ScalarEvolution.cpp revision 6c459a28ecb8d33e4b59ab2db1f9a58a2d06824b
1//===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
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.
13//
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.
20//
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.
26//
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.
31//
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.
35//
36// TODO: We should use these routines and value representations to implement
37// dependence analysis!
38//
39//===----------------------------------------------------------------------===//
40//
41// There are several good references for the techniques used in this analysis.
42//
43//  Chains of recurrences -- a method to expedite the evaluation
44//  of closed-form functions
45//  Olaf Bachmann, Paul S. Wang, Eugene V. Zima
46//
47//  On computational properties of chains of recurrences
48//  Eugene V. Zima
49//
50//  Symbolic Evaluation of Chains of Recurrences for Loop Optimization
51//  Robert A. van Engelen
52//
53//  Efficient Symbolic Analysis for Optimizing Compilers
54//  Robert A. van Engelen
55//
56//  Using the chains of recurrences algebra for data dependence testing and
57//  induction variable substitution
58//  MS Thesis, Johnie Birch
59//
60//===----------------------------------------------------------------------===//
61
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"
81#include <ostream>
82#include <algorithm>
83#include <cmath>
84using namespace llvm;
85
86STATISTIC(NumBruteForceEvaluations,
87          "Number of brute force evaluations needed to "
88          "calculate high-order polynomial exit values");
89STATISTIC(NumArrayLenItCounts,
90          "Number of trip counts computed with array length");
91STATISTIC(NumTripCountsComputed,
92          "Number of loops with predictable loop counts");
93STATISTIC(NumTripCountsNotComputed,
94          "Number of loops without predictable loop counts");
95STATISTIC(NumBruteForceTripCountsComputed,
96          "Number of loops with trip counts computed by force");
97
98static cl::opt<unsigned>
99MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
100                        cl::desc("Maximum number of iterations SCEV will "
101                                 "symbolically execute a constant derived loop"),
102                        cl::init(100));
103
104static RegisterPass<ScalarEvolution>
105R("scalar-evolution", "Scalar Evolution Analysis", false, true);
106char ScalarEvolution::ID = 0;
107
108//===----------------------------------------------------------------------===//
109//                           SCEV class definitions
110//===----------------------------------------------------------------------===//
111
112//===----------------------------------------------------------------------===//
113// Implementation of the SCEV class.
114//
115SCEV::~SCEV() {}
116void SCEV::dump() const {
117  print(cerr);
118}
119
120/// getValueRange - Return the tightest constant bounds that this value is
121/// known to have.  This method is only valid on integer SCEV objects.
122ConstantRange SCEV::getValueRange() const {
123  const Type *Ty = getType();
124  assert(Ty->isInteger() && "Can't get range for a non-integer SCEV!");
125  // Default to a full range if no better information is available.
126  return ConstantRange(getBitWidth());
127}
128
129uint32_t SCEV::getBitWidth() const {
130  if (const IntegerType* ITy = dyn_cast<IntegerType>(getType()))
131    return ITy->getBitWidth();
132  return 0;
133}
134
135bool SCEV::isZero() const {
136  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
137    return SC->getValue()->isZero();
138  return false;
139}
140
141
142SCEVCouldNotCompute::SCEVCouldNotCompute() : SCEV(scCouldNotCompute) {}
143
144bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
145  assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
146  return false;
147}
148
149const Type *SCEVCouldNotCompute::getType() const {
150  assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
151  return 0;
152}
153
154bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
155  assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
156  return false;
157}
158
159SCEVHandle SCEVCouldNotCompute::
160replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
161                                  const SCEVHandle &Conc,
162                                  ScalarEvolution &SE) const {
163  return this;
164}
165
166void SCEVCouldNotCompute::print(std::ostream &OS) const {
167  OS << "***COULDNOTCOMPUTE***";
168}
169
170bool SCEVCouldNotCompute::classof(const SCEV *S) {
171  return S->getSCEVType() == scCouldNotCompute;
172}
173
174
175// SCEVConstants - Only allow the creation of one SCEVConstant for any
176// particular value.  Don't use a SCEVHandle here, or else the object will
177// never be deleted!
178static ManagedStatic<std::map<ConstantInt*, SCEVConstant*> > SCEVConstants;
179
180
181SCEVConstant::~SCEVConstant() {
182  SCEVConstants->erase(V);
183}
184
185SCEVHandle ScalarEvolution::getConstant(ConstantInt *V) {
186  SCEVConstant *&R = (*SCEVConstants)[V];
187  if (R == 0) R = new SCEVConstant(V);
188  return R;
189}
190
191SCEVHandle ScalarEvolution::getConstant(const APInt& Val) {
192  return getConstant(ConstantInt::get(Val));
193}
194
195ConstantRange SCEVConstant::getValueRange() const {
196  return ConstantRange(V->getValue());
197}
198
199const Type *SCEVConstant::getType() const { return V->getType(); }
200
201void SCEVConstant::print(std::ostream &OS) const {
202  WriteAsOperand(OS, V, false);
203}
204
205// SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any
206// particular input.  Don't use a SCEVHandle here, or else the object will
207// never be deleted!
208static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
209                     SCEVTruncateExpr*> > SCEVTruncates;
210
211SCEVTruncateExpr::SCEVTruncateExpr(const SCEVHandle &op, const Type *ty)
212  : SCEV(scTruncate), Op(op), Ty(ty) {
213  assert(Op->getType()->isInteger() && Ty->isInteger() &&
214         "Cannot truncate non-integer value!");
215  assert(Op->getType()->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits()
216         && "This is not a truncating conversion!");
217}
218
219SCEVTruncateExpr::~SCEVTruncateExpr() {
220  SCEVTruncates->erase(std::make_pair(Op, Ty));
221}
222
223ConstantRange SCEVTruncateExpr::getValueRange() const {
224  return getOperand()->getValueRange().truncate(getBitWidth());
225}
226
227void SCEVTruncateExpr::print(std::ostream &OS) const {
228  OS << "(truncate " << *Op << " to " << *Ty << ")";
229}
230
231// SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any
232// particular input.  Don't use a SCEVHandle here, or else the object will never
233// be deleted!
234static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
235                     SCEVZeroExtendExpr*> > SCEVZeroExtends;
236
237SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty)
238  : SCEV(scZeroExtend), Op(op), Ty(ty) {
239  assert(Op->getType()->isInteger() && Ty->isInteger() &&
240         "Cannot zero extend non-integer value!");
241  assert(Op->getType()->getPrimitiveSizeInBits() < Ty->getPrimitiveSizeInBits()
242         && "This is not an extending conversion!");
243}
244
245SCEVZeroExtendExpr::~SCEVZeroExtendExpr() {
246  SCEVZeroExtends->erase(std::make_pair(Op, Ty));
247}
248
249ConstantRange SCEVZeroExtendExpr::getValueRange() const {
250  return getOperand()->getValueRange().zeroExtend(getBitWidth());
251}
252
253void SCEVZeroExtendExpr::print(std::ostream &OS) const {
254  OS << "(zeroextend " << *Op << " to " << *Ty << ")";
255}
256
257// SCEVSignExtends - Only allow the creation of one SCEVSignExtendExpr for any
258// particular input.  Don't use a SCEVHandle here, or else the object will never
259// be deleted!
260static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
261                     SCEVSignExtendExpr*> > SCEVSignExtends;
262
263SCEVSignExtendExpr::SCEVSignExtendExpr(const SCEVHandle &op, const Type *ty)
264  : SCEV(scSignExtend), Op(op), Ty(ty) {
265  assert(Op->getType()->isInteger() && Ty->isInteger() &&
266         "Cannot sign extend non-integer value!");
267  assert(Op->getType()->getPrimitiveSizeInBits() < Ty->getPrimitiveSizeInBits()
268         && "This is not an extending conversion!");
269}
270
271SCEVSignExtendExpr::~SCEVSignExtendExpr() {
272  SCEVSignExtends->erase(std::make_pair(Op, Ty));
273}
274
275ConstantRange SCEVSignExtendExpr::getValueRange() const {
276  return getOperand()->getValueRange().signExtend(getBitWidth());
277}
278
279void SCEVSignExtendExpr::print(std::ostream &OS) const {
280  OS << "(signextend " << *Op << " to " << *Ty << ")";
281}
282
283// SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any
284// particular input.  Don't use a SCEVHandle here, or else the object will never
285// be deleted!
286static ManagedStatic<std::map<std::pair<unsigned, std::vector<SCEV*> >,
287                     SCEVCommutativeExpr*> > SCEVCommExprs;
288
289SCEVCommutativeExpr::~SCEVCommutativeExpr() {
290  SCEVCommExprs->erase(std::make_pair(getSCEVType(),
291                                      std::vector<SCEV*>(Operands.begin(),
292                                                         Operands.end())));
293}
294
295void SCEVCommutativeExpr::print(std::ostream &OS) const {
296  assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
297  const char *OpStr = getOperationStr();
298  OS << "(" << *Operands[0];
299  for (unsigned i = 1, e = Operands.size(); i != e; ++i)
300    OS << OpStr << *Operands[i];
301  OS << ")";
302}
303
304SCEVHandle SCEVCommutativeExpr::
305replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
306                                  const SCEVHandle &Conc,
307                                  ScalarEvolution &SE) const {
308  for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
309    SCEVHandle H =
310      getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
311    if (H != getOperand(i)) {
312      std::vector<SCEVHandle> NewOps;
313      NewOps.reserve(getNumOperands());
314      for (unsigned j = 0; j != i; ++j)
315        NewOps.push_back(getOperand(j));
316      NewOps.push_back(H);
317      for (++i; i != e; ++i)
318        NewOps.push_back(getOperand(i)->
319                         replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
320
321      if (isa<SCEVAddExpr>(this))
322        return SE.getAddExpr(NewOps);
323      else if (isa<SCEVMulExpr>(this))
324        return SE.getMulExpr(NewOps);
325      else if (isa<SCEVSMaxExpr>(this))
326        return SE.getSMaxExpr(NewOps);
327      else if (isa<SCEVUMaxExpr>(this))
328        return SE.getUMaxExpr(NewOps);
329      else
330        assert(0 && "Unknown commutative expr!");
331    }
332  }
333  return this;
334}
335
336
337// SCEVUDivs - Only allow the creation of one SCEVUDivExpr for any particular
338// input.  Don't use a SCEVHandle here, or else the object will never be
339// deleted!
