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/LoopInfo.h"
70 #include "llvm/Assembly/Writer.h"
71 #include "llvm/Transforms/Scalar.h"
72 #include "llvm/Support/CFG.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/Compiler.h"
75 #include "llvm/Support/ConstantRange.h"
76 #include "llvm/Support/InstIterator.h"
77 #include "llvm/Support/ManagedStatic.h"
78 #include "llvm/Support/MathExtras.h"
79 #include "llvm/Support/Streams.h"
80 #include "llvm/ADT/Statistic.h"
86 STATISTIC(NumArrayLenItCounts,
87 "Number of trip counts computed with array length");
88 STATISTIC(NumTripCountsComputed,
89 "Number of loops with predictable loop counts");
90 STATISTIC(NumTripCountsNotComputed,
91 "Number of loops without predictable loop counts");
92 STATISTIC(NumBruteForceTripCountsComputed,
93 "Number of loops with trip counts computed by force");
95 static cl::opt<unsigned>
96 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
97 cl::desc("Maximum number of iterations SCEV will "
98 "symbolically execute a constant derived loop"),
101 static RegisterPass<ScalarEvolution>
102 R("scalar-evolution", "Scalar Evolution Analysis", false, true);
103 char ScalarEvolution::ID = 0;
105 //===----------------------------------------------------------------------===//
106 // SCEV class definitions
107 //===----------------------------------------------------------------------===//
109 //===----------------------------------------------------------------------===//
110 // Implementation of the SCEV class.
113 void SCEV::dump() const {
117 uint32_t SCEV::getBitWidth() const {
118 if (const IntegerType* ITy = dyn_cast<IntegerType>(getType()))
119 return ITy->getBitWidth();
123 bool SCEV::isZero() const {
124 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
125 return SC->getValue()->isZero();
130 SCEVCouldNotCompute::SCEVCouldNotCompute() : SCEV(scCouldNotCompute) {}
132 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
133 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
137 const Type *SCEVCouldNotCompute::getType() const {
138 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
142 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
143 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
147 SCEVHandle SCEVCouldNotCompute::
148 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
149 const SCEVHandle &Conc,
150 ScalarEvolution &SE) const {
154 void SCEVCouldNotCompute::print(std::ostream &OS) const {
155 OS << "***COULDNOTCOMPUTE***";
158 bool SCEVCouldNotCompute::classof(const SCEV *S) {
159 return S->getSCEVType() == scCouldNotCompute;
163 // SCEVConstants - Only allow the creation of one SCEVConstant for any
164 // particular value. Don't use a SCEVHandle here, or else the object will
166 static ManagedStatic<std::map<ConstantInt*, SCEVConstant*> > SCEVConstants;
169 SCEVConstant::~SCEVConstant() {
170 SCEVConstants->erase(V);
173 SCEVHandle ScalarEvolution::getConstant(ConstantInt *V) {
174 SCEVConstant *&R = (*SCEVConstants)[V];
175 if (R == 0) R = new SCEVConstant(V);
179 SCEVHandle ScalarEvolution::getConstant(const APInt& Val) {
180 return getConstant(ConstantInt::get(Val));
183 const Type *SCEVConstant::getType() const { return V->getType(); }
185 void SCEVConstant::print(std::ostream &OS) const {
186 WriteAsOperand(OS, V, false);
189 // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any
190 // particular input. Don't use a SCEVHandle here, or else the object will
192 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
193 SCEVTruncateExpr*> > SCEVTruncates;
195 SCEVTruncateExpr::SCEVTruncateExpr(const SCEVHandle &op, const Type *ty)
196 : SCEV(scTruncate), Op(op), Ty(ty) {
197 assert(Op->getType()->isInteger() && Ty->isInteger() &&
198 "Cannot truncate non-integer value!");
199 assert(Op->getType()->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits()
200 && "This is not a truncating conversion!");
203 SCEVTruncateExpr::~SCEVTruncateExpr() {
204 SCEVTruncates->erase(std::make_pair(Op, Ty));
207 void SCEVTruncateExpr::print(std::ostream &OS) const {
208 OS << "(truncate " << *Op << " to " << *Ty << ")";
211 // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any
212 // particular input. Don't use a SCEVHandle here, or else the object will never
214 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
215 SCEVZeroExtendExpr*> > SCEVZeroExtends;
217 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty)
218 : SCEV(scZeroExtend), Op(op), Ty(ty) {
219 assert(Op->getType()->isInteger() && Ty->isInteger() &&
220 "Cannot zero extend non-integer value!");
221 assert(Op->getType()->getPrimitiveSizeInBits() < Ty->getPrimitiveSizeInBits()
222 && "This is not an extending conversion!");
225 SCEVZeroExtendExpr::~SCEVZeroExtendExpr() {
226 SCEVZeroExtends->erase(std::make_pair(Op, Ty));
229 void SCEVZeroExtendExpr::print(std::ostream &OS) const {
230 OS << "(zeroextend " << *Op << " to " << *Ty << ")";
233 // SCEVSignExtends - Only allow the creation of one SCEVSignExtendExpr for any
234 // particular input. Don't use a SCEVHandle here, or else the object will never
236 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
237 SCEVSignExtendExpr*> > SCEVSignExtends;
239 SCEVSignExtendExpr::SCEVSignExtendExpr(const SCEVHandle &op, const Type *ty)
240 : SCEV(scSignExtend), Op(op), Ty(ty) {
241 assert(Op->getType()->isInteger() && Ty->isInteger() &&
242 "Cannot sign extend non-integer value!");
243 assert(Op->getType()->getPrimitiveSizeInBits() < Ty->getPrimitiveSizeInBits()
244 && "This is not an extending conversion!");
247 SCEVSignExtendExpr::~SCEVSignExtendExpr() {
248 SCEVSignExtends->erase(std::make_pair(Op, Ty));
251 void SCEVSignExtendExpr::print(std::ostream &OS) const {
252 OS << "(signextend " << *Op << " to " << *Ty << ")";
255 // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any
256 // particular input. Don't use a SCEVHandle here, or else the object will never
258 static ManagedStatic<std::map<std::pair<unsigned, std::vector<SCEV*> >,
259 SCEVCommutativeExpr*> > SCEVCommExprs;
261 SCEVCommutativeExpr::~SCEVCommutativeExpr() {
262 SCEVCommExprs->erase(std::make_pair(getSCEVType(),
263 std::vector<SCEV*>(Operands.begin(),
267 void SCEVCommutativeExpr::print(std::ostream &OS) const {
268 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
269 const char *OpStr = getOperationStr();
270 OS << "(" << *Operands[0];
271 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
272 OS << OpStr << *Operands[i];
276 SCEVHandle SCEVCommutativeExpr::
277 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
278 const SCEVHandle &Conc,
279 ScalarEvolution &SE) const {
280 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
282 getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
283 if (H != getOperand(i)) {
284 std::vector<SCEVHandle> NewOps;
285 NewOps.reserve(getNumOperands());
286 for (unsigned j = 0; j != i; ++j)
287 NewOps.push_back(getOperand(j));
289 for (++i; i != e; ++i)
290 NewOps.push_back(getOperand(i)->
291 replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
293 if (isa<SCEVAddExpr>(this))
294 return SE.getAddExpr(NewOps);
295 else if (isa<SCEVMulExpr>(this))
296 return SE.getMulExpr(NewOps);
297 else if (isa<SCEVSMaxExpr>(this))
298 return SE.getSMaxExpr(NewOps);
299 else if (isa<SCEVUMaxExpr>(this))
300 return SE.getUMaxExpr(NewOps);
302 assert(0 && "Unknown commutative expr!");
309 // SCEVUDivs - Only allow the creation of one SCEVUDivExpr for any particular
310 // input. Don't use a SCEVHandle here, or else the object will never be
312 static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>,
313 SCEVUDivExpr*> > SCEVUDivs;
315 SCEVUDivExpr::~SCEVUDivExpr() {
316 SCEVUDivs->erase(std::make_pair(LHS, RHS));
319 void SCEVUDivExpr::print(std::ostream &OS) const {
320 OS << "(" << *LHS << " /u " << *RHS << ")";
323 const Type *SCEVUDivExpr::getType() const {
324 return LHS->getType();
328 // SCEVSDivs - Only allow the creation of one SCEVSDivExpr for any particular
329 // input. Don't use a SCEVHandle here, or else the object will never be
331 static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>,
332 SCEVSDivExpr*> > SCEVSDivs;
334 SCEVSDivExpr::~SCEVSDivExpr() {
335 SCEVSDivs->erase(std::make_pair(LHS, RHS));
338 void SCEVSDivExpr::print(std::ostream &OS) const {
339 OS << "(" << *LHS << " /s " << *RHS << ")";
342 const Type *SCEVSDivExpr::getType() const {
343 return LHS->getType();
347 // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
348 // particular input. Don't use a SCEVHandle here, or else the object will never
350 static ManagedStatic<std::map<std::pair<const Loop *, std::vector<SCEV*> >,
351 SCEVAddRecExpr*> > SCEVAddRecExprs;
353 SCEVAddRecExpr::~SCEVAddRecExpr() {
354 SCEVAddRecExprs->erase(std::make_pair(L,
355 std::vector<SCEV*>(Operands.begin(),
359 SCEVHandle SCEVAddRecExpr::
360 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
361 const SCEVHandle &Conc,
362 ScalarEvolution &SE) const {
363 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
365 getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
366 if (H != getOperand(i)) {
367 std::vector<SCEVHandle> NewOps;
368 NewOps.reserve(getNumOperands());
369 for (unsigned j = 0; j != i; ++j)
370 NewOps.push_back(getOperand(j));
372 for (++i; i != e; ++i)
373 NewOps.push_back(getOperand(i)->
374 replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
376 return SE.getAddRecExpr(NewOps, L);
383 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
384 // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't
385 // contain L and if the start is invariant.
386 return !QueryLoop->contains(L->getHeader()) &&
387 getOperand(0)->isLoopInvariant(QueryLoop);
391 void SCEVAddRecExpr::print(std::ostream &OS) const {
392 OS << "{" << *Operands[0];
393 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
394 OS << ",+," << *Operands[i];
395 OS << "}<" << L->getHeader()->getName() + ">";
398 // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular
399 // value. Don't use a SCEVHandle here, or else the object will never be
401 static ManagedStatic<std::map<Value*, SCEVUnknown*> > SCEVUnknowns;
403 SCEVUnknown::~SCEVUnknown() { SCEVUnknowns->erase(V); }
405 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
406 // All non-instruction values are loop invariant. All instructions are loop
407 // invariant if they are not contained in the specified loop.
408 if (Instruction *I = dyn_cast<Instruction>(V))
409 return !L->contains(I->getParent());
413 const Type *SCEVUnknown::getType() const {
417 void SCEVUnknown::print(std::ostream &OS) const {
418 WriteAsOperand(OS, V, false);
421 //===----------------------------------------------------------------------===//
423 //===----------------------------------------------------------------------===//
426 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
427 /// than the complexity of the RHS. This comparator is used to canonicalize
429 struct VISIBILITY_HIDDEN SCEVComplexityCompare {
430 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
431 return LHS->getSCEVType() < RHS->getSCEVType();
436 /// GroupByComplexity - Given a list of SCEV objects, order them by their
437 /// complexity, and group objects of the same complexity together by value.
438 /// When this routine is finished, we know that any duplicates in the vector are
439 /// consecutive and that complexity is monotonically increasing.
441 /// Note that we go take special precautions to ensure that we get determinstic
442 /// results from this routine. In other words, we don't want the results of
443 /// this to depend on where the addresses of various SCEV objects happened to
446 static void GroupByComplexity(std::vector<SCEVHandle> &Ops) {
447 if (Ops.size() < 2) return; // Noop
448 if (Ops.size() == 2) {
449 // This is the common case, which also happens to be trivially simple.
451 if (SCEVComplexityCompare()(Ops[1], Ops[0]))
452 std::swap(Ops[0], Ops[1]);
456 // Do the rough sort by complexity.
457 std::sort(Ops.begin(), Ops.end(), SCEVComplexityCompare());
459 // Now that we are sorted by complexity, group elements of the same
460 // complexity. Note that this is, at worst, N^2, but the vector is likely to
461 // be extremely short in practice. Note that we take this approach because we
462 // do not want to depend on the addresses of the objects we are grouping.
463 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
465 unsigned Complexity = S->getSCEVType();
467 // If there are any objects of the same complexity and same value as this
469 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
470 if (Ops[j] == S) { // Found a duplicate.
471 // Move it to immediately after i'th element.
472 std::swap(Ops[i+1], Ops[j]);
473 ++i; // no need to rescan it.
474 if (i == e-2) return; // Done!
482 //===----------------------------------------------------------------------===//
483 // Simple SCEV method implementations
484 //===----------------------------------------------------------------------===//
486 /// getIntegerSCEV - Given an integer or FP type, create a constant for the
487 /// specified signed integer value and return a SCEV for the constant.
