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. These classes are reference counted, managed by the SCEVHandle
18 // class. We only create one SCEV of a particular shape, so pointer-comparisons
19 // for equality are legal.
21 // One important aspect of the SCEV objects is that they are never cyclic, even
22 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
23 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
24 // recurrence) then we represent it directly as a recurrence node, otherwise we
25 // represent it as a SCEVUnknown node.
27 // In addition to being able to represent expressions of various types, we also
28 // have folders that are used to build the *canonical* representation for a
29 // particular expression. These folders are capable of using a variety of
30 // rewrite rules to simplify the expressions.
32 // Once the folders are defined, we can implement the more interesting
33 // higher-level code, such as the code that recognizes PHI nodes of various
34 // types, computes the execution count of a loop, etc.
36 // TODO: We should use these routines and value representations to implement
37 // dependence analysis!
39 //===----------------------------------------------------------------------===//
41 // There are several good references for the techniques used in this analysis.
43 // Chains of recurrences -- a method to expedite the evaluation
44 // of closed-form functions
45 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
47 // On computational properties of chains of recurrences
50 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
51 // Robert A. van Engelen
53 // Efficient Symbolic Analysis for Optimizing Compilers
54 // Robert A. van Engelen
56 // Using the chains of recurrences algebra for data dependence testing and
57 // induction variable substitution
58 // MS Thesis, Johnie Birch
60 //===----------------------------------------------------------------------===//
62 #define DEBUG_TYPE "scalar-evolution"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Constants.h"
65 #include "llvm/DerivedTypes.h"
66 #include "llvm/GlobalVariable.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/Analysis/ConstantFolding.h"
69 #include "llvm/Analysis/Dominators.h"
70 #include "llvm/Analysis/LoopInfo.h"
71 #include "llvm/Assembly/Writer.h"
72 #include "llvm/Target/TargetData.h"
73 #include "llvm/Transforms/Scalar.h"
74 #include "llvm/Support/CFG.h"
75 #include "llvm/Support/CommandLine.h"
76 #include "llvm/Support/Compiler.h"
77 #include "llvm/Support/ConstantRange.h"
78 #include "llvm/Support/GetElementPtrTypeIterator.h"
79 #include "llvm/Support/InstIterator.h"
80 #include "llvm/Support/ManagedStatic.h"
81 #include "llvm/Support/MathExtras.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include "llvm/ADT/Statistic.h"
84 #include "llvm/ADT/STLExtras.h"
90 STATISTIC(NumArrayLenItCounts,
91 "Number of trip counts computed with array length");
92 STATISTIC(NumTripCountsComputed,
93 "Number of loops with predictable loop counts");
94 STATISTIC(NumTripCountsNotComputed,
95 "Number of loops without predictable loop counts");
96 STATISTIC(NumBruteForceTripCountsComputed,
97 "Number of loops with trip counts computed by force");
99 static cl::opt<unsigned>
100 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
101 cl::desc("Maximum number of iterations SCEV will "
102 "symbolically execute a constant derived loop"),
105 static RegisterPass<ScalarEvolution>
106 R("scalar-evolution", "Scalar Evolution Analysis", false, true);
107 char ScalarEvolution::ID = 0;
109 //===----------------------------------------------------------------------===//
110 // SCEV class definitions
111 //===----------------------------------------------------------------------===//
113 //===----------------------------------------------------------------------===//
114 // Implementation of the SCEV class.
117 void SCEV::dump() const {
122 void SCEV::print(std::ostream &o) const {
123 raw_os_ostream OS(o);
127 bool SCEV::isZero() const {
128 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
129 return SC->getValue()->isZero();
134 SCEVCouldNotCompute::SCEVCouldNotCompute() : SCEV(scCouldNotCompute) {}
135 SCEVCouldNotCompute::~SCEVCouldNotCompute() {}
137 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
138 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
142 const Type *SCEVCouldNotCompute::getType() const {
143 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
147 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
148 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
152 SCEVHandle SCEVCouldNotCompute::
153 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
154 const SCEVHandle &Conc,
155 ScalarEvolution &SE) const {
159 void SCEVCouldNotCompute::print(raw_ostream &OS) const {
160 OS << "***COULDNOTCOMPUTE***";
163 bool SCEVCouldNotCompute::classof(const SCEV *S) {
164 return S->getSCEVType() == scCouldNotCompute;
168 // SCEVConstants - Only allow the creation of one SCEVConstant for any
169 // particular value. Don't use a SCEVHandle here, or else the object will
171 static ManagedStatic<std::map<ConstantInt*, SCEVConstant*> > SCEVConstants;
174 SCEVConstant::~SCEVConstant() {
175 SCEVConstants->erase(V);
178 SCEVHandle ScalarEvolution::getConstant(ConstantInt *V) {
179 SCEVConstant *&R = (*SCEVConstants)[V];
180 if (R == 0) R = new SCEVConstant(V);
184 SCEVHandle ScalarEvolution::getConstant(const APInt& Val) {
185 return getConstant(ConstantInt::get(Val));
188 const Type *SCEVConstant::getType() const { return V->getType(); }
190 void SCEVConstant::print(raw_ostream &OS) const {
191 WriteAsOperand(OS, V, false);
194 SCEVCastExpr::SCEVCastExpr(unsigned SCEVTy,
195 const SCEVHandle &op, const Type *ty)
196 : SCEV(SCEVTy), Op(op), Ty(ty) {}
198 SCEVCastExpr::~SCEVCastExpr() {}
200 bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
201 return Op->dominates(BB, DT);
204 // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any
205 // particular input. Don't use a SCEVHandle here, or else the object will
207 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
208 SCEVTruncateExpr*> > SCEVTruncates;
210 SCEVTruncateExpr::SCEVTruncateExpr(const SCEVHandle &op, const Type *ty)
211 : SCEVCastExpr(scTruncate, op, ty) {
212 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
213 (Ty->isInteger() || isa<PointerType>(Ty)) &&
214 "Cannot truncate non-integer value!");
217 SCEVTruncateExpr::~SCEVTruncateExpr() {
218 SCEVTruncates->erase(std::make_pair(Op, Ty));
221 void SCEVTruncateExpr::print(raw_ostream &OS) const {
222 OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
225 // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any
226 // particular input. Don't use a SCEVHandle here, or else the object will never
228 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
229 SCEVZeroExtendExpr*> > SCEVZeroExtends;
231 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty)
232 : SCEVCastExpr(scZeroExtend, op, ty) {
233 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
234 (Ty->isInteger() || isa<PointerType>(Ty)) &&
235 "Cannot zero extend non-integer value!");
238 SCEVZeroExtendExpr::~SCEVZeroExtendExpr() {
239 SCEVZeroExtends->erase(std::make_pair(Op, Ty));
242 void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
243 OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
246 // SCEVSignExtends - Only allow the creation of one SCEVSignExtendExpr for any
247 // particular input. Don't use a SCEVHandle here, or else the object will never
249 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
250 SCEVSignExtendExpr*> > SCEVSignExtends;
252 SCEVSignExtendExpr::SCEVSignExtendExpr(const SCEVHandle &op, const Type *ty)
253 : SCEVCastExpr(scSignExtend, op, ty) {
254 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
255 (Ty->isInteger() || isa<PointerType>(Ty)) &&
256 "Cannot sign extend non-integer value!");
259 SCEVSignExtendExpr::~SCEVSignExtendExpr() {
260 SCEVSignExtends->erase(std::make_pair(Op, Ty));
263 void SCEVSignExtendExpr::print(raw_ostream &OS) const {
264 OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
267 // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any
268 // particular input. Don't use a SCEVHandle here, or else the object will never
270 static ManagedStatic<std::map<std::pair<unsigned, std::vector<SCEV*> >,
271 SCEVCommutativeExpr*> > SCEVCommExprs;
273 SCEVCommutativeExpr::~SCEVCommutativeExpr() {
274 SCEVCommExprs->erase(std::make_pair(getSCEVType(),
275 std::vector<SCEV*>(Operands.begin(),
279 void SCEVCommutativeExpr::print(raw_ostream &OS) const {
280 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
281 const char *OpStr = getOperationStr();
282 OS << "(" << *Operands[0];
283 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
284 OS << OpStr << *Operands[i];
288 SCEVHandle SCEVCommutativeExpr::
289 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
290 const SCEVHandle &Conc,
291 ScalarEvolution &SE) const {
292 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
294 getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
295 if (H != getOperand(i)) {
296 std::vector<SCEVHandle> NewOps;
297 NewOps.reserve(getNumOperands());
298 for (unsigned j = 0; j != i; ++j)
299 NewOps.push_back(getOperand(j));
301 for (++i; i != e; ++i)
302 NewOps.push_back(getOperand(i)->
303 replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
305 if (isa<SCEVAddExpr>(this))
306 return SE.getAddExpr(NewOps);
307 else if (isa<SCEVMulExpr>(this))
308 return SE.getMulExpr(NewOps);
309 else if (isa<SCEVSMaxExpr>(this))
310 return SE.getSMaxExpr(NewOps);
311 else if (isa<SCEVUMaxExpr>(this))
312 return SE.getUMaxExpr(NewOps);
314 assert(0 && "Unknown commutative expr!");
320 bool SCEVCommutativeExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
321 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
322 if (!getOperand(i)->dominates(BB, DT))
329 // SCEVUDivs - Only allow the creation of one SCEVUDivExpr for any particular
330 // input. Don't use a SCEVHandle here, or else the object will never be
332 static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>,
333 SCEVUDivExpr*> > SCEVUDivs;
335 SCEVUDivExpr::~SCEVUDivExpr() {
336 SCEVUDivs->erase(std::make_pair(LHS, RHS));
339 bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
340 return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
343 void SCEVUDivExpr::print(raw_ostream &OS) const {
344 OS << "(" << *LHS << " /u " << *RHS << ")";
347 const Type *SCEVUDivExpr::getType() const {
348 return LHS->getType();
351 // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
352 // particular input. Don't use a SCEVHandle here, or else the object will never
354 static ManagedStatic<std::map<std::pair<const Loop *, std::vector<SCEV*> >,
355 SCEVAddRecExpr*> > SCEVAddRecExprs;
357 SCEVAddRecExpr::~SCEVAddRecExpr() {
358 SCEVAddRecExprs->erase(std::make_pair(L,
359 std::vector<SCEV*>(Operands.begin(),
363 bool SCEVAddRecExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
364 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
365 if (!getOperand(i)->dominates(BB, DT))
372 SCEVHandle SCEVAddRecExpr::
373 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
374 const SCEVHandle &Conc,
375 ScalarEvolution &SE) const {
376 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
378 getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
379 if (H != getOperand(i)) {
380 std::vector<SCEVHandle> NewOps;
381 NewOps.reserve(getNumOperands());
382 for (unsigned j = 0; j != i; ++j)
383 NewOps.push_back(getOperand(j));
385 for (++i; i != e; ++i)
386 NewOps.push_back(getOperand(i)->
387 replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
389 return SE.getAddRecExpr(NewOps, L);
396 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
397 // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't
398 // contain L and if the start is invariant.
399 return !QueryLoop->contains(L->getHeader()) &&
400 getOperand(0)->isLoopInvariant(QueryLoop);
404 void SCEVAddRecExpr::print(raw_ostream &OS) const {
405 OS << "{" << *Operands[0];
406 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
407 OS << ",+," << *Operands[i];
408 OS << "}<" << L->getHeader()->getName() + ">";
411 // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular
412 // value. Don't use a SCEVHandle here, or else the object will never be
414 static ManagedStatic<std::map<Value*, SCEVUnknown*> > SCEVUnknowns;
416 SCEVUnknown::~SCEVUnknown() { SCEVUnknowns->erase(V); }
418 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
419 // All non-instruction values are loop invariant. All instructions are loop
420 // invariant if they are not contained in the specified loop.
421 if (Instruction *I = dyn_cast<Instruction>(V))
422 return !L->contains(I->getParent());
426 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
427 if (Instruction *I = dyn_cast<Instruction>(getValue()))
428 return DT->dominates(I->getParent(), BB);
432 const Type *SCEVUnknown::getType() const {
436 void SCEVUnknown::print(raw_ostream &OS) const {
437 if (isa<PointerType>(V->getType()))
438 OS << "(ptrtoint " << *V->getType() << " ";
439 WriteAsOperand(OS, V, false);
440 if (isa<PointerType>(V->getType()))
444 //===----------------------------------------------------------------------===//
446 //===----------------------------------------------------------------------===//
449 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
450 /// than the complexity of the RHS. This comparator is used to canonicalize
452 struct VISIBILITY_HIDDEN SCEVComplexityCompare {
453 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
454 return LHS->getSCEVType() < RHS->getSCEVType();
459 /// GroupByComplexity - Given a list of SCEV objects, order them by their
460 /// complexity, and group objects of the same complexity together by value.
461 /// When this routine is finished, we know that any duplicates in the vector are
462 /// consecutive and that complexity is monotonically increasing.
464 /// Note that we go take special precautions to ensure that we get determinstic
465 /// results from this routine. In other words, we don't want the results of
466 /// this to depend on where the addresses of various SCEV objects happened to
469 static void GroupByComplexity(std::vector<SCEVHandle> &Ops) {
470 if (Ops.size() < 2) return; // Noop
471 if (Ops.size() == 2) {
472 // This is the common case, which also happens to be trivially simple.
474 if (SCEVComplexityCompare()(Ops[1], Ops[0]))
475 std::swap(Ops[0], Ops[1]);
479 // Do the rough sort by complexity.
480 std::sort(Ops.begin(), Ops.end(), SCEVComplexityCompare());
482 // Now that we are sorted by complexity, group elements of the same
483 // complexity. Note that this is, at worst, N^2, but the vector is likely to
484 // be extremely short in practice. Note that we take this approach because we
485 // do not want to depend on the addresses of the objects we are grouping.