340static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>,
341                     SCEVUDivExpr*> > SCEVUDivs;
342
343SCEVUDivExpr::~SCEVUDivExpr() {
344  SCEVUDivs->erase(std::make_pair(LHS, RHS));
345}
346
347void SCEVUDivExpr::print(std::ostream &OS) const {
348  OS << "(" << *LHS << " /u " << *RHS << ")";
349}
350
351const Type *SCEVUDivExpr::getType() const {
352  return LHS->getType();
353}
354
355// SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
356// particular input.  Don't use a SCEVHandle here, or else the object will never
357// be deleted!
358static ManagedStatic<std::map<std::pair<const Loop *, std::vector<SCEV*> >,
359                     SCEVAddRecExpr*> > SCEVAddRecExprs;
360
361SCEVAddRecExpr::~SCEVAddRecExpr() {
362  SCEVAddRecExprs->erase(std::make_pair(L,
363                                        std::vector<SCEV*>(Operands.begin(),
364                                                           Operands.end())));
365}
366
367SCEVHandle SCEVAddRecExpr::
368replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
369                                  const SCEVHandle &Conc,
370                                  ScalarEvolution &SE) const {
371  for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
372    SCEVHandle H =
373      getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
374    if (H != getOperand(i)) {
375      std::vector<SCEVHandle> NewOps;
376      NewOps.reserve(getNumOperands());
377      for (unsigned j = 0; j != i; ++j)
378        NewOps.push_back(getOperand(j));
379      NewOps.push_back(H);
380      for (++i; i != e; ++i)
381        NewOps.push_back(getOperand(i)->
382                         replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
383
384      return SE.getAddRecExpr(NewOps, L);
385    }
386  }
387  return this;
388}
389
390
391bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
392  // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't
393  // contain L and if the start is invariant.
394  return !QueryLoop->contains(L->getHeader()) &&
395         getOperand(0)->isLoopInvariant(QueryLoop);
396}
397
398
399void SCEVAddRecExpr::print(std::ostream &OS) const {
400  OS << "{" << *Operands[0];
401  for (unsigned i = 1, e = Operands.size(); i != e; ++i)
402    OS << ",+," << *Operands[i];
403  OS << "}<" << L->getHeader()->getName() + ">";
404}
405
406// SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular
407// value.  Don't use a SCEVHandle here, or else the object will never be
408// deleted!
409static ManagedStatic<std::map<Value*, SCEVUnknown*> > SCEVUnknowns;
410
411SCEVUnknown::~SCEVUnknown() { SCEVUnknowns->erase(V); }
412
413bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
414  // All non-instruction values are loop invariant.  All instructions are loop
415  // invariant if they are not contained in the specified loop.
416  if (Instruction *I = dyn_cast<Instruction>(V))
417    return !L->contains(I->getParent());
418  return true;
419}
420
421const Type *SCEVUnknown::getType() const {
422  return V->getType();
423}
424
425void SCEVUnknown::print(std::ostream &OS) const {
426  WriteAsOperand(OS, V, false);
427}
428
429//===----------------------------------------------------------------------===//
430//                               SCEV Utilities
431//===----------------------------------------------------------------------===//
432
433namespace {
434  /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
435  /// than the complexity of the RHS.  This comparator is used to canonicalize
436  /// expressions.
437  struct VISIBILITY_HIDDEN SCEVComplexityCompare {
438    bool operator()(const SCEV *LHS, const SCEV *RHS) const {
439      return LHS->getSCEVType() < RHS->getSCEVType();
440    }
441  };
442}
443
444/// GroupByComplexity - Given a list of SCEV objects, order them by their
445/// complexity, and group objects of the same complexity together by value.
446/// When this routine is finished, we know that any duplicates in the vector are
447/// consecutive and that complexity is monotonically increasing.
448///
449/// Note that we go take special precautions to ensure that we get determinstic
450/// results from this routine.  In other words, we don't want the results of
451/// this to depend on where the addresses of various SCEV objects happened to
452/// land in memory.
453///
454static void GroupByComplexity(std::vector<SCEVHandle> &Ops) {
455  if (Ops.size() < 2) return;  // Noop
456  if (Ops.size() == 2) {
457    // This is the common case, which also happens to be trivially simple.
458    // Special case it.
459    if (SCEVComplexityCompare()(Ops[1], Ops[0]))
460      std::swap(Ops[0], Ops[1]);
461    return;
462  }
463
464  // Do the rough sort by complexity.
465  std::sort(Ops.begin(), Ops.end(), SCEVComplexityCompare());
466
467  // Now that we are sorted by complexity, group elements of the same
468  // complexity.  Note that this is, at worst, N^2, but the vector is likely to
469  // be extremely short in practice.  Note that we take this approach because we
470  // do not want to depend on the addresses of the objects we are grouping.
471  for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
472    SCEV *S = Ops[i];
473    unsigned Complexity = S->getSCEVType();
474
475    // If there are any objects of the same complexity and same value as this
476    // one, group them.
477    for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
478      if (Ops[j] == S) { // Found a duplicate.
479        // Move it to immediately after i'th element.
480        std::swap(Ops[i+1], Ops[j]);
481        ++i;   // no need to rescan it.
482        if (i == e-2) return;  // Done!
483      }
484    }
485  }
486}
487
488
489
490//===----------------------------------------------------------------------===//
491//                      Simple SCEV method implementations
492//===----------------------------------------------------------------------===//
493
494/// getIntegerSCEV - Given an integer or FP type, create a constant for the
495/// specified signed integer value and return a SCEV for the constant.
496SCEVHandle ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
497  Constant *C;
498  if (Val == 0)
499    C = Constant::getNullValue(Ty);
500  else if (Ty->isFloatingPoint())
501    C = ConstantFP::get(APFloat(Ty==Type::FloatTy ? APFloat::IEEEsingle :
502                                APFloat::IEEEdouble, Val));
503  else
504    C = ConstantInt::get(Ty, Val);
505  return getUnknown(C);
506}
507
508/// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
509///
510SCEVHandle ScalarEvolution::getNegativeSCEV(const SCEVHandle &V) {
511  if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
512    return getUnknown(ConstantExpr::getNeg(VC->getValue()));
513
514  return getMulExpr(V, getConstant(ConstantInt::getAllOnesValue(V->getType())));
515}
516
517/// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
518SCEVHandle ScalarEvolution::getNotSCEV(const SCEVHandle &V) {
519  if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
520    return getUnknown(ConstantExpr::getNot(VC->getValue()));
521
522  SCEVHandle AllOnes = getConstant(ConstantInt::getAllOnesValue(V->getType()));
523  return getMinusSCEV(AllOnes, V);
524}
525
526/// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
527///
528SCEVHandle ScalarEvolution::getMinusSCEV(const SCEVHandle &LHS,
529                                         const SCEVHandle &RHS) {
530  // X - Y --> X + -Y
531  return getAddExpr(LHS, getNegativeSCEV(RHS));
532}
533
534
535/// BinomialCoefficient - Compute BC(It, K).  The result is of the same type as
536/// It.  Assume, K > 0.
537static SCEVHandle BinomialCoefficient(SCEVHandle It, unsigned K,
538                                      ScalarEvolution &SE) {
539  // We are using the following formula for BC(It, K):
540  //
541  //   BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
542  //
543  // Suppose, W is the bitwidth of It (and of the return value as well).  We
544  // must be prepared for overflow.  Hence, we must assure that the result of
545  // our computation is equal to the accurate one modulo 2^W.  Unfortunately,
546  // division isn't safe in modular arithmetic.  This means we must perform the
547  // whole computation accurately and then truncate the result to W bits.
548  //
549  // The dividend of the formula is a multiplication of K integers of bitwidth
550  // W.  K*W bits suffice to compute it accurately.
551  //
552  // FIXME: We assume the divisor can be accurately computed using 16-bit
553  // unsigned integer type. It is true up to K = 8 (AddRecs of length 9). In
554  // future we may use APInt to use the minimum number of bits necessary to
555  // compute it accurately.
556  //
557  // It is safe to use unsigned division here: the dividend is nonnegative and
558  // the divisor is positive.
559
560  // Handle the simplest case efficiently.
561  if (K == 1)
562    return It;
563
564  assert(K < 9 && "We cannot handle such long AddRecs yet.");
565
566  // FIXME: A temporary hack to remove in future.  Arbitrary precision integers
567  // aren't supported by the code generator yet.  For the dividend, the bitwidth
568  // we use is the smallest power of 2 greater or equal to K*W and less or equal
569  // to 64.  Note that setting the upper bound for bitwidth may still lead to
570  // miscompilation in some cases.