488 SCEVHandle ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
491 C = Constant::getNullValue(Ty);
492 else if (Ty->isFloatingPoint())
493 C = ConstantFP::get(APFloat(Ty==Type::FloatTy ? APFloat::IEEEsingle :
494 APFloat::IEEEdouble, Val));
496 C = ConstantInt::get(Ty, Val);
497 return getUnknown(C);
500 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
502 SCEVHandle ScalarEvolution::getNegativeSCEV(const SCEVHandle &V) {
503 if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
504 return getUnknown(ConstantExpr::getNeg(VC->getValue()));
506 return getMulExpr(V, getConstant(ConstantInt::getAllOnesValue(V->getType())));
509 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
510 SCEVHandle ScalarEvolution::getNotSCEV(const SCEVHandle &V) {
511 if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
512 return getUnknown(ConstantExpr::getNot(VC->getValue()));
514 SCEVHandle AllOnes = getConstant(ConstantInt::getAllOnesValue(V->getType()));
515 return getMinusSCEV(AllOnes, V);
518 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
520 SCEVHandle ScalarEvolution::getMinusSCEV(const SCEVHandle &LHS,
521 const SCEVHandle &RHS) {
523 return getAddExpr(LHS, getNegativeSCEV(RHS));
527 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
529 static SCEVHandle BinomialCoefficient(SCEVHandle It, unsigned K,
531 const IntegerType* ResultTy) {
532 // Handle the simplest case efficiently.
534 return SE.getTruncateOrZeroExtend(It, ResultTy);
536 // We are using the following formula for BC(It, K):
538 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
540 // Suppose, W is the bitwidth of the return value. We must be prepared for
541 // overflow. Hence, we must assure that the result of our computation is
542 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
543 // safe in modular arithmetic.
545 // However, this code doesn't use exactly that formula; the formula it uses
546 // is something like the following, where T is the number of factors of 2 in
547 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
550 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
552 // This formula is trivially equivalent to the previous formula. However,
553 // this formula can be implemented much more efficiently. The trick is that
554 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
555 // arithmetic. To do exact division in modular arithmetic, all we have
556 // to do is multiply by the inverse. Therefore, this step can be done at
559 // The next issue is how to safely do the division by 2^T. The way this
560 // is done is by doing the multiplication step at a width of at least W + T
561 // bits. This way, the bottom W+T bits of the product are accurate. Then,
562 // when we perform the division by 2^T (which is equivalent to a right shift
563 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
564 // truncated out after the division by 2^T.
566 // In comparison to just directly using the first formula, this technique
567 // is much more efficient; using the first formula requires W * K bits,
568 // but this formula less than W + K bits. Also, the first formula requires
569 // a division step, whereas this formula only requires multiplies and shifts.
571 // It doesn't matter whether the subtraction step is done in the calculation
572 // width or the input iteration count's width; if the subtraction overflows,
573 // the result must be zero anyway. We prefer here to do it in the width of
574 // the induction variable because it helps a lot for certain cases; CodeGen
575 // isn't smart enough to ignore the overflow, which leads to much less
576 // efficient code if the width of the subtraction is wider than the native
579 // (It's possible to not widen at all by pulling out factors of 2 before
580 // the multiplication; for example, K=2 can be calculated as
581 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
582 // extra arithmetic, so it's not an obvious win, and it gets
583 // much more complicated for K > 3.)
585 // Protection from insane SCEVs; this bound is conservative,
586 // but it probably doesn't matter.
588 return new SCEVCouldNotCompute();
590 unsigned W = ResultTy->getBitWidth();
592 // Calculate K! / 2^T and T; we divide out the factors of two before
593 // multiplying for calculating K! / 2^T to avoid overflow.
594 // Other overflow doesn't matter because we only care about the bottom
595 // W bits of the result.
596 APInt OddFactorial(W, 1);
598 for (unsigned i = 3; i <= K; ++i) {
600 unsigned TwoFactors = Mult.countTrailingZeros();
602 Mult = Mult.lshr(TwoFactors);
603 OddFactorial *= Mult;
606 // We need at least W + T bits for the multiplication step
607 // FIXME: A temporary hack; we round up the bitwidths
608 // to the nearest power of 2 to be nice to the code generator.
609 unsigned CalculationBits = 1U << Log2_32_Ceil(W + T);
610 // FIXME: Temporary hack to avoid generating integers that are too wide.
611 // Although, it's not completely clear how to determine how much
612 // widening is safe; for example, on X86, we can't really widen
613 // beyond 64 because we need to be able to do multiplication
614 // that's CalculationBits wide, but on X86-64, we can safely widen up to
616 if (CalculationBits > 64)
617 return new SCEVCouldNotCompute();
619 // Calcuate 2^T, at width T+W.
620 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
622 // Calculate the multiplicative inverse of K! / 2^T;
623 // this multiplication factor will perform the exact division by
625 APInt Mod = APInt::getSignedMinValue(W+1);
626 APInt MultiplyFactor = OddFactorial.zext(W+1);
627 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
628 MultiplyFactor = MultiplyFactor.trunc(W);
630 // Calculate the product, at width T+W
631 const IntegerType *CalculationTy = IntegerType::get(CalculationBits);
632 SCEVHandle Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
633 for (unsigned i = 1; i != K; ++i) {
634 SCEVHandle S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
635 Dividend = SE.getMulExpr(Dividend,
636 SE.getTruncateOrZeroExtend(S, CalculationTy));
640 SCEVHandle DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
642 // Truncate the result, and divide by K! / 2^T.
644 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
645 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
648 /// evaluateAtIteration - Return the value of this chain of recurrences at
649 /// the specified iteration number. We can evaluate this recurrence by
650 /// multiplying each element in the chain by the binomial coefficient
651 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
653 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
655 /// where BC(It, k) stands for binomial coefficient.
657 SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It,
658 ScalarEvolution &SE) const {
659 SCEVHandle Result = getStart();
660 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
661 // The computation is correct in the face of overflow provided that the
662 // multiplication is performed _after_ the evaluation of the binomial
664 SCEVHandle Coeff = BinomialCoefficient(It, i, SE,
665 cast<IntegerType>(getType()));
666 if (isa<SCEVCouldNotCompute>(Coeff))
669 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
674 //===----------------------------------------------------------------------===//
675 // SCEV Expression folder implementations
676 //===----------------------------------------------------------------------===//
678 SCEVHandle ScalarEvolution::getTruncateExpr(const SCEVHandle &Op, const Type *Ty) {
679 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
681 ConstantExpr::getTrunc(SC->getValue(), Ty));
683 // If the input value is a chrec scev made out of constants, truncate
684 // all of the constants.
685 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
686 std::vector<SCEVHandle> Operands;
687 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
688 // FIXME: This should allow truncation of other expression types!
689 if (isa<SCEVConstant>(AddRec->getOperand(i)))
690 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
693 if (Operands.size() == AddRec->getNumOperands())
694 return getAddRecExpr(Operands, AddRec->getLoop());
697 SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)];
698 if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
702 SCEVHandle ScalarEvolution::getZeroExtendExpr(const SCEVHandle &Op, const Type *Ty) {
703 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
705 ConstantExpr::getZExt(SC->getValue(), Ty));
707 // FIXME: If the input value is a chrec scev, and we can prove that the value
708 // did not overflow the old, smaller, value, we can zero extend all of the
709 // operands (often constants). This would allow analysis of something like
710 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
712 SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)];
713 if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
717 SCEVHandle ScalarEvolution::getSignExtendExpr(const SCEVHandle &Op, const Type *Ty) {
718 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
720 ConstantExpr::getSExt(SC->getValue(), Ty));
722 // FIXME: If the input value is a chrec scev, and we can prove that the value
723 // did not overflow the old, smaller, value, we can sign extend all of the
724 // operands (often constants). This would allow analysis of something like
725 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
727 SCEVSignExtendExpr *&Result = (*SCEVSignExtends)[std::make_pair(Op, Ty)];
728 if (Result == 0) Result = new SCEVSignExtendExpr(Op, Ty);
732 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion
733 /// of the input value to the specified type. If the type must be
734 /// extended, it is zero extended.
735 SCEVHandle ScalarEvolution::getTruncateOrZeroExtend(const SCEVHandle &V,
737 const Type *SrcTy = V->getType();
738 assert(SrcTy->isInteger() && Ty->isInteger() &&
739 "Cannot truncate or zero extend with non-integer arguments!");
740 if (SrcTy->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
741 return V; // No conversion
742 if (SrcTy->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits())
743 return getTruncateExpr(V, Ty);
744 return getZeroExtendExpr(V, Ty);
747 // get - Get a canonical add expression, or something simpler if possible.
748 SCEVHandle ScalarEvolution::getAddExpr(std::vector<SCEVHandle> &Ops) {
749 assert(!Ops.empty() && "Cannot get empty add!");
750 if (Ops.size() == 1) return Ops[0];
752 // Sort by complexity, this groups all similar expression types together.
753 GroupByComplexity(Ops);
755 // If there are any constants, fold them together.
757 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
759 assert(Idx < Ops.size());
760 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
761 // We found two constants, fold them together!
762 ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() +
763 RHSC->getValue()->getValue());
764 Ops[0] = getConstant(Fold);
765 Ops.erase(Ops.begin()+1); // Erase the folded element
766 if (Ops.size() == 1) return Ops[0];
767 LHSC = cast<SCEVConstant>(Ops[0]);
770 // If we are left with a constant zero being added, strip it off.
771 if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
772 Ops.erase(Ops.begin());
777 if (Ops.size() == 1) return Ops[0];
779 // Okay, check to see if the same value occurs in the operand list twice. If
780 // so, merge them together into an multiply expression. Since we sorted the
781 // list, these values are required to be adjacent.
782 const Type *Ty = Ops[0]->getType();
783 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
784 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
785 // Found a match, merge the two values into a multiply, and add any
786 // remaining values to the result.
787 SCEVHandle Two = getIntegerSCEV(2, Ty);
788 SCEVHandle Mul = getMulExpr(Ops[i], Two);
791 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
793 return getAddExpr(Ops);
796 // Now we know the first non-constant operand. Skip past any cast SCEVs.
797 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
800 // If there are add operands they would be next.
801 if (Idx < Ops.size()) {
802 bool DeletedAdd = false;
803 while (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
804 // If we have an add, expand the add operands onto the end of the operands
806 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
807 Ops.erase(Ops.begin()+Idx);
811 // If we deleted at least one add, we added operands to the end of the list,
812 // and they are not necessarily sorted. Recurse to resort and resimplify
813 // any operands we just aquired.
815 return getAddExpr(Ops);
818 // Skip over the add expression until we get to a multiply.
819 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
822 // If we are adding something to a multiply expression, make sure the
823 // something is not already an operand of the multiply. If so, merge it into
825 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
826 SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
827 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
828 SCEV *MulOpSCEV = Mul->getOperand(MulOp);
829 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
830 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(MulOpSCEV)) {
831 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
832 SCEVHandle InnerMul = Mul->getOperand(MulOp == 0);
833 if (Mul->getNumOperands() != 2) {
834 // If the multiply has more than two operands, we must get the
836 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
837 MulOps.erase(MulOps.begin()+MulOp);
838 InnerMul = getMulExpr(MulOps);
840 SCEVHandle One = getIntegerSCEV(1, Ty);
841 SCEVHandle AddOne = getAddExpr(InnerMul, One);
842 SCEVHandle OuterMul = getMulExpr(AddOne, Ops[AddOp]);
843 if (Ops.size() == 2) return OuterMul;
845 Ops.erase(Ops.begin()+AddOp);
846 Ops.erase(Ops.begin()+Idx-1);
848 Ops.erase(Ops.begin()+Idx);
849 Ops.erase(Ops.begin()+AddOp-1);
851 Ops.push_back(OuterMul);
852 return getAddExpr(Ops);
855 // Check this multiply against other multiplies being added together.
856 for (unsigned OtherMulIdx = Idx+1;
857 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
859 SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
860 // If MulOp occurs in OtherMul, we can fold the two multiplies
862 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
863 OMulOp != e; ++OMulOp)
864 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
865 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
866 SCEVHandle InnerMul1 = Mul->getOperand(MulOp == 0);
867 if (Mul->getNumOperands() != 2) {
868 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
869 MulOps.erase(MulOps.begin()+MulOp);
870 InnerMul1 = getMulExpr(MulOps);
872 SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0);
873 if (OtherMul->getNumOperands() != 2) {
874 std::vector<SCEVHandle> MulOps(OtherMul->op_begin(),
876 MulOps.erase(MulOps.begin()+OMulOp);
877 InnerMul2 = getMulExpr(MulOps);
879 SCEVHandle InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
880 SCEVHandle OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
881 if (Ops.size() == 2) return OuterMul;
882 Ops.erase(Ops.begin()+Idx);
883 Ops.erase(Ops.begin()+OtherMulIdx-1);
884 Ops.push_back(OuterMul);
885 return getAddExpr(Ops);
891 // If there are any add recurrences in the operands list, see if any other
892 // added values are loop invariant. If so, we can fold them into the
894 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
897 // Scan over all recurrences, trying to fold loop invariants into them.