486 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
488 unsigned Complexity = S->getSCEVType();
490 // If there are any objects of the same complexity and same value as this
492 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
493 if (Ops[j] == S) { // Found a duplicate.
494 // Move it to immediately after i'th element.
495 std::swap(Ops[i+1], Ops[j]);
496 ++i; // no need to rescan it.
497 if (i == e-2) return; // Done!
505 //===----------------------------------------------------------------------===//
506 // Simple SCEV method implementations
507 //===----------------------------------------------------------------------===//
509 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
511 static SCEVHandle BinomialCoefficient(SCEVHandle It, unsigned K,
513 const Type* ResultTy) {
514 // Handle the simplest case efficiently.
516 return SE.getTruncateOrZeroExtend(It, ResultTy);
518 // We are using the following formula for BC(It, K):
520 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
522 // Suppose, W is the bitwidth of the return value. We must be prepared for
523 // overflow. Hence, we must assure that the result of our computation is
524 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
525 // safe in modular arithmetic.
527 // However, this code doesn't use exactly that formula; the formula it uses
528 // is something like the following, where T is the number of factors of 2 in
529 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
532 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
534 // This formula is trivially equivalent to the previous formula. However,
535 // this formula can be implemented much more efficiently. The trick is that
536 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
537 // arithmetic. To do exact division in modular arithmetic, all we have
538 // to do is multiply by the inverse. Therefore, this step can be done at
541 // The next issue is how to safely do the division by 2^T. The way this
542 // is done is by doing the multiplication step at a width of at least W + T
543 // bits. This way, the bottom W+T bits of the product are accurate. Then,
544 // when we perform the division by 2^T (which is equivalent to a right shift
545 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
546 // truncated out after the division by 2^T.
548 // In comparison to just directly using the first formula, this technique
549 // is much more efficient; using the first formula requires W * K bits,
550 // but this formula less than W + K bits. Also, the first formula requires
551 // a division step, whereas this formula only requires multiplies and shifts.
553 // It doesn't matter whether the subtraction step is done in the calculation
554 // width or the input iteration count's width; if the subtraction overflows,
555 // the result must be zero anyway. We prefer here to do it in the width of
556 // the induction variable because it helps a lot for certain cases; CodeGen
557 // isn't smart enough to ignore the overflow, which leads to much less
558 // efficient code if the width of the subtraction is wider than the native
561 // (It's possible to not widen at all by pulling out factors of 2 before
562 // the multiplication; for example, K=2 can be calculated as
563 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
564 // extra arithmetic, so it's not an obvious win, and it gets
565 // much more complicated for K > 3.)
567 // Protection from insane SCEVs; this bound is conservative,
568 // but it probably doesn't matter.
570 return SE.getCouldNotCompute();
572 unsigned W = SE.getTypeSizeInBits(ResultTy);
574 // Calculate K! / 2^T and T; we divide out the factors of two before
575 // multiplying for calculating K! / 2^T to avoid overflow.
576 // Other overflow doesn't matter because we only care about the bottom
577 // W bits of the result.
578 APInt OddFactorial(W, 1);
580 for (unsigned i = 3; i <= K; ++i) {
582 unsigned TwoFactors = Mult.countTrailingZeros();
584 Mult = Mult.lshr(TwoFactors);
585 OddFactorial *= Mult;
588 // We need at least W + T bits for the multiplication step
589 unsigned CalculationBits = W + T;
591 // Calcuate 2^T, at width T+W.
592 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
594 // Calculate the multiplicative inverse of K! / 2^T;
595 // this multiplication factor will perform the exact division by
597 APInt Mod = APInt::getSignedMinValue(W+1);
598 APInt MultiplyFactor = OddFactorial.zext(W+1);
599 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
600 MultiplyFactor = MultiplyFactor.trunc(W);
602 // Calculate the product, at width T+W
603 const IntegerType *CalculationTy = IntegerType::get(CalculationBits);
604 SCEVHandle Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
605 for (unsigned i = 1; i != K; ++i) {
606 SCEVHandle S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
607 Dividend = SE.getMulExpr(Dividend,
608 SE.getTruncateOrZeroExtend(S, CalculationTy));
612 SCEVHandle DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
614 // Truncate the result, and divide by K! / 2^T.
616 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
617 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
620 /// evaluateAtIteration - Return the value of this chain of recurrences at
621 /// the specified iteration number. We can evaluate this recurrence by
622 /// multiplying each element in the chain by the binomial coefficient
623 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
625 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
627 /// where BC(It, k) stands for binomial coefficient.
629 SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It,
630 ScalarEvolution &SE) const {
631 SCEVHandle Result = getStart();
632 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
633 // The computation is correct in the face of overflow provided that the
634 // multiplication is performed _after_ the evaluation of the binomial
636 SCEVHandle Coeff = BinomialCoefficient(It, i, SE, getType());
637 if (isa<SCEVCouldNotCompute>(Coeff))
640 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
645 //===----------------------------------------------------------------------===//
646 // SCEV Expression folder implementations
647 //===----------------------------------------------------------------------===//
649 SCEVHandle ScalarEvolution::getTruncateExpr(const SCEVHandle &Op,
651 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
652 "This is not a truncating conversion!");
653 assert(isSCEVable(Ty) &&
654 "This is not a conversion to a SCEVable type!");
655 Ty = getEffectiveSCEVType(Ty);
657 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
659 ConstantExpr::getTrunc(SC->getValue(), Ty));
661 // trunc(trunc(x)) --> trunc(x)
662 if (SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
663 return getTruncateExpr(ST->getOperand(), Ty);
665 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
666 if (SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
667 return getTruncateOrSignExtend(SS->getOperand(), Ty);
669 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
670 if (SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
671 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
673 // If the input value is a chrec scev made out of constants, truncate
674 // all of the constants.
675 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
676 std::vector<SCEVHandle> Operands;
677 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
678 // FIXME: This should allow truncation of other expression types!
679 if (isa<SCEVConstant>(AddRec->getOperand(i)))
680 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
683 if (Operands.size() == AddRec->getNumOperands())
684 return getAddRecExpr(Operands, AddRec->getLoop());
687 SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)];
688 if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
692 SCEVHandle ScalarEvolution::getZeroExtendExpr(const SCEVHandle &Op,
694 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
695 "This is not an extending conversion!");
696 assert(isSCEVable(Ty) &&
697 "This is not a conversion to a SCEVable type!");
698 Ty = getEffectiveSCEVType(Ty);
700 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
701 const Type *IntTy = getEffectiveSCEVType(Ty);
702 Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
703 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
704 return getUnknown(C);
707 // zext(zext(x)) --> zext(x)
708 if (SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
709 return getZeroExtendExpr(SZ->getOperand(), Ty);
711 // If the input value is a chrec scev, and we can prove that the value
712 // did not overflow the old, smaller, value, we can zero extend all of the
713 // operands (often constants). This allows analysis of something like
714 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
715 if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
716 if (AR->isAffine()) {
717 // Check whether the backedge-taken count is SCEVCouldNotCompute.
718 // Note that this serves two purposes: It filters out loops that are
719 // simply not analyzable, and it covers the case where this code is
720 // being called from within backedge-taken count analysis, such that
721 // attempting to ask for the backedge-taken count would likely result
722 // in infinite recursion. In the later case, the analysis code will
723 // cope with a conservative value, and it will take care to purge
724 // that value once it has finished.
725 SCEVHandle MaxBECount = getMaxBackedgeTakenCount(AR->getLoop());
726 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
727 // Manually compute the final value for AR, checking for
729 SCEVHandle Start = AR->getStart();
730 SCEVHandle Step = AR->getStepRecurrence(*this);
732 // Check whether the backedge-taken count can be losslessly casted to
733 // the addrec's type. The count is always unsigned.
734 SCEVHandle CastedMaxBECount =
735 getTruncateOrZeroExtend(MaxBECount, Start->getType());
737 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType())) {
739 IntegerType::get(getTypeSizeInBits(Start->getType()) * 2);
740 // Check whether Start+Step*MaxBECount has no unsigned overflow.
742 getMulExpr(CastedMaxBECount,
743 getTruncateOrZeroExtend(Step, Start->getType()));
744 SCEVHandle Add = getAddExpr(Start, ZMul);
745 if (getZeroExtendExpr(Add, WideTy) ==
746 getAddExpr(getZeroExtendExpr(Start, WideTy),
747 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
748 getZeroExtendExpr(Step, WideTy))))
749 // Return the expression with the addrec on the outside.
750 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
751 getZeroExtendExpr(Step, Ty),
754 // Similar to above, only this time treat the step value as signed.
755 // This covers loops that count down.
757 getMulExpr(CastedMaxBECount,
758 getTruncateOrSignExtend(Step, Start->getType()));
759 Add = getAddExpr(Start, SMul);
760 if (getZeroExtendExpr(Add, WideTy) ==
761 getAddExpr(getZeroExtendExpr(Start, WideTy),
762 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
763 getSignExtendExpr(Step, WideTy))))
764 // Return the expression with the addrec on the outside.
765 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
766 getSignExtendExpr(Step, Ty),
772 SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)];
773 if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
777 SCEVHandle ScalarEvolution::getSignExtendExpr(const SCEVHandle &Op,
779 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
780 "This is not an extending conversion!");
781 assert(isSCEVable(Ty) &&
782 "This is not a conversion to a SCEVable type!");
783 Ty = getEffectiveSCEVType(Ty);
785 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
786 const Type *IntTy = getEffectiveSCEVType(Ty);
787 Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
788 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
789 return getUnknown(C);
792 // sext(sext(x)) --> sext(x)
793 if (SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
794 return getSignExtendExpr(SS->getOperand(), Ty);
796 // If the input value is a chrec scev, and we can prove that the value
797 // did not overflow the old, smaller, value, we can sign extend all of the
798 // operands (often constants). This allows analysis of something like
799 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
800 if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
801 if (AR->isAffine()) {
802 // Check whether the backedge-taken count is SCEVCouldNotCompute.
803 // Note that this serves two purposes: It filters out loops that are
804 // simply not analyzable, and it covers the case where this code is
805 // being called from within backedge-taken count analysis, such that
806 // attempting to ask for the backedge-taken count would likely result
807 // in infinite recursion. In the later case, the analysis code will
808 // cope with a conservative value, and it will take care to purge
809 // that value once it has finished.
810 SCEVHandle MaxBECount = getMaxBackedgeTakenCount(AR->getLoop());
811 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
812 // Manually compute the final value for AR, checking for
814 SCEVHandle Start = AR->getStart();
815 SCEVHandle Step = AR->getStepRecurrence(*this);
817 // Check whether the backedge-taken count can be losslessly casted to
818 // the addrec's type. The count is always unsigned.
819 SCEVHandle CastedMaxBECount =
820 getTruncateOrZeroExtend(MaxBECount, Start->getType());
822 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType())) {
824 IntegerType::get(getTypeSizeInBits(Start->getType()) * 2);
825 // Check whether Start+Step*MaxBECount has no signed overflow.
827 getMulExpr(CastedMaxBECount,
828 getTruncateOrSignExtend(Step, Start->getType()));
829 SCEVHandle Add = getAddExpr(Start, SMul);
830 if (getSignExtendExpr(Add, WideTy) ==
831 getAddExpr(getSignExtendExpr(Start, WideTy),
832 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
833 getSignExtendExpr(Step, WideTy))))
834 // Return the expression with the addrec on the outside.
835 return getAddRecExpr(getSignExtendExpr(Start, Ty),
836 getSignExtendExpr(Step, Ty),
842 SCEVSignExtendExpr *&Result = (*SCEVSignExtends)[std::make_pair(Op, Ty)];
843 if (Result == 0) Result = new SCEVSignExtendExpr(Op, Ty);
847 // get - Get a canonical add expression, or something simpler if possible.
848 SCEVHandle ScalarEvolution::getAddExpr(std::vector<SCEVHandle> &Ops) {
849 assert(!Ops.empty() && "Cannot get empty add!");
850 if (Ops.size() == 1) return Ops[0];
852 // Sort by complexity, this groups all similar expression types together.
853 GroupByComplexity(Ops);
855 // If there are any constants, fold them together.
857 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
859 assert(Idx < Ops.size());
860 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
861 // We found two constants, fold them together!
862 ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() +
863 RHSC->getValue()->getValue());
864 Ops[0] = getConstant(Fold);
865 Ops.erase(Ops.begin()+1); // Erase the folded element
866 if (Ops.size() == 1) return Ops[0];
867 LHSC = cast<SCEVConstant>(Ops[0]);
870 // If we are left with a constant zero being added, strip it off.
871 if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
872 Ops.erase(Ops.begin());
877 if (Ops.size() == 1) return Ops[0];
879 // Okay, check to see if the same value occurs in the operand list twice. If
880 // so, merge them together into an multiply expression. Since we sorted the
881 // list, these values are required to be adjacent.
882 const Type *Ty = Ops[0]->getType();
883 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
884 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
885 // Found a match, merge the two values into a multiply, and add any
886 // remaining values to the result.
887 SCEVHandle Two = getIntegerSCEV(2, Ty);
888 SCEVHandle Mul = getMulExpr(Ops[i], Two);
891 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
893 return getAddExpr(Ops);
896 // Now we know the first non-constant operand. Skip past any cast SCEVs.
897 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
900 // If there are add operands they would be next.
901 if (Idx < Ops.size()) {
902 bool DeletedAdd = false;
903 while (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
904 // If we have an add, expand the add operands onto the end of the operands
906 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
907 Ops.erase(Ops.begin()+Idx);
911 // If we deleted at least one add, we added operands to the end of the list,
912 // and they are not necessarily sorted. Recurse to resort and resimplify
913 // any operands we just aquired.