571  unsigned DividendBits = 1U << Log2_32_Ceil(K * It->getBitWidth());
572  if (DividendBits > 64)
573    DividendBits = 64;
574#if 0 // Waiting for the APInt support in the code generator...
575  unsigned DividendBits = K * It->getBitWidth();
576#endif
577
578  const IntegerType *DividendTy = IntegerType::get(DividendBits);
579  const SCEVHandle ExIt = SE.getTruncateOrZeroExtend(It, DividendTy);
580
581  // The final number of bits we need to perform the division is the maximum of
582  // dividend and divisor bitwidths.
583  const IntegerType *DivisionTy =
584    IntegerType::get(std::max(DividendBits, 16U));
585
586  // Compute K!  We know K >= 2 here.
587  unsigned F = 2;
588  for (unsigned i = 3; i <= K; ++i)
589    F *= i;
590  APInt Divisor(DivisionTy->getBitWidth(), F);
591
592  // Handle this case efficiently, it is common to have constant iteration
593  // counts while computing loop exit values.
594  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(ExIt)) {
595    const APInt& N = SC->getValue()->getValue();
596    APInt Dividend(N.getBitWidth(), 1);
597    for (; K; --K)
598      Dividend *= N-(K-1);
599    if (DividendTy != DivisionTy)
600      Dividend = Dividend.zext(DivisionTy->getBitWidth());
601
602    APInt Result = Dividend.udiv(Divisor);
603    if (Result.getBitWidth() != It->getBitWidth())
604      Result = Result.trunc(It->getBitWidth());
605
606    return SE.getConstant(Result);
607  }
608
609  SCEVHandle Dividend = ExIt;
610  for (unsigned i = 1; i != K; ++i)
611    Dividend =
612      SE.getMulExpr(Dividend,
613                    SE.getMinusSCEV(ExIt, SE.getIntegerSCEV(i, DividendTy)));
614
615  return SE.getTruncateOrZeroExtend(
616             SE.getUDivExpr(
617                 SE.getTruncateOrZeroExtend(Dividend, DivisionTy),
618                 SE.getConstant(Divisor)
619             ), It->getType());
620}
621
622/// evaluateAtIteration - Return the value of this chain of recurrences at
623/// the specified iteration number.  We can evaluate this recurrence by
624/// multiplying each element in the chain by the binomial coefficient
625/// corresponding to it.  In other words, we can evaluate {A,+,B,+,C,+,D} as:
626///
627///   A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
628///
629/// where BC(It, k) stands for binomial coefficient.
630///
631SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It,
632                                               ScalarEvolution &SE) const {
633  SCEVHandle Result = getStart();
634  for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
635    // The computation is correct in the face of overflow provided that the
636    // multiplication is performed _after_ the evaluation of the binomial
637    // coefficient.
638    SCEVHandle Val = SE.getMulExpr(getOperand(i),
639                                   BinomialCoefficient(It, i, SE));
640    Result = SE.getAddExpr(Result, Val);
641  }
642  return Result;
643}
644
645//===----------------------------------------------------------------------===//
646//                    SCEV Expression folder implementations
647//===----------------------------------------------------------------------===//
648
649SCEVHandle ScalarEvolution::getTruncateExpr(const SCEVHandle &Op, const Type *Ty) {
650  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
651    return getUnknown(
652        ConstantExpr::getTrunc(SC->getValue(), Ty));
653
654  // If the input value is a chrec scev made out of constants, truncate
655  // all of the constants.
656  if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
657    std::vector<SCEVHandle> Operands;
658    for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
659      // FIXME: This should allow truncation of other expression types!
660      if (isa<SCEVConstant>(AddRec->getOperand(i)))
661        Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
662      else
663        break;
664    if (Operands.size() == AddRec->getNumOperands())
665      return getAddRecExpr(Operands, AddRec->getLoop());
666  }
667
668  SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)];
669  if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
670  return Result;
671}
672
673SCEVHandle ScalarEvolution::getZeroExtendExpr(const SCEVHandle &Op, const Type *Ty) {
674  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
675    return getUnknown(
676        ConstantExpr::getZExt(SC->getValue(), Ty));
677
678  // FIXME: If the input value is a chrec scev, and we can prove that the value
679  // did not overflow the old, smaller, value, we can zero extend all of the
680  // operands (often constants).  This would allow analysis of something like
681  // this:  for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
682
683  SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)];
684  if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
685  return Result;
686}
687
688SCEVHandle ScalarEvolution::getSignExtendExpr(const SCEVHandle &Op, const Type *Ty) {
689  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
690    return getUnknown(
691        ConstantExpr::getSExt(SC->getValue(), Ty));
692
693  // FIXME: If the input value is a chrec scev, and we can prove that the value
694  // did not overflow the old, smaller, value, we can sign extend all of the
695  // operands (often constants).  This would allow analysis of something like
696  // this:  for (signed char X = 0; X < 100; ++X) { int Y = X; }
697
698  SCEVSignExtendExpr *&Result = (*SCEVSignExtends)[std::make_pair(Op, Ty)];
699  if (Result == 0) Result = new SCEVSignExtendExpr(Op, Ty);
700  return Result;
701}
702
703/// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion
704/// of the input value to the specified type.  If the type must be
705/// extended, it is zero extended.
706SCEVHandle ScalarEvolution::getTruncateOrZeroExtend(const SCEVHandle &V,
707                                                    const Type *Ty) {
708  const Type *SrcTy = V->getType();
709  assert(SrcTy->isInteger() && Ty->isInteger() &&
710         "Cannot truncate or zero extend with non-integer arguments!");
711  if (SrcTy->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
712    return V;  // No conversion
713  if (SrcTy->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits())
714    return getTruncateExpr(V, Ty);
715  return getZeroExtendExpr(V, Ty);
716}
717
718// get - Get a canonical add expression, or something simpler if possible.
719SCEVHandle ScalarEvolution::getAddExpr(std::vector<SCEVHandle> &Ops) {
720  assert(!Ops.empty() && "Cannot get empty add!");
721  if (Ops.size() == 1) return Ops[0];
722
723  // Sort by complexity, this groups all similar expression types together.
724  GroupByComplexity(Ops);
725
726  // If there are any constants, fold them together.
727  unsigned Idx = 0;
728  if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
729    ++Idx;
730    assert(Idx < Ops.size());
731    while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
732      // We found two constants, fold them together!
733      ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() +
734                                           RHSC->getValue()->getValue());
735      Ops[0] = getConstant(Fold);
736      Ops.erase(Ops.begin()+1);  // Erase the folded element
737      if (Ops.size() == 1) return Ops[0];
738      LHSC = cast<SCEVConstant>(Ops[0]);
739    }
740
741    // If we are left with a constant zero being added, strip it off.
742    if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
743      Ops.erase(Ops.begin());
744      --Idx;
745    }
746  }
747
748  if (Ops.size() == 1) return Ops[0];
749
750  // Okay, check to see if the same value occurs in the operand list twice.  If
751  // so, merge them together into an multiply expression.  Since we sorted the
752  // list, these values are required to be adjacent.
753  const Type *Ty = Ops[0]->getType();
754  for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
755    if (Ops[i] == Ops[i+1]) {      //  X + Y + Y  -->  X + Y*2
756      // Found a match, merge the two values into a multiply, and add any
757      // remaining values to the result.
758      SCEVHandle Two = getIntegerSCEV(2, Ty);
759      SCEVHandle Mul = getMulExpr(Ops[i], Two);
760      if (Ops.size() == 2)
761        return Mul;
762      Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
763      Ops.push_back(Mul);
764      return getAddExpr(Ops);
765    }
766
767  // Now we know the first non-constant operand.  Skip past any cast SCEVs.
768  while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
769    ++Idx;
770
771  // If there are add operands they would be next.
772  if (Idx < Ops.size()) {
773    bool DeletedAdd = false;
774    while (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
775      // If we have an add, expand the add operands onto the end of the operands
776      // list.
777      Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
778      Ops.erase(Ops.begin()+Idx);
779      DeletedAdd = true;
780    }
781
782    // If we deleted at least one add, we added operands to the end of the list,
783    // and they are not necessarily sorted.  Recurse to resort and resimplify
784    // any operands we just aquired.
785    if (DeletedAdd)
786      return getAddExpr(Ops);
787  }
788
789  // Skip over the add expression until we get to a multiply.
790  while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
791    ++Idx;
792
793  // If we are adding something to a multiply expression, make sure the
794  // something is not already an operand of the multiply.  If so, merge it into
795  // the multiply.
796  for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
797    SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
798    for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
799      SCEV *MulOpSCEV = Mul->getOperand(MulOp);
800      for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
801        if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(MulOpSCEV)) {
802          // Fold W + X + (X * Y * Z)  -->  W + (X * ((Y*Z)+1))
803          SCEVHandle InnerMul = Mul->getOperand(MulOp == 0);
804          if (Mul->getNumOperands() != 2) {
805            // If the multiply has more than two operands, we must get the
806            // Y*Z term.
807            std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
808            MulOps.erase(MulOps.begin()+MulOp);
809            InnerMul = getMulExpr(MulOps);
810          }
811          SCEVHandle One = getIntegerSCEV(1, Ty);
812          SCEVHandle AddOne = getAddExpr(InnerMul, One);
813          SCEVHandle OuterMul = getMulExpr(AddOne, Ops[AddOp]);
814          if (Ops.size() == 2) return OuterMul;
815          if (AddOp < Idx) {
816            Ops.erase(Ops.begin()+AddOp);
817            Ops.erase(Ops.begin()+Idx-1);
818          } else {
819            Ops.erase(Ops.begin()+Idx);
820            Ops.erase(Ops.begin()+AddOp-1);
821          }
822          Ops.push_back(OuterMul);
823          return getAddExpr(Ops);
824        }
825
826      // Check this multiply against other multiplies being added together.