898 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
899 // Scan all of the other operands to this add and add them to the vector if
900 // they are loop invariant w.r.t. the recurrence.
901 std::vector<SCEVHandle> LIOps;
902 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
903 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
904 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
905 LIOps.push_back(Ops[i]);
906 Ops.erase(Ops.begin()+i);
910 // If we found some loop invariants, fold them into the recurrence.
911 if (!LIOps.empty()) {
912 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
913 LIOps.push_back(AddRec->getStart());
915 std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
916 AddRecOps[0] = getAddExpr(LIOps);
918 SCEVHandle NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
919 // If all of the other operands were loop invariant, we are done.
920 if (Ops.size() == 1) return NewRec;
922 // Otherwise, add the folded AddRec by the non-liv parts.
923 for (unsigned i = 0;; ++i)
924 if (Ops[i] == AddRec) {
928 return getAddExpr(Ops);
931 // Okay, if there weren't any loop invariants to be folded, check to see if
932 // there are multiple AddRec's with the same loop induction variable being
933 // added together. If so, we can fold them.
934 for (unsigned OtherIdx = Idx+1;
935 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
936 if (OtherIdx != Idx) {
937 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
938 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
939 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
940 std::vector<SCEVHandle> NewOps(AddRec->op_begin(), AddRec->op_end());
941 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
942 if (i >= NewOps.size()) {
943 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
944 OtherAddRec->op_end());
947 NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
949 SCEVHandle NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
951 if (Ops.size() == 2) return NewAddRec;
953 Ops.erase(Ops.begin()+Idx);
954 Ops.erase(Ops.begin()+OtherIdx-1);
955 Ops.push_back(NewAddRec);
956 return getAddExpr(Ops);
960 // Otherwise couldn't fold anything into this recurrence. Move onto the
964 // Okay, it looks like we really DO need an add expr. Check to see if we
965 // already have one, otherwise create a new one.
966 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
967 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scAddExpr,
969 if (Result == 0) Result = new SCEVAddExpr(Ops);
974 SCEVHandle ScalarEvolution::getMulExpr(std::vector<SCEVHandle> &Ops) {
975 assert(!Ops.empty() && "Cannot get empty mul!");
977 // Sort by complexity, this groups all similar expression types together.
978 GroupByComplexity(Ops);
980 // If there are any constants, fold them together.
982 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
984 // C1*(C2+V) -> C1*C2 + C1*V
986 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
987 if (Add->getNumOperands() == 2 &&
988 isa<SCEVConstant>(Add->getOperand(0)))
989 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
990 getMulExpr(LHSC, Add->getOperand(1)));
994 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
995 // We found two constants, fold them together!
996 ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() *
997 RHSC->getValue()->getValue());
998 Ops[0] = getConstant(Fold);
999 Ops.erase(Ops.begin()+1); // Erase the folded element
1000 if (Ops.size() == 1) return Ops[0];
1001 LHSC = cast<SCEVConstant>(Ops[0]);
1004 // If we are left with a constant one being multiplied, strip it off.
1005 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1006 Ops.erase(Ops.begin());
1008 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1009 // If we have a multiply of zero, it will always be zero.
1014 // Skip over the add expression until we get to a multiply.
1015 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1018 if (Ops.size() == 1)
1021 // If there are mul operands inline them all into this expression.
1022 if (Idx < Ops.size()) {
1023 bool DeletedMul = false;
1024 while (SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1025 // If we have an mul, expand the mul operands onto the end of the operands
1027 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1028 Ops.erase(Ops.begin()+Idx);
1032 // If we deleted at least one mul, we added operands to the end of the list,
1033 // and they are not necessarily sorted. Recurse to resort and resimplify
1034 // any operands we just aquired.
1036 return getMulExpr(Ops);
1039 // If there are any add recurrences in the operands list, see if any other
1040 // added values are loop invariant. If so, we can fold them into the
1042 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1045 // Scan over all recurrences, trying to fold loop invariants into them.
1046 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1047 // Scan all of the other operands to this mul and add them to the vector if
1048 // they are loop invariant w.r.t. the recurrence.
1049 std::vector<SCEVHandle> LIOps;
1050 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1051 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1052 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1053 LIOps.push_back(Ops[i]);
1054 Ops.erase(Ops.begin()+i);
1058 // If we found some loop invariants, fold them into the recurrence.
1059 if (!LIOps.empty()) {
1060 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1061 std::vector<SCEVHandle> NewOps;
1062 NewOps.reserve(AddRec->getNumOperands());
1063 if (LIOps.size() == 1) {
1064 SCEV *Scale = LIOps[0];
1065 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1066 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1068 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1069 std::vector<SCEVHandle> MulOps(LIOps);
1070 MulOps.push_back(AddRec->getOperand(i));
1071 NewOps.push_back(getMulExpr(MulOps));
1075 SCEVHandle NewRec = getAddRecExpr(NewOps, AddRec->getLoop());
1077 // If all of the other operands were loop invariant, we are done.
1078 if (Ops.size() == 1) return NewRec;
1080 // Otherwise, multiply the folded AddRec by the non-liv parts.
1081 for (unsigned i = 0;; ++i)
1082 if (Ops[i] == AddRec) {
1086 return getMulExpr(Ops);
1089 // Okay, if there weren't any loop invariants to be folded, check to see if
1090 // there are multiple AddRec's with the same loop induction variable being
1091 // multiplied together. If so, we can fold them.
1092 for (unsigned OtherIdx = Idx+1;
1093 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1094 if (OtherIdx != Idx) {
1095 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1096 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1097 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1098 SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1099 SCEVHandle NewStart = getMulExpr(F->getStart(),
1101 SCEVHandle B = F->getStepRecurrence(*this);
1102 SCEVHandle D = G->getStepRecurrence(*this);
1103 SCEVHandle NewStep = getAddExpr(getMulExpr(F, D),
1106 SCEVHandle NewAddRec = getAddRecExpr(NewStart, NewStep,
1108 if (Ops.size() == 2) return NewAddRec;
1110 Ops.erase(Ops.begin()+Idx);
1111 Ops.erase(Ops.begin()+OtherIdx-1);
1112 Ops.push_back(NewAddRec);
1113 return getMulExpr(Ops);
1117 // Otherwise couldn't fold anything into this recurrence. Move onto the
1121 // Okay, it looks like we really DO need an mul expr. Check to see if we
1122 // already have one, otherwise create a new one.
1123 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1124 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scMulExpr,
1127 Result = new SCEVMulExpr(Ops);
1131 SCEVHandle ScalarEvolution::getUDivExpr(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1133 return getIntegerSCEV(1, LHS->getType()); // X udiv X --> 1
1135 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1136 if (RHSC->getValue()->equalsInt(1))
1137 return LHS; // X udiv 1 --> X
1139 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1140 Constant *LHSCV = LHSC->getValue();
1141 Constant *RHSCV = RHSC->getValue();
1142 return getUnknown(ConstantExpr::getUDiv(LHSCV, RHSCV));
1146 SCEVUDivExpr *&Result = (*SCEVUDivs)[std::make_pair(LHS, RHS)];
1147 if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
1151 SCEVHandle ScalarEvolution::getSDivExpr(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1153 return getIntegerSCEV(1, LHS->getType()); // X sdiv X --> 1
1155 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1156 if (RHSC->getValue()->equalsInt(1))
1157 return LHS; // X sdiv 1 --> X
1159 if (RHSC->getValue()->isAllOnesValue())
1160 return getNegativeSCEV(LHS); // X sdiv -1 --> -X
1162 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1163 Constant *LHSCV = LHSC->getValue();
1164 Constant *RHSCV = RHSC->getValue();
1165 return getUnknown(ConstantExpr::getSDiv(LHSCV, RHSCV));
1169 SCEVSDivExpr *&Result = (*SCEVSDivs)[std::make_pair(LHS, RHS)];
1170 if (Result == 0) Result = new SCEVSDivExpr(LHS, RHS);
1175 /// SCEVAddRecExpr::get - Get a add recurrence expression for the
1176 /// specified loop. Simplify the expression as much as possible.
1177 SCEVHandle ScalarEvolution::getAddRecExpr(const SCEVHandle &Start,
1178 const SCEVHandle &Step, const Loop *L) {
1179 std::vector<SCEVHandle> Operands;
1180 Operands.push_back(Start);
1181 if (SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1182 if (StepChrec->getLoop() == L) {
1183 Operands.insert(Operands.end(), StepChrec->op_begin(),
1184 StepChrec->op_end());
1185 return getAddRecExpr(Operands, L);
1188 Operands.push_back(Step);
1189 return getAddRecExpr(Operands, L);
1192 /// SCEVAddRecExpr::get - Get a add recurrence expression for the
1193 /// specified loop. Simplify the expression as much as possible.
1194 SCEVHandle ScalarEvolution::getAddRecExpr(std::vector<SCEVHandle> &Operands,
1196 if (Operands.size() == 1) return Operands[0];
1198 if (Operands.back()->isZero()) {
1199 Operands.pop_back();
1200 return getAddRecExpr(Operands, L); // {X,+,0} --> X
1203 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
1204 if (SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
1205 const Loop* NestedLoop = NestedAR->getLoop();
1206 if (L->getLoopDepth() < NestedLoop->getLoopDepth()) {
1207 std::vector<SCEVHandle> NestedOperands(NestedAR->op_begin(),
1208 NestedAR->op_end());
1209 SCEVHandle NestedARHandle(NestedAR);
1210 Operands[0] = NestedAR->getStart();
1211 NestedOperands[0] = getAddRecExpr(Operands, L);
1212 return getAddRecExpr(NestedOperands, NestedLoop);
1216 SCEVAddRecExpr *&Result =
1217 (*SCEVAddRecExprs)[std::make_pair(L, std::vector<SCEV*>(Operands.begin(),
1219 if (Result == 0) Result = new SCEVAddRecExpr(Operands, L);
1223 SCEVHandle ScalarEvolution::getSMaxExpr(const SCEVHandle &LHS,
1224 const SCEVHandle &RHS) {
1225 std::vector<SCEVHandle> Ops;
1228 return getSMaxExpr(Ops);
1231 SCEVHandle ScalarEvolution::getSMaxExpr(std::vector<SCEVHandle> Ops) {
1232 assert(!Ops.empty() && "Cannot get empty smax!");
1233 if (Ops.size() == 1) return Ops[0];
1235 // Sort by complexity, this groups all similar expression types together.
1236 GroupByComplexity(Ops);
1238 // If there are any constants, fold them together.
1240 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1242 assert(Idx < Ops.size());
1243 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1244 // We found two constants, fold them together!
1245 ConstantInt *Fold = ConstantInt::get(
1246 APIntOps::smax(LHSC->getValue()->getValue(),
1247 RHSC->getValue()->getValue()));
1248 Ops[0] = getConstant(Fold);
1249 Ops.erase(Ops.begin()+1); // Erase the folded element
1250 if (Ops.size() == 1) return Ops[0];
1251 LHSC = cast<SCEVConstant>(Ops[0]);
1254 // If we are left with a constant -inf, strip it off.
1255 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
1256 Ops.erase(Ops.begin());
1261 if (Ops.size() == 1) return Ops[0];
1263 // Find the first SMax
1264 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
1267 // Check to see if one of the operands is an SMax. If so, expand its operands
1268 // onto our operand list, and recurse to simplify.
1269 if (Idx < Ops.size()) {
1270 bool DeletedSMax = false;
1271 while (SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
1272 Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
1273 Ops.erase(Ops.begin()+Idx);
1278 return getSMaxExpr(Ops);
1281 // Okay, check to see if the same value occurs in the operand list twice. If
1282 // so, delete one. Since we sorted the list, these values are required to
1284 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1285 if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
1286 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1290 if (Ops.size() == 1) return Ops[0];
1292 assert(!Ops.empty() && "Reduced smax down to nothing!");
1294 // Okay, it looks like we really DO need an smax expr. Check to see if we
1295 // already have one, otherwise create a new one.
1296 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1297 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scSMaxExpr,
1299 if (Result == 0) Result = new SCEVSMaxExpr(Ops);
1303 SCEVHandle ScalarEvolution::getUMaxExpr(const SCEVHandle &LHS,
1304 const SCEVHandle &RHS) {
1305 std::vector<SCEVHandle> Ops;
1308 return getUMaxExpr(Ops);
1311 SCEVHandle ScalarEvolution::getUMaxExpr(std::vector<SCEVHandle> Ops) {
1312 assert(!Ops.empty() && "Cannot get empty umax!");
1313 if (Ops.size() == 1) return Ops[0];
1315 // Sort by complexity, this groups all similar expression types together.
1316 GroupByComplexity(Ops);
1318 // If there are any constants, fold them together.
1320 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1322 assert(Idx < Ops.size());
1323 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1324 // We found two constants, fold them together!