915 return getAddExpr(Ops);
918 // Skip over the add expression until we get to a multiply.
919 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
922 // If we are adding something to a multiply expression, make sure the
923 // something is not already an operand of the multiply. If so, merge it into
925 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
926 SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
927 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
928 SCEV *MulOpSCEV = Mul->getOperand(MulOp);
929 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
930 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(MulOpSCEV)) {
931 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
932 SCEVHandle InnerMul = Mul->getOperand(MulOp == 0);
933 if (Mul->getNumOperands() != 2) {
934 // If the multiply has more than two operands, we must get the
936 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
937 MulOps.erase(MulOps.begin()+MulOp);
938 InnerMul = getMulExpr(MulOps);
940 SCEVHandle One = getIntegerSCEV(1, Ty);
941 SCEVHandle AddOne = getAddExpr(InnerMul, One);
942 SCEVHandle OuterMul = getMulExpr(AddOne, Ops[AddOp]);
943 if (Ops.size() == 2) return OuterMul;
945 Ops.erase(Ops.begin()+AddOp);
946 Ops.erase(Ops.begin()+Idx-1);
948 Ops.erase(Ops.begin()+Idx);
949 Ops.erase(Ops.begin()+AddOp-1);
951 Ops.push_back(OuterMul);
952 return getAddExpr(Ops);
955 // Check this multiply against other multiplies being added together.
956 for (unsigned OtherMulIdx = Idx+1;
957 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
959 SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
960 // If MulOp occurs in OtherMul, we can fold the two multiplies
962 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
963 OMulOp != e; ++OMulOp)
964 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
965 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
966 SCEVHandle InnerMul1 = Mul->getOperand(MulOp == 0);
967 if (Mul->getNumOperands() != 2) {
968 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
969 MulOps.erase(MulOps.begin()+MulOp);
970 InnerMul1 = getMulExpr(MulOps);
972 SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0);
973 if (OtherMul->getNumOperands() != 2) {
974 std::vector<SCEVHandle> MulOps(OtherMul->op_begin(),
976 MulOps.erase(MulOps.begin()+OMulOp);
977 InnerMul2 = getMulExpr(MulOps);
979 SCEVHandle InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
980 SCEVHandle OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
981 if (Ops.size() == 2) return OuterMul;
982 Ops.erase(Ops.begin()+Idx);
983 Ops.erase(Ops.begin()+OtherMulIdx-1);
984 Ops.push_back(OuterMul);
985 return getAddExpr(Ops);
991 // If there are any add recurrences in the operands list, see if any other
992 // added values are loop invariant. If so, we can fold them into the
994 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
997 // Scan over all recurrences, trying to fold loop invariants into them.
998 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
999 // Scan all of the other operands to this add and add them to the vector if
1000 // they are loop invariant w.r.t. the recurrence.
1001 std::vector<SCEVHandle> LIOps;
1002 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1003 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1004 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1005 LIOps.push_back(Ops[i]);
1006 Ops.erase(Ops.begin()+i);
1010 // If we found some loop invariants, fold them into the recurrence.
1011 if (!LIOps.empty()) {
1012 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1013 LIOps.push_back(AddRec->getStart());
1015 std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
1016 AddRecOps[0] = getAddExpr(LIOps);
1018 SCEVHandle NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
1019 // If all of the other operands were loop invariant, we are done.
1020 if (Ops.size() == 1) return NewRec;
1022 // Otherwise, add the folded AddRec by the non-liv parts.
1023 for (unsigned i = 0;; ++i)
1024 if (Ops[i] == AddRec) {
1028 return getAddExpr(Ops);
1031 // Okay, if there weren't any loop invariants to be folded, check to see if
1032 // there are multiple AddRec's with the same loop induction variable being
1033 // added together. If so, we can fold them.
1034 for (unsigned OtherIdx = Idx+1;
1035 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1036 if (OtherIdx != Idx) {
1037 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1038 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1039 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
1040 std::vector<SCEVHandle> NewOps(AddRec->op_begin(), AddRec->op_end());
1041 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
1042 if (i >= NewOps.size()) {
1043 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
1044 OtherAddRec->op_end());
1047 NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
1049 SCEVHandle NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
1051 if (Ops.size() == 2) return NewAddRec;
1053 Ops.erase(Ops.begin()+Idx);
1054 Ops.erase(Ops.begin()+OtherIdx-1);
1055 Ops.push_back(NewAddRec);
1056 return getAddExpr(Ops);
1060 // Otherwise couldn't fold anything into this recurrence. Move onto the
1064 // Okay, it looks like we really DO need an add expr. Check to see if we
1065 // already have one, otherwise create a new one.
1066 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1067 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scAddExpr,
1069 if (Result == 0) Result = new SCEVAddExpr(Ops);
1074 SCEVHandle ScalarEvolution::getMulExpr(std::vector<SCEVHandle> &Ops) {
1075 assert(!Ops.empty() && "Cannot get empty mul!");
1077 // Sort by complexity, this groups all similar expression types together.
1078 GroupByComplexity(Ops);
1080 // If there are any constants, fold them together.
1082 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1084 // C1*(C2+V) -> C1*C2 + C1*V
1085 if (Ops.size() == 2)
1086 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1087 if (Add->getNumOperands() == 2 &&
1088 isa<SCEVConstant>(Add->getOperand(0)))
1089 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1090 getMulExpr(LHSC, Add->getOperand(1)));
1094 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1095 // We found two constants, fold them together!
1096 ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() *
1097 RHSC->getValue()->getValue());
1098 Ops[0] = getConstant(Fold);
1099 Ops.erase(Ops.begin()+1); // Erase the folded element
1100 if (Ops.size() == 1) return Ops[0];
1101 LHSC = cast<SCEVConstant>(Ops[0]);
1104 // If we are left with a constant one being multiplied, strip it off.
1105 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1106 Ops.erase(Ops.begin());
1108 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1109 // If we have a multiply of zero, it will always be zero.
1114 // Skip over the add expression until we get to a multiply.
1115 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1118 if (Ops.size() == 1)
1121 // If there are mul operands inline them all into this expression.
1122 if (Idx < Ops.size()) {
1123 bool DeletedMul = false;
1124 while (SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1125 // If we have an mul, expand the mul operands onto the end of the operands
1127 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1128 Ops.erase(Ops.begin()+Idx);
1132 // If we deleted at least one mul, we added operands to the end of the list,
1133 // and they are not necessarily sorted. Recurse to resort and resimplify
1134 // any operands we just aquired.
1136 return getMulExpr(Ops);
1139 // If there are any add recurrences in the operands list, see if any other
1140 // added values are loop invariant. If so, we can fold them into the
1142 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1145 // Scan over all recurrences, trying to fold loop invariants into them.
1146 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1147 // Scan all of the other operands to this mul and add them to the vector if
1148 // they are loop invariant w.r.t. the recurrence.
1149 std::vector<SCEVHandle> LIOps;
1150 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1151 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1152 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1153 LIOps.push_back(Ops[i]);
1154 Ops.erase(Ops.begin()+i);
1158 // If we found some loop invariants, fold them into the recurrence.
1159 if (!LIOps.empty()) {
1160 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1161 std::vector<SCEVHandle> NewOps;
1162 NewOps.reserve(AddRec->getNumOperands());
1163 if (LIOps.size() == 1) {
1164 SCEV *Scale = LIOps[0];
1165 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1166 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1168 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1169 std::vector<SCEVHandle> MulOps(LIOps);
1170 MulOps.push_back(AddRec->getOperand(i));
1171 NewOps.push_back(getMulExpr(MulOps));
1175 SCEVHandle NewRec = getAddRecExpr(NewOps, AddRec->getLoop());
1177 // If all of the other operands were loop invariant, we are done.
1178 if (Ops.size() == 1) return NewRec;
1180 // Otherwise, multiply the folded AddRec by the non-liv parts.
1181 for (unsigned i = 0;; ++i)
1182 if (Ops[i] == AddRec) {
1186 return getMulExpr(Ops);
1189 // Okay, if there weren't any loop invariants to be folded, check to see if
1190 // there are multiple AddRec's with the same loop induction variable being
1191 // multiplied together. If so, we can fold them.
1192 for (unsigned OtherIdx = Idx+1;
1193 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1194 if (OtherIdx != Idx) {
1195 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1196 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1197 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1198 SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1199 SCEVHandle NewStart = getMulExpr(F->getStart(),
1201 SCEVHandle B = F->getStepRecurrence(*this);
1202 SCEVHandle D = G->getStepRecurrence(*this);
1203 SCEVHandle NewStep = getAddExpr(getMulExpr(F, D),
1206 SCEVHandle NewAddRec = getAddRecExpr(NewStart, NewStep,
1208 if (Ops.size() == 2) return NewAddRec;
1210 Ops.erase(Ops.begin()+Idx);
1211 Ops.erase(Ops.begin()+OtherIdx-1);
1212 Ops.push_back(NewAddRec);
1213 return getMulExpr(Ops);
1217 // Otherwise couldn't fold anything into this recurrence. Move onto the
1221 // Okay, it looks like we really DO need an mul expr. Check to see if we
1222 // already have one, otherwise create a new one.
1223 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1224 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scMulExpr,
1227 Result = new SCEVMulExpr(Ops);
1231 SCEVHandle ScalarEvolution::getUDivExpr(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1232 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1233 if (RHSC->getValue()->equalsInt(1))
1234 return LHS; // X udiv 1 --> x
1236 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1237 Constant *LHSCV = LHSC->getValue();
1238 Constant *RHSCV = RHSC->getValue();
1239 return getUnknown(ConstantExpr::getUDiv(LHSCV, RHSCV));
1243 // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow.
1245 SCEVUDivExpr *&Result = (*SCEVUDivs)[std::make_pair(LHS, RHS)];
1246 if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
1251 /// SCEVAddRecExpr::get - Get a add recurrence expression for the
1252 /// specified loop. Simplify the expression as much as possible.
1253 SCEVHandle ScalarEvolution::getAddRecExpr(const SCEVHandle &Start,
1254 const SCEVHandle &Step, const Loop *L) {
1255 std::vector<SCEVHandle> Operands;
1256 Operands.push_back(Start);
1257 if (SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1258 if (StepChrec->getLoop() == L) {
1259 Operands.insert(Operands.end(), StepChrec->op_begin(),
1260 StepChrec->op_end());
1261 return getAddRecExpr(Operands, L);
1264 Operands.push_back(Step);
1265 return getAddRecExpr(Operands, L);
1268 /// SCEVAddRecExpr::get - Get a add recurrence expression for the
1269 /// specified loop. Simplify the expression as much as possible.
1270 SCEVHandle ScalarEvolution::getAddRecExpr(std::vector<SCEVHandle> &Operands,
1272 if (Operands.size() == 1) return Operands[0];
1274 if (Operands.back()->isZero()) {
1275 Operands.pop_back();
1276 return getAddRecExpr(Operands, L); // {X,+,0} --> X
1279 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
1280 if (SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
1281 const Loop* NestedLoop = NestedAR->getLoop();
1282 if (L->getLoopDepth() < NestedLoop->getLoopDepth()) {
1283 std::vector<SCEVHandle> NestedOperands(NestedAR->op_begin(),
1284 NestedAR->op_end());
1285 SCEVHandle NestedARHandle(NestedAR);
1286 Operands[0] = NestedAR->getStart();
1287 NestedOperands[0] = getAddRecExpr(Operands, L);
1288 return getAddRecExpr(NestedOperands, NestedLoop);
1292 SCEVAddRecExpr *&Result =
1293 (*SCEVAddRecExprs)[std::make_pair(L, std::vector<SCEV*>(Operands.begin(),
1295 if (Result == 0) Result = new SCEVAddRecExpr(Operands, L);
1299 SCEVHandle ScalarEvolution::getSMaxExpr(const SCEVHandle &LHS,
1300 const SCEVHandle &RHS) {
1301 std::vector<SCEVHandle> Ops;
1304 return getSMaxExpr(Ops);
1307 SCEVHandle ScalarEvolution::getSMaxExpr(std::vector<SCEVHandle> Ops) {
1308 assert(!Ops.empty() && "Cannot get empty smax!");
1309 if (Ops.size() == 1) return Ops[0];
1311 // Sort by complexity, this groups all similar expression types together.
1312 GroupByComplexity(Ops);
1314 // If there are any constants, fold them together.
1316 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1318 assert(Idx < Ops.size());
1319 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1320 // We found two constants, fold them together!
1321 ConstantInt *Fold = ConstantInt::get(
1322 APIntOps::smax(LHSC->getValue()->getValue(),
1323 RHSC->getValue()->getValue()));
1324 Ops[0] = getConstant(Fold);
1325 Ops.erase(Ops.begin()+1); // Erase the folded element
1326 if (Ops.size() == 1) return Ops[0];
1327 LHSC = cast<SCEVConstant>(Ops[0]);
1330 // If we are left with a constant -inf, strip it off.
1331 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
1332 Ops.erase(Ops.begin());
1337 if (Ops.size() == 1) return Ops[0];
1339 // Find the first SMax
1340 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
1343 // Check to see if one of the operands is an SMax. If so, expand its operands
1344 // onto our operand list, and recurse to simplify.
1345 if (Idx < Ops.size()) {
1346 bool DeletedSMax = false;
1347 while (SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
1348 Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
1349 Ops.erase(Ops.begin()+Idx);
1354 return getSMaxExpr(Ops);
1357 // Okay, check to see if the same value occurs in the operand list twice. If
1358 // so, delete one. Since we sorted the list, these values are required to
1360 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1361 if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
1362 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1366 if (Ops.size() == 1) return Ops[0];
1368 assert(!Ops.empty() && "Reduced smax down to nothing!");
1370 // Okay, it looks like we really DO need an smax expr. Check to see if we
1371 // already have one, otherwise create a new one.