827      for (unsigned OtherMulIdx = Idx+1;
828           OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
829           ++OtherMulIdx) {
830        SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
831        // If MulOp occurs in OtherMul, we can fold the two multiplies
832        // together.
833        for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
834             OMulOp != e; ++OMulOp)
835          if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
836            // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
837            SCEVHandle InnerMul1 = Mul->getOperand(MulOp == 0);
838            if (Mul->getNumOperands() != 2) {
839              std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
840              MulOps.erase(MulOps.begin()+MulOp);
841              InnerMul1 = getMulExpr(MulOps);
842            }
843            SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0);
844            if (OtherMul->getNumOperands() != 2) {
845              std::vector<SCEVHandle> MulOps(OtherMul->op_begin(),
846                                             OtherMul->op_end());
847              MulOps.erase(MulOps.begin()+OMulOp);
848              InnerMul2 = getMulExpr(MulOps);
849            }
850            SCEVHandle InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
851            SCEVHandle OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
852            if (Ops.size() == 2) return OuterMul;
853            Ops.erase(Ops.begin()+Idx);
854            Ops.erase(Ops.begin()+OtherMulIdx-1);
855            Ops.push_back(OuterMul);
856            return getAddExpr(Ops);
857          }
858      }
859    }
860  }
861
862  // If there are any add recurrences in the operands list, see if any other
863  // added values are loop invariant.  If so, we can fold them into the
864  // recurrence.
865  while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
866    ++Idx;
867
868  // Scan over all recurrences, trying to fold loop invariants into them.
869  for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
870    // Scan all of the other operands to this add and add them to the vector if
871    // they are loop invariant w.r.t. the recurrence.
872    std::vector<SCEVHandle> LIOps;
873    SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
874    for (unsigned i = 0, e = Ops.size(); i != e; ++i)
875      if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
876        LIOps.push_back(Ops[i]);
877        Ops.erase(Ops.begin()+i);
878        --i; --e;
879      }
880
881    // If we found some loop invariants, fold them into the recurrence.
882    if (!LIOps.empty()) {
883      //  NLI + LI + { Start,+,Step}  -->  NLI + { LI+Start,+,Step }
884      LIOps.push_back(AddRec->getStart());
885
886      std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
887      AddRecOps[0] = getAddExpr(LIOps);
888
889      SCEVHandle NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
890      // If all of the other operands were loop invariant, we are done.
891      if (Ops.size() == 1) return NewRec;
892
893      // Otherwise, add the folded AddRec by the non-liv parts.
894      for (unsigned i = 0;; ++i)
895        if (Ops[i] == AddRec) {
896          Ops[i] = NewRec;
897          break;
898        }
899      return getAddExpr(Ops);
900    }
901
902    // Okay, if there weren't any loop invariants to be folded, check to see if
903    // there are multiple AddRec's with the same loop induction variable being
904    // added together.  If so, we can fold them.
905    for (unsigned OtherIdx = Idx+1;
906         OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
907      if (OtherIdx != Idx) {
908        SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
909        if (AddRec->getLoop() == OtherAddRec->getLoop()) {
910          // Other + {A,+,B} + {C,+,D}  -->  Other + {A+C,+,B+D}
911          std::vector<SCEVHandle> NewOps(AddRec->op_begin(), AddRec->op_end());
912          for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
913            if (i >= NewOps.size()) {
914              NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
915                            OtherAddRec->op_end());
916              break;
917            }
918            NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
919          }
920          SCEVHandle NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
921
922          if (Ops.size() == 2) return NewAddRec;
923
924          Ops.erase(Ops.begin()+Idx);
925          Ops.erase(Ops.begin()+OtherIdx-1);
926          Ops.push_back(NewAddRec);
927          return getAddExpr(Ops);
928        }
929      }
930
931    // Otherwise couldn't fold anything into this recurrence.  Move onto the
932    // next one.
933  }
934
935  // Okay, it looks like we really DO need an add expr.  Check to see if we
936  // already have one, otherwise create a new one.
937  std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
938  SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scAddExpr,
939                                                                 SCEVOps)];
940  if (Result == 0) Result = new SCEVAddExpr(Ops);
941  return Result;
942}
943
944
945SCEVHandle ScalarEvolution::getMulExpr(std::vector<SCEVHandle> &Ops) {
946  assert(!Ops.empty() && "Cannot get empty mul!");
947
948  // Sort by complexity, this groups all similar expression types together.
949  GroupByComplexity(Ops);
950
951  // If there are any constants, fold them together.
952  unsigned Idx = 0;
953  if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
954
955    // C1*(C2+V) -> C1*C2 + C1*V
956    if (Ops.size() == 2)
957      if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
958        if (Add->getNumOperands() == 2 &&
959            isa<SCEVConstant>(Add->getOperand(0)))
960          return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
961                            getMulExpr(LHSC, Add->getOperand(1)));
962
963
964    ++Idx;
965    while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
966      // We found two constants, fold them together!
967      ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() *
968                                           RHSC->getValue()->getValue());
969      Ops[0] = getConstant(Fold);
970      Ops.erase(Ops.begin()+1);  // Erase the folded element
971      if (Ops.size() == 1) return Ops[0];
972      LHSC = cast<SCEVConstant>(Ops[0]);
973    }
974
975    // If we are left with a constant one being multiplied, strip it off.
976    if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
977      Ops.erase(Ops.begin());
978      --Idx;
979    } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
980      // If we have a multiply of zero, it will always be zero.
981      return Ops[0];
982    }
983  }
984
985  // Skip over the add expression until we get to a multiply.
986  while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
987    ++Idx;
988
989  if (Ops.size() == 1)
990    return Ops[0];
991
992  // If there are mul operands inline them all into this expression.
993  if (Idx < Ops.size()) {
994    bool DeletedMul = false;
995    while (SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
996      // If we have an mul, expand the mul operands onto the end of the operands
997      // list.
998      Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
999      Ops.erase(Ops.begin()+Idx);
1000      DeletedMul = true;
1001    }
1002
1003    // If we deleted at least one mul, we added operands to the end of the list,
1004    // and they are not necessarily sorted.  Recurse to resort and resimplify
1005    // any operands we just aquired.
1006    if (DeletedMul)
1007      return getMulExpr(Ops);
1008  }
1009
1010  // If there are any add recurrences in the operands list, see if any other
1011  // added values are loop invariant.  If so, we can fold them into the
1012  // recurrence.
1013  while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1014    ++Idx;
1015
1016  // Scan over all recurrences, trying to fold loop invariants into them.
1017  for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1018    // Scan all of the other operands to this mul and add them to the vector if
1019    // they are loop invariant w.r.t. the recurrence.
1020    std::vector<SCEVHandle> LIOps;
1021    SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1022    for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1023      if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1024        LIOps.push_back(Ops[i]);
1025        Ops.erase(Ops.begin()+i);
1026        --i; --e;
1027      }
1028
1029    // If we found some loop invariants, fold them into the recurrence.
1030    if (!LIOps.empty()) {
1031      //  NLI * LI * { Start,+,Step}  -->  NLI * { LI*Start,+,LI*Step }
1032      std::vector<SCEVHandle> NewOps;
1033      NewOps.reserve(AddRec->getNumOperands());
1034      if (LIOps.size() == 1) {
1035        SCEV *Scale = LIOps[0];
1036        for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1037          NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1038      } else {
1039        for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1040          std::vector<SCEVHandle> MulOps(LIOps);
1041          MulOps.push_back(AddRec->getOperand(i));
1042          NewOps.push_back(getMulExpr(MulOps));
1043        }
1044      }
1045
1046      SCEVHandle NewRec = getAddRecExpr(NewOps, AddRec->getLoop());
1047
1048      // If all of the other operands were loop invariant, we are done.
1049      if (Ops.size() == 1) return NewRec;
1050
1051      // Otherwise, multiply the folded AddRec by the non-liv parts.
1052      for (unsigned i = 0;; ++i)
1053        if (Ops[i] == AddRec) {
1054          Ops[i] = NewRec;
1055          break;
1056        }
1057      return getMulExpr(Ops);
1058    }
1059
1060    // Okay, if there weren't any loop invariants to be folded, check to see if
1061    // there are multiple AddRec's with the same loop induction variable being
1062    // multiplied together.  If so, we can fold them.
1063    for (unsigned OtherIdx = Idx+1;
1064         OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1065      if (OtherIdx != Idx) {
1066        SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1067        if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1068          // F * G  -->  {A,+,B} * {C,+,D}  -->  {A*C,+,F*D + G*B + B*D}
1069          SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1070          SCEVHandle NewStart = getMulExpr(F->getStart(),
1071                                                 G->getStart());
1072          SCEVHandle B = F->getStepRecurrence(*this);
1073          SCEVHandle D = G->getStepRecurrence(*this);
1074          SCEVHandle NewStep = getAddExpr(getMulExpr(F, D),
1075                                          getMulExpr(G, B),
1076                                          getMulExpr(B, D));
1077          SCEVHandle NewAddRec = getAddRecExpr(NewStart, NewStep,
1078                                               F->getLoop());
1079          if (Ops.size() == 2) return NewAddRec;
1080
1081          Ops.erase(Ops.begin()+Idx);
1082          Ops.erase(Ops.begin()+OtherIdx-1);
1083          Ops.push_back(NewAddRec);
1084          return getMulExpr(Ops);
1085        }
1086      }
1087
1088    // Otherwise couldn't fold anything into this recurrence.  Move onto the
1089    // next one.