1325 ConstantInt *Fold = ConstantInt::get(
1326 APIntOps::umax(LHSC->getValue()->getValue(),
1327 RHSC->getValue()->getValue()));
1328 Ops[0] = getConstant(Fold);
1329 Ops.erase(Ops.begin()+1); // Erase the folded element
1330 if (Ops.size() == 1) return Ops[0];
1331 LHSC = cast<SCEVConstant>(Ops[0]);
1334 // If we are left with a constant zero, strip it off.
1335 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
1336 Ops.erase(Ops.begin());
1341 if (Ops.size() == 1) return Ops[0];
1343 // Find the first UMax
1344 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
1347 // Check to see if one of the operands is a UMax. If so, expand its operands
1348 // onto our operand list, and recurse to simplify.
1349 if (Idx < Ops.size()) {
1350 bool DeletedUMax = false;
1351 while (SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
1352 Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
1353 Ops.erase(Ops.begin()+Idx);
1358 return getUMaxExpr(Ops);
1361 // Okay, check to see if the same value occurs in the operand list twice. If
1362 // so, delete one. Since we sorted the list, these values are required to
1364 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1365 if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
1366 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1370 if (Ops.size() == 1) return Ops[0];
1372 assert(!Ops.empty() && "Reduced umax down to nothing!");
1374 // Okay, it looks like we really DO need a umax expr. Check to see if we
1375 // already have one, otherwise create a new one.
1376 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1377 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scUMaxExpr,
1379 if (Result == 0) Result = new SCEVUMaxExpr(Ops);
1383 SCEVHandle ScalarEvolution::getUnknown(Value *V) {
1384 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
1385 return getConstant(CI);
1386 SCEVUnknown *&Result = (*SCEVUnknowns)[V];
1387 if (Result == 0) Result = new SCEVUnknown(V);
1392 //===----------------------------------------------------------------------===//
1393 // ScalarEvolutionsImpl Definition and Implementation
1394 //===----------------------------------------------------------------------===//
1396 /// ScalarEvolutionsImpl - This class implements the main driver for the scalar
1400 struct VISIBILITY_HIDDEN ScalarEvolutionsImpl {
1401 /// SE - A reference to the public ScalarEvolution object.
1402 ScalarEvolution &SE;
1404 /// F - The function we are analyzing.
1408 /// LI - The loop information for the function we are currently analyzing.
1412 /// UnknownValue - This SCEV is used to represent unknown trip counts and
1414 SCEVHandle UnknownValue;
1416 /// Scalars - This is a cache of the scalars we have analyzed so far.
1418 std::map<Value*, SCEVHandle> Scalars;
1420 /// IterationCounts - Cache the iteration count of the loops for this
1421 /// function as they are computed.
1422 std::map<const Loop*, SCEVHandle> IterationCounts;
1424 /// ConstantEvolutionLoopExitValue - This map contains entries for all of
1425 /// the PHI instructions that we attempt to compute constant evolutions for.
1426 /// This allows us to avoid potentially expensive recomputation of these
1427 /// properties. An instruction maps to null if we are unable to compute its
1429 std::map<PHINode*, Constant*> ConstantEvolutionLoopExitValue;
1432 ScalarEvolutionsImpl(ScalarEvolution &se, Function &f, LoopInfo &li)
1433 : SE(se), F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {}
1435 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1436 /// expression and create a new one.
1437 SCEVHandle getSCEV(Value *V);
1439 /// hasSCEV - Return true if the SCEV for this value has already been
1441 bool hasSCEV(Value *V) const {
1442 return Scalars.count(V);
1445 /// setSCEV - Insert the specified SCEV into the map of current SCEVs for
1446 /// the specified value.
1447 void setSCEV(Value *V, const SCEVHandle &H) {
1448 bool isNew = Scalars.insert(std::make_pair(V, H)).second;
1449 assert(isNew && "This entry already existed!");
1454 /// getSCEVAtScope - Compute the value of the specified expression within
1455 /// the indicated loop (which may be null to indicate in no loop). If the
1456 /// expression cannot be evaluated, return UnknownValue itself.
1457 SCEVHandle getSCEVAtScope(SCEV *V, const Loop *L);
1460 /// hasLoopInvariantIterationCount - Return true if the specified loop has
1461 /// an analyzable loop-invariant iteration count.
1462 bool hasLoopInvariantIterationCount(const Loop *L);
1464 /// getIterationCount - If the specified loop has a predictable iteration
1465 /// count, return it. Note that it is not valid to call this method on a
1466 /// loop without a loop-invariant iteration count.
1467 SCEVHandle getIterationCount(const Loop *L);
1469 /// deleteValueFromRecords - This method should be called by the
1470 /// client before it removes a value from the program, to make sure
1471 /// that no dangling references are left around.
1472 void deleteValueFromRecords(Value *V);
1475 /// createSCEV - We know that there is no SCEV for the specified value.
1476 /// Analyze the expression.
1477 SCEVHandle createSCEV(Value *V);
1479 /// createNodeForPHI - Provide the special handling we need to analyze PHI
1481 SCEVHandle createNodeForPHI(PHINode *PN);
1483 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value
1484 /// for the specified instruction and replaces any references to the
1485 /// symbolic value SymName with the specified value. This is used during
1487 void ReplaceSymbolicValueWithConcrete(Instruction *I,
1488 const SCEVHandle &SymName,
1489 const SCEVHandle &NewVal);
1491 /// ComputeIterationCount - Compute the number of times the specified loop
1493 SCEVHandle ComputeIterationCount(const Loop *L);
1495 /// ComputeLoadConstantCompareIterationCount - Given an exit condition of
1496 /// 'icmp op load X, cst', try to see if we can compute the trip count.
1497 SCEVHandle ComputeLoadConstantCompareIterationCount(LoadInst *LI,
1500 ICmpInst::Predicate p);
1502 /// ComputeIterationCountExhaustively - If the trip is known to execute a
1503 /// constant number of times (the condition evolves only from constants),
1504 /// try to evaluate a few iterations of the loop until we get the exit
1505 /// condition gets a value of ExitWhen (true or false). If we cannot
1506 /// evaluate the trip count of the loop, return UnknownValue.
1507 SCEVHandle ComputeIterationCountExhaustively(const Loop *L, Value *Cond,
1510 /// HowFarToZero - Return the number of times a backedge comparing the
1511 /// specified value to zero will execute. If not computable, return
1513 SCEVHandle HowFarToZero(SCEV *V, const Loop *L);
1515 /// HowFarToNonZero - Return the number of times a backedge checking the
1516 /// specified value for nonzero will execute. If not computable, return
1518 SCEVHandle HowFarToNonZero(SCEV *V, const Loop *L);
1520 /// HowManyLessThans - Return the number of times a backedge containing the
1521 /// specified less-than comparison will execute. If not computable, return
1522 /// UnknownValue. isSigned specifies whether the less-than is signed.
1523 SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L,
1524 bool isSigned, bool trueWhenEqual);
1526 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
1527 /// (which may not be an immediate predecessor) which has exactly one
1528 /// successor from which BB is reachable, or null if no such block is
1530 BasicBlock* getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB);
1532 /// executesAtLeastOnce - Test whether entry to the loop is protected by
1533 /// a conditional between LHS and RHS.
1534 bool executesAtLeastOnce(const Loop *L, bool isSigned, bool trueWhenEqual,
1535 SCEV *LHS, SCEV *RHS);
1537 /// potentialInfiniteLoop - Test whether the loop might jump over the exit value
1538 /// due to wrapping.
1539 bool potentialInfiniteLoop(SCEV *Stride, SCEV *RHS, bool isSigned,
1540 bool trueWhenEqual);
1542 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
1543 /// in the header of its containing loop, we know the loop executes a
1544 /// constant number of times, and the PHI node is just a recurrence
1545 /// involving constants, fold it.
1546 Constant *getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& Its,
1551 //===----------------------------------------------------------------------===//
1552 // Basic SCEV Analysis and PHI Idiom Recognition Code
1555 /// deleteValueFromRecords - This method should be called by the
1556 /// client before it removes an instruction from the program, to make sure
1557 /// that no dangling references are left around.
1558 void ScalarEvolutionsImpl::deleteValueFromRecords(Value *V) {
1559 SmallVector<Value *, 16> Worklist;
1561 if (Scalars.erase(V)) {
1562 if (PHINode *PN = dyn_cast<PHINode>(V))
1563 ConstantEvolutionLoopExitValue.erase(PN);
1564 Worklist.push_back(V);
1567 while (!Worklist.empty()) {
1568 Value *VV = Worklist.back();
1569 Worklist.pop_back();
1571 for (Instruction::use_iterator UI = VV->use_begin(), UE = VV->use_end();
1573 Instruction *Inst = cast<Instruction>(*UI);
1574 if (Scalars.erase(Inst)) {
1575 if (PHINode *PN = dyn_cast<PHINode>(VV))
1576 ConstantEvolutionLoopExitValue.erase(PN);
1577 Worklist.push_back(Inst);
1584 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1585 /// expression and create a new one.
1586 SCEVHandle ScalarEvolutionsImpl::getSCEV(Value *V) {
1587 assert(V->getType() != Type::VoidTy && "Can't analyze void expressions!");
1589 std::map<Value*, SCEVHandle>::iterator I = Scalars.find(V);
1590 if (I != Scalars.end()) return I->second;
1591 SCEVHandle S = createSCEV(V);
1592 Scalars.insert(std::make_pair(V, S));
1596 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value for
1597 /// the specified instruction and replaces any references to the symbolic value
1598 /// SymName with the specified value. This is used during PHI resolution.
1599 void ScalarEvolutionsImpl::
1600 ReplaceSymbolicValueWithConcrete(Instruction *I, const SCEVHandle &SymName,
1601 const SCEVHandle &NewVal) {
1602 std::map<Value*, SCEVHandle>::iterator SI = Scalars.find(I);
1603 if (SI == Scalars.end()) return;
1606 SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal, SE);
1607 if (NV == SI->second) return; // No change.
1609 SI->second = NV; // Update the scalars map!
1611 // Any instruction values that use this instruction might also need to be
1613 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1615 ReplaceSymbolicValueWithConcrete(cast<Instruction>(*UI), SymName, NewVal);
1618 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
1619 /// a loop header, making it a potential recurrence, or it doesn't.
1621 SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) {
1622 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
1623 if (const Loop *L = LI.getLoopFor(PN->getParent()))
1624 if (L->getHeader() == PN->getParent()) {
1625 // If it lives in the loop header, it has two incoming values, one
1626 // from outside the loop, and one from inside.
1627 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1628 unsigned BackEdge = IncomingEdge^1;
1630 // While we are analyzing this PHI node, handle its value symbolically.
1631 SCEVHandle SymbolicName = SE.getUnknown(PN);
1632 assert(Scalars.find(PN) == Scalars.end() &&
1633 "PHI node already processed?");
1634 Scalars.insert(std::make_pair(PN, SymbolicName));
1636 // Using this symbolic name for the PHI, analyze the value coming around
1638 SCEVHandle BEValue = getSCEV(PN->getIncomingValue(BackEdge));
1640 // NOTE: If BEValue is loop invariant, we know that the PHI node just
1641 // has a special value for the first iteration of the loop.
1643 // If the value coming around the backedge is an add with the symbolic
1644 // value we just inserted, then we found a simple induction variable!
1645 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
1646 // If there is a single occurrence of the symbolic value, replace it
1647 // with a recurrence.
1648 unsigned FoundIndex = Add->getNumOperands();
1649 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1650 if (Add->getOperand(i) == SymbolicName)
1651 if (FoundIndex == e) {
1656 if (FoundIndex != Add->getNumOperands()) {
1657 // Create an add with everything but the specified operand.
1658 std::vector<SCEVHandle> Ops;
1659 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1660 if (i != FoundIndex)
1661 Ops.push_back(Add->getOperand(i));
1662 SCEVHandle Accum = SE.getAddExpr(Ops);
1664 // This is not a valid addrec if the step amount is varying each
1665 // loop iteration, but is not itself an addrec in this loop.
1666 if (Accum->isLoopInvariant(L) ||
1667 (isa<SCEVAddRecExpr>(Accum) &&
1668 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
1669 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1670 SCEVHandle PHISCEV = SE.getAddRecExpr(StartVal, Accum, L);
1672 // Okay, for the entire analysis of this edge we assumed the PHI
1673 // to be symbolic. We now need to go back and update all of the
1674 // entries for the scalars that use the PHI (except for the PHI
1675 // itself) to use the new analyzed value instead of the "symbolic"
1677 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1681 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(BEValue)) {
1682 // Otherwise, this could be a loop like this:
1683 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
1684 // In this case, j = {1,+,1} and BEValue is j.
1685 // Because the other in-value of i (0) fits the evolution of BEValue
1686 // i really is an addrec evolution.
1687 if (AddRec->getLoop() == L && AddRec->isAffine()) {
1688 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1690 // If StartVal = j.start - j.stride, we can use StartVal as the
1691 // initial step of the addrec evolution.
