarily sorted. Recurse to resort and resimplify
1976 // any operands we just acquired.
1978 return getMulExpr(Ops);
1981 // If there are any add recurrences in the operands list, see if any other
1982 // added values are loop invariant. If so, we can fold them into the
1984 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1987 // Scan over all recurrences, trying to fold loop invariants into them.
1988 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1989 // Scan all of the other operands to this mul and add them to the vector if
1990 // they are loop invariant w.r.t. the recurrence.
1991 SmallVector<const SCEV *, 8> LIOps;
1992 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1993 const Loop *AddRecLoop = AddRec->getLoop();
1994 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1995 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1996 LIOps.push_back(Ops[i]);
1997 Ops.erase(Ops.begin()+i);
2001 // If we found some loop invariants, fold them into the recurrence.
2002 if (!LIOps.empty()) {
2003 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
2004 SmallVector<const SCEV *, 4> NewOps;
2005 NewOps.reserve(AddRec->getNumOperands());
2006 const SCEV *Scale = getMulExpr(LIOps);
2007 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
2008 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
2010 // Build the new addrec. Propagate the NUW and NSW flags if both the
2011 // outer mul and the inner addrec are guaranteed to have no overflow.
2013 // No self-wrap cannot be guaranteed after changing the step size, but
2014 // will be inferred if either NUW or NSW is true.
2015 Flags = AddRec->getNoWrapFlags(clearFlags(Flags, SCEV::FlagNW));
2016 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop, Flags);
2018 // If all of the other operands were loop invariant, we are done.
2019 if (Ops.size() == 1) return NewRec;
2021 // Otherwise, multiply the folded AddRec by the non-invariant parts.
2022 for (unsigned i = 0;; ++i)
2023 if (Ops[i] == AddRec) {
2027 return getMulExpr(Ops);
2030 // Okay, if there weren't any loop invariants to be folded, check to see if
2031 // there are multiple AddRec's with the same loop induction variable being
2032 // multiplied together. If so, we can fold them.
2033 for (unsigned OtherIdx = Idx+1;
2034 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2036 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 const SCEVAddRecExpr *OtherAddRec =
2053 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]);
2054 if (!OtherAddRec || OtherAddRec->getLoop() != AddRecLoop)
2057 bool Overflow = false;
2058 Type *Ty = AddRec->getType();
2059 bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64;
2060 SmallVector<const SCEV*, 7> AddRecOps;
2061 for (int x = 0, xe = AddRec->getNumOperands() +
2062 OtherAddRec->getNumOperands() - 1; x != xe && !Overflow; ++x) {
2063 const SCEV *Term = getConstant(Ty, 0);
2064 for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) {
2065 uint64_t Coeff1 = Choose(x, 2*x - y, Overflow);
2066 for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1),
2067 ze = std::min(x+1, (int)OtherAddRec->getNumOperands());
2068 z < ze && !Overflow; ++z) {
2069 uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow);
2071 if (LargerThan64Bits)
2072 Coeff = umul_ov(Coeff1, Coeff2, Overflow);
2074 Coeff = Coeff1*Coeff2;
2075 const SCEV *CoeffTerm = getConstant(Ty, Coeff);
2076 const SCEV *Term1 = AddRec->getOperand(y-z);
2077 const SCEV *Term2 = OtherAddRec->getOperand(z);
2078 Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2));
2081 AddRecOps.push_back(Term);
2084 const SCEV *NewAddRec = getAddRecExpr(AddRecOps, AddRec->getLoop(),
2086 if (Ops.size() == 2) return NewAddRec;
2087 Ops[Idx] = NewAddRec;
2088 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
2090 AddRec = dyn_cast<SCEVAddRecExpr>(NewAddRec);
2096 return getMulExpr(Ops);
2099 // Otherwise couldn't fold anything into this recurrence. Move onto the
2103 // Okay, it looks like we really DO need an mul expr. Check to see if we
2104 // already have one, otherwise create a new one.
2105 FoldingSetNodeID ID;
2106 ID.AddInteger(scMulExpr);
2107 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2108 ID.AddPointer(Ops[i]);
2111 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2113 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2114 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2115 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
2117 UniqueSCEVs.InsertNode(S, IP);
2119 S->setNoWrapFlags(Flags);
2123 /// getUDivExpr - Get a canonical unsigned division expression, or something
2124 /// simpler if possible.
2125 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
2127 assert(getEffectiveSCEVType(LHS->getType()) ==
2128 getEffectiveSCEVType(RHS->getType()) &&
2129 "SCEVUDivExpr operand types don't match!");
2131 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
2132 if (RHSC->getValue()->equalsInt(1))
2133 return LHS; // X udiv 1 --> x
2134 // If the denominator is zero, the result of the udiv is undefined. Don't
2135 // try to analyze it, because the resolution chosen here may differ from
2136 // the resolution chosen in other parts of the compiler.
2137 if (!RHSC->getValue()->isZero()) {
2138 // Determine if the division can be folded into the operands of
2140 // TODO: Generalize this to non-constants by using known-bits information.
2141 Type *Ty = LHS->getType();
2142 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
2143 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
2144 // For non-power-of-two values, effectively round the value up to the
2145 // nearest power of two.
2146 if (!RHSC->getValue()->getValue().isPowerOf2())
2148 IntegerType *ExtTy =
2149 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
2150 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
2151 if (const SCEVConstant *Step =
2152 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this))) {
2153 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
2154 const APInt &StepInt = Step->getValue()->getValue();
2155 const APInt &DivInt = RHSC->getValue()->getValue();
2156 if (!StepInt.urem(DivInt) &&
2157 getZeroExtendExpr(AR, ExtTy) ==
2158 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2159 getZeroExtendExpr(Step, ExtTy),
2160 AR->getLoop(), SCEV::FlagAnyWrap)) {
2161 SmallVector<const SCEV *, 4> Operands;
2162 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
2163 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
2164 return getAddRecExpr(Operands, AR->getLoop(),
2167 /// Get a canonical UDivExpr for a recurrence.
2168 /// {X,+,N}/C => {Y,+,N}/C where Y=X-(X%N). Safe when C%N=0.
2169 // We can currently only fold X%N if X is constant.
2170 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
2171 if (StartC && !DivInt.urem(StepInt) &&
2172 getZeroExtendExpr(AR, ExtTy) ==
2173 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2174 getZeroExtendExpr(Step, ExtTy),
2175 AR->getLoop(), SCEV::FlagAnyWrap)) {
2176 const APInt &StartInt = StartC->getValue()->getValue();
2177 const APInt &StartRem = StartInt.urem(StepInt);
2179 LHS = getAddRecExpr(getConstant(StartInt - StartRem), Step,
2180 AR->getLoop(), SCEV::FlagNW);
2183 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
2184 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
2185 SmallVector<const SCEV *, 4> Operands;
2186 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
2187 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
2188 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
2189 // Find an operand that's safely divisible.
2190 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
2191 const SCEV *Op = M->getOperand(i);
2192 const SCEV *Div = getUDivExpr(Op, RHSC);
2193 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
2194 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
2197 return getMulExpr(Operands);
2201 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
2202 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(LHS)) {
2203 SmallVector<const SCEV *, 4> Operands;
2204 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
2205 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
2206 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
2208 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2209 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
2210 if (isa<SCEVUDivExpr>(Op) ||
2211 getMulExpr(Op, RHS) != A->getOperand(i))
2213 Operands.push_back(Op);
2215 if (Operands.size() == A->getNumOperands())
2216 return getAddExpr(Operands);
2220 // Fold if both operands are constant.
2221 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
2222 Constant *LHSCV = LHSC->getValue();
2223 Constant *RHSCV = RHSC->getValue();
2224 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
2230 FoldingSetNodeID ID;
2231 ID.AddInteger(scUDivExpr);
2235 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2236 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
2238 UniqueSCEVs.InsertNode(S, IP);
2243 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2244 /// Simplify the expression as much as possible.
2245 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start, const SCEV *Step,
2247 SCEV::NoWrapFlags Flags) {
2248 SmallVector<const SCEV *, 4> Operands;
2249 Operands.push_back(Start);
2250 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
2251 if (StepChrec->getLoop() == L) {
2252 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
2253 return getAddRecExpr(Operands, L, maskFlags(Flags, SCEV::FlagNW));
2256 Operands.push_back(Step);
2257 return getAddRecExpr(Operands, L, Flags);
2260 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2261 /// Simplify the expression as much as possible.
2263 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2264 const Loop *L, SCEV::NoWrapFlags Flags) {
2265 if (Operands.size() == 1) return Operands[0];
2267 Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2268 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2269 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2270 "SCEVAddRecExpr operand types don't match!");
2271 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2272 assert(isLoopInvariant(Operands[i], L) &&
2273 "SCEVAddRecExpr operand is not loop-invariant!");
2276 if (Operands.back()->isZero()) {
2277 Operands.pop_back();
2278 return getAddRecExpr(Operands, L, SCEV::FlagAnyWrap); // {X,+,0} --> X
2281 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2282 // use that information to infer NUW and NSW flags. However, computing a
2283 // BE count requires calling getAddRecExpr, so we may not yet have a
2284 // meaningful BE count at this point (and if we don't, we'd be stuck
2285 // with a SCEVCouldNotCompute as the cached BE count).
2287 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
2289 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
2290 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
2291 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
2293 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2294 E = Operands.end(); I != E; ++I)
2295 if (!isKnownNonNegative(*I)) {
2299 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
2302 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2303 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2304 const Loop *NestedLoop = NestedAR->getLoop();
2305 if (L->contains(NestedLoop) ?
2306 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2307 (!NestedLoop->contains(L) &&
2308 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2309 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2310 NestedAR->op_end());
2311 Operands[0] = NestedAR->getStart();
2312 // AddRecs require their operands be loop-invariant with respect to their
2313 // loops. Don't perform this transformation if it would break this
2315 bool AllInvariant = true;
2316 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2317 if (!isLoopInvariant(Operands[i], L)) {
2318 AllInvariant = false;
2322 // Create a recurrence for the outer loop with the same step size.
2324 // The outer recurrence keeps its NW flag but only keeps NUW/NSW if the
2325 // inner recurrence has the same property.
2326 SCEV::NoWrapFlags OuterFlags =
2327 maskFlags(Flags, SCEV::FlagNW | NestedAR->getNoWrapFlags());
2329 NestedOperands[0] = getAddRecExpr(Operands, L, OuterFlags);
2330 AllInvariant = true;
2331 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2332 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2333 AllInvariant = false;
2337 // Ok, both add recurrences are valid after the transformation.
2339 // The inner recurrence keeps its NW flag but only keeps NUW/NSW if
2340 // the outer recurrence has the same property.
2341 SCEV::NoWrapFlags InnerFlags =
2342 maskFlags(NestedAR->getNoWrapFlags(), SCEV::FlagNW | Flags);
2343 return getAddRecExpr(NestedOperands, NestedLoop, InnerFlags);
2346 // Reset Operands to its original state.
2347 Operands[0] = NestedAR;
2351 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2352 // already have one, otherwise create a new one.
2353 FoldingSetNodeID ID;
2354 ID.AddInteger(scAddRecExpr);
2355 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2356 ID.AddPointer(Operands[i]);
2360 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2362 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2363 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2364 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2365 O, Operands.size(), L);
2366 UniqueSCEVs.InsertNode(S, IP);
2368 S->setNoWrapFlags(Flags);
2372 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2374 SmallVector<const SCEV *, 2> Ops;
2377 return getSMaxExpr(Ops);
2381 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2382 assert(!Ops.empty() && "Cannot get empty smax!");
2383 if (Ops.size() == 1) return Ops[0];
2385 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2386 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2387 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2388 "SCEVSMaxExpr operand types don't match!");
2391 // Sort by complexity, this groups all similar expression types together.
2392 GroupByComplexity(Ops, LI);
2394 // If there are any constants, fold them together.
2396 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2398 assert(Idx < Ops.size());
2399 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2400 // We found two constants, fold them together!
2401 ConstantInt *Fold = ConstantInt::get(getContext(),
2402 APIntOps::smax(LHSC->getValue()->getValue(),
2403 RHSC->getValue()->getValue()));
2404 Ops[0] = getConstant(Fold);
2405 Ops.erase(Ops.begin()+1); // Erase the folded element
2406 if (Ops.size() == 1) return Ops[0];
2407 LHSC = cast<SCEVConstant>(Ops[0]);
2410 // If we are left with a constant minimum-int, strip it off.
2411 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2412 Ops.erase(Ops.begin());
2414 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2415 // If we have an smax with a constant maximum-int, it will always be
2420 if (Ops.size() == 1) return Ops[0];
2423 // Find the first SMax
2424 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2427 // Check to see if one of the operands is an SMax. If so, expand its operands
2428 // onto our operand list, and recurse to simplify.
2429 if (Idx < Ops.size()) {
2430 bool DeletedSMax = false;
2431 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2432 Ops.erase(Ops.begin()+Idx);
2433 Ops.append(SMax->op_begin(), SMax->op_end());
2438 return getSMaxExpr(Ops);
2441 // Okay, check to see if the same value occurs in the operand list twice. If
2442 // so, delete one. Since we sorted the list, these values are required to
2444 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2445 // X smax Y smax Y --> X smax Y
2446 // X smax Y --> X, if X is always greater than Y
2447 if (Ops[i] == Ops[i+1] ||
2448 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2449 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2451 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2452 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2456 if (Ops.size() == 1) return Ops[0];
2458 assert(!Ops.empty() && "Reduced smax down to nothing!");
2460 // Okay, it looks like we really DO need an smax expr. Check to see if we
2461 // already have one, otherwise create a new one.
2462 FoldingSetNodeID ID;
2463 ID.AddInteger(scSMaxExpr);
2464 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2465 ID.AddPointer(Ops[i]);
2467 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2468 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2469 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2470 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2472 UniqueSCEVs.InsertNode(S, IP);
2476 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2478 SmallVector<const SCEV *, 2> Ops;
2481 return getUMaxExpr(Ops);
2485 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2486 assert(!Ops.empty() && "Cannot get empty umax!");
2487 if (Ops.size() == 1) return Ops[0];
2489 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2490 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2491 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2492 "SCEVUMaxExpr operand types don't match!");
2495 // Sort by complexity, this groups all similar expression types together.
2496 GroupByComplexity(Ops, LI);
2498 // If there are any constants, fold them together.
2500 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2502 assert(Idx < Ops.size());
2503 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2504 // We found two constants, fold them together!
2505 ConstantInt *Fold = ConstantInt::get(getContext(),
2506 APIntOps::umax(LHSC->getValue()->getValue(),
2507 RHSC->getValue()->getValue()));
2508 Ops[0] = getConstant(Fold);
2509 Ops.erase(Ops.begin()+1); // Erase the folded element
2510 if (Ops.size() == 1) return Ops[0];
2511 LHSC = cast<SCEVConstant>(Ops[0]);
2514 // If we are left with a constant minimum-int, strip it off.
2515 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2516 Ops.erase(Ops.begin());
2518 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2519 // If we have an umax with a constant maximum-int, it will always be
2524 if (Ops.size() == 1) return Ops[0];
2527 // Find the first UMax
2528 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2531 // Check to see if one of the operands is a UMax. If so, expand its operands
2532 // onto our operand list, and recurse to simplify.
2533 if (Idx < Ops.size()) {
2534 bool DeletedUMax = false;
2535 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2536 Ops.erase(Ops.begin()+Idx);
2537 Ops.append(UMax->op_begin(), UMax->op_end());
2542 return getUMaxExpr(Ops);
2545 // Okay, check to see if the same value occurs in the operand list twice. If
2546 // so, delete one. Since we sorted the list, these values are required to
2548 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2549 // X umax Y umax Y --> X umax Y
2550 // X umax Y --> X, if X is always greater than Y
2551 if (Ops[i] == Ops[i+1] ||
2552 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2553 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2555 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2556 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2560 if (Ops.size() == 1) return Ops[0];
2562 assert(!Ops.empty() && "Reduced umax down to nothing!");
2564 // Okay, it looks like we really DO need a umax expr. Check to see if we
2565 // already have one, otherwise create a new one.
2566 FoldingSetNodeID ID;
2567 ID.AddInteger(scUMaxExpr);
2568 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2569 ID.AddPointer(Ops[i]);
2571 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2572 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2573 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2574 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2576 UniqueSCEVs.InsertNode(S, IP);
2580 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2582 // ~smax(~x, ~y) == smin(x, y).
2583 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2586 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2588 // ~umax(~x, ~y) == umin(x, y)
2589 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2592 const SCEV *ScalarEvolution::getSizeOfExpr(Type *IntTy, Type *AllocTy) {
2593 // If we have DataLayout, we can bypass creating a target-independent
2594 // constant expression and then folding it back into a ConstantInt.
2595 // This is just a compile-time optimization.
2597 return getConstant(IntTy, TD->getTypeAllocSize(AllocTy));
2599 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2600 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2601 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2603 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2604 assert(Ty == IntTy && "Effective SCEV type doesn't match");
2605 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2608 const SCEV *ScalarEvolution::getOffsetOfExpr(Type *IntTy,
2611 // If we have DataLayout, we can bypass creating a target-independent
2612 // constant expression and then folding it back into a ConstantInt.
2613 // This is just a compile-time optimization.
2615 return getConstant(IntTy,
2616 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2619 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2620 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2621 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2624 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2625 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2628 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2629 // Don't attempt to do anything other than create a SCEVUnknown object
2630 // here. createSCEV only calls getUnknown after checking for all other
2631 // interesting possibilities, and any other code that calls getUnknown
2632 // is doing so in order to hide a value from SCEV canonicalization.
2634 FoldingSetNodeID ID;
2635 ID.AddInteger(scUnknown);
2638 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2639 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2640 "Stale SCEVUnknown in uniquing map!");
2643 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2645 FirstUnknown = cast<SCEVUnknown>(S);
2646 UniqueSCEVs.InsertNode(S, IP);
2650 //===----------------------------------------------------------------------===//
2651 // Basic SCEV Analysis and PHI Idiom Recognition Code
2654 /// isSCEVable - Test if values of the given type are analyzable within
2655 /// the SCEV framework. This primarily includes integer types, and it
2656 /// can optionally include pointer types if the ScalarEvolution class
2657 /// has access to target-specific information.
2658 bool ScalarEvolution::isSCEVable(Type *Ty) const {
2659 // Integers and pointers are always SCEVable.
2660 return Ty->isIntegerTy() || Ty->isPointerTy();
2663 /// getTypeSizeInBits - Return the size in bits of the specified type,
2664 /// for which isSCEVable must return true.
2665 uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const {
2666 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2668 // If we have a DataLayout, use it!
2670 return TD->getTypeSizeInBits(Ty);
2672 // Integer types have fixed sizes.
2673 if (Ty->isIntegerTy())
2674 return Ty->getPrimitiveSizeInBits();
2676 // The only other support type is pointer. Without DataLayout, conservatively
2677 // assume pointers are 64-bit.
2678 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2682 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2683 /// the given type and which represents how SCEV will treat the given
2684 /// type, for which isSCEVable must return true. For pointer types,
2685 /// this is the pointer-sized integer type.
2686 Type *ScalarEvolution::getEffectiveSCEVType(Type *Ty) const {
2687 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2689 if (Ty->isIntegerTy()) {
2693 // The only other support type is pointer.
2694 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2697 return TD->getIntPtrType(Ty);
2699 // Without DataLayout, conservatively assume pointers are 64-bit.
2700 return Type::getInt64Ty(getContext());
2703 const SCEV *ScalarEvolution::getCouldNotCompute() {
2704 return &CouldNotCompute;
2708 // Helper class working with SCEVTraversal to figure out if a SCEV contains
2709 // a SCEVUnknown with null value-pointer. FindInvalidSCEVUnknown::FindOne
2710 // is set iff if find such SCEVUnknown.
2712 struct FindInvalidSCEVUnknown {
2714 FindInvalidSCEVUnknown() { FindOne = false; }
2715 bool follow(const SCEV *S) {
2716 switch (S->getSCEVType()) {
2720 if (!cast<SCEVUnknown>(S)->getValue())
2727 bool isDone() const { return FindOne; }
2731 bool ScalarEvolution::checkValidity(const SCEV *S) const {
2732 FindInvalidSCEVUnknown F;
2733 SCEVTraversal<FindInvalidSCEVUnknown> ST(F);
2739 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2740 /// expression and create a new one.
2741 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2742 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2744 ValueExprMapType::iterator I = ValueExprMap.find_as(V);
2745 if (I != ValueExprMap.end()) {
2746 const SCEV *S = I->second;
2747 if (checkValidity(S))
2750 ValueExprMap.erase(I);
2752 const SCEV *S = createSCEV(V);
2754 // The process of creating a SCEV for V may have caused other SCEVs
2755 // to have been created, so it's necessary to insert the new entry
2756 // from scratch, rather than trying to remember the insert position
2758 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2762 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2764 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2765 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2767 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2769 Type *Ty = V->getType();
2770 Ty = getEffectiveSCEVType(Ty);
2771 return getMulExpr(V,
2772 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2775 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2776 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2777 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2779 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2781 Type *Ty = V->getType();
2782 Ty = getEffectiveSCEVType(Ty);
2783 const SCEV *AllOnes =
2784 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2785 return getMinusSCEV(AllOnes, V);
2788 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1.
2789 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2790 SCEV::NoWrapFlags Flags) {
2791 assert(!maskFlags(Flags, SCEV::FlagNUW) && "subtraction does not have NUW");
2793 // Fast path: X - X --> 0.
2795 return getConstant(LHS->getType(), 0);
2798 return getAddExpr(LHS, getNegativeSCEV(RHS), Flags);
2801 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2802 /// input value to the specified type. If the type must be extended, it is zero
2805 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, Type *Ty) {
2806 Type *SrcTy = V->getType();
2807 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2808 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2809 "Cannot truncate or zero extend with non-integer arguments!");
2810 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2811 return V; // No conversion
2812 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2813 return getTruncateExpr(V, Ty);
2814 return getZeroExtendExpr(V, Ty);
2817 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2818 /// input value to the specified type. If the type must be extended, it is sign
2821 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2823 Type *SrcTy = V->getType();
2824 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2825 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2826 "Cannot truncate or zero extend with non-integer arguments!");
2827 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2828 return V; // No conversion
2829 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2830 return getTruncateExpr(V, Ty);
2831 return getSignExtendExpr(V, Ty);
2834 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2835 /// input value to the specified type. If the type must be extended, it is zero
2836 /// extended. The conversion must not be narrowing.