1090  }
1091
1092  // Okay, it looks like we really DO need an mul expr.  Check to see if we
1093  // already have one, otherwise create a new one.
1094  std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1095  SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scMulExpr,
1096                                                                 SCEVOps)];
1097  if (Result == 0)
1098    Result = new SCEVMulExpr(Ops);
1099  return Result;
1100}
1101
1102SCEVHandle ScalarEvolution::getUDivExpr(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1103  if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1104    if (RHSC->getValue()->equalsInt(1))
1105      return LHS;                            // X udiv 1 --> x
1106
1107    if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1108      Constant *LHSCV = LHSC->getValue();
1109      Constant *RHSCV = RHSC->getValue();
1110      return getUnknown(ConstantExpr::getUDiv(LHSCV, RHSCV));
1111    }
1112  }
1113
1114  // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow.
1115
1116  SCEVUDivExpr *&Result = (*SCEVUDivs)[std::make_pair(LHS, RHS)];
1117  if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
1118  return Result;
1119}
1120
1121
1122/// SCEVAddRecExpr::get - Get a add recurrence expression for the
1123/// specified loop.  Simplify the expression as much as possible.
1124SCEVHandle ScalarEvolution::getAddRecExpr(const SCEVHandle &Start,
1125                               const SCEVHandle &Step, const Loop *L) {
1126  std::vector<SCEVHandle> Operands;
1127  Operands.push_back(Start);
1128  if (SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1129    if (StepChrec->getLoop() == L) {
1130      Operands.insert(Operands.end(), StepChrec->op_begin(),
1131                      StepChrec->op_end());
1132      return getAddRecExpr(Operands, L);
1133    }
1134
1135  Operands.push_back(Step);
1136  return getAddRecExpr(Operands, L);
1137}
1138
1139/// SCEVAddRecExpr::get - Get a add recurrence expression for the
1140/// specified loop.  Simplify the expression as much as possible.
1141SCEVHandle ScalarEvolution::getAddRecExpr(std::vector<SCEVHandle> &Operands,
1142                               const Loop *L) {
1143  if (Operands.size() == 1) return Operands[0];
1144
1145  if (Operands.back()->isZero()) {
1146    Operands.pop_back();
1147    return getAddRecExpr(Operands, L);             // { X,+,0 }  -->  X
1148  }
1149
1150  SCEVAddRecExpr *&Result =
1151    (*SCEVAddRecExprs)[std::make_pair(L, std::vector<SCEV*>(Operands.begin(),
1152                                                            Operands.end()))];
1153  if (Result == 0) Result = new SCEVAddRecExpr(Operands, L);
1154  return Result;
1155}
1156
1157SCEVHandle ScalarEvolution::getSMaxExpr(const SCEVHandle &LHS,
1158                                        const SCEVHandle &RHS) {
1159  std::vector<SCEVHandle> Ops;
1160  Ops.push_back(LHS);
1161  Ops.push_back(RHS);
1162  return getSMaxExpr(Ops);
1163}
1164
1165SCEVHandle ScalarEvolution::getSMaxExpr(std::vector<SCEVHandle> Ops) {
1166  assert(!Ops.empty() && "Cannot get empty smax!");
1167  if (Ops.size() == 1) return Ops[0];
1168
1169  // Sort by complexity, this groups all similar expression types together.
1170  GroupByComplexity(Ops);
1171
1172  // If there are any constants, fold them together.
1173  unsigned Idx = 0;
1174  if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1175    ++Idx;
1176    assert(Idx < Ops.size());
1177    while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1178      // We found two constants, fold them together!
1179      ConstantInt *Fold = ConstantInt::get(
1180                              APIntOps::smax(LHSC->getValue()->getValue(),
1181                                             RHSC->getValue()->getValue()));
1182      Ops[0] = getConstant(Fold);
1183      Ops.erase(Ops.begin()+1);  // Erase the folded element
1184      if (Ops.size() == 1) return Ops[0];
1185      LHSC = cast<SCEVConstant>(Ops[0]);
1186    }
1187
1188    // If we are left with a constant -inf, strip it off.
1189    if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
1190      Ops.erase(Ops.begin());
1191      --Idx;
1192    }
1193  }
1194
1195  if (Ops.size() == 1) return Ops[0];
1196
1197  // Find the first SMax
1198  while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
1199    ++Idx;
1200
1201  // Check to see if one of the operands is an SMax. If so, expand its operands
1202  // onto our operand list, and recurse to simplify.
1203  if (Idx < Ops.size()) {
1204    bool DeletedSMax = false;
1205    while (SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
1206      Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
1207      Ops.erase(Ops.begin()+Idx);
1208      DeletedSMax = true;
1209    }
1210
1211    if (DeletedSMax)
1212      return getSMaxExpr(Ops);
1213  }
1214
1215  // Okay, check to see if the same value occurs in the operand list twice.  If
1216  // so, delete one.  Since we sorted the list, these values are required to
1217  // be adjacent.
1218  for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1219    if (Ops[i] == Ops[i+1]) {      //  X smax Y smax Y  -->  X smax Y
1220      Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1221      --i; --e;
1222    }
1223
1224  if (Ops.size() == 1) return Ops[0];
1225
1226  assert(!Ops.empty() && "Reduced smax down to nothing!");
1227
1228  // Okay, it looks like we really DO need an smax expr.  Check to see if we
1229  // already have one, otherwise create a new one.
1230  std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1231  SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scSMaxExpr,
1232                                                                 SCEVOps)];
1233  if (Result == 0) Result = new SCEVSMaxExpr(Ops);
1234  return Result;
1235}
1236
1237SCEVHandle ScalarEvolution::getUMaxExpr(const SCEVHandle &LHS,
1238                                        const SCEVHandle &RHS) {
1239  std::vector<SCEVHandle> Ops;
1240  Ops.push_back(LHS);
1241  Ops.push_back(RHS);
1242  return getUMaxExpr(Ops);
1243}
1244
1245SCEVHandle ScalarEvolution::getUMaxExpr(std::vector<SCEVHandle> Ops) {
1246  assert(!Ops.empty() && "Cannot get empty umax!");
1247  if (Ops.size() == 1) return Ops[0];
1248
1249  // Sort by complexity, this groups all similar expression types together.
1250  GroupByComplexity(Ops);
1251
1252  // If there are any constants, fold them together.
1253  unsigned Idx = 0;
1254  if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1255    ++Idx;
1256    assert(Idx < Ops.size());
1257    while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1258      // We found two constants, fold them together!
1259      ConstantInt *Fold = ConstantInt::get(
1260                              APIntOps::umax(LHSC->getValue()->getValue(),
1261                                             RHSC->getValue()->getValue()));
1262      Ops[0] = getConstant(Fold);
1263      Ops.erase(Ops.begin()+1);  // Erase the folded element
1264      if (Ops.size() == 1) return Ops[0];
1265      LHSC = cast<SCEVConstant>(Ops[0]);
1266    }
1267
1268    // If we are left with a constant zero, strip it off.
1269    if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
1270      Ops.erase(Ops.begin());
1271      --Idx;
1272    }
1273  }
1274
1275  if (Ops.size() == 1) return Ops[0];
1276
1277  // Find the first UMax
1278  while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
1279    ++Idx;
1280
1281  // Check to see if one of the operands is a UMax. If so, expand its operands
1282  // onto our operand list, and recurse to simplify.
1283  if (Idx < Ops.size()) {
1284    bool DeletedUMax = false;
1285    while (SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
1286      Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
1287      Ops.erase(Ops.begin()+Idx);
1288      DeletedUMax = true;
1289    }
1290
1291    if (DeletedUMax)
1292      return getUMaxExpr(Ops);
1293  }
1294
1295  // Okay, check to see if the same value occurs in the operand list twice.  If
1296  // so, delete one.  Since we sorted the list, these values are required to
1297  // be adjacent.
1298  for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1299    if (Ops[i] == Ops[i+1]) {      //  X umax Y umax Y  -->  X umax Y
1300      Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1301      --i; --e;
1302    }
1303
1304  if (Ops.size() == 1) return Ops[0];
1305
1306  assert(!Ops.empty() && "Reduced umax down to nothing!");
1307
1308  // Okay, it looks like we really DO need a umax expr.  Check to see if we
1309  // already have one, otherwise create a new one.
1310  std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1311  SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scUMaxExpr,
1312                                                                 SCEVOps)];
1313  if (Result == 0) Result = new SCEVUMaxExpr(Ops);
1314  return Result;
1315}
1316
1317SCEVHandle ScalarEvolution::getUnknown(Value *V) {
1318  if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
1319    return getConstant(CI);
1320  SCEVUnknown *&Result = (*SCEVUnknowns)[V];
1321  if (Result == 0) Result = new SCEVUnknown(V);
1322  return Result;
1323}
1324
1325
1326//===----------------------------------------------------------------------===//
1327//             ScalarEvolutionsImpl Definition and Implementation
1328//===----------------------------------------------------------------------===//
1329//
1330/// ScalarEvolutionsImpl - This class implements the main driver for the scalar
1331/// evolution code.
1332///
1333namespace {
1334  struct VISIBILITY_HIDDEN ScalarEvolutionsImpl {
1335    /// SE - A reference to the public ScalarEvolution object.