1372 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1373 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scSMaxExpr,
1375 if (Result == 0) Result = new SCEVSMaxExpr(Ops);
1379 SCEVHandle ScalarEvolution::getUMaxExpr(const SCEVHandle &LHS,
1380 const SCEVHandle &RHS) {
1381 std::vector<SCEVHandle> Ops;
1384 return getUMaxExpr(Ops);
1387 SCEVHandle ScalarEvolution::getUMaxExpr(std::vector<SCEVHandle> Ops) {
1388 assert(!Ops.empty() && "Cannot get empty umax!");
1389 if (Ops.size() == 1) return Ops[0];
1391 // Sort by complexity, this groups all similar expression types together.
1392 GroupByComplexity(Ops);
1394 // If there are any constants, fold them together.
1396 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1398 assert(Idx < Ops.size());
1399 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1400 // We found two constants, fold them together!
1401 ConstantInt *Fold = ConstantInt::get(
1402 APIntOps::umax(LHSC->getValue()->getValue(),
1403 RHSC->getValue()->getValue()));
1404 Ops[0] = getConstant(Fold);
1405 Ops.erase(Ops.begin()+1); // Erase the folded element
1406 if (Ops.size() == 1) return Ops[0];
1407 LHSC = cast<SCEVConstant>(Ops[0]);
1410 // If we are left with a constant zero, strip it off.
1411 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
1412 Ops.erase(Ops.begin());
1417 if (Ops.size() == 1) return Ops[0];
1419 // Find the first UMax
1420 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
1423 // Check to see if one of the operands is a UMax. If so, expand its operands
1424 // onto our operand list, and recurse to simplify.
1425 if (Idx < Ops.size()) {
1426 bool DeletedUMax = false;
1427 while (SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
1428 Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
1429 Ops.erase(Ops.begin()+Idx);
1434 return getUMaxExpr(Ops);
1437 // Okay, check to see if the same value occurs in the operand list twice. If
1438 // so, delete one. Since we sorted the list, these values are required to
1440 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1441 if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
1442 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1446 if (Ops.size() == 1) return Ops[0];
1448 assert(!Ops.empty() && "Reduced umax down to nothing!");
1450 // Okay, it looks like we really DO need a umax expr. Check to see if we
1451 // already have one, otherwise create a new one.
1452 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1453 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scUMaxExpr,
1455 if (Result == 0) Result = new SCEVUMaxExpr(Ops);
1459 SCEVHandle ScalarEvolution::getUnknown(Value *V) {
1460 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
1461 return getConstant(CI);
1462 if (isa<ConstantPointerNull>(V))
1463 return getIntegerSCEV(0, V->getType());
1464 SCEVUnknown *&Result = (*SCEVUnknowns)[V];
1465 if (Result == 0) Result = new SCEVUnknown(V);
1469 //===----------------------------------------------------------------------===//
1470 // Basic SCEV Analysis and PHI Idiom Recognition Code
1473 /// deleteValueFromRecords - This method should be called by the
1474 /// client before it removes an instruction from the program, to make sure
1475 /// that no dangling references are left around.
1476 void ScalarEvolution::deleteValueFromRecords(Value *V) {
1477 SmallVector<Value *, 16> Worklist;
1479 if (Scalars.erase(V)) {
1480 if (PHINode *PN = dyn_cast<PHINode>(V))
1481 ConstantEvolutionLoopExitValue.erase(PN);
1482 Worklist.push_back(V);
1485 while (!Worklist.empty()) {
1486 Value *VV = Worklist.back();
1487 Worklist.pop_back();
1489 for (Instruction::use_iterator UI = VV->use_begin(), UE = VV->use_end();
1491 Instruction *Inst = cast<Instruction>(*UI);
1492 if (Scalars.erase(Inst)) {
1493 if (PHINode *PN = dyn_cast<PHINode>(VV))
1494 ConstantEvolutionLoopExitValue.erase(PN);
1495 Worklist.push_back(Inst);
1501 /// isSCEVable - Test if values of the given type are analyzable within
1502 /// the SCEV framework. This primarily includes integer types, and it
1503 /// can optionally include pointer types if the ScalarEvolution class
1504 /// has access to target-specific information.
1505 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
1506 // Integers are always SCEVable.
1507 if (Ty->isInteger())
1510 // Pointers are SCEVable if TargetData information is available
1511 // to provide pointer size information.
1512 if (isa<PointerType>(Ty))
1515 // Otherwise it's not SCEVable.
1519 /// getTypeSizeInBits - Return the size in bits of the specified type,
1520 /// for which isSCEVable must return true.
1521 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
1522 assert(isSCEVable(Ty) && "Type is not SCEVable!");
1524 // If we have a TargetData, use it!
1526 return TD->getTypeSizeInBits(Ty);
1528 // Otherwise, we support only integer types.
1529 assert(Ty->isInteger() && "isSCEVable permitted a non-SCEVable type!");
1530 return Ty->getPrimitiveSizeInBits();
1533 /// getEffectiveSCEVType - Return a type with the same bitwidth as
1534 /// the given type and which represents how SCEV will treat the given
1535 /// type, for which isSCEVable must return true. For pointer types,
1536 /// this is the pointer-sized integer type.
1537 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
1538 assert(isSCEVable(Ty) && "Type is not SCEVable!");
1540 if (Ty->isInteger())
1543 assert(isa<PointerType>(Ty) && "Unexpected non-pointer non-integer type!");
1544 return TD->getIntPtrType();
1547 SCEVHandle ScalarEvolution::getCouldNotCompute() {
1548 return UnknownValue;
1551 // hasSCEV - Return true if the SCEV for this value has already been
1553 bool ScalarEvolution::hasSCEV(Value *V) const {
1554 return Scalars.count(V);
1557 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1558 /// expression and create a new one.
1559 SCEVHandle ScalarEvolution::getSCEV(Value *V) {
1560 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
1562 std::map<Value*, SCEVHandle>::iterator I = Scalars.find(V);
1563 if (I != Scalars.end()) return I->second;
1564 SCEVHandle S = createSCEV(V);
1565 Scalars.insert(std::make_pair(V, S));
1569 /// getIntegerSCEV - Given an integer or FP type, create a constant for the
1570 /// specified signed integer value and return a SCEV for the constant.
1571 SCEVHandle ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
1572 Ty = getEffectiveSCEVType(Ty);
1575 C = Constant::getNullValue(Ty);
1576 else if (Ty->isFloatingPoint())
1577 C = ConstantFP::get(APFloat(Ty==Type::FloatTy ? APFloat::IEEEsingle :
1578 APFloat::IEEEdouble, Val));
1580 C = ConstantInt::get(Ty, Val);
1581 return getUnknown(C);
1584 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
1586 SCEVHandle ScalarEvolution::getNegativeSCEV(const SCEVHandle &V) {
1587 if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
1588 return getUnknown(ConstantExpr::getNeg(VC->getValue()));
1590 const Type *Ty = V->getType();
1591 Ty = getEffectiveSCEVType(Ty);
1592 return getMulExpr(V, getConstant(ConstantInt::getAllOnesValue(Ty)));
1595 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
1596 SCEVHandle ScalarEvolution::getNotSCEV(const SCEVHandle &V) {
1597 if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
1598 return getUnknown(ConstantExpr::getNot(VC->getValue()));
1600 const Type *Ty = V->getType();
1601 Ty = getEffectiveSCEVType(Ty);
1602 SCEVHandle AllOnes = getConstant(ConstantInt::getAllOnesValue(Ty));
1603 return getMinusSCEV(AllOnes, V);
1606 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
1608 SCEVHandle ScalarEvolution::getMinusSCEV(const SCEVHandle &LHS,
1609 const SCEVHandle &RHS) {
1611 return getAddExpr(LHS, getNegativeSCEV(RHS));
1614 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
1615 /// input value to the specified type. If the type must be extended, it is zero
1618 ScalarEvolution::getTruncateOrZeroExtend(const SCEVHandle &V,
1620 const Type *SrcTy = V->getType();
1621 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
1622 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
1623 "Cannot truncate or zero extend with non-integer arguments!");
1624 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
1625 return V; // No conversion
1626 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
1627 return getTruncateExpr(V, Ty);
1628 return getZeroExtendExpr(V, Ty);
1631 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
1632 /// input value to the specified type. If the type must be extended, it is sign
1635 ScalarEvolution::getTruncateOrSignExtend(const SCEVHandle &V,
1637 const Type *SrcTy = V->getType();
1638 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
1639 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
1640 "Cannot truncate or zero extend with non-integer arguments!");
1641 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
1642 return V; // No conversion
1643 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
1644 return getTruncateExpr(V, Ty);
1645 return getSignExtendExpr(V, Ty);
1648 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value for
1649 /// the specified instruction and replaces any references to the symbolic value
1650 /// SymName with the specified value. This is used during PHI resolution.
1651 void ScalarEvolution::
1652 ReplaceSymbolicValueWithConcrete(Instruction *I, const SCEVHandle &SymName,
1653 const SCEVHandle &NewVal) {
1654 std::map<Value*, SCEVHandle>::iterator SI = Scalars.find(I);
1655 if (SI == Scalars.end()) return;
1658 SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal, *this);
1659 if (NV == SI->second) return; // No change.
1661 SI->second = NV; // Update the scalars map!
1663 // Any instruction values that use this instruction might also need to be
1665 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1667 ReplaceSymbolicValueWithConcrete(cast<Instruction>(*UI), SymName, NewVal);
1670 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
1671 /// a loop header, making it a potential recurrence, or it doesn't.
1673 SCEVHandle ScalarEvolution::createNodeForPHI(PHINode *PN) {
1674 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
1675 if (const Loop *L = LI->getLoopFor(PN->getParent()))
1676 if (L->getHeader() == PN->getParent()) {
1677 // If it lives in the loop header, it has two incoming values, one
1678 // from outside the loop, and one from inside.
1679 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1680 unsigned BackEdge = IncomingEdge^1;
1682 // While we are analyzing this PHI node, handle its value symbolically.
1683 SCEVHandle SymbolicName = getUnknown(PN);
1684 assert(Scalars.find(PN) == Scalars.end() &&
1685 "PHI node already processed?");
1686 Scalars.insert(std::make_pair(PN, SymbolicName));
1688 // Using this symbolic name for the PHI, analyze the value coming around
1690 SCEVHandle BEValue = getSCEV(PN->getIncomingValue(BackEdge));
1692 // NOTE: If BEValue is loop invariant, we know that the PHI node just
1693 // has a special value for the first iteration of the loop.
1695 // If the value coming around the backedge is an add with the symbolic
1696 // value we just inserted, then we found a simple induction variable!
1697 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
1698 // If there is a single occurrence of the symbolic value, replace it
1699 // with a recurrence.
1700 unsigned FoundIndex = Add->getNumOperands();
1701 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1702 if (Add->getOperand(i) == SymbolicName)
1703 if (FoundIndex == e) {
1708 if (FoundIndex != Add->getNumOperands()) {
1709 // Create an add with everything but the specified operand.
1710 std::vector<SCEVHandle> Ops;
1711 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1712 if (i != FoundIndex)
1713 Ops.push_back(Add->getOperand(i));
1714 SCEVHandle Accum = getAddExpr(Ops);
1716 // This is not a valid addrec if the step amount is varying each
1717 // loop iteration, but is not itself an addrec in this loop.
1718 if (Accum->isLoopInvariant(L) ||
1719 (isa<SCEVAddRecExpr>(Accum) &&
1720 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
1721 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1722 SCEVHandle PHISCEV = getAddRecExpr(StartVal, Accum, L);
1724 // Okay, for the entire analysis of this edge we assumed the PHI
1725 // to be symbolic. We now need to go back and update all of the
1726 // entries for the scalars that use the PHI (except for the PHI
1727 // itself) to use the new analyzed value instead of the "symbolic"
1729 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1733 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(BEValue)) {
1734 // Otherwise, this could be a loop like this:
1735 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
1736 // In this case, j = {1,+,1} and BEValue is j.
1737 // Because the other in-value of i (0) fits the evolution of BEValue
1738 // i really is an addrec evolution.
1739 if (AddRec->getLoop() == L && AddRec->isAffine()) {
1740 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1742 // If StartVal = j.start - j.stride, we can use StartVal as the
1743 // initial step of the addrec evolution.
1744 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
1745 AddRec->getOperand(1))) {
1746 SCEVHandle PHISCEV =
1747 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
1749 // Okay, for the entire analysis of this edge we assumed the PHI
1750 // to be symbolic. We now need to go back and update all of the
1751 // entries for the scalars that use the PHI (except for the PHI
1752 // itself) to use the new analyzed value instead of the "symbolic"
1754 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1760 return SymbolicName;
1763 // If it's not a loop phi, we can't handle it yet.
1764 return getUnknown(PN);
1767 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
1768 /// guaranteed to end in (at every loop iteration). It is, at the same time,
1769 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
1770 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
1771 static uint32_t GetMinTrailingZeros(SCEVHandle S, const ScalarEvolution &SE) {
1772 if (SCEVConstant *C = dyn_cast<SCEVConstant>(S))
1773 return C->getValue()->getValue().countTrailingZeros();
1775 if (SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
1776 return std::min(GetMinTrailingZeros(T->getOperand(), SE),
1777 (uint32_t)SE.getTypeSizeInBits(T->getType()));
1779 if (SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
1780 uint32_t OpRes = GetMinTrailingZeros(E->getOperand(), SE);
1781 return OpRes == SE.getTypeSizeInBits(E->getOperand()->getType()) ?
1782 SE.getTypeSizeInBits(E->getOperand()->getType()) : OpRes;
1785 if (SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
1786 uint32_t OpRes = GetMinTrailingZeros(E->getOperand(), SE);
1787 return OpRes == SE.getTypeSizeInBits(E->getOperand()->getType()) ?