2838 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, Type *Ty) {
2839 Type *SrcTy = V->getType();
2840 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2841 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2842 "Cannot noop or zero extend with non-integer arguments!");
2843 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2844 "getNoopOrZeroExtend cannot truncate!");
2845 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2846 return V; // No conversion
2847 return getZeroExtendExpr(V, Ty);
2850 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2851 /// input value to the specified type. If the type must be extended, it is sign
2852 /// extended. The conversion must not be narrowing.
2854 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, Type *Ty) {
2855 Type *SrcTy = V->getType();
2856 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2857 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2858 "Cannot noop or sign extend with non-integer arguments!");
2859 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2860 "getNoopOrSignExtend cannot truncate!");
2861 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2862 return V; // No conversion
2863 return getSignExtendExpr(V, Ty);
2866 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2867 /// the input value to the specified type. If the type must be extended,
2868 /// it is extended with unspecified bits. The conversion must not be
2871 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, Type *Ty) {
2872 Type *SrcTy = V->getType();
2873 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2874 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2875 "Cannot noop or any extend with non-integer arguments!");
2876 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2877 "getNoopOrAnyExtend cannot truncate!");
2878 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2879 return V; // No conversion
2880 return getAnyExtendExpr(V, Ty);
2883 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2884 /// input value to the specified type. The conversion must not be widening.
2886 ScalarEvolution::getTruncateOrNoop(const SCEV *V, Type *Ty) {
2887 Type *SrcTy = V->getType();
2888 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2889 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2890 "Cannot truncate or noop with non-integer arguments!");
2891 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2892 "getTruncateOrNoop cannot extend!");
2893 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2894 return V; // No conversion
2895 return getTruncateExpr(V, Ty);
2898 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2899 /// the types using zero-extension, and then perform a umax operation
2901 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2903 const SCEV *PromotedLHS = LHS;
2904 const SCEV *PromotedRHS = RHS;
2906 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2907 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2909 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2911 return getUMaxExpr(PromotedLHS, PromotedRHS);
2914 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2915 /// the types using zero-extension, and then perform a umin operation
2917 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2919 const SCEV *PromotedLHS = LHS;
2920 const SCEV *PromotedRHS = RHS;
2922 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2923 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2925 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2927 return getUMinExpr(PromotedLHS, PromotedRHS);
2930 /// getPointerBase - Transitively follow the chain of pointer-type operands
2931 /// until reaching a SCEV that does not have a single pointer operand. This
2932 /// returns a SCEVUnknown pointer for well-formed pointer-type expressions,
2933 /// but corner cases do exist.
2934 const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) {
2935 // A pointer operand may evaluate to a nonpointer expression, such as null.
2936 if (!V->getType()->isPointerTy())
2939 if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(V)) {
2940 return getPointerBase(Cast->getOperand());
2942 else if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(V)) {
2943 const SCEV *PtrOp = 0;
2944 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
2946 if ((*I)->getType()->isPointerTy()) {
2947 // Cannot find the base of an expression with multiple pointer operands.
2955 return getPointerBase(PtrOp);
2960 /// PushDefUseChildren - Push users of the given Instruction
2961 /// onto the given Worklist.
2963 PushDefUseChildren(Instruction *I,
2964 SmallVectorImpl<Instruction *> &Worklist) {
2965 // Push the def-use children onto the Worklist stack.
2966 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2968 Worklist.push_back(cast<Instruction>(*UI));
2971 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2972 /// instructions that depend on the given instruction and removes them from
2973 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2976 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2977 SmallVector<Instruction *, 16> Worklist;
2978 PushDefUseChildren(PN, Worklist);
2980 SmallPtrSet<Instruction *, 8> Visited;
2982 while (!Worklist.empty()) {
2983 Instruction *I = Worklist.pop_back_val();
2984 if (!Visited.insert(I)) continue;
2986 ValueExprMapType::iterator It =
2987 ValueExprMap.find_as(static_cast<Value *>(I));
2988 if (It != ValueExprMap.end()) {
2989 const SCEV *Old = It->second;
2991 // Short-circuit the def-use traversal if the symbolic name
2992 // ceases to appear in expressions.
2993 if (Old != SymName && !hasOperand(Old, SymName))
2996 // SCEVUnknown for a PHI either means that it has an unrecognized
2997 // structure, it's a PHI that's in the progress of being computed
2998 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2999 // additional loop trip count information isn't going to change anything.
3000 // In the second case, createNodeForPHI will perform the necessary
3001 // updates on its own when it gets to that point. In the third, we do
3002 // want to forget the SCEVUnknown.
3003 if (!isa<PHINode>(I) ||
3004 !isa<SCEVUnknown>(Old) ||
3005 (I != PN && Old == SymName)) {
3006 forgetMemoizedResults(Old);
3007 ValueExprMap.erase(It);
3011 PushDefUseChildren(I, Worklist);
3015 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
3016 /// a loop header, making it a potential recurrence, or it doesn't.
3018 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
3019 if (const Loop *L = LI->getLoopFor(PN->getParent()))
3020 if (L->getHeader() == PN->getParent()) {
3021 // The loop may have multiple entrances or multiple exits; we can analyze
3022 // this phi as an addrec if it has a unique entry value and a unique
3024 Value *BEValueV = 0, *StartValueV = 0;
3025 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
3026 Value *V = PN->getIncomingValue(i);
3027 if (L->contains(PN->getIncomingBlock(i))) {
3030 } else if (BEValueV != V) {
3034 } else if (!StartValueV) {
3036 } else if (StartValueV != V) {
3041 if (BEValueV && StartValueV) {
3042 // While we are analyzing this PHI node, handle its value symbolically.
3043 const SCEV *SymbolicName = getUnknown(PN);
3044 assert(ValueExprMap.find_as(PN) == ValueExprMap.end() &&
3045 "PHI node already processed?");
3046 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
3048 // Using this symbolic name for the PHI, analyze the value coming around
3050 const SCEV *BEValue = getSCEV(BEValueV);
3052 // NOTE: If BEValue is loop invariant, we know that the PHI node just
3053 // has a special value for the first iteration of the loop.
3055 // If the value coming around the backedge is an add with the symbolic
3056 // value we just inserted, then we found a simple induction variable!
3057 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
3058 // If there is a single occurrence of the symbolic value, replace it
3059 // with a recurrence.
3060 unsigned FoundIndex = Add->getNumOperands();
3061 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3062 if (Add->getOperand(i) == SymbolicName)
3063 if (FoundIndex == e) {
3068 if (FoundIndex != Add->getNumOperands()) {
3069 // Create an add with everything but the specified operand.
3070 SmallVector<const SCEV *, 8> Ops;
3071 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3072 if (i != FoundIndex)
3073 Ops.push_back(Add->getOperand(i));
3074 const SCEV *Accum = getAddExpr(Ops);
3076 // This is not a valid addrec if the step amount is varying each
3077 // loop iteration, but is not itself an addrec in this loop.
3078 if (isLoopInvariant(Accum, L) ||
3079 (isa<SCEVAddRecExpr>(Accum) &&
3080 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
3081 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap;
3083 // If the increment doesn't overflow, then neither the addrec nor
3084 // the post-increment will overflow.
3085 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
3086 if (OBO->hasNoUnsignedWrap())
3087 Flags = setFlags(Flags, SCEV::FlagNUW);
3088 if (OBO->hasNoSignedWrap())
3089 Flags = setFlags(Flags, SCEV::FlagNSW);
3090 } else if (GEPOperator *GEP = dyn_cast<GEPOperator>(BEValueV)) {
3091 // If the increment is an inbounds GEP, then we know the address
3092 // space cannot be wrapped around. We cannot make any guarantee
3093 // about signed or unsigned overflow because pointers are
3094 // unsigned but we may have a negative index from the base
3095 // pointer. We can guarantee that no unsigned wrap occurs if the
3096 // indices form a positive value.
3097 if (GEP->isInBounds()) {
3098 Flags = setFlags(Flags, SCEV::FlagNW);
3100 const SCEV *Ptr = getSCEV(GEP->getPointerOperand());
3101 if (isKnownPositive(getMinusSCEV(getSCEV(GEP), Ptr)))
3102 Flags = setFlags(Flags, SCEV::FlagNUW);
3104 } else if (const SubOperator *OBO =
3105 dyn_cast<SubOperator>(BEValueV)) {
3106 if (OBO->hasNoUnsignedWrap())
3107 Flags = setFlags(Flags, SCEV::FlagNUW);
3108 if (OBO->hasNoSignedWrap())
3109 Flags = setFlags(Flags, SCEV::FlagNSW);
3112 const SCEV *StartVal = getSCEV(StartValueV);
3113 const SCEV *PHISCEV = getAddRecExpr(StartVal, Accum, L, Flags);
3115 // Since the no-wrap flags are on the increment, they apply to the
3116 // post-incremented value as well.
3117 if (isLoopInvariant(Accum, L))
3118 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
3121 // Okay, for the entire analysis of this edge we assumed the PHI
3122 // to be symbolic. We now need to go back and purge all of the
3123 // entries for the scalars that use the symbolic expression.
3124 ForgetSymbolicName(PN, SymbolicName);
3125 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3129 } else if (const SCEVAddRecExpr *AddRec =
3130 dyn_cast<SCEVAddRecExpr>(BEValue)) {
3131 // Otherwise, this could be a loop like this:
3132 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
3133 // In this case, j = {1,+,1} and BEValue is j.
3134 // Because the other in-value of i (0) fits the evolution of BEValue
3135 // i really is an addrec evolution.
3136 if (AddRec->getLoop() == L && AddRec->isAffine()) {
3137 const SCEV *StartVal = getSCEV(StartValueV);
3139 // If StartVal = j.start - j.stride, we can use StartVal as the
3140 // initial step of the addrec evolution.
3141 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
3142 AddRec->getOperand(1))) {
3143 // FIXME: For constant StartVal, we should be able to infer
3145 const SCEV *PHISCEV =
3146 getAddRecExpr(StartVal, AddRec->getOperand(1), L,
3149 // Okay, for the entire analysis of this edge we assumed the PHI
3150 // to be symbolic. We now need to go back and purge all of the
3151 // entries for the scalars that use the symbolic expression.
3152 ForgetSymbolicName(PN, SymbolicName);
3153 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3161 // If the PHI has a single incoming value, follow that value, unless the
3162 // PHI's incoming blocks are in a different loop, in which case doing so
3163 // risks breaking LCSSA form. Instcombine would normally zap these, but
3164 // it doesn't have DominatorTree information, so it may miss cases.
3165 if (Value *V = SimplifyInstruction(PN, TD, TLI, DT))
3166 if (LI->replacementPreservesLCSSAForm(PN, V))
3169 // If it's not a loop phi, we can't handle it yet.
3170 return getUnknown(PN);
3173 /// createNodeForGEP - Expand GEP instructions into add and multiply
3174 /// operations. This allows them to be analyzed by regular SCEV code.
3176 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
3177 Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
3178 Value *Base = GEP->getOperand(0);
3179 // Don't attempt to analyze GEPs over unsized objects.
3180 if (!Base->getType()->getPointerElementType()->isSized())
3181 return getUnknown(GEP);
3183 // Don't blindly transfer the inbounds flag from the GEP instruction to the
3184 // Add expression, because the Instruction may be guarded by control flow
3185 // and the no-overflow bits may not be valid for the expression in any
3187 SCEV::NoWrapFlags Wrap = GEP->isInBounds() ? SCEV::FlagNSW : SCEV::FlagAnyWrap;
3189 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
3190 gep_type_iterator GTI = gep_type_begin(GEP);
3191 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
3195 // Compute the (potentially symbolic) offset in bytes for this index.
3196 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
3197 // For a struct, add the member offset.
3198 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
3199 const SCEV *FieldOffset = getOffsetOfExpr(IntPtrTy, STy, FieldNo);
3201 // Add the field offset to the running total offset.
3202 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
3204 // For an array, add the element offset, explicitly scaled.
3205 const SCEV *ElementSize = getSizeOfExpr(IntPtrTy, *GTI);
3206 const SCEV *IndexS = getSCEV(Index);
3207 // Getelementptr indices are signed.
3208 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
3210 // Multiply the index by the element size to compute the element offset.
3211 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize, Wrap);
3213 // Add the element offset to the running total offset.
3214 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
3218 // Get the SCEV for the GEP base.
3219 const SCEV *BaseS = getSCEV(Base);
3221 // Add the total offset from all the GEP indices to the base.
3222 return getAddExpr(BaseS, TotalOffset, Wrap);
3225 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
3226 /// guaranteed to end in (at every loop iteration). It is, at the same time,
3227 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
3228 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
3230 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
3231 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3232 return C->getValue()->getValue().countTrailingZeros();
3234 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
3235 return std::min(GetMinTrailingZeros(T->getOperand()),
3236 (uint32_t)getTypeSizeInBits(T->getType()));
3238 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
3239 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3240 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3241 getTypeSizeInBits(E->getType()) : OpRes;
3244 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
3245 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3246 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3247 getTypeSizeInBits(E->getType()) : OpRes;
3250 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
3251 // The result is the min of all operands results.
3252 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3253 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3254 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3258 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
3259 // The result is the sum of all operands results.
3260 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
3261 uint32_t BitWidth = getTypeSizeInBits(M->getType());
3262 for (unsigned i = 1, e = M->getNumOperands();
3263 SumOpRes != BitWidth && i != e; ++i)
3264 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
3269 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
3270 // The result is the min of all operands results.
3271 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3272 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3273 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3277 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
3278 // The result is the min of all operands results.
3279 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3280 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3281 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3285 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
3286 // The result is the min of all operands results.
3287 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3288 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3289 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3293 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3294 // For a SCEVUnknown, ask ValueTracking.
3295 unsigned BitWidth = getTypeSizeInBits(U->getType());
3296 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3297 ComputeMaskedBits(U->getValue(), Zeros, Ones);
3298 return Zeros.countTrailingOnes();
3305 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
3308 ScalarEvolution::getUnsignedRange(const SCEV *S) {
3309 // See if we've computed this range already.
3310 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
3311 if (I != UnsignedRanges.end())
3314 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3315 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
3317 unsigned BitWidth = getTypeSizeInBits(S->getType());
3318 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3320 // If the value has known zeros, the maximum unsigned value will have those
3321 // known zeros as well.
3322 uint32_t TZ = GetMinTrailingZeros(S);
3324 ConservativeResult =
3325 ConstantRange(APInt::getMinValue(BitWidth),
3326 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
3328 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3329 ConstantRange X = getUnsignedRange(Add->getOperand(0));
3330 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3331 X = X.add(getUnsignedRange(Add->getOperand(i)));
3332 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
3335 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3336 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
3337 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3338 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
3339 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
3342 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3343 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
3344 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3345 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
3346 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
3349 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3350 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3351 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3352 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3353 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
3356 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3357 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3358 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3359 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3362 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3363 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3364 return setUnsignedRange(ZExt,
3365 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3368 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3369 ConstantRange X = getUnsignedRange(SExt->getOperand());
3370 return setUnsignedRange(SExt,
3371 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3374 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3375 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3376 return setUnsignedRange(Trunc,
3377 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3380 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3381 // If there's no unsigned wrap, the value will never be less than its
3383 if (AddRec->getNoWrapFlags(SCEV::FlagNUW))
3384 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3385 if (!C->getValue()->isZero())
3386 ConservativeResult =
3387 ConservativeResult.intersectWith(
3388 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3390 // TODO: non-affine addrec
3391 if (AddRec->isAffine()) {
3392 Type *Ty = AddRec->getType();
3393 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3394 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3395 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3396 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3398 const SCEV *Start = AddRec->getStart();
3399 const SCEV *Step = AddRec->getStepRecurrence(*this);
3401 ConstantRange StartRange = getUnsignedRange(Start);
3402 ConstantRange StepRange = getSignedRange(Step);
3403 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3404 ConstantRange EndRange =
3405 StartRange.add(MaxBECountRange.multiply(StepRange));
3407 // Check for overflow. This must be done with ConstantRange arithmetic
3408 // because we could be called from within the ScalarEvolution overflow
3410 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3411 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3412 ConstantRange ExtMaxBECountRange =
3413 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3414 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3415 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3417 return setUnsignedRange(AddRec, ConservativeResult);
3419 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3420 EndRange.getUnsignedMin());
3421 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3422 EndRange.getUnsignedMax());
3423 if (Min.isMinValue() && Max.isMaxValue())
3424 return setUnsignedRange(AddRec, ConservativeResult);
3425 return setUnsignedRange(AddRec,
3426 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3430 return setUnsignedRange(AddRec, ConservativeResult);
3433 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3434 // For a SCEVUnknown, ask ValueTracking.
3435 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3436 ComputeMaskedBits(U->getValue(), Zeros, Ones, TD);
3437 if (Ones == ~Zeros + 1)
3438 return setUnsignedRange(U, ConservativeResult);
3439 return setUnsignedRange(U,
3440 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3443 return setUnsignedRange(S, ConservativeResult);
3446 /// getSignedRange - Determine the signed range for a particular SCEV.
3449 ScalarEvolution::getSignedRange(const SCEV *S) {
3450 // See if we've computed this range already.
3451 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3452 if (I != SignedRanges.end())
3455 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3456 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3458 unsigned BitWidth = getTypeSizeInBits(S->getType());
3459 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3461 // If the value has known zeros, the maximum signed value will have those
3462 // known zeros as well.
3463 uint32_t TZ = GetMinTrailingZeros(S);
3465 ConservativeResult =
3466 ConstantRange(APInt::getSignedMinValue(BitWidth),
3467 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3469 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3470 ConstantRange X = getSignedRange(Add->getOperand(0));
3471 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3472 X = X.add(getSignedRange(Add->getOperand(i)));
3473 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3476 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3477 ConstantRange X = getSignedRange(Mul->getOperand(0));
3478 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3479 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3480 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3483 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3484 ConstantRange X = getSignedRange(SMax->getOperand(0));
3485 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3486 X = X.smax(getSignedRange(SMax->getOperand(i)));
3487 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3490 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3491 ConstantRange X = getSignedRange(UMax->getOperand(0));
3492 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3493 X = X.umax(getSignedRange(UMax->getOperand(i)));
3494 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3497 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3498 ConstantRange X = getSignedRange(UDiv->getLHS());
3499 ConstantRange Y = getSignedRange(UDiv->getRHS());
3500 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3503 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3504 ConstantRange X = getSignedRange(ZExt->getOperand());
3505 return setSignedRange(ZExt,
3506 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3509 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3510 ConstantRange X = getSignedRange(SExt->getOperand());
3511 return setSignedRange(SExt,
3512 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3515 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3516 ConstantRange X = getSignedRange(Trunc->getOperand());
3517 return setSignedRange(Trunc,
3518 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3521 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3522 // If there's no signed wrap, and all the operands have the same sign or
3523 // zero, the value won't ever change sign.
3524 if (AddRec->getNoWrapFlags(SCEV::FlagNSW)) {
3525 bool AllNonNeg = true;
3526 bool AllNonPos = true;
3527 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3528 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3529 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3532 ConservativeResult = ConservativeResult.intersectWith(
3533 ConstantRange(APInt(BitWidth, 0),
3534 APInt::getSignedMinValue(BitWidth)));
3536 ConservativeResult = ConservativeResult.intersectWith(
3537 ConstantRange(APInt::getSignedMinValue(BitWidth),
3538 APInt(BitWidth, 1)));
3541 // TODO: non-affine addrec
3542 if (AddRec->isAffine()) {
3543 Type *Ty = AddRec->getType();
3544 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3545 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3546 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3547 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3549 const SCEV *Start = AddRec->getStart();
3550 const SCEV *Step = AddRec->getStepRecurrence(*this);
3552 ConstantRange StartRange = getSignedRange(Start);
3553 ConstantRange StepRange = getSignedRange(Step);
3554 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3555 ConstantRange EndRange =
3556 StartRange.add(MaxBECountRange.multiply(StepRange));
3558 // Check for overflow. This must be done with ConstantRange arithmetic
3559 // because we could be called from within the ScalarEvolution overflow
3561 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3562 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3563 ConstantRange ExtMaxBECountRange =
3564 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3565 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3566 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3568 return setSignedRange(AddRec, ConservativeResult);
3570 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3571 EndRange.getSignedMin());
3572 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3573 EndRange.getSignedMax());
3574 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3575 return setSignedRange(AddRec, ConservativeResult);
3576 return setSignedRange(AddRec,
3577 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3581 return setSignedRange(AddRec, ConservativeResult);
3584 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3585 // For a SCEVUnknown, ask ValueTracking.
3586 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3587 return setSignedRange(U, ConservativeResult);
3588 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3590 return setSignedRange(U, ConservativeResult);
3591 return setSignedRange(U, ConservativeResult.intersectWith(
3592 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3593 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3596 return setSignedRange(S, ConservativeResult);
3599 /// createSCEV - We know that there is no SCEV for the specified value.
3600 /// Analyze the expression.
3602 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3603 if (!isSCEVable(V->getType()))
3604 return getUnknown(V);
3606 unsigned Opcode = Instruction::UserOp1;
3607 if (Instruction *I = dyn_cast<Instruction>(V)) {
3608 Opcode = I->getOpcode();
3610 // Don't attempt to analyze instructions in blocks that aren't
3611 // reachable. Such instructions don't matter, and they aren't required
3612 // to obey basic rules for definitions dominating uses which this
3613 // analysis depends on.
3614 if (!DT->isReachableFromEntry(I->getParent()))
3615 return getUnknown(V);
3616 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3617 Opcode = CE->getOpcode();
3618 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3619 return getConstant(CI);
3620 else if (isa<ConstantPointerNull>(V))
3621 return getConstant(V->getType(), 0);
3622 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3623 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3625 return getUnknown(V);
3627 Operator *U = cast<Operator>(V);
3629 case Instruction::Add: {
3630 // The simple thing to do would be to just call getSCEV on both operands
3631 // and call getAddExpr with the result. However if we're looking at a
3632 // bunch of things all added together, this can be quite inefficient,
3633 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3634 // Instead, gather up all the operands and make a single getAddExpr call.