1336    ScalarEvolution &SE;
1337
1338    /// F - The function we are analyzing.
1339    ///
1340    Function &F;
1341
1342    /// LI - The loop information for the function we are currently analyzing.
1343    ///
1344    LoopInfo &LI;
1345
1346    /// UnknownValue - This SCEV is used to represent unknown trip counts and
1347    /// things.
1348    SCEVHandle UnknownValue;
1349
1350    /// Scalars - This is a cache of the scalars we have analyzed so far.
1351    ///
1352    std::map<Value*, SCEVHandle> Scalars;
1353
1354    /// IterationCounts - Cache the iteration count of the loops for this
1355    /// function as they are computed.
1356    std::map<const Loop*, SCEVHandle> IterationCounts;
1357
1358    /// ConstantEvolutionLoopExitValue - This map contains entries for all of
1359    /// the PHI instructions that we attempt to compute constant evolutions for.
1360    /// This allows us to avoid potentially expensive recomputation of these
1361    /// properties.  An instruction maps to null if we are unable to compute its
1362    /// exit value.
1363    std::map<PHINode*, Constant*> ConstantEvolutionLoopExitValue;
1364
1365  public:
1366    ScalarEvolutionsImpl(ScalarEvolution &se, Function &f, LoopInfo &li)
1367      : SE(se), F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {}
1368
1369    /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1370    /// expression and create a new one.
1371    SCEVHandle getSCEV(Value *V);
1372
1373    /// hasSCEV - Return true if the SCEV for this value has already been
1374    /// computed.
1375    bool hasSCEV(Value *V) const {
1376      return Scalars.count(V);
1377    }
1378
1379    /// setSCEV - Insert the specified SCEV into the map of current SCEVs for
1380    /// the specified value.
1381    void setSCEV(Value *V, const SCEVHandle &H) {
1382      bool isNew = Scalars.insert(std::make_pair(V, H)).second;
1383      assert(isNew && "This entry already existed!");
1384    }
1385
1386
1387    /// getSCEVAtScope - Compute the value of the specified expression within
1388    /// the indicated loop (which may be null to indicate in no loop).  If the
1389    /// expression cannot be evaluated, return UnknownValue itself.
1390    SCEVHandle getSCEVAtScope(SCEV *V, const Loop *L);
1391
1392
1393    /// hasLoopInvariantIterationCount - Return true if the specified loop has
1394    /// an analyzable loop-invariant iteration count.
1395    bool hasLoopInvariantIterationCount(const Loop *L);
1396
1397    /// getIterationCount - If the specified loop has a predictable iteration
1398    /// count, return it.  Note that it is not valid to call this method on a
1399    /// loop without a loop-invariant iteration count.
1400    SCEVHandle getIterationCount(const Loop *L);
1401
1402    /// deleteValueFromRecords - This method should be called by the
1403    /// client before it removes a value from the program, to make sure
1404    /// that no dangling references are left around.
1405    void deleteValueFromRecords(Value *V);
1406
1407  private:
1408    /// createSCEV - We know that there is no SCEV for the specified value.
1409    /// Analyze the expression.
1410    SCEVHandle createSCEV(Value *V);
1411
1412    /// createNodeForPHI - Provide the special handling we need to analyze PHI
1413    /// SCEVs.
1414    SCEVHandle createNodeForPHI(PHINode *PN);
1415
1416    /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value
1417    /// for the specified instruction and replaces any references to the
1418    /// symbolic value SymName with the specified value.  This is used during
1419    /// PHI resolution.
1420    void ReplaceSymbolicValueWithConcrete(Instruction *I,
1421                                          const SCEVHandle &SymName,
1422                                          const SCEVHandle &NewVal);
1423
1424    /// ComputeIterationCount - Compute the number of times the specified loop
1425    /// will iterate.
1426    SCEVHandle ComputeIterationCount(const Loop *L);
1427
1428    /// ComputeLoadConstantCompareIterationCount - Given an exit condition of
1429    /// 'icmp op load X, cst', try to see if we can compute the trip count.
1430    SCEVHandle ComputeLoadConstantCompareIterationCount(LoadInst *LI,
1431                                                        Constant *RHS,
1432                                                        const Loop *L,
1433                                                        ICmpInst::Predicate p);
1434
1435    /// ComputeIterationCountExhaustively - If the trip is known to execute a
1436    /// constant number of times (the condition evolves only from constants),
1437    /// try to evaluate a few iterations of the loop until we get the exit
1438    /// condition gets a value of ExitWhen (true or false).  If we cannot
1439    /// evaluate the trip count of the loop, return UnknownValue.
1440    SCEVHandle ComputeIterationCountExhaustively(const Loop *L, Value *Cond,
1441                                                 bool ExitWhen);
1442
1443    /// HowFarToZero - Return the number of times a backedge comparing the
1444    /// specified value to zero will execute.  If not computable, return
1445    /// UnknownValue.
1446    SCEVHandle HowFarToZero(SCEV *V, const Loop *L);
1447
1448    /// HowFarToNonZero - Return the number of times a backedge checking the
1449    /// specified value for nonzero will execute.  If not computable, return
1450    /// UnknownValue.
1451    SCEVHandle HowFarToNonZero(SCEV *V, const Loop *L);
1452
1453    /// HowManyLessThans - Return the number of times a backedge containing the
1454    /// specified less-than comparison will execute.  If not computable, return
1455    /// UnknownValue. isSigned specifies whether the less-than is signed.
1456    SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L,
1457                                bool isSigned);
1458
1459    /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
1460    /// in the header of its containing loop, we know the loop executes a
1461    /// constant number of times, and the PHI node is just a recurrence
1462    /// involving constants, fold it.
1463    Constant *getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& Its,
1464                                                const Loop *L);
1465  };
1466}
1467
1468//===----------------------------------------------------------------------===//
1469//            Basic SCEV Analysis and PHI Idiom Recognition Code
1470//
1471
1472/// deleteValueFromRecords - This method should be called by the
1473/// client before it removes an instruction from the program, to make sure
1474/// that no dangling references are left around.
1475void ScalarEvolutionsImpl::deleteValueFromRecords(Value *V) {
1476  SmallVector<Value *, 16> Worklist;
1477
1478  if (Scalars.erase(V)) {
1479    if (PHINode *PN = dyn_cast<PHINode>(V))
1480      ConstantEvolutionLoopExitValue.erase(PN);
1481    Worklist.push_back(V);
1482  }
1483
1484  while (!Worklist.empty()) {
1485    Value *VV = Worklist.back();
1486    Worklist.pop_back();
1487
1488    for (Instruction::use_iterator UI = VV->use_begin(), UE = VV->use_end();
1489         UI != UE; ++UI) {
1490      Instruction *Inst = cast<Instruction>(*UI);
1491      if (Scalars.erase(Inst)) {
1492        if (PHINode *PN = dyn_cast<PHINode>(VV))
1493          ConstantEvolutionLoopExitValue.erase(PN);
1494        Worklist.push_back(Inst);
1495      }
1496    }
1497  }
1498}
1499
1500
1501/// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1502/// expression and create a new one.
1503SCEVHandle ScalarEvolutionsImpl::getSCEV(Value *V) {
1504  assert(V->getType() != Type::VoidTy && "Can't analyze void expressions!");
1505
1506  std::map<Value*, SCEVHandle>::iterator I = Scalars.find(V);
1507  if (I != Scalars.end()) return I->second;
1508  SCEVHandle S = createSCEV(V);
1509  Scalars.insert(std::make_pair(V, S));
1510  return S;
1511}
1512
1513/// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value for
1514/// the specified instruction and replaces any references to the symbolic value
1515/// SymName with the specified value.  This is used during PHI resolution.
1516void ScalarEvolutionsImpl::
1517ReplaceSymbolicValueWithConcrete(Instruction *I, const SCEVHandle &SymName,
1518                                 const SCEVHandle &NewVal) {
1519  std::map<Value*, SCEVHandle>::iterator SI = Scalars.find(I);
1520  if (SI == Scalars.end()) return;
1521
1522  SCEVHandle NV =
1523    SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal, SE);
1524  if (NV == SI->second) return;  // No change.
1525
1526  SI->second = NV;       // Update the scalars map!
1527
1528  // Any instruction values that use this instruction might also need to be
1529  // updated!
1530  for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1531       UI != E; ++UI)
1532    ReplaceSymbolicValueWithConcrete(cast<Instruction>(*UI), SymName, NewVal);
1533}
1534
1535/// createNodeForPHI - PHI nodes have two cases.  Either the PHI node exists in
1536/// a loop header, making it a potential recurrence, or it doesn't.
1537///
1538SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) {
1539  if (PN->getNumIncomingValues() == 2)  // The loops have been canonicalized.
1540    if (const Loop *L = LI.getLoopFor(PN->getParent()))
1541      if (L->getHeader() == PN->getParent()) {
1542        // If it lives in the loop header, it has two incoming values, one
1543        // from outside the loop, and one from inside.
1544        unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1545        unsigned BackEdge     = IncomingEdge^1;
1546
1547        // While we are analyzing this PHI node, handle its value symbolically.
1548        SCEVHandle SymbolicName = SE.getUnknown(PN);
1549        assert(Scalars.find(PN) == Scalars.end() &&
1550               "PHI node already processed?");
1551        Scalars.insert(std::make_pair(PN, SymbolicName));
1552
1553        // Using this symbolic name for the PHI, analyze the value coming around
1554        // the back-edge.