1692 if (StartVal == SE.getMinusSCEV(AddRec->getOperand(0),
1693 AddRec->getOperand(1))) {
1694 SCEVHandle PHISCEV =
1695 SE.getAddRecExpr(StartVal, AddRec->getOperand(1), L);
1697 // Okay, for the entire analysis of this edge we assumed the PHI
1698 // to be symbolic. We now need to go back and update all of the
1699 // entries for the scalars that use the PHI (except for the PHI
1700 // itself) to use the new analyzed value instead of the "symbolic"
1702 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1708 return SymbolicName;
1711 // If it's not a loop phi, we can't handle it yet.
1712 return SE.getUnknown(PN);
1715 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
1716 /// guaranteed to end in (at every loop iteration). It is, at the same time,
1717 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
1718 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
1719 static uint32_t GetMinTrailingZeros(SCEVHandle S) {
1720 if (SCEVConstant *C = dyn_cast<SCEVConstant>(S))
1721 return C->getValue()->getValue().countTrailingZeros();
1723 if (SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
1724 return std::min(GetMinTrailingZeros(T->getOperand()), T->getBitWidth());
1726 if (SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
1727 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
1728 return OpRes == E->getOperand()->getBitWidth() ? E->getBitWidth() : OpRes;
1731 if (SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
1732 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
1733 return OpRes == E->getOperand()->getBitWidth() ? E->getBitWidth() : OpRes;
1736 if (SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
1737 // The result is the min of all operands results.
1738 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
1739 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
1740 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
1744 if (SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
1745 // The result is the sum of all operands results.
1746 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
1747 uint32_t BitWidth = M->getBitWidth();
1748 for (unsigned i = 1, e = M->getNumOperands();
1749 SumOpRes != BitWidth && i != e; ++i)
1750 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
1755 if (SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
1756 // The result is the min of all operands results.
1757 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
1758 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
1759 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
1763 if (SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
1764 // The result is the min of all operands results.
1765 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
1766 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
1767 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
1771 if (SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
1772 // The result is the min of all operands results.
1773 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
1774 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
1775 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
1779 // SCEVUDivExpr, SCEVSDivExpr, SCEVUnknown
1783 /// createSCEV - We know that there is no SCEV for the specified value.
1784 /// Analyze the expression.
1786 SCEVHandle ScalarEvolutionsImpl::createSCEV(Value *V) {
1787 if (!isa<IntegerType>(V->getType()))
1788 return SE.getUnknown(V);
1790 unsigned Opcode = Instruction::UserOp1;
1791 if (Instruction *I = dyn_cast<Instruction>(V))
1792 Opcode = I->getOpcode();
1793 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1794 Opcode = CE->getOpcode();
1796 return SE.getUnknown(V);
1798 User *U = cast<User>(V);
1800 case Instruction::Add:
1801 return SE.getAddExpr(getSCEV(U->getOperand(0)),
1802 getSCEV(U->getOperand(1)));
1803 case Instruction::Mul:
1804 return SE.getMulExpr(getSCEV(U->getOperand(0)),
1805 getSCEV(U->getOperand(1)));
1806 case Instruction::UDiv:
1807 return SE.getUDivExpr(getSCEV(U->getOperand(0)),
1808 getSCEV(U->getOperand(1)));
1809 case Instruction::SDiv:
1810 return SE.getSDivExpr(getSCEV(U->getOperand(0)),
1811 getSCEV(U->getOperand(1)));
1812 case Instruction::Sub:
1813 return SE.getMinusSCEV(getSCEV(U->getOperand(0)),
1814 getSCEV(U->getOperand(1)));
1815 case Instruction::Or:
1816 // If the RHS of the Or is a constant, we may have something like:
1817 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
1818 // optimizations will transparently handle this case.
1820 // In order for this transformation to be safe, the LHS must be of the
1821 // form X*(2^n) and the Or constant must be less than 2^n.
1822 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
1823 SCEVHandle LHS = getSCEV(U->getOperand(0));
1824 const APInt &CIVal = CI->getValue();
1825 if (GetMinTrailingZeros(LHS) >=
1826 (CIVal.getBitWidth() - CIVal.countLeadingZeros()))
1827 return SE.getAddExpr(LHS, getSCEV(U->getOperand(1)));
1830 case Instruction::Xor:
1831 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
1832 // If the RHS of the xor is a signbit, then this is just an add.
1833 // Instcombine turns add of signbit into xor as a strength reduction step.
1834 if (CI->getValue().isSignBit())
1835 return SE.getAddExpr(getSCEV(U->getOperand(0)),
1836 getSCEV(U->getOperand(1)));
1838 // If the RHS of xor is -1, then this is a not operation.
1839 else if (CI->isAllOnesValue())
1840 return SE.getNotSCEV(getSCEV(U->getOperand(0)));
1844 case Instruction::Shl:
1845 // Turn shift left of a constant amount into a multiply.
1846 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
1847 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
1848 Constant *X = ConstantInt::get(
1849 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
1850 return SE.getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
1854 case Instruction::LShr:
1855 // Turn logical shift right of a constant into an unsigned divide.
1856 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
1857 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
1858 Constant *X = ConstantInt::get(
1859 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
1860 return SE.getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
1864 case Instruction::Trunc:
1865 return SE.getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
1867 case Instruction::ZExt:
1868 return SE.getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
1870 case Instruction::SExt:
1871 return SE.getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
1873 case Instruction::BitCast:
1874 // BitCasts are no-op casts so we just eliminate the cast.
1875 if (U->getType()->isInteger() &&
1876 U->getOperand(0)->getType()->isInteger())
1877 return getSCEV(U->getOperand(0));
1880 case Instruction::PHI:
1881 return createNodeForPHI(cast<PHINode>(U));
1883 case Instruction::Select:
1884 // This could be a smax or umax that was lowered earlier.
1885 // Try to recover it.
1886 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
1887 Value *LHS = ICI->getOperand(0);
1888 Value *RHS = ICI->getOperand(1);
1889 switch (ICI->getPredicate()) {
1890 case ICmpInst::ICMP_SLT:
1891 case ICmpInst::ICMP_SLE:
1892 std::swap(LHS, RHS);
1894 case ICmpInst::ICMP_SGT:
1895 case ICmpInst::ICMP_SGE:
1896 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
1897 return SE.getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
1898 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
1899 // ~smax(~x, ~y) == smin(x, y).
1900 return SE.getNotSCEV(SE.getSMaxExpr(
1901 SE.getNotSCEV(getSCEV(LHS)),
1902 SE.getNotSCEV(getSCEV(RHS))));
1904 case ICmpInst::ICMP_ULT:
1905 case ICmpInst::ICMP_ULE:
1906 std::swap(LHS, RHS);
1908 case ICmpInst::ICMP_UGT:
1909 case ICmpInst::ICMP_UGE:
1910 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
1911 return SE.getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
1912 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
1913 // ~umax(~x, ~y) == umin(x, y)
1914 return SE.getNotSCEV(SE.getUMaxExpr(SE.getNotSCEV(getSCEV(LHS)),
1915 SE.getNotSCEV(getSCEV(RHS))));
1922 default: // We cannot analyze this expression.
1926 return SE.getUnknown(V);
1931 //===----------------------------------------------------------------------===//
1932 // Iteration Count Computation Code
1935 /// getIterationCount - If the specified loop has a predictable iteration
1936 /// count, return it. Note that it is not valid to call this method on a
1937 /// loop without a loop-invariant iteration count.
1938 SCEVHandle ScalarEvolutionsImpl::getIterationCount(const Loop *L) {
1939 std::map<const Loop*, SCEVHandle>::iterator I = IterationCounts.find(L);
1940 if (I == IterationCounts.end()) {
1941 SCEVHandle ItCount = ComputeIterationCount(L);
1942 I = IterationCounts.insert(std::make_pair(L, ItCount)).first;
1943 if (ItCount != UnknownValue) {
1944 assert(ItCount->isLoopInvariant(L) &&
1945 "Computed trip count isn't loop invariant for loop!");
1946 ++NumTripCountsComputed;
1947 } else if (isa<PHINode>(L->getHeader()->begin())) {
1948 // Only count loops that have phi nodes as not being computable.
1949 ++NumTripCountsNotComputed;
1955 /// ComputeIterationCount - Compute the number of times the specified loop
1957 SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) {
1958 // If the loop has a non-one exit block count, we can't analyze it.
1959 SmallVector<BasicBlock*, 8> ExitBlocks;
1960 L->getExitBlocks(ExitBlocks);
1961 if (ExitBlocks.size() != 1) return UnknownValue;
1963 // Okay, there is one exit block. Try to find the condition that causes the
1964 // loop to be exited.
1965 BasicBlock *ExitBlock = ExitBlocks[0];
1967 BasicBlock *ExitingBlock = 0;
1968 for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock);
1970 if (L->contains(*PI)) {
1971 if (ExitingBlock == 0)
1974 return UnknownValue; // More than one block exiting!
1976 assert(ExitingBlock && "No exits from loop, something is broken!");
1978 // Okay, we've computed the exiting block. See what condition causes us to
1981 // FIXME: we should be able to handle switch instructions (with a single exit)
1982 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1983 if (ExitBr == 0) return UnknownValue;
1984 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
1986 // At this point, we know we have a conditional branch that determines whether
1987 // the loop is exited. However, we don't know if the branch is executed each
1988 // time through the loop. If not, then the execution count of the branch will
1989 // not be equal to the trip count of the loop.
1991 // Currently we check for this by checking to see if the Exit branch goes to
1992 // the loop header. If so, we know it will always execute the same number of
1993 // times as the loop. We also handle the case where the exit block *is* the
1994 // loop header. This is common for un-rotated loops. More extensive analysis
1995 // could be done to handle more cases here.
1996 if (ExitBr->getSuccessor(0) != L->getHeader() &&
1997 ExitBr->getSuccessor(1) != L->getHeader() &&
1998 ExitBr->getParent() != L->getHeader())
1999 return UnknownValue;
2001 ICmpInst *ExitCond = dyn_cast<ICmpInst>(ExitBr->getCondition());
2003 // If it's not an integer comparison then compute it the hard way.
2004 // Note that ICmpInst deals with pointer comparisons too so we must check
2005 // the type of the operand.
2006 if (ExitCond == 0 || isa<PointerType>(ExitCond->getOperand(0)->getType()))
2007 return ComputeIterationCountExhaustively(L, ExitBr->getCondition(),
2008 ExitBr->getSuccessor(0) == ExitBlock);
2010 // If the condition was exit on true, convert the condition to exit on false
2011 ICmpInst::Predicate Cond;
2012 if (ExitBr->getSuccessor(1) == ExitBlock)
2013 Cond = ExitCond->getPredicate();
2015 Cond = ExitCond->getInversePredicate();
2017 // Handle common loops like: for (X = "string"; *X; ++X)
2018 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
2019 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
2021 ComputeLoadConstantCompareIterationCount(LI, RHS, L, Cond);
2022 if (!isa<SCEVCouldNotCompute>(ItCnt)) return ItCnt;
2025 SCEVHandle LHS = getSCEV(ExitCond->getOperand(0));
2026 SCEVHandle RHS = getSCEV(ExitCond->getOperand(1));
2028 // Try to evaluate any dependencies out of the loop.
2029 SCEVHandle Tmp = getSCEVAtScope(LHS, L);
2030 if (!isa<SCEVCouldNotCompute>(Tmp)) LHS = Tmp;
2031 Tmp = getSCEVAtScope(RHS, L);
2032 if (!isa<SCEVCouldNotCompute>(Tmp)) RHS = Tmp;
2034 // At this point, we would like to compute how many iterations of the
2035 // loop the predicate will return true for these inputs.
2036 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
2037 // If there is a loop-invariant, force it into the RHS.
2038 std::swap(LHS, RHS);
2039 Cond = ICmpInst::getSwappedPredicate(Cond);
2042 // FIXME: think about handling pointer comparisons! i.e.:
2043 // while (P != P+100) ++P;
2045 // If we have a comparison of a chrec against a constant, try to use value
2046 // ranges to answer this query.
2047 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
2048 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
2049 if (AddRec->getLoop() == L) {
2050 // Form the comparison range using the constant of the correct type so
2051 // that the ConstantRange class knows to do a signed or unsigned
2053 ConstantInt *CompVal = RHSC->getValue();
2054 const Type *RealTy = ExitCond->getOperand(0)->getType();
2055 CompVal = dyn_cast<ConstantInt>(
2056 ConstantExpr::getBitCast(CompVal, RealTy));
2058 // Form the constant range.