3635 // LLVM IR canonical form means we need only traverse the left operands.
3637 // Don't apply this instruction's NSW or NUW flags to the new
3638 // expression. The instruction may be guarded by control flow that the
3639 // no-wrap behavior depends on. Non-control-equivalent instructions can be
3640 // mapped to the same SCEV expression, and it would be incorrect to transfer
3641 // NSW/NUW semantics to those operations.
3642 SmallVector<const SCEV *, 4> AddOps;
3643 AddOps.push_back(getSCEV(U->getOperand(1)));
3644 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3645 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3646 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3648 U = cast<Operator>(Op);
3649 const SCEV *Op1 = getSCEV(U->getOperand(1));
3650 if (Opcode == Instruction::Sub)
3651 AddOps.push_back(getNegativeSCEV(Op1));
3653 AddOps.push_back(Op1);
3655 AddOps.push_back(getSCEV(U->getOperand(0)));
3656 return getAddExpr(AddOps);
3658 case Instruction::Mul: {
3659 // Don't transfer NSW/NUW for the same reason as AddExpr.
3660 SmallVector<const SCEV *, 4> MulOps;
3661 MulOps.push_back(getSCEV(U->getOperand(1)));
3662 for (Value *Op = U->getOperand(0);
3663 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3664 Op = U->getOperand(0)) {
3665 U = cast<Operator>(Op);
3666 MulOps.push_back(getSCEV(U->getOperand(1)));
3668 MulOps.push_back(getSCEV(U->getOperand(0)));
3669 return getMulExpr(MulOps);
3671 case Instruction::UDiv:
3672 return getUDivExpr(getSCEV(U->getOperand(0)),
3673 getSCEV(U->getOperand(1)));
3674 case Instruction::Sub:
3675 return getMinusSCEV(getSCEV(U->getOperand(0)),
3676 getSCEV(U->getOperand(1)));
3677 case Instruction::And:
3678 // For an expression like x&255 that merely masks off the high bits,
3679 // use zext(trunc(x)) as the SCEV expression.
3680 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3681 if (CI->isNullValue())
3682 return getSCEV(U->getOperand(1));
3683 if (CI->isAllOnesValue())
3684 return getSCEV(U->getOperand(0));
3685 const APInt &A = CI->getValue();
3687 // Instcombine's ShrinkDemandedConstant may strip bits out of
3688 // constants, obscuring what would otherwise be a low-bits mask.
3689 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3690 // knew about to reconstruct a low-bits mask value.
3691 unsigned LZ = A.countLeadingZeros();
3692 unsigned BitWidth = A.getBitWidth();
3693 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3694 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD);
3696 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3698 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3700 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3701 IntegerType::get(getContext(), BitWidth - LZ)),
3706 case Instruction::Or:
3707 // If the RHS of the Or is a constant, we may have something like:
3708 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3709 // optimizations will transparently handle this case.
3711 // In order for this transformation to be safe, the LHS must be of the
3712 // form X*(2^n) and the Or constant must be less than 2^n.
3713 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3714 const SCEV *LHS = getSCEV(U->getOperand(0));
3715 const APInt &CIVal = CI->getValue();
3716 if (GetMinTrailingZeros(LHS) >=
3717 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3718 // Build a plain add SCEV.
3719 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3720 // If the LHS of the add was an addrec and it has no-wrap flags,
3721 // transfer the no-wrap flags, since an or won't introduce a wrap.
3722 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3723 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3724 const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags(
3725 OldAR->getNoWrapFlags());
3731 case Instruction::Xor:
3732 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3733 // If the RHS of the xor is a signbit, then this is just an add.
3734 // Instcombine turns add of signbit into xor as a strength reduction step.
3735 if (CI->getValue().isSignBit())
3736 return getAddExpr(getSCEV(U->getOperand(0)),
3737 getSCEV(U->getOperand(1)));
3739 // If the RHS of xor is -1, then this is a not operation.
3740 if (CI->isAllOnesValue())
3741 return getNotSCEV(getSCEV(U->getOperand(0)));
3743 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3744 // This is a variant of the check for xor with -1, and it handles
3745 // the case where instcombine has trimmed non-demanded bits out
3746 // of an xor with -1.
3747 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3748 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3749 if (BO->getOpcode() == Instruction::And &&
3750 LCI->getValue() == CI->getValue())
3751 if (const SCEVZeroExtendExpr *Z =
3752 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3753 Type *UTy = U->getType();
3754 const SCEV *Z0 = Z->getOperand();
3755 Type *Z0Ty = Z0->getType();
3756 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3758 // If C is a low-bits mask, the zero extend is serving to
3759 // mask off the high bits. Complement the operand and
3760 // re-apply the zext.
3761 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3762 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3764 // If C is a single bit, it may be in the sign-bit position
3765 // before the zero-extend. In this case, represent the xor
3766 // using an add, which is equivalent, and re-apply the zext.
3767 APInt Trunc = CI->getValue().trunc(Z0TySize);
3768 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3770 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3776 case Instruction::Shl:
3777 // Turn shift left of a constant amount into a multiply.
3778 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3779 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3781 // If the shift count is not less than the bitwidth, the result of
3782 // the shift is undefined. Don't try to analyze it, because the
3783 // resolution chosen here may differ from the resolution chosen in
3784 // other parts of the compiler.
3785 if (SA->getValue().uge(BitWidth))
3788 Constant *X = ConstantInt::get(getContext(),
3789 APInt::getOneBitSet(BitWidth, SA->getZExtValue()));
3790 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3794 case Instruction::LShr:
3795 // Turn logical shift right of a constant into a unsigned divide.
3796 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3797 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3799 // If the shift count is not less than the bitwidth, the result of
3800 // the shift is undefined. Don't try to analyze it, because the
3801 // resolution chosen here may differ from the resolution chosen in
3802 // other parts of the compiler.
3803 if (SA->getValue().uge(BitWidth))
3806 Constant *X = ConstantInt::get(getContext(),
3807 APInt::getOneBitSet(BitWidth, SA->getZExtValue()));
3808 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3812 case Instruction::AShr:
3813 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3814 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3815 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3816 if (L->getOpcode() == Instruction::Shl &&
3817 L->getOperand(1) == U->getOperand(1)) {
3818 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3820 // If the shift count is not less than the bitwidth, the result of
3821 // the shift is undefined. Don't try to analyze it, because the
3822 // resolution chosen here may differ from the resolution chosen in
3823 // other parts of the compiler.
3824 if (CI->getValue().uge(BitWidth))
3827 uint64_t Amt = BitWidth - CI->getZExtValue();
3828 if (Amt == BitWidth)
3829 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3831 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3832 IntegerType::get(getContext(),
3838 case Instruction::Trunc:
3839 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3841 case Instruction::ZExt:
3842 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3844 case Instruction::SExt:
3845 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3847 case Instruction::BitCast:
3848 // BitCasts are no-op casts so we just eliminate the cast.
3849 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3850 return getSCEV(U->getOperand(0));
3853 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3854 // lead to pointer expressions which cannot safely be expanded to GEPs,
3855 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3856 // simplifying integer expressions.
3858 case Instruction::GetElementPtr:
3859 return createNodeForGEP(cast<GEPOperator>(U));
3861 case Instruction::PHI:
3862 return createNodeForPHI(cast<PHINode>(U));
3864 case Instruction::Select:
3865 // This could be a smax or umax that was lowered earlier.
3866 // Try to recover it.
3867 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3868 Value *LHS = ICI->getOperand(0);
3869 Value *RHS = ICI->getOperand(1);
3870 switch (ICI->getPredicate()) {
3871 case ICmpInst::ICMP_SLT:
3872 case ICmpInst::ICMP_SLE:
3873 std::swap(LHS, RHS);
3875 case ICmpInst::ICMP_SGT:
3876 case ICmpInst::ICMP_SGE:
3877 // a >s b ? a+x : b+x -> smax(a, b)+x
3878 // a >s b ? b+x : a+x -> smin(a, b)+x
3879 if (LHS->getType() == U->getType()) {
3880 const SCEV *LS = getSCEV(LHS);
3881 const SCEV *RS = getSCEV(RHS);
3882 const SCEV *LA = getSCEV(U->getOperand(1));
3883 const SCEV *RA = getSCEV(U->getOperand(2));
3884 const SCEV *LDiff = getMinusSCEV(LA, LS);
3885 const SCEV *RDiff = getMinusSCEV(RA, RS);
3887 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3888 LDiff = getMinusSCEV(LA, RS);
3889 RDiff = getMinusSCEV(RA, LS);
3891 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3894 case ICmpInst::ICMP_ULT:
3895 case ICmpInst::ICMP_ULE:
3896 std::swap(LHS, RHS);
3898 case ICmpInst::ICMP_UGT:
3899 case ICmpInst::ICMP_UGE:
3900 // a >u b ? a+x : b+x -> umax(a, b)+x
3901 // a >u b ? b+x : a+x -> umin(a, b)+x
3902 if (LHS->getType() == U->getType()) {
3903 const SCEV *LS = getSCEV(LHS);
3904 const SCEV *RS = getSCEV(RHS);
3905 const SCEV *LA = getSCEV(U->getOperand(1));
3906 const SCEV *RA = getSCEV(U->getOperand(2));
3907 const SCEV *LDiff = getMinusSCEV(LA, LS);
3908 const SCEV *RDiff = getMinusSCEV(RA, RS);
3910 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3911 LDiff = getMinusSCEV(LA, RS);
3912 RDiff = getMinusSCEV(RA, LS);
3914 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3917 case ICmpInst::ICMP_NE:
3918 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3919 if (LHS->getType() == U->getType() &&
3920 isa<ConstantInt>(RHS) &&
3921 cast<ConstantInt>(RHS)->isZero()) {
3922 const SCEV *One = getConstant(LHS->getType(), 1);
3923 const SCEV *LS = getSCEV(LHS);
3924 const SCEV *LA = getSCEV(U->getOperand(1));
3925 const SCEV *RA = getSCEV(U->getOperand(2));
3926 const SCEV *LDiff = getMinusSCEV(LA, LS);
3927 const SCEV *RDiff = getMinusSCEV(RA, One);
3929 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3932 case ICmpInst::ICMP_EQ:
3933 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3934 if (LHS->getType() == U->getType() &&
3935 isa<ConstantInt>(RHS) &&
3936 cast<ConstantInt>(RHS)->isZero()) {
3937 const SCEV *One = getConstant(LHS->getType(), 1);
3938 const SCEV *LS = getSCEV(LHS);
3939 const SCEV *LA = getSCEV(U->getOperand(1));
3940 const SCEV *RA = getSCEV(U->getOperand(2));
3941 const SCEV *LDiff = getMinusSCEV(LA, One);
3942 const SCEV *RDiff = getMinusSCEV(RA, LS);
3944 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3952 default: // We cannot analyze this expression.
3956 return getUnknown(V);
3961 //===----------------------------------------------------------------------===//
3962 // Iteration Count Computation Code
3965 /// getSmallConstantTripCount - Returns the maximum trip count of this loop as a
3966 /// normal unsigned value. Returns 0 if the trip count is unknown or not
3967 /// constant. Will also return 0 if the maximum trip count is very large (>=
3970 /// This "trip count" assumes that control exits via ExitingBlock. More
3971 /// precisely, it is the number of times that control may reach ExitingBlock
3972 /// before taking the branch. For loops with multiple exits, it may not be the
3973 /// number times that the loop header executes because the loop may exit
3974 /// prematurely via another branch.
3976 /// FIXME: We conservatively call getBackedgeTakenCount(L) instead of
3977 /// getExitCount(L, ExitingBlock) to compute a safe trip count considering all
3978 /// loop exits. getExitCount() may return an exact count for this branch
3979 /// assuming no-signed-wrap. The number of well-defined iterations may actually
3980 /// be higher than this trip count if this exit test is skipped and the loop
3981 /// exits via a different branch. Ideally, getExitCount() would know whether it
3982 /// depends on a NSW assumption, and we would only fall back to a conservative
3983 /// trip count in that case.
3984 unsigned ScalarEvolution::
3985 getSmallConstantTripCount(Loop *L, BasicBlock * /*ExitingBlock*/) {
3986 const SCEVConstant *ExitCount =
3987 dyn_cast<SCEVConstant>(getBackedgeTakenCount(L));
3991 ConstantInt *ExitConst = ExitCount->getValue();
3993 // Guard against huge trip counts.
3994 if (ExitConst->getValue().getActiveBits() > 32)
3997 // In case of integer overflow, this returns 0, which is correct.
3998 return ((unsigned)ExitConst->getZExtValue()) + 1;
4001 /// getSmallConstantTripMultiple - Returns the largest constant divisor of the
4002 /// trip count of this loop as a normal unsigned value, if possible. This
4003 /// means that the actual trip count is always a multiple of the returned
4004 /// value (don't forget the trip count could very well be zero as well!).
4006 /// Returns 1 if the trip count is unknown or not guaranteed to be the
4007 /// multiple of a constant (which is also the case if the trip count is simply
4008 /// constant, use getSmallConstantTripCount for that case), Will also return 1
4009 /// if the trip count is very large (>= 2^32).
4011 /// As explained in the comments for getSmallConstantTripCount, this assumes
4012 /// that control exits the loop via ExitingBlock.
4013 unsigned ScalarEvolution::
4014 getSmallConstantTripMultiple(Loop *L, BasicBlock * /*ExitingBlock*/) {
4015 const SCEV *ExitCount = getBackedgeTakenCount(L);
4016 if (ExitCount == getCouldNotCompute())
4019 // Get the trip count from the BE count by adding 1.
4020 const SCEV *TCMul = getAddExpr(ExitCount,
4021 getConstant(ExitCount->getType(), 1));
4022 // FIXME: SCEV distributes multiplication as V1*C1 + V2*C1. We could attempt
4023 // to factor simple cases.
4024 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(TCMul))
4025 TCMul = Mul->getOperand(0);
4027 const SCEVConstant *MulC = dyn_cast<SCEVConstant>(TCMul);
4031 ConstantInt *Result = MulC->getValue();
4033 // Guard against huge trip counts (this requires checking
4034 // for zero to handle the case where the trip count == -1 and the
4036 if (!Result || Result->getValue().getActiveBits() > 32 ||
4037 Result->getValue().getActiveBits() == 0)
4040 return (unsigned)Result->getZExtValue();
4043 // getExitCount - Get the expression for the number of loop iterations for which
4044 // this loop is guaranteed not to exit via ExitingBlock. Otherwise return
4045 // SCEVCouldNotCompute.
4046 const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) {
4047 return getBackedgeTakenInfo(L).getExact(ExitingBlock, this);
4050 /// getBackedgeTakenCount - If the specified loop has a predictable
4051 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
4052 /// object. The backedge-taken count is the number of times the loop header
4053 /// will be branched to from within the loop. This is one less than the
4054 /// trip count of the loop, since it doesn't count the first iteration,
4055 /// when the header is branched to from outside the loop.
4057 /// Note that it is not valid to call this method on a loop without a
4058 /// loop-invariant backedge-taken count (see
4059 /// hasLoopInvariantBackedgeTakenCount).
4061 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
4062 return getBackedgeTakenInfo(L).getExact(this);
4065 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
4066 /// return the least SCEV value that is known never to be less than the
4067 /// actual backedge taken count.
4068 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
4069 return getBackedgeTakenInfo(L).getMax(this);
4072 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
4073 /// onto the given Worklist.
4075 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
4076 BasicBlock *Header = L->getHeader();
4078 // Push all Loop-header PHIs onto the Worklist stack.
4079 for (BasicBlock::iterator I = Header->begin();
4080 PHINode *PN = dyn_cast<PHINode>(I); ++I)
4081 Worklist.push_back(PN);
4084 const ScalarEvolution::BackedgeTakenInfo &
4085 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
4086 // Initially insert an invalid entry for this loop. If the insertion
4087 // succeeds, proceed to actually compute a backedge-taken count and
4088 // update the value. The temporary CouldNotCompute value tells SCEV
4089 // code elsewhere that it shouldn't attempt to request a new
4090 // backedge-taken count, which could result in infinite recursion.
4091 std::pair<DenseMap<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
4092 BackedgeTakenCounts.insert(std::make_pair(L, BackedgeTakenInfo()));
4094 return Pair.first->second;
4096 // ComputeBackedgeTakenCount may allocate memory for its result. Inserting it
4097 // into the BackedgeTakenCounts map transfers ownership. Otherwise, the result
4098 // must be cleared in this scope.
4099 BackedgeTakenInfo Result = ComputeBackedgeTakenCount(L);
4101 if (Result.getExact(this) != getCouldNotCompute()) {
4102 assert(isLoopInvariant(Result.getExact(this), L) &&
4103 isLoopInvariant(Result.getMax(this), L) &&
4104 "Computed backedge-taken count isn't loop invariant for loop!");
4105 ++NumTripCountsComputed;
4107 else if (Result.getMax(this) == getCouldNotCompute() &&
4108 isa<PHINode>(L->getHeader()->begin())) {
4109 // Only count loops that have phi nodes as not being computable.
4110 ++NumTripCountsNotComputed;
4113 // Now that we know more about the trip count for this loop, forget any
4114 // existing SCEV values for PHI nodes in this loop since they are only
4115 // conservative estimates made without the benefit of trip count
4116 // information. This is similar to the code in forgetLoop, except that
4117 // it handles SCEVUnknown PHI nodes specially.
4118 if (Result.hasAnyInfo()) {
4119 SmallVector<Instruction *, 16> Worklist;
4120 PushLoopPHIs(L, Worklist);
4122 SmallPtrSet<Instruction *, 8> Visited;
4123 while (!Worklist.empty()) {
4124 Instruction *I = Worklist.pop_back_val();
4125 if (!Visited.insert(I)) continue;
4127 ValueExprMapType::iterator It =
4128 ValueExprMap.find_as(static_cast<Value *>(I));
4129 if (It != ValueExprMap.end()) {
4130 const SCEV *Old = It->second;
4132 // SCEVUnknown for a PHI either means that it has an unrecognized
4133 // structure, or it's a PHI that's in the progress of being computed
4134 // by createNodeForPHI. In the former case, additional loop trip
4135 // count information isn't going to change anything. In the later
4136 // case, createNodeForPHI will perform the necessary updates on its
4137 // own when it gets to that point.
4138 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
4139 forgetMemoizedResults(Old);
4140 ValueExprMap.erase(It);
4142 if (PHINode *PN = dyn_cast<PHINode>(I))
4143 ConstantEvolutionLoopExitValue.erase(PN);
4146 PushDefUseChildren(I, Worklist);
4150 // Re-lookup the insert position, since the call to
4151 // ComputeBackedgeTakenCount above could result in a
4152 // recusive call to getBackedgeTakenInfo (on a different
4153 // loop), which would invalidate the iterator computed
4155 return BackedgeTakenCounts.find(L)->second = Result;
4158 /// forgetLoop - This method should be called by the client when it has
4159 /// changed a loop in a way that may effect ScalarEvolution's ability to
4160 /// compute a trip count, or if the loop is deleted.
4161 void ScalarEvolution::forgetLoop(const Loop *L) {
4162 // Drop any stored trip count value.
4163 DenseMap<const Loop*, BackedgeTakenInfo>::iterator BTCPos =
4164 BackedgeTakenCounts.find(L);
4165 if (BTCPos != BackedgeTakenCounts.end()) {
4166 BTCPos->second.clear();
4167 BackedgeTakenCounts.erase(BTCPos);
4170 // Drop information about expressions based on loop-header PHIs.
4171 SmallVector<Instruction *, 16> Worklist;
4172 PushLoopPHIs(L, Worklist);
4174 SmallPtrSet<Instruction *, 8> Visited;
4175 while (!Worklist.empty()) {
4176 Instruction *I = Worklist.pop_back_val();
4177 if (!Visited.insert(I)) continue;
4179 ValueExprMapType::iterator It =
4180 ValueExprMap.find_as(static_cast<Value *>(I));
4181 if (It != ValueExprMap.end()) {
4182 forgetMemoizedResults(It->second);
4183 ValueExprMap.erase(It);
4184 if (PHINode *PN = dyn_cast<PHINode>(I))
4185 ConstantEvolutionLoopExitValue.erase(PN);
4188 PushDefUseChildren(I, Worklist);
4191 // Forget all contained loops too, to avoid dangling entries in the
4192 // ValuesAtScopes map.
4193 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
4197 /// forgetValue - This method should be called by the client when it has
4198 /// changed a value in a way that may effect its value, or which may
4199 /// disconnect it from a def-use chain linking it to a loop.
4200 void ScalarEvolution::forgetValue(Value *V) {
4201 Instruction *I = dyn_cast<Instruction>(V);
4204 // Drop information about expressions based on loop-header PHIs.
4205 SmallVector<Instruction *, 16> Worklist;
4206 Worklist.push_back(I);
4208 SmallPtrSet<Instruction *, 8> Visited;
4209 while (!Worklist.empty()) {
4210 I = Worklist.pop_back_val();
4211 if (!Visited.insert(I)) continue;
4213 ValueExprMapType::iterator It =
4214 ValueExprMap.find_as(static_cast<Value *>(I));
4215 if (It != ValueExprMap.end()) {
4216 forgetMemoizedResults(It->second);
4217 ValueExprMap.erase(It);
4218 if (PHINode *PN = dyn_cast<PHINode>(I))
4219 ConstantEvolutionLoopExitValue.erase(PN);
4222 PushDefUseChildren(I, Worklist);
4226 /// getExact - Get the exact loop backedge taken count considering all loop
4227 /// exits. A computable result can only be return for loops with a single exit.
4228 /// Returning the minimum taken count among all exits is incorrect because one
4229 /// of the loop's exit limit's may have been skipped. HowFarToZero assumes that
4230 /// the limit of each loop test is never skipped. This is a valid assumption as
4231 /// long as the loop exits via that test. For precise results, it is the
4232 /// caller's responsibility to specify the relevant loop exit using
4233 /// getExact(ExitingBlock, SE).