1555        SCEVHandle BEValue = getSCEV(PN->getIncomingValue(BackEdge));
1556
1557        // NOTE: If BEValue is loop invariant, we know that the PHI node just
1558        // has a special value for the first iteration of the loop.
1559
1560        // If the value coming around the backedge is an add with the symbolic
1561        // value we just inserted, then we found a simple induction variable!
1562        if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
1563          // If there is a single occurrence of the symbolic value, replace it
1564          // with a recurrence.
1565          unsigned FoundIndex = Add->getNumOperands();
1566          for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1567            if (Add->getOperand(i) == SymbolicName)
1568              if (FoundIndex == e) {
1569                FoundIndex = i;
1570                break;
1571              }
1572
1573          if (FoundIndex != Add->getNumOperands()) {
1574            // Create an add with everything but the specified operand.
1575            std::vector<SCEVHandle> Ops;
1576            for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1577              if (i != FoundIndex)
1578                Ops.push_back(Add->getOperand(i));
1579            SCEVHandle Accum = SE.getAddExpr(Ops);
1580
1581            // This is not a valid addrec if the step amount is varying each
1582            // loop iteration, but is not itself an addrec in this loop.
1583            if (Accum->isLoopInvariant(L) ||
1584                (isa<SCEVAddRecExpr>(Accum) &&
1585                 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
1586              SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1587              SCEVHandle PHISCEV  = SE.getAddRecExpr(StartVal, Accum, L);
1588
1589              // Okay, for the entire analysis of this edge we assumed the PHI
1590              // to be symbolic.  We now need to go back and update all of the
1591              // entries for the scalars that use the PHI (except for the PHI
1592              // itself) to use the new analyzed value instead of the "symbolic"
1593              // value.
1594              ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1595              return PHISCEV;
1596            }
1597          }
1598        } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(BEValue)) {
1599          // Otherwise, this could be a loop like this:
1600          //     i = 0;  for (j = 1; ..; ++j) { ....  i = j; }
1601          // In this case, j = {1,+,1}  and BEValue is j.
1602          // Because the other in-value of i (0) fits the evolution of BEValue
1603          // i really is an addrec evolution.
1604          if (AddRec->getLoop() == L && AddRec->isAffine()) {
1605            SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1606
1607            // If StartVal = j.start - j.stride, we can use StartVal as the
1608            // initial step of the addrec evolution.
1609            if (StartVal == SE.getMinusSCEV(AddRec->getOperand(0),
1610                                            AddRec->getOperand(1))) {
1611              SCEVHandle PHISCEV =
1612                 SE.getAddRecExpr(StartVal, AddRec->getOperand(1), L);
1613
1614              // Okay, for the entire analysis of this edge we assumed the PHI
1615              // to be symbolic.  We now need to go back and update all of the
1616              // entries for the scalars that use the PHI (except for the PHI
1617              // itself) to use the new analyzed value instead of the "symbolic"
1618              // value.
1619              ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1620              return PHISCEV;
1621            }
1622          }
1623        }
1624
1625        return SymbolicName;
1626      }
1627
1628  // If it's not a loop phi, we can't handle it yet.
1629  return SE.getUnknown(PN);
1630}
1631
1632/// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
1633/// guaranteed to end in (at every loop iteration).  It is, at the same time,
1634/// the minimum number of times S is divisible by 2.  For example, given {4,+,8}
1635/// it returns 2.  If S is guaranteed to be 0, it returns the bitwidth of S.
1636static uint32_t GetMinTrailingZeros(SCEVHandle S) {
1637  if (SCEVConstant *C = dyn_cast<SCEVConstant>(S))
1638    return C->getValue()->getValue().countTrailingZeros();
1639
1640  if (SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
1641    return std::min(GetMinTrailingZeros(T->getOperand()), T->getBitWidth());
1642
1643  if (SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
1644    uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
1645    return OpRes == E->getOperand()->getBitWidth() ? E->getBitWidth() : OpRes;
1646  }
1647
1648  if (SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
1649    uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
1650    return OpRes == E->getOperand()->getBitWidth() ? E->getBitWidth() : OpRes;
1651  }
1652
1653  if (SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
1654    // The result is the min of all operands results.
1655    uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
1656    for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
1657      MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
1658    return MinOpRes;
1659  }
1660
1661  if (SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
1662    // The result is the sum of all operands results.
1663    uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
1664    uint32_t BitWidth = M->getBitWidth();
1665    for (unsigned i = 1, e = M->getNumOperands();
1666         SumOpRes != BitWidth && i != e; ++i)
1667      SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
1668                          BitWidth);
1669    return SumOpRes;
1670  }
1671
1672  if (SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
1673    // The result is the min of all operands results.
1674    uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
1675    for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
1676      MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
1677    return MinOpRes;
1678  }
1679
1680  if (SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
1681    // The result is the min of all operands results.
1682    uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
1683    for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
1684      MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
1685    return MinOpRes;
1686  }
1687
1688  if (SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
1689    // The result is the min of all operands results.
1690    uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
1691    for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
1692      MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
1693    return MinOpRes;
1694  }
1695
1696  // SCEVUDivExpr, SCEVUnknown
1697  return 0;
1698}
1699
1700/// createSCEV - We know that there is no SCEV for the specified value.
1701/// Analyze the expression.
1702///
1703SCEVHandle ScalarEvolutionsImpl::createSCEV(Value *V) {
1704  if (!isa<IntegerType>(V->getType()))
1705    return SE.getUnknown(V);
1706
1707  unsigned Opcode = Instruction::UserOp1;
1708  if (Instruction *I = dyn_cast<Instruction>(V))
1709    Opcode = I->getOpcode();
1710  else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1711    Opcode = CE->getOpcode();
1712  else
1713    return SE.getUnknown(V);
1714
1715  User *U = cast<User>(V);
1716  switch (Opcode) {
1717  case Instruction::Add:
1718    return SE.getAddExpr(getSCEV(U->getOperand(0)),
1719                         getSCEV(U->getOperand(1)));
1720  case Instruction::Mul:
1721    return SE.getMulExpr(getSCEV(U->getOperand(0)),
1722                         getSCEV(U->getOperand(1)));
1723  case Instruction::UDiv:
1724    return SE.getUDivExpr(getSCEV(U->getOperand(0)),
1725                          getSCEV(U->getOperand(1)));
1726  case Instruction::Sub:
1727    return SE.getMinusSCEV(getSCEV(U->getOperand(0)),
1728                           getSCEV(U->getOperand(1)));
1729  case Instruction::Or:
1730    // If the RHS of the Or is a constant, we may have something like:
1731    // X*4+1 which got turned into X*4|1.  Handle this as an Add so loop
1732    // optimizations will transparently handle this case.
1733    //
1734    // In order for this transformation to be safe, the LHS must be of the
1735    // form X*(2^n) and the Or constant must be less than 2^n.
1736    if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
1737      SCEVHandle LHS = getSCEV(U->getOperand(0));
1738      const APInt &CIVal = CI->getValue();
1739      if (GetMinTrailingZeros(LHS) >=
1740          (CIVal.getBitWidth() - CIVal.countLeadingZeros()))
1741        return SE.getAddExpr(LHS, getSCEV(U->getOperand(1)));
1742    }
1743    break;
1744  case Instruction::Xor:
1745    // If the RHS of the xor is a signbit, then this is just an add.
1746    // Instcombine turns add of signbit into xor as a strength reduction step.
1747    if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
1748      if (CI->getValue().isSignBit())
1749        return SE.getAddExpr(getSCEV(U->getOperand(0)),
1750                             getSCEV(U->getOperand(1)));
1751      else if (CI->isAllOnesValue())
1752        return SE.getNotSCEV(getSCEV(U->getOperand(0)));
1753    }
1754    break;
1755
1756  case Instruction::Shl:
1757    // Turn shift left of a constant amount into a multiply.
1758    if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
1759      uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
1760      Constant *X = ConstantInt::get(
1761        APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
1762      return SE.getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
1763    }
1764    break;
1765
1766  case Instruction::Trunc:
1767    return SE.getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
1768
1769  case Instruction::ZExt:
1770    return SE.getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
1771
1772  case Instruction::SExt:
1773    return SE.getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
1774
1775  case Instruction::BitCast:
1776    // BitCasts are no-op casts so we just eliminate the cast.
1777    if (U->getType()->isInteger() &&
1778        U->getOperand(0)->getType()->isInteger())
1779      return getSCEV(U->getOperand(0));
1780    break;
1781
1782  case Instruction::PHI:
1783    return createNodeForPHI(cast<PHINode>(U));
1784
1785  case Instruction::Select:
1786    // This could be a smax or umax that was lowered earlier.
1787    // Try to recover it.
1788    if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
1789      Value *LHS = ICI->getOperand(0);
1790      Value *RHS = ICI->getOperand(1);
1791      switch (ICI->getPredicate()) {
1792      case ICmpInst::ICMP_SLT:
1793      case ICmpInst::ICMP_SLE:
1794        std::swap(LHS, RHS);
1795        // fall through
1796      case ICmpInst::ICMP_SGT:
1797      case ICmpInst::ICMP_SGE:
1798        if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
1799          return SE.getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
1800        else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
1801          // -smax(-x, -y) == smin(x, y).