2059 ConstantRange CompRange(
2060 ICmpInst::makeConstantRange(Cond, CompVal->getValue()));
2062 SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange, SE);
2063 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
2068 case ICmpInst::ICMP_NE: { // while (X != Y)
2069 // Convert to: while (X-Y != 0)
2070 SCEVHandle TC = HowFarToZero(SE.getMinusSCEV(LHS, RHS), L);
2071 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2074 case ICmpInst::ICMP_EQ: {
2075 // Convert to: while (X-Y == 0) // while (X == Y)
2076 SCEVHandle TC = HowFarToNonZero(SE.getMinusSCEV(LHS, RHS), L);
2077 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2080 case ICmpInst::ICMP_SLT: {
2081 SCEVHandle TC = HowManyLessThans(LHS, RHS, L, true, false);
2082 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2085 case ICmpInst::ICMP_SGT: {
2086 SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
2087 SE.getNotSCEV(RHS), L, true, false);
2088 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2091 case ICmpInst::ICMP_ULT: {
2092 SCEVHandle TC = HowManyLessThans(LHS, RHS, L, false, false);
2093 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2096 case ICmpInst::ICMP_UGT: {
2097 SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
2098 SE.getNotSCEV(RHS), L, false, false);
2099 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2102 case ICmpInst::ICMP_SLE: {
2103 SCEVHandle TC = HowManyLessThans(LHS, RHS, L, true, true);
2104 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2107 case ICmpInst::ICMP_SGE: {
2108 SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
2109 SE.getNotSCEV(RHS), L, true, true);
2110 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2113 case ICmpInst::ICMP_ULE: {
2114 SCEVHandle TC = HowManyLessThans(LHS, RHS, L, false, true);
2115 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2118 case ICmpInst::ICMP_UGE: {
2119 SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
2120 SE.getNotSCEV(RHS), L, false, true);
2121 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2126 cerr << "ComputeIterationCount ";
2127 if (ExitCond->getOperand(0)->getType()->isUnsigned())
2128 cerr << "[unsigned] ";
2130 << Instruction::getOpcodeName(Instruction::ICmp)
2131 << " " << *RHS << "\n";
2135 return ComputeIterationCountExhaustively(L, ExitCond,
2136 ExitBr->getSuccessor(0) == ExitBlock);
2139 static ConstantInt *
2140 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
2141 ScalarEvolution &SE) {
2142 SCEVHandle InVal = SE.getConstant(C);
2143 SCEVHandle Val = AddRec->evaluateAtIteration(InVal, SE);
2144 assert(isa<SCEVConstant>(Val) &&
2145 "Evaluation of SCEV at constant didn't fold correctly?");
2146 return cast<SCEVConstant>(Val)->getValue();
2149 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
2150 /// and a GEP expression (missing the pointer index) indexing into it, return
2151 /// the addressed element of the initializer or null if the index expression is
2154 GetAddressedElementFromGlobal(GlobalVariable *GV,
2155 const std::vector<ConstantInt*> &Indices) {
2156 Constant *Init = GV->getInitializer();
2157 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
2158 uint64_t Idx = Indices[i]->getZExtValue();
2159 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2160 assert(Idx < CS->getNumOperands() && "Bad struct index!");
2161 Init = cast<Constant>(CS->getOperand(Idx));
2162 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2163 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
2164 Init = cast<Constant>(CA->getOperand(Idx));
2165 } else if (isa<ConstantAggregateZero>(Init)) {
2166 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2167 assert(Idx < STy->getNumElements() && "Bad struct index!");
2168 Init = Constant::getNullValue(STy->getElementType(Idx));
2169 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
2170 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
2171 Init = Constant::getNullValue(ATy->getElementType());
2173 assert(0 && "Unknown constant aggregate type!");
2177 return 0; // Unknown initializer type
2183 /// ComputeLoadConstantCompareIterationCount - Given an exit condition of
2184 /// 'icmp op load X, cst', try to see if we can compute the trip count.
2185 SCEVHandle ScalarEvolutionsImpl::
2186 ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS,
2188 ICmpInst::Predicate predicate) {
2189 if (LI->isVolatile()) return UnknownValue;
2191 // Check to see if the loaded pointer is a getelementptr of a global.
2192 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
2193 if (!GEP) return UnknownValue;
2195 // Make sure that it is really a constant global we are gepping, with an
2196 // initializer, and make sure the first IDX is really 0.
2197 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
2198 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
2199 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
2200 !cast<Constant>(GEP->getOperand(1))->isNullValue())
2201 return UnknownValue;
2203 // Okay, we allow one non-constant index into the GEP instruction.
2205 std::vector<ConstantInt*> Indexes;
2206 unsigned VarIdxNum = 0;
2207 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
2208 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
2209 Indexes.push_back(CI);
2210 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
2211 if (VarIdx) return UnknownValue; // Multiple non-constant idx's.
2212 VarIdx = GEP->getOperand(i);
2214 Indexes.push_back(0);
2217 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
2218 // Check to see if X is a loop variant variable value now.
2219 SCEVHandle Idx = getSCEV(VarIdx);
2220 SCEVHandle Tmp = getSCEVAtScope(Idx, L);
2221 if (!isa<SCEVCouldNotCompute>(Tmp)) Idx = Tmp;
2223 // We can only recognize very limited forms of loop index expressions, in
2224 // particular, only affine AddRec's like {C1,+,C2}.
2225 SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
2226 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
2227 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
2228 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
2229 return UnknownValue;
2231 unsigned MaxSteps = MaxBruteForceIterations;
2232 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
2233 ConstantInt *ItCst =
2234 ConstantInt::get(IdxExpr->getType(), IterationNum);
2235 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, SE);
2237 // Form the GEP offset.
2238 Indexes[VarIdxNum] = Val;
2240 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
2241 if (Result == 0) break; // Cannot compute!
2243 // Evaluate the condition for this iteration.
2244 Result = ConstantExpr::getICmp(predicate, Result, RHS);
2245 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
2246 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
2248 cerr << "\n***\n*** Computed loop count " << *ItCst
2249 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
2252 ++NumArrayLenItCounts;
2253 return SE.getConstant(ItCst); // Found terminating iteration!
2256 return UnknownValue;
2260 /// CanConstantFold - Return true if we can constant fold an instruction of the
2261 /// specified type, assuming that all operands were constants.
2262 static bool CanConstantFold(const Instruction *I) {
2263 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
2264 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
2267 if (const CallInst *CI = dyn_cast<CallInst>(I))
2268 if (const Function *F = CI->getCalledFunction())
2269 return canConstantFoldCallTo(F);
2273 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
2274 /// in the loop that V is derived from. We allow arbitrary operations along the
2275 /// way, but the operands of an operation must either be constants or a value
2276 /// derived from a constant PHI. If this expression does not fit with these
2277 /// constraints, return null.
2278 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
2279 // If this is not an instruction, or if this is an instruction outside of the
2280 // loop, it can't be derived from a loop PHI.
2281 Instruction *I = dyn_cast<Instruction>(V);
2282 if (I == 0 || !L->contains(I->getParent())) return 0;
2284 if (PHINode *PN = dyn_cast<PHINode>(I)) {
2285 if (L->getHeader() == I->getParent())
2288 // We don't currently keep track of the control flow needed to evaluate
2289 // PHIs, so we cannot handle PHIs inside of loops.
2293 // If we won't be able to constant fold this expression even if the operands
2294 // are constants, return early.
2295 if (!CanConstantFold(I)) return 0;
2297 // Otherwise, we can evaluate this instruction if all of its operands are
2298 // constant or derived from a PHI node themselves.
2300 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
2301 if (!(isa<Constant>(I->getOperand(Op)) ||
2302 isa<GlobalValue>(I->getOperand(Op)))) {
2303 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
2304 if (P == 0) return 0; // Not evolving from PHI
2308 return 0; // Evolving from multiple different PHIs.
2311 // This is a expression evolving from a constant PHI!
2315 /// EvaluateExpression - Given an expression that passes the
2316 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
2317 /// in the loop has the value PHIVal. If we can't fold this expression for some
2318 /// reason, return null.
2319 static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
2320 if (isa<PHINode>(V)) return PHIVal;
2321 if (Constant *C = dyn_cast<Constant>(V)) return C;
2322 Instruction *I = cast<Instruction>(V);
2324 std::vector<Constant*> Operands;
2325 Operands.resize(I->getNumOperands());
2327 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2328 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal);
2329 if (Operands[i] == 0) return 0;
2332 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
2333 return ConstantFoldCompareInstOperands(CI->getPredicate(),
2334 &Operands[0], Operands.size());
2336 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
2337 &Operands[0], Operands.size());
2340 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
2341 /// in the header of its containing loop, we know the loop executes a
2342 /// constant number of times, and the PHI node is just a recurrence
2343 /// involving constants, fold it.
2344 Constant *ScalarEvolutionsImpl::
2345 getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& Its, const Loop *L){
2346 std::map<PHINode*, Constant*>::iterator I =
2347 ConstantEvolutionLoopExitValue.find(PN);
2348 if (I != ConstantEvolutionLoopExitValue.end())
2351 if (Its.ugt(APInt(Its.getBitWidth(),MaxBruteForceIterations)))
2352 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
2354 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
2356 // Since the loop is canonicalized, the PHI node must have two entries. One
2357 // entry must be a constant (coming in from outside of the loop), and the
2358 // second must be derived from the same PHI.
2359 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
2360 Constant *StartCST =
2361 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
2363 return RetVal = 0; // Must be a constant.
2365 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
2366 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
2368 return RetVal = 0; // Not derived from same PHI.
2370 // Execute the loop symbolically to determine the exit value.
2371 if (Its.getActiveBits() >= 32)
2372 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
2374 unsigned NumIterations = Its.getZExtValue(); // must be in range
2375 unsigned IterationNum = 0;
2376 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
2377 if (IterationNum == NumIterations)
2378 return RetVal = PHIVal; // Got exit value!
2380 // Compute the value of the PHI node for the next iteration.
2381 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
2382 if (NextPHI == PHIVal)
2383 return RetVal = NextPHI; // Stopped evolving!
2385 return 0; // Couldn't evaluate!
2390 /// ComputeIterationCountExhaustively - If the trip is known to execute a
2391 /// constant number of times (the condition evolves only from constants),
2392 /// try to evaluate a few iterations of the loop until we get the exit
2393 /// condition gets a value of ExitWhen (true or false). If we cannot
2394 /// evaluate the trip count of the loop, return UnknownValue.
2395 SCEVHandle ScalarEvolutionsImpl::
2396 ComputeIterationCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) {
2397 PHINode *PN = getConstantEvolvingPHI(Cond, L);
2398 if (PN == 0) return UnknownValue;
2400 // Since the loop is canonicalized, the PHI node must have two entries. One
2401 // entry must be a constant (coming in from outside of the loop), and the
2402 // second must be derived from the same PHI.
2403 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
2404 Constant *StartCST =
2405 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
2406 if (StartCST == 0) return UnknownValue; // Must be a constant.
2408 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
2409 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
2410 if (PN2 != PN) return UnknownValue; // Not derived from same PHI.
2412 // Okay, we find a PHI node that defines the trip count of this loop. Execute
2413 // the loop symbolically to determine when the condition gets a value of
2415 unsigned IterationNum = 0;
2416 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
2417 for (Constant *PHIVal = StartCST;
2418 IterationNum != MaxIterations; ++IterationNum) {
2419 ConstantInt *CondVal =
2420 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal));
2422 // Couldn't symbolically evaluate.
2423 if (!CondVal) return UnknownValue;
2425 if (CondVal->getValue() == uint64_t(ExitWhen)) {
2426 ConstantEvolutionLoopExitValue[PN] = PHIVal;
2427 ++NumBruteForceTripCountsComputed;
2428 return SE.getConstant(ConstantInt::get(Type::Int32Ty, IterationNum));
2431 // Compute the value of the PHI node for the next iteration.
2432 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
2433 if (NextPHI == 0 || NextPHI == PHIVal)
2434 return UnknownValue; // Couldn't evaluate or not making progress...
2438 // Too many iterations were needed to evaluate.
2439 return UnknownValue;
2442 /// getSCEVAtScope - Compute the value of the specified expression within the
2443 /// indicated loop (which may be null to indicate in no loop). If the
2444 /// expression cannot be evaluated, return UnknownValue.
2445 SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) {
2446 // FIXME: this should be turned into a virtual method on SCEV!
2448 if (isa<SCEVConstant>(V)) return V;
2450 // If this instruction is evolved from a constant-evolving PHI, compute the
2451 // exit value from the loop without using SCEVs.
2452 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
2453 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
2454 const Loop *LI = this->LI[I->getParent()];
2455 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
2456 if (PHINode *PN = dyn_cast<PHINode>(I))
2457 if (PN->getParent() == LI->getHeader()) {
2458 // Okay, there is no closed form solution for the PHI node. Check
2459 // to see if the loop that contains it has a known iteration count.
2460 // If so, we may be able to force computation of the exit value.
2461 SCEVHandle IterationCount = getIterationCount(LI);
2462 if (SCEVConstant *ICC = dyn_cast<SCEVConstant>(IterationCount)) {
2463 // Okay, we know how many times the containing loop executes. If
2464 // this is a constant evolving PHI node, get the final value at
2465 // the specified iteration number.
2466 Constant *RV = getConstantEvolutionLoopExitValue(PN,
2467 ICC->getValue()->getValue(),
2469 if (RV) return SE.getUnknown(RV);
2473 // Okay, this is an expression that we cannot symbolically evaluate
2474 // into a SCEV. Check to see if it's possible to symbolically evaluate
2475 // the arguments into constants, and if so, try to constant propagate the
2476 // result. This is particularly useful for computing loop exit values.