4235 ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
4236 // If any exits were not computable, the loop is not computable.
4237 if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute();
4239 // We need exactly one computable exit.
4240 if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
4241 assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
4243 const SCEV *BECount = 0;
4244 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4245 ENT != 0; ENT = ENT->getNextExit()) {
4247 assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV");
4250 BECount = ENT->ExactNotTaken;
4251 else if (BECount != ENT->ExactNotTaken)
4252 return SE->getCouldNotCompute();
4254 assert(BECount && "Invalid not taken count for loop exit");
4258 /// getExact - Get the exact not taken count for this loop exit.
4260 ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock,
4261 ScalarEvolution *SE) const {
4262 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4263 ENT != 0; ENT = ENT->getNextExit()) {
4265 if (ENT->ExitingBlock == ExitingBlock)
4266 return ENT->ExactNotTaken;
4268 return SE->getCouldNotCompute();
4271 /// getMax - Get the max backedge taken count for the loop.
4273 ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const {
4274 return Max ? Max : SE->getCouldNotCompute();
4277 bool ScalarEvolution::BackedgeTakenInfo::hasOperand(const SCEV *S,
4278 ScalarEvolution *SE) const {
4279 if (Max && Max != SE->getCouldNotCompute() && SE->hasOperand(Max, S))
4282 if (!ExitNotTaken.ExitingBlock)
4285 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4286 ENT != 0; ENT = ENT->getNextExit()) {
4288 if (ENT->ExactNotTaken != SE->getCouldNotCompute()
4289 && SE->hasOperand(ENT->ExactNotTaken, S)) {
4296 /// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each
4297 /// computable exit into a persistent ExitNotTakenInfo array.
4298 ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo(
4299 SmallVectorImpl< std::pair<BasicBlock *, const SCEV *> > &ExitCounts,
4300 bool Complete, const SCEV *MaxCount) : Max(MaxCount) {
4303 ExitNotTaken.setIncomplete();
4305 unsigned NumExits = ExitCounts.size();
4306 if (NumExits == 0) return;
4308 ExitNotTaken.ExitingBlock = ExitCounts[0].first;
4309 ExitNotTaken.ExactNotTaken = ExitCounts[0].second;
4310 if (NumExits == 1) return;
4312 // Handle the rare case of multiple computable exits.
4313 ExitNotTakenInfo *ENT = new ExitNotTakenInfo[NumExits-1];
4315 ExitNotTakenInfo *PrevENT = &ExitNotTaken;
4316 for (unsigned i = 1; i < NumExits; ++i, PrevENT = ENT, ++ENT) {
4317 PrevENT->setNextExit(ENT);
4318 ENT->ExitingBlock = ExitCounts[i].first;
4319 ENT->ExactNotTaken = ExitCounts[i].second;
4323 /// clear - Invalidate this result and free the ExitNotTakenInfo array.
4324 void ScalarEvolution::BackedgeTakenInfo::clear() {
4325 ExitNotTaken.ExitingBlock = 0;
4326 ExitNotTaken.ExactNotTaken = 0;
4327 delete[] ExitNotTaken.getNextExit();
4330 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
4331 /// of the specified loop will execute.
4332 ScalarEvolution::BackedgeTakenInfo
4333 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
4334 SmallVector<BasicBlock *, 8> ExitingBlocks;
4335 L->getExitingBlocks(ExitingBlocks);
4337 // Examine all exits and pick the most conservative values.
4338 const SCEV *MaxBECount = getCouldNotCompute();
4339 bool CouldComputeBECount = true;
4340 BasicBlock *Latch = L->getLoopLatch(); // may be NULL.
4341 const SCEV *LatchMaxCount = 0;
4342 SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts;
4343 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
4344 ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]);
4345 if (EL.Exact == getCouldNotCompute())
4346 // We couldn't compute an exact value for this exit, so
4347 // we won't be able to compute an exact value for the loop.
4348 CouldComputeBECount = false;
4350 ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact));
4352 if (MaxBECount == getCouldNotCompute())
4353 MaxBECount = EL.Max;
4354 else if (EL.Max != getCouldNotCompute()) {
4355 // We cannot take the "min" MaxBECount, because non-unit stride loops may
4356 // skip some loop tests. Taking the max over the exits is sufficiently
4357 // conservative. TODO: We could do better taking into consideration
4358 // non-latch exits that dominate the latch.
4359 if (EL.MustExit && ExitingBlocks[i] == Latch)
4360 LatchMaxCount = EL.Max;
4362 MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
4365 // Be more precise in the easy case of a loop latch that must exit.
4366 if (LatchMaxCount) {
4367 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, LatchMaxCount);
4369 return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
4372 /// ComputeExitLimit - Compute the number of times the backedge of the specified
4373 /// loop will execute if it exits via the specified block.
4374 ScalarEvolution::ExitLimit
4375 ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) {
4377 // Okay, we've chosen an exiting block. See what condition causes us to
4378 // exit at this block.
4380 // FIXME: we should be able to handle switch instructions (with a single exit)
4381 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
4382 if (ExitBr == 0) return getCouldNotCompute();
4383 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
4385 // At this point, we know we have a conditional branch that determines whether
4386 // the loop is exited. However, we don't know if the branch is executed each
4387 // time through the loop. If not, then the execution count of the branch will
4388 // not be equal to the trip count of the loop.
4390 // Currently we check for this by checking to see if the Exit branch goes to
4391 // the loop header. If so, we know it will always execute the same number of
4392 // times as the loop. We also handle the case where the exit block *is* the
4393 // loop header. This is common for un-rotated loops.
4395 // If both of those tests fail, walk up the unique predecessor chain to the
4396 // header, stopping if there is an edge that doesn't exit the loop. If the
4397 // header is reached, the execution count of the branch will be equal to the
4398 // trip count of the loop.
4400 // More extensive analysis could be done to handle more cases here.
4402 if (ExitBr->getSuccessor(0) != L->getHeader() &&
4403 ExitBr->getSuccessor(1) != L->getHeader() &&
4404 ExitBr->getParent() != L->getHeader()) {
4405 // The simple checks failed, try climbing the unique predecessor chain
4406 // up to the header.
4408 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
4409 BasicBlock *Pred = BB->getUniquePredecessor();
4411 return getCouldNotCompute();
4412 TerminatorInst *PredTerm = Pred->getTerminator();
4413 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
4414 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
4417 // If the predecessor has a successor that isn't BB and isn't
4418 // outside the loop, assume the worst.
4419 if (L->contains(PredSucc))
4420 return getCouldNotCompute();
4422 if (Pred == L->getHeader()) {
4429 return getCouldNotCompute();
4432 // Proceed to the next level to examine the exit condition expression.
4433 return ComputeExitLimitFromCond(L, ExitBr->getCondition(),
4434 ExitBr->getSuccessor(0),
4435 ExitBr->getSuccessor(1),
4436 /*IsSubExpr=*/false);
4439 /// ComputeExitLimitFromCond - Compute the number of times the
4440 /// backedge of the specified loop will execute if its exit condition
4441 /// were a conditional branch of ExitCond, TBB, and FBB.
4443 /// @param IsSubExpr is true if ExitCond does not directly control the exit
4444 /// branch. In this case, we cannot assume that the loop only exits when the
4445 /// condition is true and cannot infer that failing to meet the condition prior
4446 /// to integer wraparound results in undefined behavior.
4447 ScalarEvolution::ExitLimit
4448 ScalarEvolution::ComputeExitLimitFromCond(const Loop *L,
4453 // Check if the controlling expression for this loop is an And or Or.
4454 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
4455 if (BO->getOpcode() == Instruction::And) {
4456 // Recurse on the operands of the and.
4457 bool EitherMayExit = L->contains(TBB);
4458 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB,
4459 IsSubExpr || EitherMayExit);
4460 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB,
4461 IsSubExpr || EitherMayExit);
4462 const SCEV *BECount = getCouldNotCompute();
4463 const SCEV *MaxBECount = getCouldNotCompute();
4464 bool MustExit = false;
4465 if (EitherMayExit) {
4466 // Both conditions must be true for the loop to continue executing.
4467 // Choose the less conservative count.
4468 if (EL0.Exact == getCouldNotCompute() ||
4469 EL1.Exact == getCouldNotCompute())
4470 BECount = getCouldNotCompute();
4472 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4473 if (EL0.Max == getCouldNotCompute())
4474 MaxBECount = EL1.Max;
4475 else if (EL1.Max == getCouldNotCompute())
4476 MaxBECount = EL0.Max;
4478 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4479 MustExit = EL0.MustExit || EL1.MustExit;
4481 // Both conditions must be true at the same time for the loop to exit.
4482 // For now, be conservative.
4483 assert(L->contains(FBB) && "Loop block has no successor in loop!");
4484 if (EL0.Max == EL1.Max)
4485 MaxBECount = EL0.Max;
4486 if (EL0.Exact == EL1.Exact)
4487 BECount = EL0.Exact;
4488 MustExit = EL0.MustExit && EL1.MustExit;
4491 return ExitLimit(BECount, MaxBECount, MustExit);
4493 if (BO->getOpcode() == Instruction::Or) {
4494 // Recurse on the operands of the or.
4495 bool EitherMayExit = L->contains(FBB);
4496 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB,
4497 IsSubExpr || EitherMayExit);
4498 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB,
4499 IsSubExpr || EitherMayExit);
4500 const SCEV *BECount = getCouldNotCompute();
4501 const SCEV *MaxBECount = getCouldNotCompute();
4502 bool MustExit = false;
4503 if (EitherMayExit) {
4504 // Both conditions must be false for the loop to continue executing.
4505 // Choose the less conservative count.
4506 if (EL0.Exact == getCouldNotCompute() ||
4507 EL1.Exact == getCouldNotCompute())
4508 BECount = getCouldNotCompute();
4510 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4511 if (EL0.Max == getCouldNotCompute())
4512 MaxBECount = EL1.Max;
4513 else if (EL1.Max == getCouldNotCompute())
4514 MaxBECount = EL0.Max;
4516 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4517 MustExit = EL0.MustExit || EL1.MustExit;
4519 // Both conditions must be false at the same time for the loop to exit.
4520 // For now, be conservative.
4521 assert(L->contains(TBB) && "Loop block has no successor in loop!");
4522 if (EL0.Max == EL1.Max)
4523 MaxBECount = EL0.Max;
4524 if (EL0.Exact == EL1.Exact)
4525 BECount = EL0.Exact;
4526 MustExit = EL0.MustExit && EL1.MustExit;
4529 return ExitLimit(BECount, MaxBECount, MustExit);
4533 // With an icmp, it may be feasible to compute an exact backedge-taken count.
4534 // Proceed to the next level to examine the icmp.
4535 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
4536 return ComputeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB, IsSubExpr);
4538 // Check for a constant condition. These are normally stripped out by
4539 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
4540 // preserve the CFG and is temporarily leaving constant conditions
4542 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
4543 if (L->contains(FBB) == !CI->getZExtValue())
4544 // The backedge is always taken.
4545 return getCouldNotCompute();
4547 // The backedge is never taken.
4548 return getConstant(CI->getType(), 0);
4551 // If it's not an integer or pointer comparison then compute it the hard way.
4552 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4555 /// ComputeExitLimitFromICmp - Compute the number of times the
4556 /// backedge of the specified loop will execute if its exit condition
4557 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4558 ScalarEvolution::ExitLimit
4559 ScalarEvolution::ComputeExitLimitFromICmp(const Loop *L,
4565 // If the condition was exit on true, convert the condition to exit on false
4566 ICmpInst::Predicate Cond;
4567 if (!L->contains(FBB))
4568 Cond = ExitCond->getPredicate();
4570 Cond = ExitCond->getInversePredicate();
4572 // Handle common loops like: for (X = "string"; *X; ++X)
4573 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4574 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4576 ComputeLoadConstantCompareExitLimit(LI, RHS, L, Cond);
4577 if (ItCnt.hasAnyInfo())
4581 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4582 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4584 // Try to evaluate any dependencies out of the loop.
4585 LHS = getSCEVAtScope(LHS, L);
4586 RHS = getSCEVAtScope(RHS, L);
4588 // At this point, we would like to compute how many iterations of the
4589 // loop the predicate will return true for these inputs.
4590 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4591 // If there is a loop-invariant, force it into the RHS.
4592 std::swap(LHS, RHS);
4593 Cond = ICmpInst::getSwappedPredicate(Cond);
4596 // Simplify the operands before analyzing them.
4597 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4599 // If we have a comparison of a chrec against a constant, try to use value
4600 // ranges to answer this query.
4601 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4602 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4603 if (AddRec->getLoop() == L) {
4604 // Form the constant range.
4605 ConstantRange CompRange(
4606 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4608 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4609 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4613 case ICmpInst::ICMP_NE: { // while (X != Y)
4614 // Convert to: while (X-Y != 0)
4615 ExitLimit EL = HowFarToZero(getMinusSCEV(LHS, RHS), L, IsSubExpr);
4616 if (EL.hasAnyInfo()) return EL;
4619 case ICmpInst::ICMP_EQ: { // while (X == Y)
4620 // Convert to: while (X-Y == 0)
4621 ExitLimit EL = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4622 if (EL.hasAnyInfo()) return EL;
4625 case ICmpInst::ICMP_SLT:
4626 case ICmpInst::ICMP_ULT: { // while (X < Y)
4627 bool IsSigned = Cond == ICmpInst::ICMP_SLT;
4628 ExitLimit EL = HowManyLessThans(LHS, RHS, L, IsSigned, IsSubExpr);
4629 if (EL.hasAnyInfo()) return EL;
4632 case ICmpInst::ICMP_SGT:
4633 case ICmpInst::ICMP_UGT: { // while (X > Y)
4634 bool IsSigned = Cond == ICmpInst::ICMP_SGT;
4635 ExitLimit EL = HowManyGreaterThans(LHS, RHS, L, IsSigned, IsSubExpr);
4636 if (EL.hasAnyInfo()) return EL;
4641 dbgs() << "ComputeBackedgeTakenCount ";
4642 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4643 dbgs() << "[unsigned] ";
4644 dbgs() << *LHS << " "
4645 << Instruction::getOpcodeName(Instruction::ICmp)
4646 << " " << *RHS << "\n";
4650 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4653 static ConstantInt *
4654 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4655 ScalarEvolution &SE) {
4656 const SCEV *InVal = SE.getConstant(C);
4657 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4658 assert(isa<SCEVConstant>(Val) &&
4659 "Evaluation of SCEV at constant didn't fold correctly?");
4660 return cast<SCEVConstant>(Val)->getValue();
4663 /// ComputeLoadConstantCompareExitLimit - Given an exit condition of
4664 /// 'icmp op load X, cst', try to see if we can compute the backedge
4665 /// execution count.
4666 ScalarEvolution::ExitLimit
4667 ScalarEvolution::ComputeLoadConstantCompareExitLimit(
4671 ICmpInst::Predicate predicate) {
4673 if (LI->isVolatile()) return getCouldNotCompute();
4675 // Check to see if the loaded pointer is a getelementptr of a global.
4676 // TODO: Use SCEV instead of manually grubbing with GEPs.
4677 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4678 if (!GEP) return getCouldNotCompute();
4680 // Make sure that it is really a constant global we are gepping, with an
4681 // initializer, and make sure the first IDX is really 0.
4682 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4683 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4684 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4685 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4686 return getCouldNotCompute();
4688 // Okay, we allow one non-constant index into the GEP instruction.
4690 std::vector<Constant*> Indexes;
4691 unsigned VarIdxNum = 0;
4692 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4693 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4694 Indexes.push_back(CI);
4695 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4696 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4697 VarIdx = GEP->getOperand(i);
4699 Indexes.push_back(0);
4702 // Loop-invariant loads may be a byproduct of loop optimization. Skip them.
4704 return getCouldNotCompute();
4706 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4707 // Check to see if X is a loop variant variable value now.
4708 const SCEV *Idx = getSCEV(VarIdx);
4709 Idx = getSCEVAtScope(Idx, L);
4711 // We can only recognize very limited forms of loop index expressions, in
4712 // particular, only affine AddRec's like {C1,+,C2}.
4713 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4714 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4715 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4716 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4717 return getCouldNotCompute();
4719 unsigned MaxSteps = MaxBruteForceIterations;
4720 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4721 ConstantInt *ItCst = ConstantInt::get(
4722 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4723 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4725 // Form the GEP offset.
4726 Indexes[VarIdxNum] = Val;
4728 Constant *Result = ConstantFoldLoadThroughGEPIndices(GV->getInitializer(),
4730 if (Result == 0) break; // Cannot compute!
4732 // Evaluate the condition for this iteration.
4733 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4734 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4735 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4737 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4738 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4741 ++NumArrayLenItCounts;
4742 return getConstant(ItCst); // Found terminating iteration!
4745 return getCouldNotCompute();
4749 /// CanConstantFold - Return true if we can constant fold an instruction of the
4750 /// specified type, assuming that all operands were constants.
4751 static bool CanConstantFold(const Instruction *I) {
4752 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4753 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) ||
4757 if (const CallInst *CI = dyn_cast<CallInst>(I))
4758 if (const Function *F = CI->getCalledFunction())
4759 return canConstantFoldCallTo(F);
4763 /// Determine whether this instruction can constant evolve within this loop
4764 /// assuming its operands can all constant evolve.
4765 static bool canConstantEvolve(Instruction *I, const Loop *L) {
4766 // An instruction outside of the loop can't be derived from a loop PHI.
4767 if (!L->contains(I)) return false;
4769 if (isa<PHINode>(I)) {
4770 if (L->getHeader() == I->getParent())
4773 // We don't currently keep track of the control flow needed to evaluate
4774 // PHIs, so we cannot handle PHIs inside of loops.
4778 // If we won't be able to constant fold this expression even if the operands
4779 // are constants, bail early.
4780 return CanConstantFold(I);
4783 /// getConstantEvolvingPHIOperands - Implement getConstantEvolvingPHI by
4784 /// recursing through each instruction operand until reaching a loop header phi.
4786 getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L,
4787 DenseMap<Instruction *, PHINode *> &PHIMap) {
4789 // Otherwise, we can evaluate this instruction if all of its operands are
4790 // constant or derived from a PHI node themselves.
4792 for (Instruction::op_iterator OpI = UseInst->op_begin(),
4793 OpE = UseInst->op_end(); OpI != OpE; ++OpI) {
4795 if (isa<Constant>(*OpI)) continue;
4797 Instruction *OpInst = dyn_cast<Instruction>(*OpI);
4798 if (!OpInst || !canConstantEvolve(OpInst, L)) return 0;
4800 PHINode *P = dyn_cast<PHINode>(OpInst);
4802 // If this operand is already visited, reuse the prior result.
4803 // We may have P != PHI if this is the deepest point at which the
4804 // inconsistent paths meet.
4805 P = PHIMap.lookup(OpInst);
4807 // Recurse and memoize the results, whether a phi is found or not.
4808 // This recursive call invalidates pointers into PHIMap.
4809 P = getConstantEvolvingPHIOperands(OpInst, L, PHIMap);
4812 if (P == 0) return 0; // Not evolving from PHI
4813 if (PHI && PHI != P) return 0; // Evolving from multiple different PHIs.
4816 // This is a expression evolving from a constant PHI!
4820 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4821 /// in the loop that V is derived from. We allow arbitrary operations along the
4822 /// way, but the operands of an operation must either be constants or a value
4823 /// derived from a constant PHI. If this expression does not fit with these
4824 /// constraints, return null.
4825 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4826 Instruction *I = dyn_cast<Instruction>(V);
4827 if (I == 0 || !canConstantEvolve(I, L)) return 0;
4829 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4833 // Record non-constant instructions contained by the loop.
4834 DenseMap<Instruction *, PHINode *> PHIMap;
4835 return getConstantEvolvingPHIOperands(I, L, PHIMap);
4838 /// EvaluateExpression - Given an expression that passes the
4839 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4840 /// in the loop has the value PHIVal. If we can't fold this expression for some
4841 /// reason, return null.
4842 static Constant *EvaluateExpression(Value *V, const Loop *L,
4843 DenseMap<Instruction *, Constant *> &Vals,
4844 const DataLayout *TD,
4845 const TargetLibraryInfo *TLI) {
4846 // Convenient constant check, but redundant for recursive calls.
4847 if (Constant *C = dyn_cast<Constant>(V)) return C;
4848 Instruction *I = dyn_cast<Instruction>(V);
4851 if (Constant *C = Vals.lookup(I)) return C;
4853 // An instruction inside the loop depends on a value outside the loop that we
4854 // weren't given a mapping for, or a value such as a call inside the loop.
4855 if (!canConstantEvolve(I, L)) return 0;
4857 // An unmapped PHI can be due to a branch or another loop inside this loop,
4858 // or due to this not being the initial iteration through a loop where we
4859 // couldn't compute the evolution of this particular PHI last time.
4860 if (isa<PHINode>(I)) return 0;
4862 std::vector<Constant*> Operands(I->getNumOperands());
4864 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4865 Instruction *Operand = dyn_cast<Instruction>(I->getOperand(i));
4867 Operands[i] = dyn_cast<Constant>(I->getOperand(i));
4868 if (!Operands[i]) return 0;
4871 Constant *C = EvaluateExpression(Operand, L, Vals, TD, TLI);
4877 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4878 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4879 Operands[1], TD, TLI);
4880 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4881 if (!LI->isVolatile())
4882 return ConstantFoldLoadFromConstPtr(Operands[0], TD);
4884 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Operands, TD,
4888 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4889 /// in the header of its containing loop, we know the loop executes a
4890 /// constant number of times, and the PHI node is just a recurrence
4891 /// involving constants, fold it.
4893 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4896 DenseMap<PHINode*, Constant*>::const_iterator I =
4897 ConstantEvolutionLoopExitValue.find(PN);
4898 if (I != ConstantEvolutionLoopExitValue.end())
4901 if (BEs.ugt(MaxBruteForceIterations))
4902 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4904 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4906 DenseMap<Instruction *, Constant *> CurrentIterVals;
4907 BasicBlock *Header = L->getHeader();
4908 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
4910 // Since the loop is canonicalized, the PHI node must have two entries. One
4911 // entry must be a constant (coming in from outside of the loop), and the
4912 // second must be derived from the same PHI.