1788 SE.getTypeSizeInBits(E->getOperand()->getType()) : OpRes;
1791 if (SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
1792 // The result is the min of all operands results.
1793 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0), SE);
1794 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
1795 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i), SE));
1799 if (SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
1800 // The result is the sum of all operands results.
1801 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0), SE);
1802 uint32_t BitWidth = SE.getTypeSizeInBits(M->getType());
1803 for (unsigned i = 1, e = M->getNumOperands();
1804 SumOpRes != BitWidth && i != e; ++i)
1805 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i), SE),
1810 if (SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
1811 // The result is the min of all operands results.
1812 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0), SE);
1813 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
1814 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i), SE));
1818 if (SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
1819 // The result is the min of all operands results.
1820 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0), SE);
1821 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
1822 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i), SE));
1826 if (SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
1827 // The result is the min of all operands results.
1828 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0), SE);
1829 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
1830 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i), SE));
1834 // SCEVUDivExpr, SCEVUnknown
1838 /// createSCEV - We know that there is no SCEV for the specified value.
1839 /// Analyze the expression.
1841 SCEVHandle ScalarEvolution::createSCEV(Value *V) {
1842 if (!isSCEVable(V->getType()))
1843 return getUnknown(V);
1845 unsigned Opcode = Instruction::UserOp1;
1846 if (Instruction *I = dyn_cast<Instruction>(V))
1847 Opcode = I->getOpcode();
1848 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1849 Opcode = CE->getOpcode();
1851 return getUnknown(V);
1853 User *U = cast<User>(V);
1855 case Instruction::Add:
1856 return getAddExpr(getSCEV(U->getOperand(0)),
1857 getSCEV(U->getOperand(1)));
1858 case Instruction::Mul:
1859 return getMulExpr(getSCEV(U->getOperand(0)),
1860 getSCEV(U->getOperand(1)));
1861 case Instruction::UDiv:
1862 return getUDivExpr(getSCEV(U->getOperand(0)),
1863 getSCEV(U->getOperand(1)));
1864 case Instruction::Sub:
1865 return getMinusSCEV(getSCEV(U->getOperand(0)),
1866 getSCEV(U->getOperand(1)));
1867 case Instruction::And:
1868 // For an expression like x&255 that merely masks off the high bits,
1869 // use zext(trunc(x)) as the SCEV expression.
1870 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
1871 if (CI->isNullValue())
1872 return getSCEV(U->getOperand(1));
1873 if (CI->isAllOnesValue())
1874 return getSCEV(U->getOperand(0));
1875 const APInt &A = CI->getValue();
1876 unsigned Ones = A.countTrailingOnes();
1877 if (APIntOps::isMask(Ones, A))
1879 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
1880 IntegerType::get(Ones)),
1884 case Instruction::Or:
1885 // If the RHS of the Or is a constant, we may have something like:
1886 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
1887 // optimizations will transparently handle this case.
1889 // In order for this transformation to be safe, the LHS must be of the
1890 // form X*(2^n) and the Or constant must be less than 2^n.
1891 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
1892 SCEVHandle LHS = getSCEV(U->getOperand(0));
1893 const APInt &CIVal = CI->getValue();
1894 if (GetMinTrailingZeros(LHS, *this) >=
1895 (CIVal.getBitWidth() - CIVal.countLeadingZeros()))
1896 return getAddExpr(LHS, getSCEV(U->getOperand(1)));
1899 case Instruction::Xor:
1900 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
1901 // If the RHS of the xor is a signbit, then this is just an add.
1902 // Instcombine turns add of signbit into xor as a strength reduction step.
1903 if (CI->getValue().isSignBit())
1904 return getAddExpr(getSCEV(U->getOperand(0)),
1905 getSCEV(U->getOperand(1)));
1907 // If the RHS of xor is -1, then this is a not operation.
1908 else if (CI->isAllOnesValue())
1909 return getNotSCEV(getSCEV(U->getOperand(0)));
1913 case Instruction::Shl:
1914 // Turn shift left of a constant amount into a multiply.
1915 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
1916 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
1917 Constant *X = ConstantInt::get(
1918 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
1919 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
1923 case Instruction::LShr:
1924 // Turn logical shift right of a constant into a unsigned divide.
1925 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
1926 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
1927 Constant *X = ConstantInt::get(
1928 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
1929 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
1933 case Instruction::AShr:
1934 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
1935 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
1936 if (Instruction *L = dyn_cast<Instruction>(U->getOperand(0)))
1937 if (L->getOpcode() == Instruction::Shl &&
1938 L->getOperand(1) == U->getOperand(1)) {
1939 unsigned BitWidth = getTypeSizeInBits(U->getType());
1940 uint64_t Amt = BitWidth - CI->getZExtValue();
1941 if (Amt == BitWidth)
1942 return getSCEV(L->getOperand(0)); // shift by zero --> noop
1944 return getIntegerSCEV(0, U->getType()); // value is undefined
1946 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
1947 IntegerType::get(Amt)),
1952 case Instruction::Trunc:
1953 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
1955 case Instruction::ZExt:
1956 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
1958 case Instruction::SExt:
1959 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
1961 case Instruction::BitCast:
1962 // BitCasts are no-op casts so we just eliminate the cast.
1963 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
1964 return getSCEV(U->getOperand(0));
1967 case Instruction::IntToPtr:
1968 if (!TD) break; // Without TD we can't analyze pointers.
1969 return getTruncateOrZeroExtend(getSCEV(U->getOperand(0)),
1970 TD->getIntPtrType());
1972 case Instruction::PtrToInt:
1973 if (!TD) break; // Without TD we can't analyze pointers.
1974 return getTruncateOrZeroExtend(getSCEV(U->getOperand(0)),
1977 case Instruction::GetElementPtr: {
1978 if (!TD) break; // Without TD we can't analyze pointers.
1979 const Type *IntPtrTy = TD->getIntPtrType();
1980 Value *Base = U->getOperand(0);
1981 SCEVHandle TotalOffset = getIntegerSCEV(0, IntPtrTy);
1982 gep_type_iterator GTI = gep_type_begin(U);
1983 for (GetElementPtrInst::op_iterator I = next(U->op_begin()),
1987 // Compute the (potentially symbolic) offset in bytes for this index.
1988 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
1989 // For a struct, add the member offset.
1990 const StructLayout &SL = *TD->getStructLayout(STy);
1991 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
1992 uint64_t Offset = SL.getElementOffset(FieldNo);
1993 TotalOffset = getAddExpr(TotalOffset,
1994 getIntegerSCEV(Offset, IntPtrTy));
1996 // For an array, add the element offset, explicitly scaled.
1997 SCEVHandle LocalOffset = getSCEV(Index);
1998 if (!isa<PointerType>(LocalOffset->getType()))
1999 // Getelementptr indicies are signed.
2000 LocalOffset = getTruncateOrSignExtend(LocalOffset,
2003 getMulExpr(LocalOffset,
2004 getIntegerSCEV(TD->getTypePaddedSize(*GTI),
2006 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2009 return getAddExpr(getSCEV(Base), TotalOffset);
2012 case Instruction::PHI:
2013 return createNodeForPHI(cast<PHINode>(U));
2015 case Instruction::Select:
2016 // This could be a smax or umax that was lowered earlier.
2017 // Try to recover it.
2018 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
2019 Value *LHS = ICI->getOperand(0);
2020 Value *RHS = ICI->getOperand(1);
2021 switch (ICI->getPredicate()) {
2022 case ICmpInst::ICMP_SLT:
2023 case ICmpInst::ICMP_SLE:
2024 std::swap(LHS, RHS);
2026 case ICmpInst::ICMP_SGT:
2027 case ICmpInst::ICMP_SGE:
2028 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
2029 return getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
2030 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
2031 // ~smax(~x, ~y) == smin(x, y).
2032 return getNotSCEV(getSMaxExpr(
2033 getNotSCEV(getSCEV(LHS)),
2034 getNotSCEV(getSCEV(RHS))));
2036 case ICmpInst::ICMP_ULT:
2037 case ICmpInst::ICMP_ULE:
2038 std::swap(LHS, RHS);
2040 case ICmpInst::ICMP_UGT:
2041 case ICmpInst::ICMP_UGE:
2042 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
2043 return getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
2044 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
2045 // ~umax(~x, ~y) == umin(x, y)
2046 return getNotSCEV(getUMaxExpr(getNotSCEV(getSCEV(LHS)),
2047 getNotSCEV(getSCEV(RHS))));
2054 default: // We cannot analyze this expression.
2058 return getUnknown(V);
2063 //===----------------------------------------------------------------------===//
2064 // Iteration Count Computation Code
2067 /// getBackedgeTakenCount - If the specified loop has a predictable
2068 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
2069 /// object. The backedge-taken count is the number of times the loop header
2070 /// will be branched to from within the loop. This is one less than the
2071 /// trip count of the loop, since it doesn't count the first iteration,
2072 /// when the header is branched to from outside the loop.
2074 /// Note that it is not valid to call this method on a loop without a
2075 /// loop-invariant backedge-taken count (see
2076 /// hasLoopInvariantBackedgeTakenCount).
2078 SCEVHandle ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
2079 return getBackedgeTakenInfo(L).Exact;
2082 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
2083 /// return the least SCEV value that is known never to be less than the
2084 /// actual backedge taken count.
2085 SCEVHandle ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
2086 return getBackedgeTakenInfo(L).Max;
2089 const ScalarEvolution::BackedgeTakenInfo &
2090 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
2091 // Initially insert a CouldNotCompute for this loop. If the insertion
2092 // succeeds, procede to actually compute a backedge-taken count and
2093 // update the value. The temporary CouldNotCompute value tells SCEV
2094 // code elsewhere that it shouldn't attempt to request a new
2095 // backedge-taken count, which could result in infinite recursion.
2096 std::pair<std::map<const Loop*, BackedgeTakenInfo>::iterator, bool> Pair =
2097 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
2099 BackedgeTakenInfo ItCount = ComputeBackedgeTakenCount(L);
2100 if (ItCount.Exact != UnknownValue) {
2101 assert(ItCount.Exact->isLoopInvariant(L) &&
2102 ItCount.Max->isLoopInvariant(L) &&
2103 "Computed trip count isn't loop invariant for loop!");
2104 ++NumTripCountsComputed;
2106 // Update the value in the map.
2107 Pair.first->second = ItCount;
2108 } else if (isa<PHINode>(L->getHeader()->begin())) {
2109 // Only count loops that have phi nodes as not being computable.
2110 ++NumTripCountsNotComputed;
2113 // Now that we know more about the trip count for this loop, forget any
2114 // existing SCEV values for PHI nodes in this loop since they are only
2115 // conservative estimates made without the benefit
2116 // of trip count information.
2117 if (ItCount.hasAnyInfo())
2118 for (BasicBlock::iterator I = L->getHeader()->begin();
2119 PHINode *PN = dyn_cast<PHINode>(I); ++I)
2120 deleteValueFromRecords(PN);
2122 return Pair.first->second;
2125 /// forgetLoopBackedgeTakenCount - This method should be called by the
2126 /// client when it has changed a loop in a way that may effect
2127 /// ScalarEvolution's ability to compute a trip count, or if the loop
2129 void ScalarEvolution::forgetLoopBackedgeTakenCount(const Loop *L) {
2130 BackedgeTakenCounts.erase(L);
2133 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
2134 /// of the specified loop will execute.
2135 ScalarEvolution::BackedgeTakenInfo
2136 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
2137 // If the loop has a non-one exit block count, we can't analyze it.
2138 SmallVector<BasicBlock*, 8> ExitBlocks;
2139 L->getExitBlocks(ExitBlocks);
2140 if (ExitBlocks.size() != 1) return UnknownValue;
2142 // Okay, there is one exit block. Try to find the condition that causes the
2143 // loop to be exited.
2144 BasicBlock *ExitBlock = ExitBlocks[0];
2146 BasicBlock *ExitingBlock = 0;
2147 for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock);
2149 if (L->contains(*PI)) {
2150 if (ExitingBlock == 0)
2153 return UnknownValue; // More than one block exiting!
2155 assert(ExitingBlock && "No exits from loop, something is broken!");
2157 // Okay, we've computed the exiting block. See what condition causes us to
2160 // FIXME: we should be able to handle switch instructions (with a single exit)
2161 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2162 if (ExitBr == 0) return UnknownValue;
2163 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
2165 // At this point, we know we have a conditional branch that determines whether
2166 // the loop is exited. However, we don't know if the branch is executed each
2167 // time through the loop. If not, then the execution count of the branch will
2168 // not be equal to the trip count of the loop.
2170 // Currently we check for this by checking to see if the Exit branch goes to
2171 // the loop header. If so, we know it will always execute the same number of
2172 // times as the loop. We also handle the case where the exit block *is* the
2173 // loop header. This is common for un-rotated loops. More extensive analysis
2174 // could be done to handle more cases here.
2175 if (ExitBr->getSuccessor(0) != L->getHeader() &&
2176 ExitBr->getSuccessor(1) != L->getHeader() &&
2177 ExitBr->getParent() != L->getHeader())
2178 return UnknownValue;
2180 ICmpInst *ExitCond = dyn_cast<ICmpInst>(ExitBr->getCondition());
2182 // If it's not an integer comparison then compute it the hard way.
2183 // Note that ICmpInst deals with pointer comparisons too so we must check
2184 // the type of the operand.