2477 if (CanConstantFold(I)) {
2478 std::vector<Constant*> Operands;
2479 Operands.reserve(I->getNumOperands());
2480 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2481 Value *Op = I->getOperand(i);
2482 if (Constant *C = dyn_cast<Constant>(Op)) {
2483 Operands.push_back(C);
2485 // If any of the operands is non-constant and if they are
2486 // non-integer, don't even try to analyze them with scev techniques.
2487 if (!isa<IntegerType>(Op->getType()))
2490 SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L);
2491 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
2492 Operands.push_back(ConstantExpr::getIntegerCast(SC->getValue(),
2495 else if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
2496 if (Constant *C = dyn_cast<Constant>(SU->getValue()))
2497 Operands.push_back(ConstantExpr::getIntegerCast(C,
2509 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
2510 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
2511 &Operands[0], Operands.size());
2513 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
2514 &Operands[0], Operands.size());
2515 return SE.getUnknown(C);
2519 // This is some other type of SCEVUnknown, just return it.
2523 if (SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
2524 // Avoid performing the look-up in the common case where the specified
2525 // expression has no loop-variant portions.
2526 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
2527 SCEVHandle OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
2528 if (OpAtScope != Comm->getOperand(i)) {
2529 if (OpAtScope == UnknownValue) return UnknownValue;
2530 // Okay, at least one of these operands is loop variant but might be
2531 // foldable. Build a new instance of the folded commutative expression.
2532 std::vector<SCEVHandle> NewOps(Comm->op_begin(), Comm->op_begin()+i);
2533 NewOps.push_back(OpAtScope);
2535 for (++i; i != e; ++i) {
2536 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
2537 if (OpAtScope == UnknownValue) return UnknownValue;
2538 NewOps.push_back(OpAtScope);
2540 if (isa<SCEVAddExpr>(Comm))
2541 return SE.getAddExpr(NewOps);
2542 if (isa<SCEVMulExpr>(Comm))
2543 return SE.getMulExpr(NewOps);
2544 if (isa<SCEVSMaxExpr>(Comm))
2545 return SE.getSMaxExpr(NewOps);
2546 if (isa<SCEVUMaxExpr>(Comm))
2547 return SE.getUMaxExpr(NewOps);
2548 assert(0 && "Unknown commutative SCEV type!");
2551 // If we got here, all operands are loop invariant.
2555 if (SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(V)) {
2556 SCEVHandle LHS = getSCEVAtScope(UDiv->getLHS(), L);
2557 if (LHS == UnknownValue) return LHS;
2558 SCEVHandle RHS = getSCEVAtScope(UDiv->getRHS(), L);
2559 if (RHS == UnknownValue) return RHS;
2560 if (LHS == UDiv->getLHS() && RHS == UDiv->getRHS())
2561 return UDiv; // must be loop invariant
2562 return SE.getUDivExpr(LHS, RHS);
2565 if (SCEVSDivExpr *SDiv = dyn_cast<SCEVSDivExpr>(V)) {
2566 SCEVHandle LHS = getSCEVAtScope(SDiv->getLHS(), L);
2567 if (LHS == UnknownValue) return LHS;
2568 SCEVHandle RHS = getSCEVAtScope(SDiv->getRHS(), L);
2569 if (RHS == UnknownValue) return RHS;
2570 if (LHS == SDiv->getLHS() && RHS == SDiv->getRHS())
2571 return SDiv; // must be loop invariant
2572 return SE.getSDivExpr(LHS, RHS);
2575 // If this is a loop recurrence for a loop that does not contain L, then we
2576 // are dealing with the final value computed by the loop.
2577 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
2578 if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
2579 // To evaluate this recurrence, we need to know how many times the AddRec
2580 // loop iterates. Compute this now.
2581 SCEVHandle IterationCount = getIterationCount(AddRec->getLoop());
2582 if (IterationCount == UnknownValue) return UnknownValue;
2584 // Then, evaluate the AddRec.
2585 return AddRec->evaluateAtIteration(IterationCount, SE);
2587 return UnknownValue;
2590 //assert(0 && "Unknown SCEV type!");
2591 return UnknownValue;
2594 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
2595 /// following equation:
2597 /// A * X = B (mod N)
2599 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
2600 /// A and B isn't important.
2602 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
2603 static SCEVHandle SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
2604 ScalarEvolution &SE) {
2605 uint32_t BW = A.getBitWidth();
2606 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
2607 assert(A != 0 && "A must be non-zero.");
2611 // The gcd of A and N may have only one prime factor: 2. The number of
2612 // trailing zeros in A is its multiplicity
2613 uint32_t Mult2 = A.countTrailingZeros();
2616 // 2. Check if B is divisible by D.
2618 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
2619 // is not less than multiplicity of this prime factor for D.
2620 if (B.countTrailingZeros() < Mult2)
2621 return new SCEVCouldNotCompute();
2623 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
2626 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
2627 // bit width during computations.
2628 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
2629 APInt Mod(BW + 1, 0);
2630 Mod.set(BW - Mult2); // Mod = N / D
2631 APInt I = AD.multiplicativeInverse(Mod);
2633 // 4. Compute the minimum unsigned root of the equation:
2634 // I * (B / D) mod (N / D)
2635 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
2637 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
2639 return SE.getConstant(Result.trunc(BW));
2642 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
2643 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
2644 /// might be the same) or two SCEVCouldNotCompute objects.
2646 static std::pair<SCEVHandle,SCEVHandle>
2647 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
2648 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
2649 SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
2650 SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
2651 SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
2653 // We currently can only solve this if the coefficients are constants.
2654 if (!LC || !MC || !NC) {
2655 SCEV *CNC = new SCEVCouldNotCompute();
2656 return std::make_pair(CNC, CNC);
2659 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
2660 const APInt &L = LC->getValue()->getValue();
2661 const APInt &M = MC->getValue()->getValue();
2662 const APInt &N = NC->getValue()->getValue();
2663 APInt Two(BitWidth, 2);
2664 APInt Four(BitWidth, 4);
2667 using namespace APIntOps;
2669 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
2670 // The B coefficient is M-N/2
2674 // The A coefficient is N/2
2675 APInt A(N.sdiv(Two));
2677 // Compute the B^2-4ac term.
2680 SqrtTerm -= Four * (A * C);
2682 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
2683 // integer value or else APInt::sqrt() will assert.
2684 APInt SqrtVal(SqrtTerm.sqrt());
2686 // Compute the two solutions for the quadratic formula.
2687 // The divisions must be performed as signed divisions.
2689 APInt TwoA( A << 1 );
2690 if (TwoA.isMinValue()) {
2691 SCEV *CNC = new SCEVCouldNotCompute();
2692 return std::make_pair(CNC, CNC);
2695 ConstantInt *Solution1 = ConstantInt::get((NegB + SqrtVal).sdiv(TwoA));
2696 ConstantInt *Solution2 = ConstantInt::get((NegB - SqrtVal).sdiv(TwoA));
2698 return std::make_pair(SE.getConstant(Solution1),
2699 SE.getConstant(Solution2));
2700 } // end APIntOps namespace
2703 /// HowFarToZero - Return the number of times a backedge comparing the specified
2704 /// value to zero will execute. If not computable, return UnknownValue
2705 SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) {
2706 // If the value is a constant
2707 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2708 // If the value is already zero, the branch will execute zero times.
2709 if (C->getValue()->isZero()) return C;
2710 return UnknownValue; // Otherwise it will loop infinitely.
2713 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
2714 if (!AddRec || AddRec->getLoop() != L)
2715 return UnknownValue;
2717 if (AddRec->isAffine()) {
2718 // If this is an affine expression, the execution count of this branch is
2719 // the minimum unsigned root of the following equation:
2721 // Start + Step*N = 0 (mod 2^BW)
2725 // Step*N = -Start (mod 2^BW)
2727 // where BW is the common bit width of Start and Step.
2729 // Get the initial value for the loop.
2730 SCEVHandle Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
2731 if (isa<SCEVCouldNotCompute>(Start)) return UnknownValue;
2733 SCEVHandle Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
2735 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
2736 // For now we handle only constant steps.
2738 // First, handle unitary steps.
2739 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
2740 return SE.getNegativeSCEV(Start); // N = -Start (as unsigned)
2741 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
2742 return Start; // N = Start (as unsigned)
2744 // Then, try to solve the above equation provided that Start is constant.
2745 if (SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
2746 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
2747 -StartC->getValue()->getValue(),SE);
2749 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
2750 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
2751 // the quadratic equation to solve it.
2752 std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec, SE);
2753 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2754 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2757 cerr << "HFTZ: " << *V << " - sol#1: " << *R1
2758 << " sol#2: " << *R2 << "\n";
2760 // Pick the smallest positive root value.
2761 if (ConstantInt *CB =
2762 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
2763 R1->getValue(), R2->getValue()))) {
2764 if (CB->getZExtValue() == false)
2765 std::swap(R1, R2); // R1 is the minimum root now.
2767 // We can only use this value if the chrec ends up with an exact zero
2768 // value at this index. When solving for "X*X != 5", for example, we
2769 // should not accept a root of 2.
2770 SCEVHandle Val = AddRec->evaluateAtIteration(R1, SE);
2772 return R1; // We found a quadratic root!
2777 return UnknownValue;
2780 /// HowFarToNonZero - Return the number of times a backedge checking the
2781 /// specified value for nonzero will execute. If not computable, return
2783 SCEVHandle ScalarEvolutionsImpl::HowFarToNonZero(SCEV *V, const Loop *L) {
2784 // Loops that look like: while (X == 0) are very strange indeed. We don't
2785 // handle them yet except for the trivial case. This could be expanded in the
2786 // future as needed.
2788 // If the value is a constant, check to see if it is known to be non-zero
2789 // already. If so, the backedge will execute zero times.
2790 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2791 if (!C->getValue()->isNullValue())
2792 return SE.getIntegerSCEV(0, C->getType());
2793 return UnknownValue; // Otherwise it will loop infinitely.
2796 // We could implement others, but I really doubt anyone writes loops like
2797 // this, and if they did, they would already be constant folded.
2798 return UnknownValue;
2801 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
2802 /// (which may not be an immediate predecessor) which has exactly one
2803 /// successor from which BB is reachable, or null if no such block is
2807 ScalarEvolutionsImpl::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
2808 // If the block has a unique predecessor, the predecessor must have
2809 // no other successors from which BB is reachable.
2810 if (BasicBlock *Pred = BB->getSinglePredecessor())
2813 // A loop's header is defined to be a block that dominates the loop.
2814 // If the loop has a preheader, it must be a block that has exactly
2815 // one successor that can reach BB. This is slightly more strict
2816 // than necessary, but works if critical edges are split.
2817 if (Loop *L = LI.getLoopFor(BB))
2818 return L->getLoopPreheader();
2823 /// executesAtLeastOnce - Test whether entry to the loop is protected by
2824 /// a conditional between LHS and RHS.
2825 bool ScalarEvolutionsImpl::executesAtLeastOnce(const Loop *L, bool isSigned,
2827 SCEV *LHS, SCEV *RHS) {
2828 BasicBlock *Preheader = L->getLoopPreheader();
2829 BasicBlock *PreheaderDest = L->getHeader();
2831 // Starting at the preheader, climb up the predecessor chain, as long as
2832 // there are predecessors that can be found that have unique successors
2833 // leading to the original header.
2835 PreheaderDest = Preheader,
2836 Preheader = getPredecessorWithUniqueSuccessorForBB(Preheader)) {
2838 BranchInst *LoopEntryPredicate =
2839 dyn_cast<BranchInst>(Preheader->getTerminator());
2840 if (!LoopEntryPredicate ||
2841 LoopEntryPredicate->isUnconditional())
2844 ICmpInst *ICI = dyn_cast<ICmpInst>(LoopEntryPredicate->getCondition());
2847 // Now that we found a conditional branch that dominates the loop, check to
2848 // see if it is the comparison we are looking for.