4913 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4915 for (BasicBlock::iterator I = Header->begin();
4916 (PHI = dyn_cast<PHINode>(I)); ++I) {
4917 Constant *StartCST =
4918 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
4919 if (StartCST == 0) continue;
4920 CurrentIterVals[PHI] = StartCST;
4922 if (!CurrentIterVals.count(PN))
4925 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4927 // Execute the loop symbolically to determine the exit value.
4928 if (BEs.getActiveBits() >= 32)
4929 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4931 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4932 unsigned IterationNum = 0;
4933 for (; ; ++IterationNum) {
4934 if (IterationNum == NumIterations)
4935 return RetVal = CurrentIterVals[PN]; // Got exit value!
4937 // Compute the value of the PHIs for the next iteration.
4938 // EvaluateExpression adds non-phi values to the CurrentIterVals map.
4939 DenseMap<Instruction *, Constant *> NextIterVals;
4940 Constant *NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD,
4943 return 0; // Couldn't evaluate!
4944 NextIterVals[PN] = NextPHI;
4946 bool StoppedEvolving = NextPHI == CurrentIterVals[PN];
4948 // Also evaluate the other PHI nodes. However, we don't get to stop if we
4949 // cease to be able to evaluate one of them or if they stop evolving,
4950 // because that doesn't necessarily prevent us from computing PN.
4951 SmallVector<std::pair<PHINode *, Constant *>, 8> PHIsToCompute;
4952 for (DenseMap<Instruction *, Constant *>::const_iterator
4953 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
4954 PHINode *PHI = dyn_cast<PHINode>(I->first);
4955 if (!PHI || PHI == PN || PHI->getParent() != Header) continue;
4956 PHIsToCompute.push_back(std::make_pair(PHI, I->second));
4958 // We use two distinct loops because EvaluateExpression may invalidate any
4959 // iterators into CurrentIterVals.
4960 for (SmallVectorImpl<std::pair<PHINode *, Constant*> >::const_iterator
4961 I = PHIsToCompute.begin(), E = PHIsToCompute.end(); I != E; ++I) {
4962 PHINode *PHI = I->first;
4963 Constant *&NextPHI = NextIterVals[PHI];
4964 if (!NextPHI) { // Not already computed.
4965 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
4966 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
4968 if (NextPHI != I->second)
4969 StoppedEvolving = false;
4972 // If all entries in CurrentIterVals == NextIterVals then we can stop
4973 // iterating, the loop can't continue to change.
4974 if (StoppedEvolving)
4975 return RetVal = CurrentIterVals[PN];
4977 CurrentIterVals.swap(NextIterVals);
4981 /// ComputeExitCountExhaustively - If the loop is known to execute a
4982 /// constant number of times (the condition evolves only from constants),
4983 /// try to evaluate a few iterations of the loop until we get the exit
4984 /// condition gets a value of ExitWhen (true or false). If we cannot
4985 /// evaluate the trip count of the loop, return getCouldNotCompute().
4986 const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L,
4989 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4990 if (PN == 0) return getCouldNotCompute();
4992 // If the loop is canonicalized, the PHI will have exactly two entries.
4993 // That's the only form we support here.
4994 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4996 DenseMap<Instruction *, Constant *> CurrentIterVals;
4997 BasicBlock *Header = L->getHeader();
4998 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
5000 // One entry must be a constant (coming in from outside of the loop), and the
5001 // second must be derived from the same PHI.
5002 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
5004 for (BasicBlock::iterator I = Header->begin();
5005 (PHI = dyn_cast<PHINode>(I)); ++I) {
5006 Constant *StartCST =
5007 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
5008 if (StartCST == 0) continue;
5009 CurrentIterVals[PHI] = StartCST;
5011 if (!CurrentIterVals.count(PN))
5012 return getCouldNotCompute();
5014 // Okay, we find a PHI node that defines the trip count of this loop. Execute
5015 // the loop symbolically to determine when the condition gets a value of
5018 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
5019 for (unsigned IterationNum = 0; IterationNum != MaxIterations;++IterationNum){
5020 ConstantInt *CondVal =
5021 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, L, CurrentIterVals,
5024 // Couldn't symbolically evaluate.
5025 if (!CondVal) return getCouldNotCompute();
5027 if (CondVal->getValue() == uint64_t(ExitWhen)) {
5028 ++NumBruteForceTripCountsComputed;
5029 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
5032 // Update all the PHI nodes for the next iteration.
5033 DenseMap<Instruction *, Constant *> NextIterVals;
5035 // Create a list of which PHIs we need to compute. We want to do this before
5036 // calling EvaluateExpression on them because that may invalidate iterators
5037 // into CurrentIterVals.
5038 SmallVector<PHINode *, 8> PHIsToCompute;
5039 for (DenseMap<Instruction *, Constant *>::const_iterator
5040 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
5041 PHINode *PHI = dyn_cast<PHINode>(I->first);
5042 if (!PHI || PHI->getParent() != Header) continue;
5043 PHIsToCompute.push_back(PHI);
5045 for (SmallVectorImpl<PHINode *>::const_iterator I = PHIsToCompute.begin(),
5046 E = PHIsToCompute.end(); I != E; ++I) {
5048 Constant *&NextPHI = NextIterVals[PHI];
5049 if (NextPHI) continue; // Already computed!
5051 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
5052 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
5054 CurrentIterVals.swap(NextIterVals);
5057 // Too many iterations were needed to evaluate.
5058 return getCouldNotCompute();
5061 /// getSCEVAtScope - Return a SCEV expression for the specified value
5062 /// at the specified scope in the program. The L value specifies a loop
5063 /// nest to evaluate the expression at, where null is the top-level or a
5064 /// specified loop is immediately inside of the loop.
5066 /// This method can be used to compute the exit value for a variable defined
5067 /// in a loop by querying what the value will hold in the parent loop.
5069 /// In the case that a relevant loop exit value cannot be computed, the
5070 /// original value V is returned.
5071 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
5072 // Check to see if we've folded this expression at this loop before.
5073 SmallVector<std::pair<const Loop *, const SCEV *>, 2> &Values = ValuesAtScopes[V];
5074 for (unsigned u = 0; u < Values.size(); u++) {
5075 if (Values[u].first == L)
5076 return Values[u].second ? Values[u].second : V;
5078 Values.push_back(std::make_pair(L, static_cast<const SCEV *>(0)));
5079 // Otherwise compute it.
5080 const SCEV *C = computeSCEVAtScope(V, L);
5081 SmallVector<std::pair<const Loop *, const SCEV *>, 2> &Values2 = ValuesAtScopes[V];
5082 for (unsigned u = Values2.size(); u > 0; u--) {
5083 if (Values2[u - 1].first == L) {
5084 Values2[u - 1].second = C;
5091 /// This builds up a Constant using the ConstantExpr interface. That way, we
5092 /// will return Constants for objects which aren't represented by a
5093 /// SCEVConstant, because SCEVConstant is restricted to ConstantInt.
5094 /// Returns NULL if the SCEV isn't representable as a Constant.
5095 static Constant *BuildConstantFromSCEV(const SCEV *V) {
5096 switch (V->getSCEVType()) {
5097 default: // TODO: smax, umax.
5098 case scCouldNotCompute:
5102 return cast<SCEVConstant>(V)->getValue();
5104 return dyn_cast<Constant>(cast<SCEVUnknown>(V)->getValue());
5105 case scSignExtend: {
5106 const SCEVSignExtendExpr *SS = cast<SCEVSignExtendExpr>(V);
5107 if (Constant *CastOp = BuildConstantFromSCEV(SS->getOperand()))
5108 return ConstantExpr::getSExt(CastOp, SS->getType());
5111 case scZeroExtend: {
5112 const SCEVZeroExtendExpr *SZ = cast<SCEVZeroExtendExpr>(V);
5113 if (Constant *CastOp = BuildConstantFromSCEV(SZ->getOperand()))
5114 return ConstantExpr::getZExt(CastOp, SZ->getType());
5118 const SCEVTruncateExpr *ST = cast<SCEVTruncateExpr>(V);
5119 if (Constant *CastOp = BuildConstantFromSCEV(ST->getOperand()))
5120 return ConstantExpr::getTrunc(CastOp, ST->getType());
5124 const SCEVAddExpr *SA = cast<SCEVAddExpr>(V);
5125 if (Constant *C = BuildConstantFromSCEV(SA->getOperand(0))) {
5126 if (PointerType *PTy = dyn_cast<PointerType>(C->getType())) {
5127 unsigned AS = PTy->getAddressSpace();
5128 Type *DestPtrTy = Type::getInt8PtrTy(C->getContext(), AS);
5129 C = ConstantExpr::getBitCast(C, DestPtrTy);
5131 for (unsigned i = 1, e = SA->getNumOperands(); i != e; ++i) {
5132 Constant *C2 = BuildConstantFromSCEV(SA->getOperand(i));
5136 if (!C->getType()->isPointerTy() && C2->getType()->isPointerTy()) {
5137 unsigned AS = C2->getType()->getPointerAddressSpace();
5139 Type *DestPtrTy = Type::getInt8PtrTy(C->getContext(), AS);
5140 // The offsets have been converted to bytes. We can add bytes to an
5141 // i8* by GEP with the byte count in the first index.
5142 C = ConstantExpr::getBitCast(C, DestPtrTy);
5145 // Don't bother trying to sum two pointers. We probably can't
5146 // statically compute a load that results from it anyway.
5147 if (C2->getType()->isPointerTy())
5150 if (PointerType *PTy = dyn_cast<PointerType>(C->getType())) {
5151 if (PTy->getElementType()->isStructTy())
5152 C2 = ConstantExpr::getIntegerCast(
5153 C2, Type::getInt32Ty(C->getContext()), true);
5154 C = ConstantExpr::getGetElementPtr(C, C2);
5156 C = ConstantExpr::getAdd(C, C2);
5163 const SCEVMulExpr *SM = cast<SCEVMulExpr>(V);
5164 if (Constant *C = BuildConstantFromSCEV(SM->getOperand(0))) {
5165 // Don't bother with pointers at all.
5166 if (C->getType()->isPointerTy()) return 0;
5167 for (unsigned i = 1, e = SM->getNumOperands(); i != e; ++i) {
5168 Constant *C2 = BuildConstantFromSCEV(SM->getOperand(i));
5169 if (!C2 || C2->getType()->isPointerTy()) return 0;
5170 C = ConstantExpr::getMul(C, C2);
5177 const SCEVUDivExpr *SU = cast<SCEVUDivExpr>(V);
5178 if (Constant *LHS = BuildConstantFromSCEV(SU->getLHS()))
5179 if (Constant *RHS = BuildConstantFromSCEV(SU->getRHS()))
5180 if (LHS->getType() == RHS->getType())
5181 return ConstantExpr::getUDiv(LHS, RHS);
5188 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
5189 if (isa<SCEVConstant>(V)) return V;
5191 // If this instruction is evolved from a constant-evolving PHI, compute the
5192 // exit value from the loop without using SCEVs.
5193 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
5194 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
5195 const Loop *LI = (*this->LI)[I->getParent()];
5196 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
5197 if (PHINode *PN = dyn_cast<PHINode>(I))
5198 if (PN->getParent() == LI->getHeader()) {
5199 // Okay, there is no closed form solution for the PHI node. Check
5200 // to see if the loop that contains it has a known backedge-taken
5201 // count. If so, we may be able to force computation of the exit
5203 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
5204 if (const SCEVConstant *BTCC =
5205 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
5206 // Okay, we know how many times the containing loop executes. If
5207 // this is a constant evolving PHI node, get the final value at
5208 // the specified iteration number.
5209 Constant *RV = getConstantEvolutionLoopExitValue(PN,
5210 BTCC->getValue()->getValue(),
5212 if (RV) return getSCEV(RV);
5216 // Okay, this is an expression that we cannot symbolically evaluate
5217 // into a SCEV. Check to see if it's possible to symbolically evaluate
5218 // the arguments into constants, and if so, try to constant propagate the
5219 // result. This is particularly useful for computing loop exit values.
5220 if (CanConstantFold(I)) {
5221 SmallVector<Constant *, 4> Operands;
5222 bool MadeImprovement = false;
5223 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
5224 Value *Op = I->getOperand(i);
5225 if (Constant *C = dyn_cast<Constant>(Op)) {
5226 Operands.push_back(C);
5230 // If any of the operands is non-constant and if they are
5231 // non-integer and non-pointer, don't even try to analyze them
5232 // with scev techniques.
5233 if (!isSCEVable(Op->getType()))
5236 const SCEV *OrigV = getSCEV(Op);
5237 const SCEV *OpV = getSCEVAtScope(OrigV, L);
5238 MadeImprovement |= OrigV != OpV;
5240 Constant *C = BuildConstantFromSCEV(OpV);
5242 if (C->getType() != Op->getType())
5243 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5247 Operands.push_back(C);
5250 // Check to see if getSCEVAtScope actually made an improvement.
5251 if (MadeImprovement) {
5253 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
5254 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
5255 Operands[0], Operands[1], TD,
5257 else if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
5258 if (!LI->isVolatile())
5259 C = ConstantFoldLoadFromConstPtr(Operands[0], TD);
5261 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
5269 // This is some other type of SCEVUnknown, just return it.
5273 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
5274 // Avoid performing the look-up in the common case where the specified
5275 // expression has no loop-variant portions.
5276 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
5277 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5278 if (OpAtScope != Comm->getOperand(i)) {
5279 // Okay, at least one of these operands is loop variant but might be
5280 // foldable. Build a new instance of the folded commutative expression.
5281 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
5282 Comm->op_begin()+i);
5283 NewOps.push_back(OpAtScope);
5285 for (++i; i != e; ++i) {
5286 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5287 NewOps.push_back(OpAtScope);
5289 if (isa<SCEVAddExpr>(Comm))
5290 return getAddExpr(NewOps);
5291 if (isa<SCEVMulExpr>(Comm))
5292 return getMulExpr(NewOps);
5293 if (isa<SCEVSMaxExpr>(Comm))
5294 return getSMaxExpr(NewOps);
5295 if (isa<SCEVUMaxExpr>(Comm))
5296 return getUMaxExpr(NewOps);
5297 llvm_unreachable("Unknown commutative SCEV type!");
5300 // If we got here, all operands are loop invariant.
5304 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
5305 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
5306 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
5307 if (LHS == Div->getLHS() && RHS == Div->getRHS())
5308 return Div; // must be loop invariant
5309 return getUDivExpr(LHS, RHS);
5312 // If this is a loop recurrence for a loop that does not contain L, then we
5313 // are dealing with the final value computed by the loop.
5314 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
5315 // First, attempt to evaluate each operand.
5316 // Avoid performing the look-up in the common case where the specified
5317 // expression has no loop-variant portions.
5318 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
5319 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
5320 if (OpAtScope == AddRec->getOperand(i))
5323 // Okay, at least one of these operands is loop variant but might be
5324 // foldable. Build a new instance of the folded commutative expression.
5325 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
5326 AddRec->op_begin()+i);
5327 NewOps.push_back(OpAtScope);
5328 for (++i; i != e; ++i)
5329 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
5331 const SCEV *FoldedRec =
5332 getAddRecExpr(NewOps, AddRec->getLoop(),
5333 AddRec->getNoWrapFlags(SCEV::FlagNW));
5334 AddRec = dyn_cast<SCEVAddRecExpr>(FoldedRec);
5335 // The addrec may be folded to a nonrecurrence, for example, if the
5336 // induction variable is multiplied by zero after constant folding. Go
5337 // ahead and return the folded value.
5343 // If the scope is outside the addrec's loop, evaluate it by using the
5344 // loop exit value of the addrec.
5345 if (!AddRec->getLoop()->contains(L)) {
5346 // To evaluate this recurrence, we need to know how many times the AddRec
5347 // loop iterates. Compute this now.
5348 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
5349 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
5351 // Then, evaluate the AddRec.
5352 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
5358 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
5359 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5360 if (Op == Cast->getOperand())
5361 return Cast; // must be loop invariant
5362 return getZeroExtendExpr(Op, Cast->getType());
5365 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
5366 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5367 if (Op == Cast->getOperand())
5368 return Cast; // must be loop invariant
5369 return getSignExtendExpr(Op, Cast->getType());
5372 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
5373 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5374 if (Op == Cast->getOperand())
5375 return Cast; // must be loop invariant
5376 return getTruncateExpr(Op, Cast->getType());
5379 llvm_unreachable("Unknown SCEV type!");
5382 /// getSCEVAtScope - This is a convenience function which does
5383 /// getSCEVAtScope(getSCEV(V), L).
5384 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
5385 return getSCEVAtScope(getSCEV(V), L);
5388 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
5389 /// following equation:
5391 /// A * X = B (mod N)
5393 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
5394 /// A and B isn't important.
5396 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
5397 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
5398 ScalarEvolution &SE) {
5399 uint32_t BW = A.getBitWidth();
5400 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
5401 assert(A != 0 && "A must be non-zero.");
5405 // The gcd of A and N may have only one prime factor: 2. The number of
5406 // trailing zeros in A is its multiplicity
5407 uint32_t Mult2 = A.countTrailingZeros();
5410 // 2. Check if B is divisible by D.
5412 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
5413 // is not less than multiplicity of this prime factor for D.
5414 if (B.countTrailingZeros() < Mult2)
5415 return SE.getCouldNotCompute();
5417 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
5420 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
5421 // bit width during computations.
5422 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
5423 APInt Mod(BW + 1, 0);
5424 Mod.setBit(BW - Mult2); // Mod = N / D
5425 APInt I = AD.multiplicativeInverse(Mod);
5427 // 4. Compute the minimum unsigned root of the equation:
5428 // I * (B / D) mod (N / D)
5429 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
5431 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
5433 return SE.getConstant(Result.trunc(BW));
5436 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
5437 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
5438 /// might be the same) or two SCEVCouldNotCompute objects.
5440 static std::pair<const SCEV *,const SCEV *>
5441 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
5442 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
5443 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
5444 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
5445 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
5447 // We currently can only solve this if the coefficients are constants.
5448 if (!LC || !MC || !NC) {
5449 const SCEV *CNC = SE.getCouldNotCompute();
5450 return std::make_pair(CNC, CNC);
5453 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
5454 const APInt &L = LC->getValue()->getValue();
5455 const APInt &M = MC->getValue()->getValue();
5456 const APInt &N = NC->getValue()->getValue();
5457 APInt Two(BitWidth, 2);
5458 APInt Four(BitWidth, 4);
5461 using namespace APIntOps;
5463 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
5464 // The B coefficient is M-N/2
5468 // The A coefficient is N/2
5469 APInt A(N.sdiv(Two));
5471 // Compute the B^2-4ac term.
5474 SqrtTerm -= Four * (A * C);
5476 if (SqrtTerm.isNegative()) {
5477 // The loop is provably infinite.
5478 const SCEV *CNC = SE.getCouldNotCompute();
5479 return std::make_pair(CNC, CNC);
5482 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
5483 // integer value or else APInt::sqrt() will assert.
5484 APInt SqrtVal(SqrtTerm.sqrt());
5486 // Compute the two solutions for the quadratic formula.
5487 // The divisions must be performed as signed divisions.
5490 if (TwoA.isMinValue()) {
5491 const SCEV *CNC = SE.getCouldNotCompute();
5492 return std::make_pair(CNC, CNC);
5495 LLVMContext &Context = SE.getContext();
5497 ConstantInt *Solution1 =
5498 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
5499 ConstantInt *Solution2 =
5500 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
5502 return std::make_pair(SE.getConstant(Solution1),
5503 SE.getConstant(Solution2));
5504 } // end APIntOps namespace
5507 /// HowFarToZero - Return the number of times a backedge comparing the specified
5508 /// value to zero will execute. If not computable, return CouldNotCompute.
5510 /// This is only used for loops with a "x != y" exit test. The exit condition is
5511 /// now expressed as a single expression, V = x-y. So the exit test is
5512 /// effectively V != 0. We know and take advantage of the fact that this
5513 /// expression only being used in a comparison by zero context.
5514 ScalarEvolution::ExitLimit
5515 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L, bool IsSubExpr) {
5516 // If the value is a constant
5517 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5518 // If the value is already zero, the branch will execute zero times.
5519 if (C->getValue()->isZero()) return C;
5520 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5523 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
5524 if (!AddRec || AddRec->getLoop() != L)
5525 return getCouldNotCompute();
5527 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
5528 // the quadratic equation to solve it.
5529 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
5530 std::pair<const SCEV *,const SCEV *> Roots =
5531 SolveQuadraticEquation(AddRec, *this);
5532 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5533 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5536 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
5537 << " sol#2: " << *R2 << "\n";
5539 // Pick the smallest positive root value.
5540 if (ConstantInt *CB =
5541 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
5544 if (CB->getZExtValue() == false)
5545 std::swap(R1, R2); // R1 is the minimum root now.
5547 // We can only use this value if the chrec ends up with an exact zero
5548 // value at this index. When solving for "X*X != 5", for example, we
5549 // should not accept a root of 2.
5550 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
5552 return R1; // We found a quadratic root!
5555 return getCouldNotCompute();
5558 // Otherwise we can only handle this if it is affine.
5559 if (!AddRec->isAffine())
5560 return getCouldNotCompute();
5562 // If this is an affine expression, the execution count of this branch is
5563 // the minimum unsigned root of the following equation:
5565 // Start + Step*N = 0 (mod 2^BW)
5569 // Step*N = -Start (mod 2^BW)
5571 // where BW is the common bit width of Start and Step.
5573 // Get the initial value for the loop.
5574 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
5575 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
5577 // For now we handle only constant steps.
5579 // TODO: Handle a nonconstant Step given AddRec<NUW>. If the
5580 // AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap
5581 // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step.
5582 // We have not yet seen any such cases.