2185 if (ExitCond == 0 || isa<PointerType>(ExitCond->getOperand(0)->getType()))
2186 return ComputeBackedgeTakenCountExhaustively(L, ExitBr->getCondition(),
2187 ExitBr->getSuccessor(0) == ExitBlock);
2189 // If the condition was exit on true, convert the condition to exit on false
2190 ICmpInst::Predicate Cond;
2191 if (ExitBr->getSuccessor(1) == ExitBlock)
2192 Cond = ExitCond->getPredicate();
2194 Cond = ExitCond->getInversePredicate();
2196 // Handle common loops like: for (X = "string"; *X; ++X)
2197 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
2198 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
2200 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
2201 if (!isa<SCEVCouldNotCompute>(ItCnt)) return ItCnt;
2204 SCEVHandle LHS = getSCEV(ExitCond->getOperand(0));
2205 SCEVHandle RHS = getSCEV(ExitCond->getOperand(1));
2207 // Try to evaluate any dependencies out of the loop.
2208 SCEVHandle Tmp = getSCEVAtScope(LHS, L);
2209 if (!isa<SCEVCouldNotCompute>(Tmp)) LHS = Tmp;
2210 Tmp = getSCEVAtScope(RHS, L);
2211 if (!isa<SCEVCouldNotCompute>(Tmp)) RHS = Tmp;
2213 // At this point, we would like to compute how many iterations of the
2214 // loop the predicate will return true for these inputs.
2215 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
2216 // If there is a loop-invariant, force it into the RHS.
2217 std::swap(LHS, RHS);
2218 Cond = ICmpInst::getSwappedPredicate(Cond);
2221 // If we have a comparison of a chrec against a constant, try to use value
2222 // ranges to answer this query.
2223 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
2224 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
2225 if (AddRec->getLoop() == L) {
2226 // Form the comparison range using the constant of the correct type so
2227 // that the ConstantRange class knows to do a signed or unsigned
2229 ConstantInt *CompVal = RHSC->getValue();
2230 const Type *RealTy = ExitCond->getOperand(0)->getType();
2231 CompVal = dyn_cast<ConstantInt>(
2232 ConstantExpr::getBitCast(CompVal, RealTy));
2234 // Form the constant range.
2235 ConstantRange CompRange(
2236 ICmpInst::makeConstantRange(Cond, CompVal->getValue()));
2238 SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange, *this);
2239 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
2244 case ICmpInst::ICMP_NE: { // while (X != Y)
2245 // Convert to: while (X-Y != 0)
2246 SCEVHandle TC = HowFarToZero(getMinusSCEV(LHS, RHS), L);
2247 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2250 case ICmpInst::ICMP_EQ: {
2251 // Convert to: while (X-Y == 0) // while (X == Y)
2252 SCEVHandle TC = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
2253 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2256 case ICmpInst::ICMP_SLT: {
2257 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
2258 if (BTI.hasAnyInfo()) return BTI;
2261 case ICmpInst::ICMP_SGT: {
2262 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
2263 getNotSCEV(RHS), L, true);
2264 if (BTI.hasAnyInfo()) return BTI;
2267 case ICmpInst::ICMP_ULT: {
2268 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
2269 if (BTI.hasAnyInfo()) return BTI;
2272 case ICmpInst::ICMP_UGT: {
2273 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
2274 getNotSCEV(RHS), L, false);
2275 if (BTI.hasAnyInfo()) return BTI;
2280 errs() << "ComputeBackedgeTakenCount ";
2281 if (ExitCond->getOperand(0)->getType()->isUnsigned())
2282 errs() << "[unsigned] ";
2283 errs() << *LHS << " "
2284 << Instruction::getOpcodeName(Instruction::ICmp)
2285 << " " << *RHS << "\n";
2290 ComputeBackedgeTakenCountExhaustively(L, ExitCond,
2291 ExitBr->getSuccessor(0) == ExitBlock);
2294 static ConstantInt *
2295 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
2296 ScalarEvolution &SE) {
2297 SCEVHandle InVal = SE.getConstant(C);
2298 SCEVHandle Val = AddRec->evaluateAtIteration(InVal, SE);
2299 assert(isa<SCEVConstant>(Val) &&
2300 "Evaluation of SCEV at constant didn't fold correctly?");
2301 return cast<SCEVConstant>(Val)->getValue();
2304 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
2305 /// and a GEP expression (missing the pointer index) indexing into it, return
2306 /// the addressed element of the initializer or null if the index expression is
2309 GetAddressedElementFromGlobal(GlobalVariable *GV,
2310 const std::vector<ConstantInt*> &Indices) {
2311 Constant *Init = GV->getInitializer();
2312 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
2313 uint64_t Idx = Indices[i]->getZExtValue();
2314 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2315 assert(Idx < CS->getNumOperands() && "Bad struct index!");
2316 Init = cast<Constant>(CS->getOperand(Idx));
2317 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2318 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
2319 Init = cast<Constant>(CA->getOperand(Idx));
2320 } else if (isa<ConstantAggregateZero>(Init)) {
2321 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2322 assert(Idx < STy->getNumElements() && "Bad struct index!");
2323 Init = Constant::getNullValue(STy->getElementType(Idx));
2324 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
2325 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
2326 Init = Constant::getNullValue(ATy->getElementType());
2328 assert(0 && "Unknown constant aggregate type!");
2332 return 0; // Unknown initializer type
2338 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
2339 /// 'icmp op load X, cst', try to see if we can compute the backedge
2340 /// execution count.
2341 SCEVHandle ScalarEvolution::
2342 ComputeLoadConstantCompareBackedgeTakenCount(LoadInst *LI, Constant *RHS,
2344 ICmpInst::Predicate predicate) {
2345 if (LI->isVolatile()) return UnknownValue;
2347 // Check to see if the loaded pointer is a getelementptr of a global.
2348 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
2349 if (!GEP) return UnknownValue;
2351 // Make sure that it is really a constant global we are gepping, with an
2352 // initializer, and make sure the first IDX is really 0.
2353 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
2354 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
2355 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
2356 !cast<Constant>(GEP->getOperand(1))->isNullValue())
2357 return UnknownValue;
2359 // Okay, we allow one non-constant index into the GEP instruction.
2361 std::vector<ConstantInt*> Indexes;
2362 unsigned VarIdxNum = 0;
2363 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
2364 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
2365 Indexes.push_back(CI);
2366 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
2367 if (VarIdx) return UnknownValue; // Multiple non-constant idx's.
2368 VarIdx = GEP->getOperand(i);
2370 Indexes.push_back(0);
2373 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
2374 // Check to see if X is a loop variant variable value now.
2375 SCEVHandle Idx = getSCEV(VarIdx);
2376 SCEVHandle Tmp = getSCEVAtScope(Idx, L);
2377 if (!isa<SCEVCouldNotCompute>(Tmp)) Idx = Tmp;
2379 // We can only recognize very limited forms of loop index expressions, in
2380 // particular, only affine AddRec's like {C1,+,C2}.
2381 SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
2382 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
2383 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
2384 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
2385 return UnknownValue;
2387 unsigned MaxSteps = MaxBruteForceIterations;
2388 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
2389 ConstantInt *ItCst =
2390 ConstantInt::get(IdxExpr->getType(), IterationNum);
2391 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
2393 // Form the GEP offset.
2394 Indexes[VarIdxNum] = Val;
2396 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
2397 if (Result == 0) break; // Cannot compute!
2399 // Evaluate the condition for this iteration.
2400 Result = ConstantExpr::getICmp(predicate, Result, RHS);
2401 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
2402 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
2404 errs() << "\n***\n*** Computed loop count " << *ItCst
2405 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
2408 ++NumArrayLenItCounts;
2409 return getConstant(ItCst); // Found terminating iteration!
2412 return UnknownValue;
2416 /// CanConstantFold - Return true if we can constant fold an instruction of the
2417 /// specified type, assuming that all operands were constants.
2418 static bool CanConstantFold(const Instruction *I) {
2419 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
2420 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
2423 if (const CallInst *CI = dyn_cast<CallInst>(I))
2424 if (const Function *F = CI->getCalledFunction())
2425 return canConstantFoldCallTo(F);
2429 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
2430 /// in the loop that V is derived from. We allow arbitrary operations along the
2431 /// way, but the operands of an operation must either be constants or a value
2432 /// derived from a constant PHI. If this expression does not fit with these
2433 /// constraints, return null.
2434 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
2435 // If this is not an instruction, or if this is an instruction outside of the
2436 // loop, it can't be derived from a loop PHI.
2437 Instruction *I = dyn_cast<Instruction>(V);
2438 if (I == 0 || !L->contains(I->getParent())) return 0;
2440 if (PHINode *PN = dyn_cast<PHINode>(I)) {
2441 if (L->getHeader() == I->getParent())
2444 // We don't currently keep track of the control flow needed to evaluate
2445 // PHIs, so we cannot handle PHIs inside of loops.
2449 // If we won't be able to constant fold this expression even if the operands
2450 // are constants, return early.
2451 if (!CanConstantFold(I)) return 0;
2453 // Otherwise, we can evaluate this instruction if all of its operands are
2454 // constant or derived from a PHI node themselves.
2456 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
2457 if (!(isa<Constant>(I->getOperand(Op)) ||
2458 isa<GlobalValue>(I->getOperand(Op)))) {
2459 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
2460 if (P == 0) return 0; // Not evolving from PHI
2464 return 0; // Evolving from multiple different PHIs.
2467 // This is a expression evolving from a constant PHI!
2471 /// EvaluateExpression - Given an expression that passes the
2472 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
2473 /// in the loop has the value PHIVal. If we can't fold this expression for some
2474 /// reason, return null.
2475 static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
2476 if (isa<PHINode>(V)) return PHIVal;
2477 if (Constant *C = dyn_cast<Constant>(V)) return C;
2478 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
2479 Instruction *I = cast<Instruction>(V);
2481 std::vector<Constant*> Operands;
2482 Operands.resize(I->getNumOperands());
2484 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2485 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal);
2486 if (Operands[i] == 0) return 0;
2489 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
2490 return ConstantFoldCompareInstOperands(CI->getPredicate(),
2491 &Operands[0], Operands.size());
2493 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
2494 &Operands[0], Operands.size());
2497 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
2498 /// in the header of its containing loop, we know the loop executes a
2499 /// constant number of times, and the PHI node is just a recurrence
2500 /// involving constants, fold it.
2501 Constant *ScalarEvolution::
2502 getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& BEs, const Loop *L){
2503 std::map<PHINode*, Constant*>::iterator I =
2504 ConstantEvolutionLoopExitValue.find(PN);
2505 if (I != ConstantEvolutionLoopExitValue.end())
2508 if (BEs.ugt(APInt(BEs.getBitWidth(),MaxBruteForceIterations)))
2509 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
2511 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
2513 // Since the loop is canonicalized, the PHI node must have two entries. One
2514 // entry must be a constant (coming in from outside of the loop), and the
2515 // second must be derived from the same PHI.
2516 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
2517 Constant *StartCST =
2518 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
2520 return RetVal = 0; // Must be a constant.
2522 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
2523 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
2525 return RetVal = 0; // Not derived from same PHI.
2527 // Execute the loop symbolically to determine the exit value.
2528 if (BEs.getActiveBits() >= 32)
2529 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
2531 unsigned NumIterations = BEs.getZExtValue(); // must be in range
2532 unsigned IterationNum = 0;
2533 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
2534 if (IterationNum == NumIterations)
2535 return RetVal = PHIVal; // Got exit value!
2537 // Compute the value of the PHI node for the next iteration.
2538 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
2539 if (NextPHI == PHIVal)
2540 return RetVal = NextPHI; // Stopped evolving!
2542 return 0; // Couldn't evaluate!
2547 /// ComputeBackedgeTakenCountExhaustively - If the trip is known to execute a
2548 /// constant number of times (the condition evolves only from constants),
2549 /// try to evaluate a few iterations of the loop until we get the exit
2550 /// condition gets a value of ExitWhen (true or false). If we cannot
2551 /// evaluate the trip count of the loop, return UnknownValue.
2552 SCEVHandle ScalarEvolution::
2553 ComputeBackedgeTakenCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) {
2554 PHINode *PN = getConstantEvolvingPHI(Cond, L);
2555 if (PN == 0) return UnknownValue;
2557 // Since the loop is canonicalized, the PHI node must have two entries. One
2558 // entry must be a constant (coming in from outside of the loop), and the
2559 // second must be derived from the same PHI.
2560 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
2561 Constant *StartCST =
2562 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
2563 if (StartCST == 0) return UnknownValue; // Must be a constant.
2565 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
2566 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
2567 if (PN2 != PN) return UnknownValue; // Not derived from same PHI.
2569 // Okay, we find a PHI node that defines the trip count of this loop. Execute
2570 // the loop symbolically to determine when the condition gets a value of
2572 unsigned IterationNum = 0;
2573 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
2574 for (Constant *PHIVal = StartCST;
2575 IterationNum != MaxIterations; ++IterationNum) {
2576 ConstantInt *CondVal =
2577 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal));
2579 // Couldn't symbolically evaluate.
2580 if (!CondVal) return UnknownValue;
2582 if (CondVal->getValue() == uint64_t(ExitWhen)) {
2583 ConstantEvolutionLoopExitValue[PN] = PHIVal;
2584 ++NumBruteForceTripCountsComputed;
2585 return getConstant(ConstantInt::get(Type::Int32Ty, IterationNum));
2588 // Compute the value of the PHI node for the next iteration.
2589 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
2590 if (NextPHI == 0 || NextPHI == PHIVal)
2591 return UnknownValue; // Couldn't evaluate or not making progress...
2595 // Too many iterations were needed to evaluate.
2596 return UnknownValue;
2599 /// getSCEVAtScope - Compute the value of the specified expression within the
2600 /// indicated loop (which may be null to indicate in no loop). If the
2601 /// expression cannot be evaluated, return UnknownValue.
2602 SCEVHandle ScalarEvolution::getSCEVAtScope(SCEV *V, const Loop *L) {
2603 // FIXME: this should be turned into a virtual method on SCEV!