2849 Value *PreCondLHS = ICI->getOperand(0);
2850 Value *PreCondRHS = ICI->getOperand(1);
2851 ICmpInst::Predicate Cond;
2852 if (LoopEntryPredicate->getSuccessor(0) == PreheaderDest)
2853 Cond = ICI->getPredicate();
2855 Cond = ICI->getInversePredicate();
2858 case ICmpInst::ICMP_UGT:
2859 if (isSigned || trueWhenEqual) continue;
2860 std::swap(PreCondLHS, PreCondRHS);
2861 Cond = ICmpInst::ICMP_ULT;
2863 case ICmpInst::ICMP_SGT:
2864 if (!isSigned || trueWhenEqual) continue;
2865 std::swap(PreCondLHS, PreCondRHS);
2866 Cond = ICmpInst::ICMP_SLT;
2868 case ICmpInst::ICMP_ULT:
2869 if (isSigned || trueWhenEqual) continue;
2871 case ICmpInst::ICMP_SLT:
2872 if (!isSigned || trueWhenEqual) continue;
2874 case ICmpInst::ICMP_UGE:
2875 if (isSigned || !trueWhenEqual) continue;
2876 std::swap(PreCondLHS, PreCondRHS);
2877 Cond = ICmpInst::ICMP_ULE;
2879 case ICmpInst::ICMP_SGE:
2880 if (!isSigned || !trueWhenEqual) continue;
2881 std::swap(PreCondLHS, PreCondRHS);
2882 Cond = ICmpInst::ICMP_SLE;
2884 case ICmpInst::ICMP_ULE:
2885 if (isSigned || !trueWhenEqual) continue;
2887 case ICmpInst::ICMP_SLE:
2888 if (!isSigned || !trueWhenEqual) continue;
2894 if (!PreCondLHS->getType()->isInteger()) continue;
2896 SCEVHandle PreCondLHSSCEV = getSCEV(PreCondLHS);
2897 SCEVHandle PreCondRHSSCEV = getSCEV(PreCondRHS);
2898 if ((LHS == PreCondLHSSCEV && RHS == PreCondRHSSCEV) ||
2899 (LHS == SE.getNotSCEV(PreCondRHSSCEV) &&
2900 RHS == SE.getNotSCEV(PreCondLHSSCEV)))
2907 /// potentialInfiniteLoop - Test whether the loop might jump over the exit value
2908 /// due to wrapping around 2^n.
2909 bool ScalarEvolutionsImpl::potentialInfiniteLoop(SCEV *Stride, SCEV *RHS,
2910 bool isSigned, bool trueWhenEqual) {
2911 // Return true when the distance from RHS to maxint > Stride.
2913 if (!isa<SCEVConstant>(Stride))
2915 SCEVConstant *SC = cast<SCEVConstant>(Stride);
2917 if (SC->getValue()->isZero())
2919 if (!trueWhenEqual && SC->getValue()->isOne())
2922 if (!isa<SCEVConstant>(RHS))
2924 SCEVConstant *R = cast<SCEVConstant>(RHS);
2927 return true; // XXX: because we don't have an sdiv scev.
2929 // If negative, it wraps around every iteration, but we don't care about that.
2930 APInt S = SC->getValue()->getValue().abs();
2932 APInt Dist = APInt::getMaxValue(R->getValue()->getBitWidth()) -
2933 R->getValue()->getValue();
2936 return !S.ult(Dist);
2938 return !S.ule(Dist);
2941 /// HowManyLessThans - Return the number of times a backedge containing the
2942 /// specified less-than comparison will execute. If not computable, return
2944 SCEVHandle ScalarEvolutionsImpl::
2945 HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L,
2946 bool isSigned, bool trueWhenEqual) {
2947 // Only handle: "ADDREC < LoopInvariant".
2948 if (!RHS->isLoopInvariant(L)) return UnknownValue;
2950 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
2951 if (!AddRec || AddRec->getLoop() != L)
2952 return UnknownValue;
2954 if (AddRec->isAffine()) {
2955 SCEVHandle Stride = AddRec->getOperand(1);
2956 if (potentialInfiniteLoop(Stride, RHS, isSigned, trueWhenEqual))
2957 return UnknownValue;
2959 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
2960 // m. So, we count the number of iterations in which {n,+,s} < m is true.
2961 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
2962 // treat m-n as signed nor unsigned due to overflow possibility.
2964 // First, we get the value of the LHS in the first iteration: n
2965 SCEVHandle Start = AddRec->getOperand(0);
2967 SCEVHandle One = SE.getIntegerSCEV(1, RHS->getType());
2969 // Assuming that the loop will run at least once, we know that it will
2970 // run (m-n)/s times.
2971 SCEVHandle End = RHS;
2973 if (!executesAtLeastOnce(L, isSigned, trueWhenEqual,
2974 SE.getMinusSCEV(Start, One), RHS)) {
2975 // If not, we get the value of the LHS in the first iteration in which
2976 // the above condition doesn't hold. This equals to max(m,n).
2977 End = isSigned ? SE.getSMaxExpr(RHS, Start)
2978 : SE.getUMaxExpr(RHS, Start);
2981 // If the expression is less-than-or-equal to, we need to extend the
2982 // loop by one iteration.
2984 // The loop won't actually run (m-n)/s times because the loop iterations
2985 // won't divide evenly. For example, if you have {2,+,5} u< 10 the
2986 // division would equal one, but the loop runs twice putting the
2987 // induction variable at 12.
2990 // (Stride - 1) is correct only because we know it's unsigned.
2991 // What we really want is to decrease the magnitude of Stride by one.
2992 Start = SE.getMinusSCEV(Start, SE.getMinusSCEV(Stride, One));
2994 Start = SE.getMinusSCEV(Start, Stride);
2996 // Finally, we subtract these two values to get the number of times the
2997 // backedge is executed: max(m,n)-n.
2998 return SE.getUDivExpr(SE.getMinusSCEV(End, Start), Stride);
3001 return UnknownValue;
3004 /// getNumIterationsInRange - Return the number of iterations of this loop that
3005 /// produce values in the specified constant range. Another way of looking at
3006 /// this is that it returns the first iteration number where the value is not in
3007 /// the condition, thus computing the exit count. If the iteration count can't
3008 /// be computed, an instance of SCEVCouldNotCompute is returned.
3009 SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
3010 ScalarEvolution &SE) const {
3011 if (Range.isFullSet()) // Infinite loop.
3012 return new SCEVCouldNotCompute();
3014 // If the start is a non-zero constant, shift the range to simplify things.
3015 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
3016 if (!SC->getValue()->isZero()) {
3017 std::vector<SCEVHandle> Operands(op_begin(), op_end());
3018 Operands[0] = SE.getIntegerSCEV(0, SC->getType());
3019 SCEVHandle Shifted = SE.getAddRecExpr(Operands, getLoop());
3020 if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast<SCEVAddRecExpr>(Shifted))
3021 return ShiftedAddRec->getNumIterationsInRange(
3022 Range.subtract(SC->getValue()->getValue()), SE);
3023 // This is strange and shouldn't happen.
3024 return new SCEVCouldNotCompute();
3027 // The only time we can solve this is when we have all constant indices.
3028 // Otherwise, we cannot determine the overflow conditions.
3029 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
3030 if (!isa<SCEVConstant>(getOperand(i)))
3031 return new SCEVCouldNotCompute();
3034 // Okay at this point we know that all elements of the chrec are constants and
3035 // that the start element is zero.
3037 // First check to see if the range contains zero. If not, the first
3039 if (!Range.contains(APInt(getBitWidth(),0)))
3040 return SE.getConstant(ConstantInt::get(getType(),0));
3043 // If this is an affine expression then we have this situation:
3044 // Solve {0,+,A} in Range === Ax in Range
3046 // We know that zero is in the range. If A is positive then we know that
3047 // the upper value of the range must be the first possible exit value.
3048 // If A is negative then the lower of the range is the last possible loop
3049 // value. Also note that we already checked for a full range.
3050 APInt One(getBitWidth(),1);
3051 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
3052 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
3054 // The exit value should be (End+A)/A.
3055 APInt ExitVal = (End + A).udiv(A);
3056 ConstantInt *ExitValue = ConstantInt::get(ExitVal);
3058 // Evaluate at the exit value. If we really did fall out of the valid
3059 // range, then we computed our trip count, otherwise wrap around or other
3060 // things must have happened.
3061 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
3062 if (Range.contains(Val->getValue()))
3063 return new SCEVCouldNotCompute(); // Something strange happened
3065 // Ensure that the previous value is in the range. This is a sanity check.
3066 assert(Range.contains(
3067 EvaluateConstantChrecAtConstant(this,
3068 ConstantInt::get(ExitVal - One), SE)->getValue()) &&
3069 "Linear scev computation is off in a bad way!");
3070 return SE.getConstant(ExitValue);
3071 } else if (isQuadratic()) {
3072 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
3073 // quadratic equation to solve it. To do this, we must frame our problem in
3074 // terms of figuring out when zero is crossed, instead of when
3075 // Range.getUpper() is crossed.
3076 std::vector<SCEVHandle> NewOps(op_begin(), op_end());
3077 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
3078 SCEVHandle NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
3080 // Next, solve the constructed addrec
3081 std::pair<SCEVHandle,SCEVHandle> Roots =
3082 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
3083 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
3084 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
3086 // Pick the smallest positive root value.
3087 if (ConstantInt *CB =
3088 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
3089 R1->getValue(), R2->getValue()))) {
3090 if (CB->getZExtValue() == false)
3091 std::swap(R1, R2); // R1 is the minimum root now.
3093 // Make sure the root is not off by one. The returned iteration should
3094 // not be in the range, but the previous one should be. When solving
3095 // for "X*X < 5", for example, we should not return a root of 2.
3096 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
3099 if (Range.contains(R1Val->getValue())) {
3100 // The next iteration must be out of the range...
3101 ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()+1);
3103 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
3104 if (!Range.contains(R1Val->getValue()))
3105 return SE.getConstant(NextVal);
3106 return new SCEVCouldNotCompute(); // Something strange happened
3109 // If R1 was not in the range, then it is a good return value. Make
3110 // sure that R1-1 WAS in the range though, just in case.
3111 ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()-1);
3112 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
3113 if (Range.contains(R1Val->getValue()))
3115 return new SCEVCouldNotCompute(); // Something strange happened
3120 return new SCEVCouldNotCompute();
3125 //===----------------------------------------------------------------------===//
3126 // ScalarEvolution Class Implementation
3127 //===----------------------------------------------------------------------===//
3129 bool ScalarEvolution::runOnFunction(Function &F) {
3130 Impl = new ScalarEvolutionsImpl(*this, F, getAnalysis<LoopInfo>());
3134 void ScalarEvolution::releaseMemory() {
3135 delete (ScalarEvolutionsImpl*)Impl;
3139 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
3140 AU.setPreservesAll();
3141 AU.addRequiredTransitive<LoopInfo>();
3144 SCEVHandle ScalarEvolution::getSCEV(Value *V) const {
3145 return ((ScalarEvolutionsImpl*)Impl)->getSCEV(V);
3148 /// hasSCEV - Return true if the SCEV for this value has already been
3150 bool ScalarEvolution::hasSCEV(Value *V) const {
3151 return ((ScalarEvolutionsImpl*)Impl)->hasSCEV(V);
3155 /// setSCEV - Insert the specified SCEV into the map of current SCEVs for
3156 /// the specified value.
3157 void ScalarEvolution::setSCEV(Value *V, const SCEVHandle &H) {
3158 ((ScalarEvolutionsImpl*)Impl)->setSCEV(V, H);
3162 SCEVHandle ScalarEvolution::getIterationCount(const Loop *L) const {
3163 return ((ScalarEvolutionsImpl*)Impl)->getIterationCount(L);
3166 bool ScalarEvolution::hasLoopInvariantIterationCount(const Loop *L) const {
3167 return !isa<SCEVCouldNotCompute>(getIterationCount(L));
3170 SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) const {
3171 return ((ScalarEvolutionsImpl*)Impl)->getSCEVAtScope(getSCEV(V), L);
3174 void ScalarEvolution::deleteValueFromRecords(Value *V) const {
3175 return ((ScalarEvolutionsImpl*)Impl)->deleteValueFromRecords(V);
3178 static void PrintLoopInfo(std::ostream &OS, const ScalarEvolution *SE,
3180 // Print all inner loops first
3181 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
3182 PrintLoopInfo(OS, SE, *I);
3184 OS << "Loop " << L->getHeader()->getName() << ": ";
3186 SmallVector<BasicBlock*, 8> ExitBlocks;
3187 L->getExitBlocks(ExitBlocks);
3188 if (ExitBlocks.size() != 1)
3189 OS << "<multiple exits> ";
3191 if (SE->hasLoopInvariantIterationCount(L)) {
3192 OS << *SE->getIterationCount(L) << " iterations! ";
3194 OS << "Unpredictable iteration count. ";
3200 void ScalarEvolution::print(std::ostream &OS, const Module* ) const {
3201 Function &F = ((ScalarEvolutionsImpl*)Impl)->F;
3202 LoopInfo &LI = ((ScalarEvolutionsImpl*)Impl)->LI;
3204 OS << "Classifying expressions for: " << F.getName() << "\n";
3205 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
3206 if (I->getType()->isInteger()) {
3209 SCEVHandle SV = getSCEV(&*I);
3213 if (const Loop *L = LI.getLoopFor((*I).getParent())) {
3215 SCEVHandle ExitValue = getSCEVAtScope(&*I, L->getParentLoop());
3216 if (isa<SCEVCouldNotCompute>(ExitValue)) {
3217 OS << "<<Unknown>>";
3227 OS << "Determining loop execution counts for: " << F.getName() << "\n";
3228 for (LoopInfo::iterator I = LI.begin(), E = LI.end(); I != E; ++I)
3229 PrintLoopInfo(OS, this, *I);