5583 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
5584 if (StepC == 0 || StepC->getValue()->equalsInt(0))
5585 return getCouldNotCompute();
5587 // For positive steps (counting up until unsigned overflow):
5588 // N = -Start/Step (as unsigned)
5589 // For negative steps (counting down to zero):
5591 // First compute the unsigned distance from zero in the direction of Step.
5592 bool CountDown = StepC->getValue()->getValue().isNegative();
5593 const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
5595 // Handle unitary steps, which cannot wraparound.
5596 // 1*N = -Start; -1*N = Start (mod 2^BW), so:
5597 // N = Distance (as unsigned)
5598 if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) {
5599 ConstantRange CR = getUnsignedRange(Start);
5600 const SCEV *MaxBECount;
5601 if (!CountDown && CR.getUnsignedMin().isMinValue())
5602 // When counting up, the worst starting value is 1, not 0.
5603 MaxBECount = CR.getUnsignedMax().isMinValue()
5604 ? getConstant(APInt::getMinValue(CR.getBitWidth()))
5605 : getConstant(APInt::getMaxValue(CR.getBitWidth()));
5607 MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
5608 : -CR.getUnsignedMin());
5609 return ExitLimit(Distance, MaxBECount, /*MustExit=*/true);
5612 // If the recurrence is known not to wraparound, unsigned divide computes the
5613 // back edge count. (Ideally we would have an "isexact" bit for udiv). We know
5614 // that the value will either become zero (and thus the loop terminates), that
5615 // the loop will terminate through some other exit condition first, or that
5616 // the loop has undefined behavior. This means we can't "miss" the exit
5617 // value, even with nonunit stride, and exit later via the same branch. Note
5618 // that we can skip this exit if loop later exits via a different
5619 // branch. Hence MustExit=false.
5621 // This is only valid for expressions that directly compute the loop exit. It
5622 // is invalid for subexpressions in which the loop may exit through this
5623 // branch even if this subexpression is false. In that case, the trip count
5624 // computed by this udiv could be smaller than the number of well-defined
5626 if (!IsSubExpr && AddRec->getNoWrapFlags(SCEV::FlagNW)) {
5628 getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
5629 return ExitLimit(Exact, Exact, /*MustExit=*/false);
5631 // Then, try to solve the above equation provided that Start is constant.
5632 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
5633 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
5634 -StartC->getValue()->getValue(),
5636 return getCouldNotCompute();
5639 /// HowFarToNonZero - Return the number of times a backedge checking the
5640 /// specified value for nonzero will execute. If not computable, return
5642 ScalarEvolution::ExitLimit
5643 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
5644 // Loops that look like: while (X == 0) are very strange indeed. We don't
5645 // handle them yet except for the trivial case. This could be expanded in the
5646 // future as needed.
5648 // If the value is a constant, check to see if it is known to be non-zero
5649 // already. If so, the backedge will execute zero times.
5650 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5651 if (!C->getValue()->isNullValue())
5652 return getConstant(C->getType(), 0);
5653 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5656 // We could implement others, but I really doubt anyone writes loops like
5657 // this, and if they did, they would already be constant folded.
5658 return getCouldNotCompute();
5661 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
5662 /// (which may not be an immediate predecessor) which has exactly one
5663 /// successor from which BB is reachable, or null if no such block is
5666 std::pair<BasicBlock *, BasicBlock *>
5667 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
5668 // If the block has a unique predecessor, then there is no path from the
5669 // predecessor to the block that does not go through the direct edge
5670 // from the predecessor to the block.
5671 if (BasicBlock *Pred = BB->getSinglePredecessor())
5672 return std::make_pair(Pred, BB);
5674 // A loop's header is defined to be a block that dominates the loop.
5675 // If the header has a unique predecessor outside the loop, it must be
5676 // a block that has exactly one successor that can reach the loop.
5677 if (Loop *L = LI->getLoopFor(BB))
5678 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
5680 return std::pair<BasicBlock *, BasicBlock *>();
5683 /// HasSameValue - SCEV structural equivalence is usually sufficient for
5684 /// testing whether two expressions are equal, however for the purposes of
5685 /// looking for a condition guarding a loop, it can be useful to be a little
5686 /// more general, since a front-end may have replicated the controlling
5689 static bool HasSameValue(const SCEV *A, const SCEV *B) {
5690 // Quick check to see if they are the same SCEV.
5691 if (A == B) return true;
5693 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
5694 // two different instructions with the same value. Check for this case.
5695 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
5696 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
5697 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
5698 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
5699 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
5702 // Otherwise assume they may have a different value.
5706 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
5707 /// predicate Pred. Return true iff any changes were made.
5709 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
5710 const SCEV *&LHS, const SCEV *&RHS,
5712 bool Changed = false;
5714 // If we hit the max recursion limit bail out.
5718 // Canonicalize a constant to the right side.
5719 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
5720 // Check for both operands constant.
5721 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
5722 if (ConstantExpr::getICmp(Pred,
5724 RHSC->getValue())->isNullValue())
5725 goto trivially_false;
5727 goto trivially_true;
5729 // Otherwise swap the operands to put the constant on the right.
5730 std::swap(LHS, RHS);
5731 Pred = ICmpInst::getSwappedPredicate(Pred);
5735 // If we're comparing an addrec with a value which is loop-invariant in the
5736 // addrec's loop, put the addrec on the left. Also make a dominance check,
5737 // as both operands could be addrecs loop-invariant in each other's loop.
5738 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
5739 const Loop *L = AR->getLoop();
5740 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
5741 std::swap(LHS, RHS);
5742 Pred = ICmpInst::getSwappedPredicate(Pred);
5747 // If there's a constant operand, canonicalize comparisons with boundary
5748 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
5749 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
5750 const APInt &RA = RC->getValue()->getValue();
5752 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5753 case ICmpInst::ICMP_EQ:
5754 case ICmpInst::ICMP_NE:
5755 // Fold ((-1) * %a) + %b == 0 (equivalent to %b-%a == 0) into %a == %b.
5757 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(LHS))
5758 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(AE->getOperand(0)))
5759 if (AE->getNumOperands() == 2 && ME->getNumOperands() == 2 &&
5760 ME->getOperand(0)->isAllOnesValue()) {
5761 RHS = AE->getOperand(1);
5762 LHS = ME->getOperand(1);
5766 case ICmpInst::ICMP_UGE:
5767 if ((RA - 1).isMinValue()) {
5768 Pred = ICmpInst::ICMP_NE;
5769 RHS = getConstant(RA - 1);
5773 if (RA.isMaxValue()) {
5774 Pred = ICmpInst::ICMP_EQ;
5778 if (RA.isMinValue()) goto trivially_true;
5780 Pred = ICmpInst::ICMP_UGT;
5781 RHS = getConstant(RA - 1);
5784 case ICmpInst::ICMP_ULE:
5785 if ((RA + 1).isMaxValue()) {
5786 Pred = ICmpInst::ICMP_NE;
5787 RHS = getConstant(RA + 1);
5791 if (RA.isMinValue()) {
5792 Pred = ICmpInst::ICMP_EQ;
5796 if (RA.isMaxValue()) goto trivially_true;
5798 Pred = ICmpInst::ICMP_ULT;
5799 RHS = getConstant(RA + 1);
5802 case ICmpInst::ICMP_SGE:
5803 if ((RA - 1).isMinSignedValue()) {
5804 Pred = ICmpInst::ICMP_NE;
5805 RHS = getConstant(RA - 1);
5809 if (RA.isMaxSignedValue()) {
5810 Pred = ICmpInst::ICMP_EQ;
5814 if (RA.isMinSignedValue()) goto trivially_true;
5816 Pred = ICmpInst::ICMP_SGT;
5817 RHS = getConstant(RA - 1);
5820 case ICmpInst::ICMP_SLE:
5821 if ((RA + 1).isMaxSignedValue()) {
5822 Pred = ICmpInst::ICMP_NE;
5823 RHS = getConstant(RA + 1);
5827 if (RA.isMinSignedValue()) {
5828 Pred = ICmpInst::ICMP_EQ;
5832 if (RA.isMaxSignedValue()) goto trivially_true;
5834 Pred = ICmpInst::ICMP_SLT;
5835 RHS = getConstant(RA + 1);
5838 case ICmpInst::ICMP_UGT:
5839 if (RA.isMinValue()) {
5840 Pred = ICmpInst::ICMP_NE;
5844 if ((RA + 1).isMaxValue()) {
5845 Pred = ICmpInst::ICMP_EQ;
5846 RHS = getConstant(RA + 1);
5850 if (RA.isMaxValue()) goto trivially_false;
5852 case ICmpInst::ICMP_ULT:
5853 if (RA.isMaxValue()) {
5854 Pred = ICmpInst::ICMP_NE;
5858 if ((RA - 1).isMinValue()) {
5859 Pred = ICmpInst::ICMP_EQ;
5860 RHS = getConstant(RA - 1);
5864 if (RA.isMinValue()) goto trivially_false;
5866 case ICmpInst::ICMP_SGT:
5867 if (RA.isMinSignedValue()) {
5868 Pred = ICmpInst::ICMP_NE;
5872 if ((RA + 1).isMaxSignedValue()) {
5873 Pred = ICmpInst::ICMP_EQ;
5874 RHS = getConstant(RA + 1);
5878 if (RA.isMaxSignedValue()) goto trivially_false;
5880 case ICmpInst::ICMP_SLT:
5881 if (RA.isMaxSignedValue()) {
5882 Pred = ICmpInst::ICMP_NE;
5886 if ((RA - 1).isMinSignedValue()) {
5887 Pred = ICmpInst::ICMP_EQ;
5888 RHS = getConstant(RA - 1);
5892 if (RA.isMinSignedValue()) goto trivially_false;
5897 // Check for obvious equality.
5898 if (HasSameValue(LHS, RHS)) {
5899 if (ICmpInst::isTrueWhenEqual(Pred))
5900 goto trivially_true;
5901 if (ICmpInst::isFalseWhenEqual(Pred))
5902 goto trivially_false;
5905 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5906 // adding or subtracting 1 from one of the operands.
5908 case ICmpInst::ICMP_SLE:
5909 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5910 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5912 Pred = ICmpInst::ICMP_SLT;
5914 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5915 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5917 Pred = ICmpInst::ICMP_SLT;
5921 case ICmpInst::ICMP_SGE:
5922 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5923 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5925 Pred = ICmpInst::ICMP_SGT;
5927 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5928 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5930 Pred = ICmpInst::ICMP_SGT;
5934 case ICmpInst::ICMP_ULE:
5935 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5936 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5938 Pred = ICmpInst::ICMP_ULT;
5940 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5941 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5943 Pred = ICmpInst::ICMP_ULT;
5947 case ICmpInst::ICMP_UGE:
5948 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5949 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5951 Pred = ICmpInst::ICMP_UGT;
5953 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5954 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5956 Pred = ICmpInst::ICMP_UGT;
5964 // TODO: More simplifications are possible here.
5966 // Recursively simplify until we either hit a recursion limit or nothing
5969 return SimplifyICmpOperands(Pred, LHS, RHS, Depth+1);
5975 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5976 Pred = ICmpInst::ICMP_EQ;
5981 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5982 Pred = ICmpInst::ICMP_NE;
5986 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5987 return getSignedRange(S).getSignedMax().isNegative();
5990 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5991 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5994 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5995 return !getSignedRange(S).getSignedMin().isNegative();
5998 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5999 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
6002 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
6003 return isKnownNegative(S) || isKnownPositive(S);
6006 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
6007 const SCEV *LHS, const SCEV *RHS) {
6008 // Canonicalize the inputs first.
6009 (void)SimplifyICmpOperands(Pred, LHS, RHS);
6011 // If LHS or RHS is an addrec, check to see if the condition is true in
6012 // every iteration of the loop.
6013 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
6014 if (isLoopEntryGuardedByCond(
6015 AR->getLoop(), Pred, AR->getStart(), RHS) &&
6016 isLoopBackedgeGuardedByCond(
6017 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
6019 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
6020 if (isLoopEntryGuardedByCond(
6021 AR->getLoop(), Pred, LHS, AR->getStart()) &&
6022 isLoopBackedgeGuardedByCond(
6023 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
6026 // Otherwise see what can be done with known constant ranges.
6027 return isKnownPredicateWithRanges(Pred, LHS, RHS);
6031 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
6032 const SCEV *LHS, const SCEV *RHS) {
6033 if (HasSameValue(LHS, RHS))
6034 return ICmpInst::isTrueWhenEqual(Pred);
6036 // This code is split out from isKnownPredicate because it is called from
6037 // within isLoopEntryGuardedByCond.
6040 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
6041 case ICmpInst::ICMP_SGT:
6042 Pred = ICmpInst::ICMP_SLT;
6043 std::swap(LHS, RHS);
6044 case ICmpInst::ICMP_SLT: {
6045 ConstantRange LHSRange = getSignedRange(LHS);
6046 ConstantRange RHSRange = getSignedRange(RHS);
6047 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
6049 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
6053 case ICmpInst::ICMP_SGE:
6054 Pred = ICmpInst::ICMP_SLE;
6055 std::swap(LHS, RHS);
6056 case ICmpInst::ICMP_SLE: {
6057 ConstantRange LHSRange = getSignedRange(LHS);
6058 ConstantRange RHSRange = getSignedRange(RHS);
6059 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
6061 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
6065 case ICmpInst::ICMP_UGT:
6066 Pred = ICmpInst::ICMP_ULT;
6067 std::swap(LHS, RHS);
6068 case ICmpInst::ICMP_ULT: {
6069 ConstantRange LHSRange = getUnsignedRange(LHS);
6070 ConstantRange RHSRange = getUnsignedRange(RHS);
6071 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
6073 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
6077 case ICmpInst::ICMP_UGE:
6078 Pred = ICmpInst::ICMP_ULE;
6079 std::swap(LHS, RHS);
6080 case ICmpInst::ICMP_ULE: {
6081 ConstantRange LHSRange = getUnsignedRange(LHS);
6082 ConstantRange RHSRange = getUnsignedRange(RHS);
6083 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
6085 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
6089 case ICmpInst::ICMP_NE: {
6090 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
6092 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
6095 const SCEV *Diff = getMinusSCEV(LHS, RHS);
6096 if (isKnownNonZero(Diff))
6100 case ICmpInst::ICMP_EQ:
6101 // The check at the top of the function catches the case where
6102 // the values are known to be equal.
6108 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
6109 /// protected by a conditional between LHS and RHS. This is used to
6110 /// to eliminate casts.
6112 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
6113 ICmpInst::Predicate Pred,
6114 const SCEV *LHS, const SCEV *RHS) {
6115 // Interpret a null as meaning no loop, where there is obviously no guard
6116 // (interprocedural conditions notwithstanding).
6117 if (!L) return true;
6119 BasicBlock *Latch = L->getLoopLatch();
6123 BranchInst *LoopContinuePredicate =
6124 dyn_cast<BranchInst>(Latch->getTerminator());
6125 if (!LoopContinuePredicate ||
6126 LoopContinuePredicate->isUnconditional())
6129 return isImpliedCond(Pred, LHS, RHS,
6130 LoopContinuePredicate->getCondition(),
6131 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
6134 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
6135 /// by a conditional between LHS and RHS. This is used to help avoid max
6136 /// expressions in loop trip counts, and to eliminate casts.
6138 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
6139 ICmpInst::Predicate Pred,
6140 const SCEV *LHS, const SCEV *RHS) {
6141 // Interpret a null as meaning no loop, where there is obviously no guard
6142 // (interprocedural conditions notwithstanding).
6143 if (!L) return false;
6145 // Starting at the loop predecessor, climb up the predecessor chain, as long
6146 // as there are predecessors that can be found that have unique successors
6147 // leading to the original header.
6148 for (std::pair<BasicBlock *, BasicBlock *>
6149 Pair(L->getLoopPredecessor(), L->getHeader());
6151 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
6153 BranchInst *LoopEntryPredicate =
6154 dyn_cast<BranchInst>(Pair.first->getTerminator());
6155 if (!LoopEntryPredicate ||
6156 LoopEntryPredicate->isUnconditional())
6159 if (isImpliedCond(Pred, LHS, RHS,
6160 LoopEntryPredicate->getCondition(),
6161 LoopEntryPredicate->getSuccessor(0) != Pair.second))
6168 /// RAII wrapper to prevent recursive application of isImpliedCond.
6169 /// ScalarEvolution's PendingLoopPredicates set must be empty unless we are
6170 /// currently evaluating isImpliedCond.
6171 struct MarkPendingLoopPredicate {
6173 DenseSet<Value*> &LoopPreds;
6176 MarkPendingLoopPredicate(Value *C, DenseSet<Value*> &LP)
6177 : Cond(C), LoopPreds(LP) {
6178 Pending = !LoopPreds.insert(Cond).second;
6180 ~MarkPendingLoopPredicate() {
6182 LoopPreds.erase(Cond);
6186 /// isImpliedCond - Test whether the condition described by Pred, LHS,
6187 /// and RHS is true whenever the given Cond value evaluates to true.
6188 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
6189 const SCEV *LHS, const SCEV *RHS,
6190 Value *FoundCondValue,
6192 MarkPendingLoopPredicate Mark(FoundCondValue, PendingLoopPredicates);
6196 // Recursively handle And and Or conditions.
6197 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
6198 if (BO->getOpcode() == Instruction::And) {
6200 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6201 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6202 } else if (BO->getOpcode() == Instruction::Or) {
6204 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6205 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6209 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
6210 if (!ICI) return false;
6212 // Bail if the ICmp's operands' types are wider than the needed type
6213 // before attempting to call getSCEV on them. This avoids infinite
6214 // recursion, since the analysis of widening casts can require loop
6215 // exit condition information for overflow checking, which would
6217 if (getTypeSizeInBits(LHS->getType()) <
6218 getTypeSizeInBits(ICI->getOperand(0)->getType()))
6221 // Now that we found a conditional branch that dominates the loop or controls
6222 // the loop latch. Check to see if it is the comparison we are looking for.
6223 ICmpInst::Predicate FoundPred;
6225 FoundPred = ICI->getInversePredicate();
6227 FoundPred = ICI->getPredicate();
6229 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
6230 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
6232 // Balance the types. The case where FoundLHS' type is wider than
6233 // LHS' type is checked for above.
6234 if (getTypeSizeInBits(LHS->getType()) >
6235 getTypeSizeInBits(FoundLHS->getType())) {
6236 if (CmpInst::isSigned(FoundPred)) {
6237 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
6238 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
6240 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
6241 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
6245 // Canonicalize the query to match the way instcombine will have
6246 // canonicalized the comparison.
6247 if (SimplifyICmpOperands(Pred, LHS, RHS))
6249 return CmpInst::isTrueWhenEqual(Pred);
6250 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
6251 if (FoundLHS == FoundRHS)
6252 return CmpInst::isFalseWhenEqual(FoundPred);
6254 // Check to see if we can make the LHS or RHS match.
6255 if (LHS == FoundRHS || RHS == FoundLHS) {
6256 if (isa<SCEVConstant>(RHS)) {
6257 std::swap(FoundLHS, FoundRHS);
6258 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
6260 std::swap(LHS, RHS);
6261 Pred = ICmpInst::getSwappedPredicate(Pred);
6265 // Check whether the found predicate is the same as the desired predicate.
6266 if (FoundPred == Pred)
6267 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
6269 // Check whether swapping the found predicate makes it the same as the
6270 // desired predicate.
6271 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
6272 if (isa<SCEVConstant>(RHS))
6273 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
6275 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
6276 RHS, LHS, FoundLHS, FoundRHS);
6279 // Check whether the actual condition is beyond sufficient.
6280 if (FoundPred == ICmpInst::ICMP_EQ)
6281 if (ICmpInst::isTrueWhenEqual(Pred))
6282 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
6284 if (Pred == ICmpInst::ICMP_NE)
6285 if (!ICmpInst::isTrueWhenEqual(FoundPred))
6286 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
6289 // Otherwise assume the worst.
6293 /// isImpliedCondOperands - Test whether the condition described by Pred,
6294 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
6295 /// and FoundRHS is true.
6296 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
6297 const SCEV *LHS, const SCEV *RHS,
6298 const SCEV *FoundLHS,
6299 const SCEV *FoundRHS) {
6300 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
6301 FoundLHS, FoundRHS) ||
6302 // ~x < ~y --> x > y
6303 isImpliedCondOperandsHelper(Pred, LHS, RHS,
6304 getNotSCEV(FoundRHS),
6305 getNotSCEV(FoundLHS));
6308 /// isImpliedCondOperandsHelper - Test whether the condition described by
6309 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
6310 /// FoundLHS, and FoundRHS is true.
6312 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
6313 const SCEV *LHS, const SCEV *RHS,
6314 const SCEV *FoundLHS,
6315 const SCEV *FoundRHS) {
6317 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
6318 case ICmpInst::ICMP_EQ:
6319 case ICmpInst::ICMP_NE:
6320 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
6323 case ICmpInst::ICMP_SLT:
6324 case ICmpInst::ICMP_SLE:
6325 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
6326 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
6329 case ICmpInst::ICMP_SGT:
6330 case ICmpInst::ICMP_SGE:
6331 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
6332 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
6335 case ICmpInst::ICMP_ULT:
6336 case ICmpInst::ICMP_ULE:
6337 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
6338 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
6341 case ICmpInst::ICMP_UGT:
6342 case ICmpInst::ICMP_UGE:
6343 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
6344 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
6352 // Verify if an linear IV with positive stride can overflow when in a
6353 // less-than comparison, knowing the invariant term of the comparison, the
6354 // stride and the knowledge of NSW/NUW flags on the recurrence.
6355 bool ScalarEvolution::doesIVOverflowOnLT(const SCEV *RHS, const SCEV *Stride,
6356 bool IsSigned, bool NoWrap) {
6357 if (NoWrap) return false;
6359 unsigned BitWidth = getTypeSizeInBits(RHS->getType());
6360 const SCEV *One = getConstant(Stride->getType(), 1);
6363 APInt MaxRHS = getSignedRange(RHS).getSignedMax();
6364 APInt MaxValue = APInt::getSignedMaxValue(BitWidth);
6365 APInt MaxStrideMinusOne = getSignedRange(getMinusSCEV(Stride, One))
6368 // SMaxRHS + SMaxStrideMinusOne > SMaxValue => overflow!