2605 if (isa<SCEVConstant>(V)) return V;
2607 // If this instruction is evolved from a constant-evolving PHI, compute the
2608 // exit value from the loop without using SCEVs.
2609 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
2610 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
2611 const Loop *LI = (*this->LI)[I->getParent()];
2612 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
2613 if (PHINode *PN = dyn_cast<PHINode>(I))
2614 if (PN->getParent() == LI->getHeader()) {
2615 // Okay, there is no closed form solution for the PHI node. Check
2616 // to see if the loop that contains it has a known backedge-taken
2617 // count. If so, we may be able to force computation of the exit
2619 SCEVHandle BackedgeTakenCount = getBackedgeTakenCount(LI);
2620 if (SCEVConstant *BTCC =
2621 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
2622 // Okay, we know how many times the containing loop executes. If
2623 // this is a constant evolving PHI node, get the final value at
2624 // the specified iteration number.
2625 Constant *RV = getConstantEvolutionLoopExitValue(PN,
2626 BTCC->getValue()->getValue(),
2628 if (RV) return getUnknown(RV);
2632 // Okay, this is an expression that we cannot symbolically evaluate
2633 // into a SCEV. Check to see if it's possible to symbolically evaluate
2634 // the arguments into constants, and if so, try to constant propagate the
2635 // result. This is particularly useful for computing loop exit values.
2636 if (CanConstantFold(I)) {
2637 std::vector<Constant*> Operands;
2638 Operands.reserve(I->getNumOperands());
2639 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2640 Value *Op = I->getOperand(i);
2641 if (Constant *C = dyn_cast<Constant>(Op)) {
2642 Operands.push_back(C);
2644 // If any of the operands is non-constant and if they are
2645 // non-integer and non-pointer, don't even try to analyze them
2646 // with scev techniques.
2647 if (!isSCEVable(Op->getType()))
2650 SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L);
2651 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
2652 Constant *C = SC->getValue();
2653 if (C->getType() != Op->getType())
2654 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
2658 Operands.push_back(C);
2659 } else if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
2660 if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
2661 if (C->getType() != Op->getType())
2663 ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
2667 Operands.push_back(C);
2677 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
2678 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
2679 &Operands[0], Operands.size());
2681 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
2682 &Operands[0], Operands.size());
2683 return getUnknown(C);
2687 // This is some other type of SCEVUnknown, just return it.
2691 if (SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
2692 // Avoid performing the look-up in the common case where the specified
2693 // expression has no loop-variant portions.
2694 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
2695 SCEVHandle OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
2696 if (OpAtScope != Comm->getOperand(i)) {
2697 if (OpAtScope == UnknownValue) return UnknownValue;
2698 // Okay, at least one of these operands is loop variant but might be
2699 // foldable. Build a new instance of the folded commutative expression.
2700 std::vector<SCEVHandle> NewOps(Comm->op_begin(), Comm->op_begin()+i);
2701 NewOps.push_back(OpAtScope);
2703 for (++i; i != e; ++i) {
2704 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
2705 if (OpAtScope == UnknownValue) return UnknownValue;
2706 NewOps.push_back(OpAtScope);
2708 if (isa<SCEVAddExpr>(Comm))
2709 return getAddExpr(NewOps);
2710 if (isa<SCEVMulExpr>(Comm))
2711 return getMulExpr(NewOps);
2712 if (isa<SCEVSMaxExpr>(Comm))
2713 return getSMaxExpr(NewOps);
2714 if (isa<SCEVUMaxExpr>(Comm))
2715 return getUMaxExpr(NewOps);
2716 assert(0 && "Unknown commutative SCEV type!");
2719 // If we got here, all operands are loop invariant.
2723 if (SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
2724 SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L);
2725 if (LHS == UnknownValue) return LHS;
2726 SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L);
2727 if (RHS == UnknownValue) return RHS;
2728 if (LHS == Div->getLHS() && RHS == Div->getRHS())
2729 return Div; // must be loop invariant
2730 return getUDivExpr(LHS, RHS);
2733 // If this is a loop recurrence for a loop that does not contain L, then we
2734 // are dealing with the final value computed by the loop.
2735 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
2736 if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
2737 // To evaluate this recurrence, we need to know how many times the AddRec
2738 // loop iterates. Compute this now.
2739 SCEVHandle BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
2740 if (BackedgeTakenCount == UnknownValue) return UnknownValue;
2742 // Then, evaluate the AddRec.
2743 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
2745 return UnknownValue;
2748 if (SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
2749 SCEVHandle Op = getSCEVAtScope(Cast->getOperand(), L);
2750 if (Op == UnknownValue) return Op;
2751 if (Op == Cast->getOperand())
2752 return Cast; // must be loop invariant
2753 return getZeroExtendExpr(Op, Cast->getType());
2756 if (SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
2757 SCEVHandle Op = getSCEVAtScope(Cast->getOperand(), L);
2758 if (Op == UnknownValue) return Op;
2759 if (Op == Cast->getOperand())
2760 return Cast; // must be loop invariant
2761 return getSignExtendExpr(Op, Cast->getType());
2764 if (SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
2765 SCEVHandle Op = getSCEVAtScope(Cast->getOperand(), L);
2766 if (Op == UnknownValue) return Op;
2767 if (Op == Cast->getOperand())
2768 return Cast; // must be loop invariant
2769 return getTruncateExpr(Op, Cast->getType());
2772 assert(0 && "Unknown SCEV type!");
2775 /// getSCEVAtScope - Return a SCEV expression handle for the specified value
2776 /// at the specified scope in the program. The L value specifies a loop
2777 /// nest to evaluate the expression at, where null is the top-level or a
2778 /// specified loop is immediately inside of the loop.
2780 /// This method can be used to compute the exit value for a variable defined
2781 /// in a loop by querying what the value will hold in the parent loop.
2783 /// If this value is not computable at this scope, a SCEVCouldNotCompute
2784 /// object is returned.
2785 SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
2786 return getSCEVAtScope(getSCEV(V), L);
2789 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
2790 /// following equation:
2792 /// A * X = B (mod N)
2794 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
2795 /// A and B isn't important.
2797 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
2798 static SCEVHandle SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
2799 ScalarEvolution &SE) {
2800 uint32_t BW = A.getBitWidth();
2801 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
2802 assert(A != 0 && "A must be non-zero.");
2806 // The gcd of A and N may have only one prime factor: 2. The number of
2807 // trailing zeros in A is its multiplicity
2808 uint32_t Mult2 = A.countTrailingZeros();
2811 // 2. Check if B is divisible by D.
2813 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
2814 // is not less than multiplicity of this prime factor for D.
2815 if (B.countTrailingZeros() < Mult2)
2816 return SE.getCouldNotCompute();
2818 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
2821 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
2822 // bit width during computations.
2823 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
2824 APInt Mod(BW + 1, 0);
2825 Mod.set(BW - Mult2); // Mod = N / D
2826 APInt I = AD.multiplicativeInverse(Mod);
2828 // 4. Compute the minimum unsigned root of the equation:
2829 // I * (B / D) mod (N / D)
2830 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
2832 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
2834 return SE.getConstant(Result.trunc(BW));
2837 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
2838 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
2839 /// might be the same) or two SCEVCouldNotCompute objects.
2841 static std::pair<SCEVHandle,SCEVHandle>
2842 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
2843 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
2844 SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
2845 SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
2846 SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
2848 // We currently can only solve this if the coefficients are constants.
2849 if (!LC || !MC || !NC) {
2850 SCEV *CNC = SE.getCouldNotCompute();
2851 return std::make_pair(CNC, CNC);
2854 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
2855 const APInt &L = LC->getValue()->getValue();
2856 const APInt &M = MC->getValue()->getValue();
2857 const APInt &N = NC->getValue()->getValue();
2858 APInt Two(BitWidth, 2);
2859 APInt Four(BitWidth, 4);
2862 using namespace APIntOps;
2864 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
2865 // The B coefficient is M-N/2
2869 // The A coefficient is N/2
2870 APInt A(N.sdiv(Two));
2872 // Compute the B^2-4ac term.
2875 SqrtTerm -= Four * (A * C);
2877 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
2878 // integer value or else APInt::sqrt() will assert.
2879 APInt SqrtVal(SqrtTerm.sqrt());
2881 // Compute the two solutions for the quadratic formula.
2882 // The divisions must be performed as signed divisions.
2884 APInt TwoA( A << 1 );
2885 if (TwoA.isMinValue()) {
2886 SCEV *CNC = SE.getCouldNotCompute();
2887 return std::make_pair(CNC, CNC);
2890 ConstantInt *Solution1 = ConstantInt::get((NegB + SqrtVal).sdiv(TwoA));
2891 ConstantInt *Solution2 = ConstantInt::get((NegB - SqrtVal).sdiv(TwoA));
2893 return std::make_pair(SE.getConstant(Solution1),
2894 SE.getConstant(Solution2));
2895 } // end APIntOps namespace
2898 /// HowFarToZero - Return the number of times a backedge comparing the specified
2899 /// value to zero will execute. If not computable, return UnknownValue
2900 SCEVHandle ScalarEvolution::HowFarToZero(SCEV *V, const Loop *L) {
2901 // If the value is a constant
2902 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2903 // If the value is already zero, the branch will execute zero times.
2904 if (C->getValue()->isZero()) return C;
2905 return UnknownValue; // Otherwise it will loop infinitely.
2908 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
2909 if (!AddRec || AddRec->getLoop() != L)
2910 return UnknownValue;
2912 if (AddRec->isAffine()) {
2913 // If this is an affine expression, the execution count of this branch is
2914 // the minimum unsigned root of the following equation:
2916 // Start + Step*N = 0 (mod 2^BW)
2920 // Step*N = -Start (mod 2^BW)
2922 // where BW is the common bit width of Start and Step.
2924 // Get the initial value for the loop.
2925 SCEVHandle Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
2926 if (isa<SCEVCouldNotCompute>(Start)) return UnknownValue;
2928 SCEVHandle Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
2930 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
2931 // For now we handle only constant steps.
2933 // First, handle unitary steps.
2934 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
2935 return getNegativeSCEV(Start); // N = -Start (as unsigned)
2936 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
2937 return Start; // N = Start (as unsigned)
2939 // Then, try to solve the above equation provided that Start is constant.
2940 if (SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
2941 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
2942 -StartC->getValue()->getValue(),
2945 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
2946 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
2947 // the quadratic equation to solve it.
2948 std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec,
2950 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2951 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2954 errs() << "HFTZ: " << *V << " - sol#1: " << *R1
2955 << " sol#2: " << *R2 << "\n";
2957 // Pick the smallest positive root value.
2958 if (ConstantInt *CB =
2959 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
2960 R1->getValue(), R2->getValue()))) {
2961 if (CB->getZExtValue() == false)
2962 std::swap(R1, R2); // R1 is the minimum root now.
2964 // We can only use this value if the chrec ends up with an exact zero
2965 // value at this index. When solving for "X*X != 5", for example, we
2966 // should not accept a root of 2.
2967 SCEVHandle Val = AddRec->evaluateAtIteration(R1, *this);
2969 return R1; // We found a quadratic root!
2974 return UnknownValue;
2977 /// HowFarToNonZero - Return the number of times a backedge checking the
2978 /// specified value for nonzero will execute. If not computable, return
2980 SCEVHandle ScalarEvolution::HowFarToNonZero(SCEV *V, const Loop *L) {
2981 // Loops that look like: while (X == 0) are very strange indeed. We don't
2982 // handle them yet except for the trivial case. This could be expanded in the
2983 // future as needed.
2985 // If the value is a constant, check to see if it is known to be non-zero
2986 // already. If so, the backedge will execute zero times.
2987 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2988 if (!C->getValue()->isNullValue())
2989 return getIntegerSCEV(0, C->getType());
2990 return UnknownValue; // Otherwise it will loop infinitely.
2993 // We could implement others, but I really doubt anyone writes loops like
2994 // this, and if they did, they would already be constant folded.
2995 return UnknownValue;
2998 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
2999 /// (which may not be an immediate predecessor) which has exactly one
3000 /// successor from which BB is reachable, or null if no such block is
3004 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
3005 // If the block has a unique predecessor, then there is no path from the
3006 // predecessor to the block that does not go through the direct edge
3007 // from the predecessor to the block.
3008 if (BasicBlock *Pred = BB->getSinglePredecessor())
3011 // A loop's header is defined to be a block that dominates the loop.
3012 // If the loop has a preheader, it must be a block that has exactly
3013 // one successor that can reach BB. This is slightly more strict
3014 // than necessary, but works if critical edges are split.
3015 if (Loop *L = LI->getLoopFor(BB))
3016 return L->getLoopPreheader();
3021 /// isLoopGuardedByCond - Test whether entry to the loop is protected by
3022 /// a conditional between LHS and RHS. This is used to help avoid max
3023 /// expressions in loop trip counts.
3024 bool ScalarEvolution::isLoopGuardedByCond(const Loop *L,
3025 ICmpInst::Predicate Pred,
3026 SCEV *LHS, SCEV *RHS) {
3027 BasicBlock *Preheader = L->getLoopPreheader();
3028 BasicBlock *PreheaderDest = L->getHeader();
3030 // Starting at the preheader, climb up the predecessor chain, as long as
3031 // there are predecessors that can be found that have unique successors
3032 // leading to the original header.
3034 PreheaderDest = Preheader,
3035 Preheader = getPredecessorWithUniqueSuccessorForBB(Preheader)) {
3037 BranchInst *LoopEntryPredicate =
3038 dyn_cast<BranchInst>(Preheader->getTerminator());
3039 if (!LoopEntryPredicate ||
3040 LoopEntryPredicate->isUnconditional())
3043 ICmpInst *ICI = dyn_cast<ICmpInst>(LoopEntryPredicate->getCondition());
3046 // Now that we found a conditional branch that dominates the loop, check to
3047 // see if it is the comparison we are looking for.