6369 return (MaxValue - MaxStrideMinusOne).slt(MaxRHS);
6372 APInt MaxRHS = getUnsignedRange(RHS).getUnsignedMax();
6373 APInt MaxValue = APInt::getMaxValue(BitWidth);
6374 APInt MaxStrideMinusOne = getUnsignedRange(getMinusSCEV(Stride, One))
6377 // UMaxRHS + UMaxStrideMinusOne > UMaxValue => overflow!
6378 return (MaxValue - MaxStrideMinusOne).ult(MaxRHS);
6381 // Verify if an linear IV with negative stride can overflow when in a
6382 // greater-than comparison, knowing the invariant term of the comparison,
6383 // the stride and the knowledge of NSW/NUW flags on the recurrence.
6384 bool ScalarEvolution::doesIVOverflowOnGT(const SCEV *RHS, const SCEV *Stride,
6385 bool IsSigned, bool NoWrap) {
6386 if (NoWrap) return false;
6388 unsigned BitWidth = getTypeSizeInBits(RHS->getType());
6389 const SCEV *One = getConstant(Stride->getType(), 1);
6392 APInt MinRHS = getSignedRange(RHS).getSignedMin();
6393 APInt MinValue = APInt::getSignedMinValue(BitWidth);
6394 APInt MaxStrideMinusOne = getSignedRange(getMinusSCEV(Stride, One))
6397 // SMinRHS - SMaxStrideMinusOne < SMinValue => overflow!
6398 return (MinValue + MaxStrideMinusOne).sgt(MinRHS);
6401 APInt MinRHS = getUnsignedRange(RHS).getUnsignedMin();
6402 APInt MinValue = APInt::getMinValue(BitWidth);
6403 APInt MaxStrideMinusOne = getUnsignedRange(getMinusSCEV(Stride, One))
6406 // UMinRHS - UMaxStrideMinusOne < UMinValue => overflow!
6407 return (MinValue + MaxStrideMinusOne).ugt(MinRHS);
6410 // Compute the backedge taken count knowing the interval difference, the
6411 // stride and presence of the equality in the comparison.
6412 const SCEV *ScalarEvolution::computeBECount(const SCEV *Delta, const SCEV *Step,
6414 const SCEV *One = getConstant(Step->getType(), 1);
6415 Delta = Equality ? getAddExpr(Delta, Step)
6416 : getAddExpr(Delta, getMinusSCEV(Step, One));
6417 return getUDivExpr(Delta, Step);
6420 /// HowManyLessThans - Return the number of times a backedge containing the
6421 /// specified less-than comparison will execute. If not computable, return
6422 /// CouldNotCompute.
6424 /// @param IsSubExpr is true when the LHS < RHS condition does not directly
6425 /// control the branch. In this case, we can only compute an iteration count for
6426 /// a subexpression that cannot overflow before evaluating true.
6427 ScalarEvolution::ExitLimit
6428 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
6429 const Loop *L, bool IsSigned,
6431 // We handle only IV < Invariant
6432 if (!isLoopInvariant(RHS, L))
6433 return getCouldNotCompute();
6435 const SCEVAddRecExpr *IV = dyn_cast<SCEVAddRecExpr>(LHS);
6437 // Avoid weird loops
6438 if (!IV || IV->getLoop() != L || !IV->isAffine())
6439 return getCouldNotCompute();
6441 bool NoWrap = !IsSubExpr &&
6442 IV->getNoWrapFlags(IsSigned ? SCEV::FlagNSW : SCEV::FlagNUW);
6444 const SCEV *Stride = IV->getStepRecurrence(*this);
6446 // Avoid negative or zero stride values
6447 if (!isKnownPositive(Stride))
6448 return getCouldNotCompute();
6450 // Avoid proven overflow cases: this will ensure that the backedge taken count
6451 // will not generate any unsigned overflow. Relaxed no-overflow conditions
6452 // exploit NoWrapFlags, allowing to optimize in presence of undefined
6453 // behaviors like the case of C language.
6454 if (!Stride->isOne() && doesIVOverflowOnLT(RHS, Stride, IsSigned, NoWrap))
6455 return getCouldNotCompute();
6457 ICmpInst::Predicate Cond = IsSigned ? ICmpInst::ICMP_SLT
6458 : ICmpInst::ICMP_ULT;
6459 const SCEV *Start = IV->getStart();
6460 const SCEV *End = RHS;
6461 if (!isLoopEntryGuardedByCond(L, Cond, getMinusSCEV(Start, Stride), RHS))
6462 End = IsSigned ? getSMaxExpr(RHS, Start)
6463 : getUMaxExpr(RHS, Start);
6465 const SCEV *BECount = computeBECount(getMinusSCEV(End, Start), Stride, false);
6467 APInt MinStart = IsSigned ? getSignedRange(Start).getSignedMin()
6468 : getUnsignedRange(Start).getUnsignedMin();
6470 APInt MinStride = IsSigned ? getSignedRange(Stride).getSignedMin()
6471 : getUnsignedRange(Stride).getUnsignedMin();
6473 unsigned BitWidth = getTypeSizeInBits(LHS->getType());
6474 APInt Limit = IsSigned ? APInt::getSignedMaxValue(BitWidth) - (MinStride - 1)
6475 : APInt::getMaxValue(BitWidth) - (MinStride - 1);
6477 // Although End can be a MAX expression we estimate MaxEnd considering only
6478 // the case End = RHS. This is safe because in the other case (End - Start)
6479 // is zero, leading to a zero maximum backedge taken count.
6481 IsSigned ? APIntOps::smin(getSignedRange(RHS).getSignedMax(), Limit)
6482 : APIntOps::umin(getUnsignedRange(RHS).getUnsignedMax(), Limit);
6484 const SCEV *MaxBECount = getCouldNotCompute();
6485 if (isa<SCEVConstant>(BECount))
6486 MaxBECount = BECount;
6488 MaxBECount = computeBECount(getConstant(MaxEnd - MinStart),
6489 getConstant(MinStride), false);
6491 if (isa<SCEVCouldNotCompute>(MaxBECount))
6492 MaxBECount = BECount;
6494 return ExitLimit(BECount, MaxBECount, /*MustExit=*/true);
6497 ScalarEvolution::ExitLimit
6498 ScalarEvolution::HowManyGreaterThans(const SCEV *LHS, const SCEV *RHS,
6499 const Loop *L, bool IsSigned,
6501 // We handle only IV > Invariant
6502 if (!isLoopInvariant(RHS, L))
6503 return getCouldNotCompute();
6505 const SCEVAddRecExpr *IV = dyn_cast<SCEVAddRecExpr>(LHS);
6507 // Avoid weird loops
6508 if (!IV || IV->getLoop() != L || !IV->isAffine())
6509 return getCouldNotCompute();
6511 bool NoWrap = !IsSubExpr &&
6512 IV->getNoWrapFlags(IsSigned ? SCEV::FlagNSW : SCEV::FlagNUW);
6514 const SCEV *Stride = getNegativeSCEV(IV->getStepRecurrence(*this));
6516 // Avoid negative or zero stride values
6517 if (!isKnownPositive(Stride))
6518 return getCouldNotCompute();
6520 // Avoid proven overflow cases: this will ensure that the backedge taken count
6521 // will not generate any unsigned overflow. Relaxed no-overflow conditions
6522 // exploit NoWrapFlags, allowing to optimize in presence of undefined
6523 // behaviors like the case of C language.
6524 if (!Stride->isOne() && doesIVOverflowOnGT(RHS, Stride, IsSigned, NoWrap))
6525 return getCouldNotCompute();
6527 ICmpInst::Predicate Cond = IsSigned ? ICmpInst::ICMP_SGT
6528 : ICmpInst::ICMP_UGT;
6530 const SCEV *Start = IV->getStart();
6531 const SCEV *End = RHS;
6532 if (!isLoopEntryGuardedByCond(L, Cond, getAddExpr(Start, Stride), RHS))
6533 End = IsSigned ? getSMinExpr(RHS, Start)
6534 : getUMinExpr(RHS, Start);
6536 const SCEV *BECount = computeBECount(getMinusSCEV(Start, End), Stride, false);
6538 APInt MaxStart = IsSigned ? getSignedRange(Start).getSignedMax()
6539 : getUnsignedRange(Start).getUnsignedMax();
6541 APInt MinStride = IsSigned ? getSignedRange(Stride).getSignedMin()
6542 : getUnsignedRange(Stride).getUnsignedMin();
6544 unsigned BitWidth = getTypeSizeInBits(LHS->getType());
6545 APInt Limit = IsSigned ? APInt::getSignedMinValue(BitWidth) + (MinStride - 1)
6546 : APInt::getMinValue(BitWidth) + (MinStride - 1);
6548 // Although End can be a MIN expression we estimate MinEnd considering only
6549 // the case End = RHS. This is safe because in the other case (Start - End)
6550 // is zero, leading to a zero maximum backedge taken count.
6552 IsSigned ? APIntOps::smax(getSignedRange(RHS).getSignedMin(), Limit)
6553 : APIntOps::umax(getUnsignedRange(RHS).getUnsignedMin(), Limit);
6556 const SCEV *MaxBECount = getCouldNotCompute();
6557 if (isa<SCEVConstant>(BECount))
6558 MaxBECount = BECount;
6560 MaxBECount = computeBECount(getConstant(MaxStart - MinEnd),
6561 getConstant(MinStride), false);
6563 if (isa<SCEVCouldNotCompute>(MaxBECount))
6564 MaxBECount = BECount;
6566 return ExitLimit(BECount, MaxBECount, /*MustExit=*/true);
6569 /// getNumIterationsInRange - Return the number of iterations of this loop that
6570 /// produce values in the specified constant range. Another way of looking at
6571 /// this is that it returns the first iteration number where the value is not in
6572 /// the condition, thus computing the exit count. If the iteration count can't
6573 /// be computed, an instance of SCEVCouldNotCompute is returned.
6574 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
6575 ScalarEvolution &SE) const {
6576 if (Range.isFullSet()) // Infinite loop.
6577 return SE.getCouldNotCompute();
6579 // If the start is a non-zero constant, shift the range to simplify things.
6580 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
6581 if (!SC->getValue()->isZero()) {
6582 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
6583 Operands[0] = SE.getConstant(SC->getType(), 0);
6584 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop(),
6585 getNoWrapFlags(FlagNW));
6586 if (const SCEVAddRecExpr *ShiftedAddRec =
6587 dyn_cast<SCEVAddRecExpr>(Shifted))
6588 return ShiftedAddRec->getNumIterationsInRange(
6589 Range.subtract(SC->getValue()->getValue()), SE);
6590 // This is strange and shouldn't happen.
6591 return SE.getCouldNotCompute();
6594 // The only time we can solve this is when we have all constant indices.
6595 // Otherwise, we cannot determine the overflow conditions.
6596 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
6597 if (!isa<SCEVConstant>(getOperand(i)))
6598 return SE.getCouldNotCompute();
6601 // Okay at this point we know that all elements of the chrec are constants and
6602 // that the start element is zero.
6604 // First check to see if the range contains zero. If not, the first
6606 unsigned BitWidth = SE.getTypeSizeInBits(getType());
6607 if (!Range.contains(APInt(BitWidth, 0)))
6608 return SE.getConstant(getType(), 0);
6611 // If this is an affine expression then we have this situation:
6612 // Solve {0,+,A} in Range === Ax in Range
6614 // We know that zero is in the range. If A is positive then we know that
6615 // the upper value of the range must be the first possible exit value.
6616 // If A is negative then the lower of the range is the last possible loop
6617 // value. Also note that we already checked for a full range.
6618 APInt One(BitWidth,1);
6619 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
6620 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
6622 // The exit value should be (End+A)/A.
6623 APInt ExitVal = (End + A).udiv(A);
6624 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
6626 // Evaluate at the exit value. If we really did fall out of the valid
6627 // range, then we computed our trip count, otherwise wrap around or other
6628 // things must have happened.
6629 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
6630 if (Range.contains(Val->getValue()))
6631 return SE.getCouldNotCompute(); // Something strange happened
6633 // Ensure that the previous value is in the range. This is a sanity check.
6634 assert(Range.contains(
6635 EvaluateConstantChrecAtConstant(this,
6636 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
6637 "Linear scev computation is off in a bad way!");
6638 return SE.getConstant(ExitValue);
6639 } else if (isQuadratic()) {
6640 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
6641 // quadratic equation to solve it. To do this, we must frame our problem in
6642 // terms of figuring out when zero is crossed, instead of when
6643 // Range.getUpper() is crossed.
6644 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
6645 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
6646 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop(),
6647 // getNoWrapFlags(FlagNW)
6650 // Next, solve the constructed addrec
6651 std::pair<const SCEV *,const SCEV *> Roots =
6652 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
6653 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
6654 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
6656 // Pick the smallest positive root value.
6657 if (ConstantInt *CB =
6658 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
6659 R1->getValue(), R2->getValue()))) {
6660 if (CB->getZExtValue() == false)
6661 std::swap(R1, R2); // R1 is the minimum root now.
6663 // Make sure the root is not off by one. The returned iteration should
6664 // not be in the range, but the previous one should be. When solving
6665 // for "X*X < 5", for example, we should not return a root of 2.
6666 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
6669 if (Range.contains(R1Val->getValue())) {
6670 // The next iteration must be out of the range...
6671 ConstantInt *NextVal =
6672 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
6674 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6675 if (!Range.contains(R1Val->getValue()))
6676 return SE.getConstant(NextVal);
6677 return SE.getCouldNotCompute(); // Something strange happened
6680 // If R1 was not in the range, then it is a good return value. Make
6681 // sure that R1-1 WAS in the range though, just in case.
6682 ConstantInt *NextVal =
6683 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
6684 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6685 if (Range.contains(R1Val->getValue()))
6687 return SE.getCouldNotCompute(); // Something strange happened
6692 return SE.getCouldNotCompute();
6695 static const APInt gcd(const SCEVConstant *C1, const SCEVConstant *C2) {
6696 APInt A = C1->getValue()->getValue().abs();
6697 APInt B = C2->getValue()->getValue().abs();
6698 uint32_t ABW = A.getBitWidth();
6699 uint32_t BBW = B.getBitWidth();
6706 return APIntOps::GreatestCommonDivisor(A, B);
6709 static const APInt srem(const SCEVConstant *C1, const SCEVConstant *C2) {
6710 APInt A = C1->getValue()->getValue();
6711 APInt B = C2->getValue()->getValue();
6712 uint32_t ABW = A.getBitWidth();
6713 uint32_t BBW = B.getBitWidth();
6720 return APIntOps::srem(A, B);
6723 static const APInt sdiv(const SCEVConstant *C1, const SCEVConstant *C2) {
6724 APInt A = C1->getValue()->getValue();
6725 APInt B = C2->getValue()->getValue();
6726 uint32_t ABW = A.getBitWidth();
6727 uint32_t BBW = B.getBitWidth();
6734 return APIntOps::sdiv(A, B);
6738 struct SCEVGCD : public SCEVVisitor<SCEVGCD, const SCEV *> {
6740 // Pattern match Step into Start. When Step is a multiply expression, find
6741 // the largest subexpression of Step that appears in Start. When Start is an
6742 // add expression, try to match Step in the subexpressions of Start, non
6743 // matching subexpressions are returned under Remainder.
6744 static const SCEV *findGCD(ScalarEvolution &SE, const SCEV *Start,
6745 const SCEV *Step, const SCEV **Remainder) {
6746 assert(Remainder && "Remainder should not be NULL");
6747 SCEVGCD R(SE, Step, SE.getConstant(Step->getType(), 0));
6748 const SCEV *Res = R.visit(Start);
6749 *Remainder = R.Remainder;
6753 SCEVGCD(ScalarEvolution &S, const SCEV *G, const SCEV *R)
6754 : SE(S), GCD(G), Remainder(R) {
6755 Zero = SE.getConstant(GCD->getType(), 0);
6756 One = SE.getConstant(GCD->getType(), 1);
6759 const SCEV *visitConstant(const SCEVConstant *Constant) {
6760 if (GCD == Constant || Constant == Zero)
6763 if (const SCEVConstant *CGCD = dyn_cast<SCEVConstant>(GCD)) {
6764 const SCEV *Res = SE.getConstant(gcd(Constant, CGCD));
6768 Remainder = SE.getConstant(srem(Constant, CGCD));
6769 Constant = cast<SCEVConstant>(SE.getMinusSCEV(Constant, Remainder));
6770 Res = SE.getConstant(gcd(Constant, CGCD));
6774 // When GCD is not a constant, it could be that the GCD is an Add, Mul,
6775 // AddRec, etc., in which case we want to find out how many times the
6776 // Constant divides the GCD: we then return that as the new GCD.
6777 const SCEV *Rem = Zero;
6778 const SCEV *Res = findGCD(SE, GCD, Constant, &Rem);
6780 if (Res == One || Rem != Zero) {
6781 Remainder = Constant;
6785 assert(isa<SCEVConstant>(Res) && "Res should be a constant");
6786 Remainder = SE.getConstant(srem(Constant, cast<SCEVConstant>(Res)));
6790 const SCEV *visitTruncateExpr(const SCEVTruncateExpr *Expr) {
6796 const SCEV *visitZeroExtendExpr(const SCEVZeroExtendExpr *Expr) {
6802 const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *Expr) {
6808 const SCEV *visitAddExpr(const SCEVAddExpr *Expr) {
6812 for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
6813 const SCEV *Rem = Zero;
6814 const SCEV *Res = findGCD(SE, Expr->getOperand(e - 1 - i), GCD, &Rem);
6816 // FIXME: There may be ambiguous situations: for instance,
6817 // GCD(-4 + (3 * %m), 2 * %m) where 2 divides -4 and %m divides (3 * %m).
6818 // The order in which the AddExpr is traversed computes a different GCD
6823 Remainder = SE.getAddExpr(Remainder, Rem);
6829 const SCEV *visitMulExpr(const SCEVMulExpr *Expr) {
6833 for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
6834 if (Expr->getOperand(i) == GCD)
6838 // If we have not returned yet, it means that GCD is not part of Expr.
6839 const SCEV *PartialGCD = One;
6840 for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
6841 const SCEV *Rem = Zero;
6842 const SCEV *Res = findGCD(SE, Expr->getOperand(i), GCD, &Rem);
6844 // GCD does not divide Expr->getOperand(i).
6849 PartialGCD = SE.getMulExpr(PartialGCD, Res);
6850 if (PartialGCD == GCD)
6854 if (PartialGCD != One)
6858 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(GCD);
6862 // When the GCD is a multiply expression, try to decompose it:
6863 // this occurs when Step does not divide the Start expression
6864 // as in: {(-4 + (3 * %m)),+,(2 * %m)}
6865 for (int i = 0, e = Mul->getNumOperands(); i < e; ++i) {
6866 const SCEV *Rem = Zero;
6867 const SCEV *Res = findGCD(SE, Expr, Mul->getOperand(i), &Rem);
6877 const SCEV *visitUDivExpr(const SCEVUDivExpr *Expr) {
6883 const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) {
6887 if (!Expr->isAffine()) {
6892 const SCEV *Rem = Zero;
6893 const SCEV *Res = findGCD(SE, Expr->getOperand(0), GCD, &Rem);
6895 Remainder = SE.getAddExpr(Remainder, Rem);
6898 Res = findGCD(SE, Expr->getOperand(1), Res, &Rem);
6907 const SCEV *visitSMaxExpr(const SCEVSMaxExpr *Expr) {
6913 const SCEV *visitUMaxExpr(const SCEVUMaxExpr *Expr) {
6919 const SCEV *visitUnknown(const SCEVUnknown *Expr) {
6925 const SCEV *visitCouldNotCompute(const SCEVCouldNotCompute *Expr) {
6930 ScalarEvolution &SE;
6931 const SCEV *GCD, *Remainder, *Zero, *One;
6934 struct SCEVDivision : public SCEVVisitor<SCEVDivision, const SCEV *> {
6936 // Remove from Start all multiples of Step.
6937 static const SCEV *divide(ScalarEvolution &SE, const SCEV *Start,
6939 SCEVDivision D(SE, Step);
6940 const SCEV *Rem = D.Zero;
6942 // The division is guaranteed to succeed: Step should divide Start with no
6944 assert(Step == SCEVGCD::findGCD(SE, Start, Step, &Rem) && Rem == D.Zero &&
6945 "Step should divide Start with no remainder.");
6946 return D.visit(Start);
6949 SCEVDivision(ScalarEvolution &S, const SCEV *G) : SE(S), GCD(G) {
6950 Zero = SE.getConstant(GCD->getType(), 0);
6951 One = SE.getConstant(GCD->getType(), 1);
6954 const SCEV *visitConstant(const SCEVConstant *Constant) {
6955 if (GCD == Constant)
6958 if (const SCEVConstant *CGCD = dyn_cast<SCEVConstant>(GCD))
6959 return SE.getConstant(sdiv(Constant, CGCD));
6963 const SCEV *visitTruncateExpr(const SCEVTruncateExpr *Expr) {
6969 const SCEV *visitZeroExtendExpr(const SCEVZeroExtendExpr *Expr) {
6975 const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *Expr) {
6981 const SCEV *visitAddExpr(const SCEVAddExpr *Expr) {
6985 SmallVector<const SCEV *, 2> Operands;
6986 for (int i = 0, e = Expr->getNumOperands(); i < e; ++i)
6987 Operands.push_back(divide(SE, Expr->getOperand(i), GCD));
6989 if (Operands.size() == 1)
6991 return SE.getAddExpr(Operands);
6994 const SCEV *visitMulExpr(const SCEVMulExpr *Expr) {
6998 bool FoundGCDTerm = false;
6999 for (int i = 0, e = Expr->getNumOperands(); i < e; ++i)
7000 if (Expr->getOperand(i) == GCD)
7001 FoundGCDTerm = true;
7003 SmallVector<const SCEV *, 2> Operands;
7005 FoundGCDTerm = false;
7006 for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
7008 Operands.push_back(Expr->getOperand(i));
7009 else if (Expr->getOperand(i) == GCD)
7010 FoundGCDTerm = true;
7012 Operands.push_back(Expr->getOperand(i));
7015 FoundGCDTerm = false;
7016 const SCEV *PartialGCD = One;
7017 for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
7018 if (PartialGCD == GCD) {
7019 Operands.push_back(Expr->getOperand(i));
7023 const SCEV *Rem = Zero;
7024 const SCEV *Res = SCEVGCD::findGCD(SE, Expr->getOperand(i), GCD, &Rem);
7026 PartialGCD = SE.getMulExpr(PartialGCD, Res);
7027 Operands.push_back(divide(SE, Expr->getOperand(i), GCD));
7029 Operands.push_back(Expr->getOperand(i));
7034 if (Operands.size() == 1)
7036 return SE.getMulExpr(Operands);
7039 const SCEV *visitUDivExpr(const SCEVUDivExpr *Expr) {
7045 const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) {
7049 assert(Expr->isAffine() && "Expr should be affine");
7051 const SCEV *Start = divide(SE, Expr->getStart(), GCD);
7052 const SCEV *Step = divide(SE, Expr->getStepRecurrence(SE), GCD);
7054 return SE.getAddRecExpr(Start, Step, Expr->getLoop(),
7055 Expr->getNoWrapFlags());
7058 const SCEV *visitSMaxExpr(const SCEVSMaxExpr *Expr) {
7064 const SCEV *visitUMaxExpr(const SCEVUMaxExpr *Expr) {
7070 const SCEV *visitUnknown(const SCEVUnknown *Expr) {
7076 const SCEV *visitCouldNotCompute(const SCEVCouldNotCompute *Expr) {
7081 ScalarEvolution &SE;
7082 const SCEV *GCD, *Zero, *One;
7086 /// Splits the SCEV into two vectors of SCEVs representing the subscripts and
7087 /// sizes of an array access. Returns the remainder of the delinearization that
7088 /// is the offset start of the array. The SCEV->delinearize algorithm computes
7089 /// the multiples of SCEV coefficients: that is a pattern matching of sub
7090 /// expressions in the stride and base of a SCEV corresponding to the
7091 /// computation of a GCD (greatest common divisor) of base and stride. When
7092 /// SCEV->delinearize fails, it returns the SCEV unchanged.
7094 /// For example: when analyzing the memory access A[i][j][k] in this loop nest
7096 /// void foo(long n, long m, long o, double A[n][m][o]) {
7098 /// for (long i = 0; i < n; i++)
7099 /// for (long j = 0; j < m; j++)
7100 /// for (long k = 0; k < o; k++)
7101 /// A[i][j][k] = 1.0;
7104 /// the delinearization input is the following AddRec SCEV:
7106 /// AddRec: {{{%A,+,(8 * %m * %o)}<%for.i>,+,(8 * %o)}<%for.j>,+,8}<%for.k>
7108 /// From this SCEV, we are able to say that the base offset of the access is %A
7109 /// because it appears as an offset that does not divide any of the strides in
7112 /// CHECK: Base offset: %A
7114 /// and then SCEV->delinearize determines the size of some of the dimensions of
7115 /// the array as these are the multiples by which the strides are happening:
7117 /// CHECK: ArrayDecl[UnknownSize][%m][%o] with elements of sizeof(double) bytes.
7119 /// Note that the outermost dimension remains of UnknownSize because there are
7120 /// no strides that would help identifying the size of the last dimension: when
7121 /// the array has been statically allocated, one could compute the size of that
7122 /// dimension by dividing the overall size of the array by the size of the known
7123 /// dimensions: %m * %o * 8.
7125 /// Finally delinearize provides the access functions for the array reference
7126 /// that does correspond to A[i][j][k] of the above C testcase:
7128 /// CHECK: ArrayRef[{0,+,1}<%for.i>][{0,+,1}<%for.j>][{0,+,1}<%for.k>]
7130 /// The testcases are checking the output of a function pass:
7131 /// DelinearizationPass that walks through all loads and stores of a function
7132 /// asking for the SCEV of the memory access with respect to all enclosing
7133 /// loops, calling SCEV->delinearize on that and printing the results.
7136 SCEVAddRecExpr::delinearize(ScalarEvolution &SE,
7137 SmallVectorImpl<const SCEV *> &Subscripts,
7138 SmallVectorImpl<const SCEV *> &Sizes) const {
7139 // Early exit in case this SCEV is not an affine multivariate function.
7140 if (!this->isAffine())
7143 const SCEV *Start = this->getStart();
7144 const SCEV *Step = this->getStepRecurrence(SE);
7146 // Build the SCEV representation of the cannonical induction variable in the
7147 // loop of this SCEV.
7148 const SCEV *Zero = SE.getConstant(this->getType(), 0);
7149 const SCEV *One = SE.getConstant(this->getType(), 1);
7151 SE.getAddRecExpr(Zero, One, this->getLoop(), this->getNoWrapFlags());
7153 DEBUG(dbgs() << "(delinearize: " << *this << "\n");
7155 // Currently we fail to delinearize when the stride of this SCEV is 1. We
7156 // could decide to not fail in this case: we could just return 1 for the size
7157 // of the subscript, and this same SCEV for the access function.
7159 DEBUG(dbgs() << "failed to delinearize " << *this << "\n)\n");
7163 // Find the GCD and Remainder of the Start and Step coefficients of this SCEV.
7164 const SCEV *Remainder = NULL;
7165 const SCEV *GCD = SCEVGCD::findGCD(SE, Start, Step, &Remainder);
7167 DEBUG(dbgs() << "GCD: " << *GCD << "\n");
7168 DEBUG(dbgs() << "Remainder: " << *Remainder << "\n");
7170 // Same remark as above: we currently fail the delinearization, although we
7171 // can very well handle this special case.
7173 DEBUG(dbgs() << "failed to delinearize " << *this << "\n)\n");
7177 // As findGCD computed Remainder, GCD divides "Start - Remainder." The
7178 // Quotient is then this SCEV without Remainder, scaled down by the GCD. The
7179 // Quotient is what will be used in the next subscript delinearization.
7180 const SCEV *Quotient =
7181 SCEVDivision::divide(SE, SE.getMinusSCEV(Start, Remainder), GCD);
7182 DEBUG(dbgs() << "Quotient: " << *Quotient << "\n");
7185 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Quotient))
7186 // Recursively call delinearize on the Quotient until there are no more
7187 // multiples that can be recognized.
7188 Rem = AR->delinearize(SE, Subscripts, Sizes);
7192 // Scale up the cannonical induction variable IV by whatever remains from the
7193 // Step after division by the GCD: the GCD is the size of all the sub-array.
7195 Step = SCEVDivision::divide(SE, Step, GCD);
7196 IV = SE.getMulExpr(IV, Step);
7198 // The access function in the current subscript is computed as the cannonical
7199 // induction variable IV (potentially scaled up by the step) and offset by
7200 // Rem, the offset of delinearization in the sub-array.
7201 const SCEV *Index = SE.getAddExpr(IV, Rem);
7203 // Record the access function and the size of the current subscript.
7204 Subscripts.push_back(Index);
7205 Sizes.push_back(GCD);
7208 int Size = Sizes.size();
7209 DEBUG(dbgs() << "succeeded to delinearize " << *this << "\n");
7210 DEBUG(dbgs() << "ArrayDecl[UnknownSize]");
7211 for (int i = 0; i < Size - 1; i++)
7212 DEBUG(dbgs() << "[" << *Sizes[i] << "]");
7213 DEBUG(dbgs() << " with elements of " << *Sizes[Size - 1] << " bytes.\n");
7215 DEBUG(dbgs() << "ArrayRef");
7216 for (int i = 0; i < Size; i++)
7217 DEBUG(dbgs() << "[" << *Subscripts[i] << "]");
7218 DEBUG(dbgs() << "\n)\n");
7224 //===----------------------------------------------------------------------===//
7225 // SCEVCallbackVH Class Implementation
7226 //===----------------------------------------------------------------------===//
7228 void ScalarEvolution::SCEVCallbackVH::deleted() {
7229 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
7230 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
7231 SE->ConstantEvolutionLoopExitValue.erase(PN);
7232 SE->ValueExprMap.erase(getValPtr());
7233 // this now dangles!
7236 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
7237 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
7239 // Forget all the expressions associated with users of the old value,
7240 // so that future queries will recompute the expressions using the new
7242 Value *Old = getValPtr();
7243 SmallVector<User *, 16> Worklist;
7244 SmallPtrSet<User *, 8> Visited;
7245 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
7247 Worklist.push_back(*UI);
7248 while (!Worklist.empty()) {
7249 User *U = Worklist.pop_back_val();
7250 // Deleting the Old value will cause this to dangle. Postpone
7251 // that until everything else is done.
7254 if (!Visited.insert(U))
7256 if (PHINode *PN = dyn_cast<PHINode>(U))
7257 SE->ConstantEvolutionLoopExitValue.erase(PN);
7258 SE->ValueExprMap.erase(U);
7259 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
7261 Worklist.push_back(*UI);
7263 // Delete the Old value.
7264 if (PHINode *PN = dyn_cast<PHINode>(Old))
7265 SE->ConstantEvolutionLoopExitValue.erase(PN);
7266 SE->ValueExprMap.erase(Old);
7267 // this now dangles!
7270 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
7271 : CallbackVH(V), SE(se) {}
7273 //===----------------------------------------------------------------------===//
7274 // ScalarEvolution Class Implementation
7275 //===----------------------------------------------------------------------===//
7277 ScalarEvolution::ScalarEvolution()
7278 : FunctionPass(ID), ValuesAtScopes(64), LoopDispositions(64), BlockDispositions(64), FirstUnknown(0) {
7279 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
7282 bool ScalarEvolution::runOnFunction(Function &F) {
7284 LI = &getAnalysis<LoopInfo>();
7285 TD = getAnalysisIfAvailable<DataLayout>();
7286 TLI = &getAnalysis<TargetLibraryInfo>();
7287 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
7291 void ScalarEvolution::releaseMemory() {
7292 // Iterate through all the SCEVUnknown instances and call their
7293 // destructors, so that they release their references to their values.
7294 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
7298 ValueExprMap.clear();
7300 // Free any extra memory created for ExitNotTakenInfo in the unlikely event
7301 // that a loop had multiple computable exits.
7302 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
7303 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end();
7308 assert(PendingLoopPredicates.empty() && "isImpliedCond garbage");
7310 BackedgeTakenCounts.clear();
7311 ConstantEvolutionLoopExitValue.clear();
7312 ValuesAtScopes.clear();
7313 LoopDispositions.clear();
7314 BlockDispositions.clear();
7315 UnsignedRanges.clear();
7316 SignedRanges.clear();
7317 UniqueSCEVs.clear();
7318 SCEVAllocator.Reset();
7321 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
7322 AU.setPreservesAll();
7323 AU.addRequiredTransitive<LoopInfo>();
7324 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
7325 AU.addRequired<TargetLibraryInfo>();
7328 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
7329 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
7332 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
7334 // Print all inner loops first
7335 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
7336 PrintLoopInfo(OS, SE, *I);
7339 L->getHeader()->printAsOperand(OS, /*PrintType=*/false);
7342 SmallVector<BasicBlock *, 8> ExitBlocks;
7343 L->getExitBlocks(ExitBlocks);
7344 if (ExitBlocks.size() != 1)
7345 OS << "<multiple exits> ";
7347 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
7348 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
7350 OS << "Unpredictable backedge-taken count. ";
7355 L->getHeader()->printAsOperand(OS, /*PrintType=*/false);
7358 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
7359 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
7361 OS << "Unpredictable max backedge-taken count. ";
7367 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
7368 // ScalarEvolution's implementation of the print method is to print
7369 // out SCEV values of all instructions that are interesting. Doing
7370 // this potentially causes it to create new SCEV objects though,
7371 // which technically conflicts with the const qualifier. This isn't
7372 // observable from outside the class though, so casting away the
7373 // const isn't dangerous.
7374 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
7376 OS << "Classifying expressions for: ";
7377 F->printAsOperand(OS, /*PrintType=*/false);
7379 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
7380 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
7383 const SCEV *SV = SE.getSCEV(&*I);
7386 const Loop *L = LI->getLoopFor((*I).getParent());
7388 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
7395 OS << "\t\t" "Exits: ";
7396 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
7397 if (!SE.isLoopInvariant(ExitValue, L)) {
7398 OS << "<<Unknown>>";
7407 OS << "Determining loop execution counts for: ";
7408 F->printAsOperand(OS, /*PrintType=*/false);
7410 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
7411 PrintLoopInfo(OS, &SE, *I);
7414 ScalarEvolution::LoopDisposition
7415 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
7416 SmallVector<std::pair<const Loop *, LoopDisposition>, 2> &Values = LoopDispositions[S];
7417 for (unsigned u = 0; u < Values.size(); u++) {
7418 if (Values[u].first == L)
7419 return Values[u].second;
7421 Values.push_back(std::make_pair(L, LoopVariant));
7422 LoopDisposition D = computeLoopDisposition(S, L);
7423 SmallVector<std::pair<const Loop *, LoopDisposition>, 2> &Values2 = LoopDispositions[S];
7424 for (unsigned u = Values2.size(); u > 0; u--) {
7425 if (Values2[u - 1].first == L) {
7426 Values2[u - 1].second = D;
7433 ScalarEvolution::LoopDisposition
7434 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
7435 switch (S->getSCEVType()) {
7437 return LoopInvariant;
7441 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
7442 case scAddRecExpr: {
7443 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
7445 // If L is the addrec's loop, it's computable.
7446 if (AR->getLoop() == L)
7447 return LoopComputable;
7449 // Add recurrences are never invariant in the function-body (null loop).
7453 // This recurrence is variant w.r.t. L if L contains AR's loop.
7454 if (L->contains(AR->getLoop()))
7457 // This recurrence is invariant w.r.t. L if AR's loop contains L.
7458 if (AR->getLoop()->contains(L))
7459 return LoopInvariant;
7461 // This recurrence is variant w.r.t. L if any of its operands
7463 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
7465 if (!isLoopInvariant(*I, L))
7468 // Otherwise it's loop-invariant.
7469 return LoopInvariant;
7475 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
7476 bool HasVarying = false;
7477 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
7479 LoopDisposition D = getLoopDisposition(*I, L);
7480 if (D == LoopVariant)
7482 if (D == LoopComputable)
7485 return HasVarying ? LoopComputable : LoopInvariant;
7488 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
7489 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
7490 if (LD == LoopVariant)
7492 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
7493 if (RD == LoopVariant)
7495 return (LD == LoopInvariant && RD == LoopInvariant) ?
7496 LoopInvariant : LoopComputable;
7499 // All non-instruction values are loop invariant. All instructions are loop
7500 // invariant if they are not contained in the specified loop.
7501 // Instructions are never considered invariant in the function body
7502 // (null loop) because they are defined within the "loop".
7503 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
7504 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
7505 return LoopInvariant;
7506 case scCouldNotCompute:
7507 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
7508 default: llvm_unreachable("Unknown SCEV kind!");
7512 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
7513 return getLoopDisposition(S, L) == LoopInvariant;
7516 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
7517 return getLoopDisposition(S, L) == LoopComputable;
7520 ScalarEvolution::BlockDisposition
7521 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
7522 SmallVector<std::pair<const BasicBlock *, BlockDisposition>, 2> &Values = BlockDispositions[S];
7523 for (unsigned u = 0; u < Values.size(); u++) {
7524 if (Values[u].first == BB)
7525 return Values[u].second;
7527 Values.push_back(std::make_pair(BB, DoesNotDominateBlock));
7528 BlockDisposition D = computeBlockDisposition(S, BB);
7529 SmallVector<std::pair<const BasicBlock *, BlockDisposition>, 2> &Values2 = BlockDispositions[S];
7530 for (unsigned u = Values2.size(); u > 0; u--) {
7531 if (Values2[u - 1].first == BB) {
7532 Values2[u - 1].second = D;
7539 ScalarEvolution::BlockDisposition
7540 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
7541 switch (S->getSCEVType()) {
7543 return ProperlyDominatesBlock;
7547 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
7548 case scAddRecExpr: {
7549 // This uses a "dominates" query instead of "properly dominates" query
7550 // to test for proper dominance too, because the instruction which
7551 // produces the addrec's value is a PHI, and a PHI effectively properly
7552 // dominates its entire containing block.
7553 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
7554 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
7555 return DoesNotDominateBlock;
7557 // FALL THROUGH into SCEVNAryExpr handling.
7562 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
7564 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
7566 BlockDisposition D = getBlockDisposition(*I, BB);
7567 if (D == DoesNotDominateBlock)
7568 return DoesNotDominateBlock;
7569 if (D == DominatesBlock)
7572 return Proper ? ProperlyDominatesBlock : DominatesBlock;
7575 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
7576 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
7577 BlockDisposition LD = getBlockDisposition(LHS, BB);
7578 if (LD == DoesNotDominateBlock)
7579 return DoesNotDominateBlock;
7580 BlockDisposition RD = getBlockDisposition(RHS, BB);
7581 if (RD == DoesNotDominateBlock)
7582 return DoesNotDominateBlock;
7583 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
7584 ProperlyDominatesBlock : DominatesBlock;
7587 if (Instruction *I =
7588 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
7589 if (I->getParent() == BB)
7590 return DominatesBlock;
7591 if (DT->properlyDominates(I->getParent(), BB))
7592 return ProperlyDominatesBlock;
7593 return DoesNotDominateBlock;
7595 return ProperlyDominatesBlock;
7596 case scCouldNotCompute:
7597 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
7599 llvm_unreachable("Unknown SCEV kind!");
7603 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
7604 return getBlockDisposition(S, BB) >= DominatesBlock;
7607 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
7608 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
7612 // Search for a SCEV expression node within an expression tree.
7613 // Implements SCEVTraversal::Visitor.
7618 SCEVSearch(const SCEV *N): Node(N), IsFound(false) {}
7620 bool follow(const SCEV *S) {
7621 IsFound |= (S == Node);
7624 bool isDone() const { return IsFound; }
7628 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
7629 SCEVSearch Search(Op);
7630 visitAll(S, Search);
7631 return Search.IsFound;
7634 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
7635 ValuesAtScopes.erase(S);
7636 LoopDispositions.erase(S);
7637 BlockDispositions.erase(S);
7638 UnsignedRanges.erase(S);
7639 SignedRanges.erase(S);
7641 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
7642 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end(); I != E; ) {
7643 BackedgeTakenInfo &BEInfo = I->second;
7644 if (BEInfo.hasOperand(S, this)) {
7646 BackedgeTakenCounts.erase(I++);
7653 typedef DenseMap<const Loop *, std::string> VerifyMap;
7655 /// replaceSubString - Replaces all occurences of From in Str with To.
7656 static void replaceSubString(std::string &Str, StringRef From, StringRef To) {
7658 while ((Pos = Str.find(From, Pos)) != std::string::npos) {
7659 Str.replace(Pos, From.size(), To.data(), To.size());
7664 /// getLoopBackedgeTakenCounts - Helper method for verifyAnalysis.
7666 getLoopBackedgeTakenCounts(Loop *L, VerifyMap &Map, ScalarEvolution &SE) {
7667 for (Loop::reverse_iterator I = L->rbegin(), E = L->rend(); I != E; ++I) {
7668 getLoopBackedgeTakenCounts(*I, Map, SE); // recurse.
7670 std::string &S = Map[L];
7672 raw_string_ostream OS(S);
7673 SE.getBackedgeTakenCount(L)->print(OS);
7675 // false and 0 are semantically equivalent. This can happen in dead loops.
7676 replaceSubString(OS.str(), "false", "0");
7677 // Remove wrap flags, their use in SCEV is highly fragile.
7678 // FIXME: Remove this when SCEV gets smarter about them.
7679 replaceSubString(OS.str(), "<nw>", "");
7680 replaceSubString(OS.str(), "<nsw>", "");
7681 replaceSubString(OS.str(), "<nuw>", "");
7686 void ScalarEvolution::verifyAnalysis() const {
7690 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
7692 // Gather stringified backedge taken counts for all loops using SCEV's caches.
7693 // FIXME: It would be much better to store actual values instead of strings,
7694 // but SCEV pointers will change if we drop the caches.
7695 VerifyMap BackedgeDumpsOld, BackedgeDumpsNew;
7696 for (LoopInfo::reverse_iterator I = LI->rbegin(), E = LI->rend(); I != E; ++I)
7697 getLoopBackedgeTakenCounts(*I, BackedgeDumpsOld, SE);
7699 // Gather stringified backedge taken counts for all loops without using
7702 for (LoopInfo::reverse_iterator I = LI->rbegin(), E = LI->rend(); I != E; ++I)
7703 getLoopBackedgeTakenCounts(*I, BackedgeDumpsNew, SE);
7705 // Now compare whether they're the same with and without caches. This allows
7706 // verifying that no pass changed the cache.
7707 assert(BackedgeDumpsOld.size() == BackedgeDumpsNew.size() &&
7708 "New loops suddenly appeared!");
7710 for (VerifyMap::iterator OldI = BackedgeDumpsOld.begin(),
7711 OldE = BackedgeDumpsOld.end(),
7712 NewI = BackedgeDumpsNew.begin();
7713 OldI != OldE; ++OldI, ++NewI) {
7714 assert(OldI->first == NewI->first && "Loop order changed!");
7716 // Compare the stringified SCEVs. We don't care if undef backedgetaken count
7718 // FIXME: We currently ignore SCEV changes from/to CouldNotCompute. This
7719 // means that a pass is buggy or SCEV has to learn a new pattern but is
7720 // usually not harmful.
7721 if (OldI->second != NewI->second &&
7722 OldI->second.find("undef") == std::string::npos &&
7723 NewI->second.find("undef") == std::string::npos &&
7724 OldI->second != "***COULDNOTCOMPUTE***" &&
7725 NewI->second != "***COULDNOTCOMPUTE***") {
7726 dbgs() << "SCEVValidator: SCEV for loop '"
7727 << OldI->first->getHeader()->getName()
7728 << "' changed from '" << OldI->second
7729 << "' to '" << NewI->second << "'!\n";
7734 // TODO: Verify more things.