3048 Value *PreCondLHS = ICI->getOperand(0);
3049 Value *PreCondRHS = ICI->getOperand(1);
3050 ICmpInst::Predicate Cond;
3051 if (LoopEntryPredicate->getSuccessor(0) == PreheaderDest)
3052 Cond = ICI->getPredicate();
3054 Cond = ICI->getInversePredicate();
3057 ; // An exact match.
3058 else if (!ICmpInst::isTrueWhenEqual(Cond) && Pred == ICmpInst::ICMP_NE)
3059 ; // The actual condition is beyond sufficient.
3061 // Check a few special cases.
3063 case ICmpInst::ICMP_UGT:
3064 if (Pred == ICmpInst::ICMP_ULT) {
3065 std::swap(PreCondLHS, PreCondRHS);
3066 Cond = ICmpInst::ICMP_ULT;
3070 case ICmpInst::ICMP_SGT:
3071 if (Pred == ICmpInst::ICMP_SLT) {
3072 std::swap(PreCondLHS, PreCondRHS);
3073 Cond = ICmpInst::ICMP_SLT;
3077 case ICmpInst::ICMP_NE:
3078 // Expressions like (x >u 0) are often canonicalized to (x != 0),
3079 // so check for this case by checking if the NE is comparing against
3080 // a minimum or maximum constant.
3081 if (!ICmpInst::isTrueWhenEqual(Pred))
3082 if (ConstantInt *CI = dyn_cast<ConstantInt>(PreCondRHS)) {
3083 const APInt &A = CI->getValue();
3085 case ICmpInst::ICMP_SLT:
3086 if (A.isMaxSignedValue()) break;
3088 case ICmpInst::ICMP_SGT:
3089 if (A.isMinSignedValue()) break;
3091 case ICmpInst::ICMP_ULT:
3092 if (A.isMaxValue()) break;
3094 case ICmpInst::ICMP_UGT:
3095 if (A.isMinValue()) break;
3100 Cond = ICmpInst::ICMP_NE;
3101 // NE is symmetric but the original comparison may not be. Swap
3102 // the operands if necessary so that they match below.
3103 if (isa<SCEVConstant>(LHS))
3104 std::swap(PreCondLHS, PreCondRHS);
3109 // We weren't able to reconcile the condition.
3113 if (!PreCondLHS->getType()->isInteger()) continue;
3115 SCEVHandle PreCondLHSSCEV = getSCEV(PreCondLHS);
3116 SCEVHandle PreCondRHSSCEV = getSCEV(PreCondRHS);
3117 if ((LHS == PreCondLHSSCEV && RHS == PreCondRHSSCEV) ||
3118 (LHS == getNotSCEV(PreCondRHSSCEV) &&
3119 RHS == getNotSCEV(PreCondLHSSCEV)))
3126 /// HowManyLessThans - Return the number of times a backedge containing the
3127 /// specified less-than comparison will execute. If not computable, return
3129 ScalarEvolution::BackedgeTakenInfo ScalarEvolution::
3130 HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L, bool isSigned) {
3131 // Only handle: "ADDREC < LoopInvariant".
3132 if (!RHS->isLoopInvariant(L)) return UnknownValue;
3134 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
3135 if (!AddRec || AddRec->getLoop() != L)
3136 return UnknownValue;
3138 if (AddRec->isAffine()) {
3139 // FORNOW: We only support unit strides.
3140 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
3141 SCEVHandle Step = AddRec->getStepRecurrence(*this);
3142 SCEVHandle NegOne = getIntegerSCEV(-1, AddRec->getType());
3144 // TODO: handle non-constant strides.
3145 const SCEVConstant *CStep = dyn_cast<SCEVConstant>(Step);
3146 if (!CStep || CStep->isZero())
3147 return UnknownValue;
3148 if (CStep->getValue()->getValue() == 1) {
3149 // With unit stride, the iteration never steps past the limit value.
3150 } else if (CStep->getValue()->getValue().isStrictlyPositive()) {
3151 if (const SCEVConstant *CLimit = dyn_cast<SCEVConstant>(RHS)) {
3152 // Test whether a positive iteration iteration can step past the limit
3153 // value and past the maximum value for its type in a single step.
3155 APInt Max = APInt::getSignedMaxValue(BitWidth);
3156 if ((Max - CStep->getValue()->getValue())
3157 .slt(CLimit->getValue()->getValue()))
3158 return UnknownValue;
3160 APInt Max = APInt::getMaxValue(BitWidth);
3161 if ((Max - CStep->getValue()->getValue())
3162 .ult(CLimit->getValue()->getValue()))
3163 return UnknownValue;
3166 // TODO: handle non-constant limit values below.
3167 return UnknownValue;
3169 // TODO: handle negative strides below.
3170 return UnknownValue;
3172 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
3173 // m. So, we count the number of iterations in which {n,+,s} < m is true.
3174 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
3175 // treat m-n as signed nor unsigned due to overflow possibility.
3177 // First, we get the value of the LHS in the first iteration: n
3178 SCEVHandle Start = AddRec->getOperand(0);
3180 // Determine the minimum constant start value.
3181 SCEVHandle MinStart = isa<SCEVConstant>(Start) ? Start :
3182 getConstant(isSigned ? APInt::getSignedMinValue(BitWidth) :
3183 APInt::getMinValue(BitWidth));
3185 // If we know that the condition is true in order to enter the loop,
3186 // then we know that it will run exactly (m-n)/s times. Otherwise, we
3187 // only know if will execute (max(m,n)-n)/s times. In both cases, the
3188 // division must round up.
3189 SCEVHandle End = RHS;
3190 if (!isLoopGuardedByCond(L,
3191 isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
3192 getMinusSCEV(Start, Step), RHS))
3193 End = isSigned ? getSMaxExpr(RHS, Start)
3194 : getUMaxExpr(RHS, Start);
3196 // Determine the maximum constant end value.
3197 SCEVHandle MaxEnd = isa<SCEVConstant>(End) ? End :
3198 getConstant(isSigned ? APInt::getSignedMaxValue(BitWidth) :
3199 APInt::getMaxValue(BitWidth));
3201 // Finally, we subtract these two values and divide, rounding up, to get
3202 // the number of times the backedge is executed.
3203 SCEVHandle BECount = getUDivExpr(getAddExpr(getMinusSCEV(End, Start),
3204 getAddExpr(Step, NegOne)),
3207 // The maximum backedge count is similar, except using the minimum start
3208 // value and the maximum end value.
3209 SCEVHandle MaxBECount = getUDivExpr(getAddExpr(getMinusSCEV(MaxEnd,
3211 getAddExpr(Step, NegOne)),
3214 return BackedgeTakenInfo(BECount, MaxBECount);
3217 return UnknownValue;
3220 /// getNumIterationsInRange - Return the number of iterations of this loop that
3221 /// produce values in the specified constant range. Another way of looking at
3222 /// this is that it returns the first iteration number where the value is not in
3223 /// the condition, thus computing the exit count. If the iteration count can't
3224 /// be computed, an instance of SCEVCouldNotCompute is returned.
3225 SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
3226 ScalarEvolution &SE) const {
3227 if (Range.isFullSet()) // Infinite loop.
3228 return SE.getCouldNotCompute();
3230 // If the start is a non-zero constant, shift the range to simplify things.
3231 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
3232 if (!SC->getValue()->isZero()) {
3233 std::vector<SCEVHandle> Operands(op_begin(), op_end());
3234 Operands[0] = SE.getIntegerSCEV(0, SC->getType());
3235 SCEVHandle Shifted = SE.getAddRecExpr(Operands, getLoop());
3236 if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast<SCEVAddRecExpr>(Shifted))
3237 return ShiftedAddRec->getNumIterationsInRange(
3238 Range.subtract(SC->getValue()->getValue()), SE);
3239 // This is strange and shouldn't happen.
3240 return SE.getCouldNotCompute();
3243 // The only time we can solve this is when we have all constant indices.
3244 // Otherwise, we cannot determine the overflow conditions.
3245 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
3246 if (!isa<SCEVConstant>(getOperand(i)))
3247 return SE.getCouldNotCompute();
3250 // Okay at this point we know that all elements of the chrec are constants and
3251 // that the start element is zero.
3253 // First check to see if the range contains zero. If not, the first
3255 unsigned BitWidth = SE.getTypeSizeInBits(getType());
3256 if (!Range.contains(APInt(BitWidth, 0)))
3257 return SE.getConstant(ConstantInt::get(getType(),0));
3260 // If this is an affine expression then we have this situation:
3261 // Solve {0,+,A} in Range === Ax in Range
3263 // We know that zero is in the range. If A is positive then we know that
3264 // the upper value of the range must be the first possible exit value.
3265 // If A is negative then the lower of the range is the last possible loop
3266 // value. Also note that we already checked for a full range.
3267 APInt One(BitWidth,1);
3268 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
3269 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
3271 // The exit value should be (End+A)/A.
3272 APInt ExitVal = (End + A).udiv(A);
3273 ConstantInt *ExitValue = ConstantInt::get(ExitVal);
3275 // Evaluate at the exit value. If we really did fall out of the valid
3276 // range, then we computed our trip count, otherwise wrap around or other
3277 // things must have happened.
3278 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
3279 if (Range.contains(Val->getValue()))
3280 return SE.getCouldNotCompute(); // Something strange happened
3282 // Ensure that the previous value is in the range. This is a sanity check.
3283 assert(Range.contains(
3284 EvaluateConstantChrecAtConstant(this,
3285 ConstantInt::get(ExitVal - One), SE)->getValue()) &&
3286 "Linear scev computation is off in a bad way!");
3287 return SE.getConstant(ExitValue);
3288 } else if (isQuadratic()) {
3289 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
3290 // quadratic equation to solve it. To do this, we must frame our problem in
3291 // terms of figuring out when zero is crossed, instead of when
3292 // Range.getUpper() is crossed.
3293 std::vector<SCEVHandle> NewOps(op_begin(), op_end());
3294 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
3295 SCEVHandle NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
3297 // Next, solve the constructed addrec
3298 std::pair<SCEVHandle,SCEVHandle> Roots =
3299 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
3300 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
3301 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
3303 // Pick the smallest positive root value.
3304 if (ConstantInt *CB =
3305 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
3306 R1->getValue(), R2->getValue()))) {
3307 if (CB->getZExtValue() == false)
3308 std::swap(R1, R2); // R1 is the minimum root now.
3310 // Make sure the root is not off by one. The returned iteration should
3311 // not be in the range, but the previous one should be. When solving
3312 // for "X*X < 5", for example, we should not return a root of 2.
3313 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
3316 if (Range.contains(R1Val->getValue())) {
3317 // The next iteration must be out of the range...
3318 ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()+1);
3320 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
3321 if (!Range.contains(R1Val->getValue()))
3322 return SE.getConstant(NextVal);
3323 return SE.getCouldNotCompute(); // Something strange happened
3326 // If R1 was not in the range, then it is a good return value. Make
3327 // sure that R1-1 WAS in the range though, just in case.
3328 ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()-1);
3329 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
3330 if (Range.contains(R1Val->getValue()))
3332 return SE.getCouldNotCompute(); // Something strange happened
3337 return SE.getCouldNotCompute();
3342 //===----------------------------------------------------------------------===//
3343 // ScalarEvolution Class Implementation
3344 //===----------------------------------------------------------------------===//
3346 ScalarEvolution::ScalarEvolution()
3347 : FunctionPass(&ID), UnknownValue(new SCEVCouldNotCompute()) {
3350 bool ScalarEvolution::runOnFunction(Function &F) {
3352 LI = &getAnalysis<LoopInfo>();
3353 TD = getAnalysisIfAvailable<TargetData>();
3357 void ScalarEvolution::releaseMemory() {
3359 BackedgeTakenCounts.clear();
3360 ConstantEvolutionLoopExitValue.clear();
3363 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
3364 AU.setPreservesAll();
3365 AU.addRequiredTransitive<LoopInfo>();
3368 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
3369 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
3372 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
3374 // Print all inner loops first
3375 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
3376 PrintLoopInfo(OS, SE, *I);
3378 OS << "Loop " << L->getHeader()->getName() << ": ";
3380 SmallVector<BasicBlock*, 8> ExitBlocks;
3381 L->getExitBlocks(ExitBlocks);
3382 if (ExitBlocks.size() != 1)
3383 OS << "<multiple exits> ";
3385 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
3386 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
3388 OS << "Unpredictable backedge-taken count. ";
3394 void ScalarEvolution::print(raw_ostream &OS, const Module* ) const {
3395 // ScalarEvolution's implementaiton of the print method is to print
3396 // out SCEV values of all instructions that are interesting. Doing
3397 // this potentially causes it to create new SCEV objects though,
3398 // which technically conflicts with the const qualifier. This isn't
3399 // observable from outside the class though (the hasSCEV function
3400 // notwithstanding), so casting away the const isn't dangerous.
3401 ScalarEvolution &SE = *const_cast<ScalarEvolution*>(this);
3403 OS << "Classifying expressions for: " << F->getName() << "\n";
3404 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
3405 if (isSCEVable(I->getType())) {
3408 SCEVHandle SV = SE.getSCEV(&*I);
3412 if (const Loop *L = LI->getLoopFor((*I).getParent())) {
3414 SCEVHandle ExitValue = SE.getSCEVAtScope(&*I, L->getParentLoop());
3415 if (isa<SCEVCouldNotCompute>(ExitValue)) {
3416 OS << "<<Unknown>>";
3426 OS << "Determining loop execution counts for: " << F->getName() << "\n";
3427 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
3428 PrintLoopInfo(OS, &SE, *I);
3431 void ScalarEvolution::print(std::ostream &o, const Module *M) const {
3432 raw_os_ostream OS(o);