1802          return SE.getNegativeSCEV(SE.getSMaxExpr(
1803                                        SE.getNegativeSCEV(getSCEV(LHS)),
1804                                        SE.getNegativeSCEV(getSCEV(RHS))));
1805        break;
1806      case ICmpInst::ICMP_ULT:
1807      case ICmpInst::ICMP_ULE:
1808        std::swap(LHS, RHS);
1809        // fall through
1810      case ICmpInst::ICMP_UGT:
1811      case ICmpInst::ICMP_UGE:
1812        if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
1813          return SE.getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
1814        else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
1815          // ~umax(~x, ~y) == umin(x, y)
1816          return SE.getNotSCEV(SE.getUMaxExpr(SE.getNotSCEV(getSCEV(LHS)),
1817                                              SE.getNotSCEV(getSCEV(RHS))));
1818        break;
1819      default:
1820        break;
1821      }
1822    }
1823
1824  default: // We cannot analyze this expression.
1825    break;
1826  }
1827
1828  return SE.getUnknown(V);
1829}
1830
1831
1832
1833//===----------------------------------------------------------------------===//
1834//                   Iteration Count Computation Code
1835//
1836
1837/// getIterationCount - If the specified loop has a predictable iteration
1838/// count, return it.  Note that it is not valid to call this method on a
1839/// loop without a loop-invariant iteration count.
1840SCEVHandle ScalarEvolutionsImpl::getIterationCount(const Loop *L) {
1841  std::map<const Loop*, SCEVHandle>::iterator I = IterationCounts.find(L);
1842  if (I == IterationCounts.end()) {
1843    SCEVHandle ItCount = ComputeIterationCount(L);
1844    I = IterationCounts.insert(std::make_pair(L, ItCount)).first;
1845    if (ItCount != UnknownValue) {
1846      assert(ItCount->isLoopInvariant(L) &&
1847             "Computed trip count isn't loop invariant for loop!");
1848      ++NumTripCountsComputed;
1849    } else if (isa<PHINode>(L->getHeader()->begin())) {
1850      // Only count loops that have phi nodes as not being computable.
1851      ++NumTripCountsNotComputed;
1852    }
1853  }
1854  return I->second;
1855}
1856
1857/// ComputeIterationCount - Compute the number of times the specified loop
1858/// will iterate.
1859SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) {
1860  // If the loop has a non-one exit block count, we can't analyze it.
1861  SmallVector<BasicBlock*, 8> ExitBlocks;
1862  L->getExitBlocks(ExitBlocks);
1863  if (ExitBlocks.size() != 1) return UnknownValue;
1864
1865  // Okay, there is one exit block.  Try to find the condition that causes the
1866  // loop to be exited.
1867  BasicBlock *ExitBlock = ExitBlocks[0];
1868
1869  BasicBlock *ExitingBlock = 0;
1870  for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock);
1871       PI != E; ++PI)
1872    if (L->contains(*PI)) {
1873      if (ExitingBlock == 0)
1874        ExitingBlock = *PI;
1875      else
1876        return UnknownValue;   // More than one block exiting!
1877    }
1878  assert(ExitingBlock && "No exits from loop, something is broken!");
1879
1880  // Okay, we've computed the exiting block.  See what condition causes us to
1881  // exit.
1882  //
1883  // FIXME: we should be able to handle switch instructions (with a single exit)
1884  BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1885  if (ExitBr == 0) return UnknownValue;
1886  assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
1887
1888  // At this point, we know we have a conditional branch that determines whether
1889  // the loop is exited.  However, we don't know if the branch is executed each
1890  // time through the loop.  If not, then the execution count of the branch will
1891  // not be equal to the trip count of the loop.
1892  //
1893  // Currently we check for this by checking to see if the Exit branch goes to
1894  // the loop header.  If so, we know it will always execute the same number of
1895  // times as the loop.  We also handle the case where the exit block *is* the
1896  // loop header.  This is common for un-rotated loops.  More extensive analysis
1897  // could be done to handle more cases here.
1898  if (ExitBr->getSuccessor(0) != L->getHeader() &&
1899      ExitBr->getSuccessor(1) != L->getHeader() &&
1900      ExitBr->getParent() != L->getHeader())
1901    return UnknownValue;
1902
1903  ICmpInst *ExitCond = dyn_cast<ICmpInst>(ExitBr->getCondition());
1904
1905  // If it's not an integer comparison then compute it the hard way.
1906  // Note that ICmpInst deals with pointer comparisons too so we must check
1907  // the type of the operand.
1908  if (ExitCond == 0 || isa<PointerType>(ExitCond->getOperand(0)->getType()))
1909    return ComputeIterationCountExhaustively(L, ExitBr->getCondition(),
1910                                          ExitBr->getSuccessor(0) == ExitBlock);
1911
1912  // If the condition was exit on true, convert the condition to exit on false
1913  ICmpInst::Predicate Cond;
1914  if (ExitBr->getSuccessor(1) == ExitBlock)
1915    Cond = ExitCond->getPredicate();
1916  else
1917    Cond = ExitCond->getInversePredicate();
1918
1919  // Handle common loops like: for (X = "string"; *X; ++X)
1920  if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
1921    if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
1922      SCEVHandle ItCnt =
1923        ComputeLoadConstantCompareIterationCount(LI, RHS, L, Cond);
1924      if (!isa<SCEVCouldNotCompute>(ItCnt)) return ItCnt;
1925    }
1926
1927  SCEVHandle LHS = getSCEV(ExitCond->getOperand(0));
1928  SCEVHandle RHS = getSCEV(ExitCond->getOperand(1));
1929
1930  // Try to evaluate any dependencies out of the loop.
1931  SCEVHandle Tmp = getSCEVAtScope(LHS, L);
1932  if (!isa<SCEVCouldNotCompute>(Tmp)) LHS = Tmp;
1933  Tmp = getSCEVAtScope(RHS, L);
1934  if (!isa<SCEVCouldNotCompute>(Tmp)) RHS = Tmp;
1935
1936  // At this point, we would like to compute how many iterations of the
1937  // loop the predicate will return true for these inputs.
1938  if (isa<SCEVConstant>(LHS) && !isa<SCEVConstant>(RHS)) {
1939    // If there is a constant, force it into the RHS.
1940    std::swap(LHS, RHS);
1941    Cond = ICmpInst::getSwappedPredicate(Cond);
1942  }
1943
1944  // FIXME: think about handling pointer comparisons!  i.e.:
1945  // while (P != P+100) ++P;
1946
1947  // If we have a comparison of a chrec against a constant, try to use value
1948  // ranges to answer this query.
1949  if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
1950    if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
1951      if (AddRec->getLoop() == L) {
1952        // Form the comparison range using the constant of the correct type so
1953        // that the ConstantRange class knows to do a signed or unsigned
1954        // comparison.
1955        ConstantInt *CompVal = RHSC->getValue();
1956        const Type *RealTy = ExitCond->getOperand(0)->getType();
1957        CompVal = dyn_cast<ConstantInt>(
1958          ConstantExpr::getBitCast(CompVal, RealTy));
1959        if (CompVal) {
1960          // Form the constant range.
1961          ConstantRange CompRange(
1962              ICmpInst::makeConstantRange(Cond, CompVal->getValue()));
1963
1964          SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange, SE);
1965          if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
1966        }
1967      }
1968
1969  switch (Cond) {
1970  case ICmpInst::ICMP_NE: {                     // while (X != Y)
1971    // Convert to: while (X-Y != 0)
1972    SCEVHandle TC = HowFarToZero(SE.getMinusSCEV(LHS, RHS), L);
1973    if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1974    break;
1975  }
1976  case ICmpInst::ICMP_EQ: {
1977    // Convert to: while (X-Y == 0)           // while (X == Y)
1978    SCEVHandle TC = HowFarToNonZero(SE.getMinusSCEV(LHS, RHS), L);
1979    if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1980    break;
1981  }
1982  case ICmpInst::ICMP_SLT: {
1983    SCEVHandle TC = HowManyLessThans(LHS, RHS, L, true);
1984    if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1985    break;
1986  }
1987  case ICmpInst::ICMP_SGT: {
1988    SCEVHandle TC = HowManyLessThans(SE.getNegativeSCEV(LHS),
1989                                     SE.getNegativeSCEV(RHS), L, true);
1990    if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1991    break;
1992  }
1993  case ICmpInst::ICMP_ULT: {
1994    SCEVHandle TC = HowManyLessThans(LHS, RHS, L, false);
1995    if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1996    break;
1997  }
1998  case ICmpInst::ICMP_UGT: {
1999    SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
2000                                     SE.getNotSCEV(RHS), L, false);
2001    if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2002    break;
2003  }
2004  default:
2005#if 0
2006    cerr << "ComputeIterationCount ";
2007    if (ExitCond->getOperand(0)->getType()->isUnsigned())
2008      cerr << "[unsigned] ";
2009    cerr << *LHS << "   "
2010         << Instruction::getOpcodeName(Instruction::ICmp)
2011         << "   " << *RHS << "\n";
2012#endif
2013    break;
2014  }
2015  return ComputeIterationCountExhaustively(L, ExitCond,
2016                                       ExitBr->getSuccessor(0) == ExitBlock);
2017}
2018
2019static ConstantInt *
2020EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
2021                                ScalarEvolution &SE) {
2022  SCEVHandle InVal = SE.getConstant(C);
2023  SCEVHandle Val = AddRec->evaluateAtIteration(InVal, SE);
2024  assert(isa<SCEVConstant>(Val) &&
2025         "Evaluation of SCEV at constant didn't fold correctly?");
2026  return cast<SCEVConstant>(Val)->getValue();
2027}
2028
2029/// GetAddressedElementFromGlobal - Given a global variable with an initializer
2030/// and a GEP expression (missing the pointer index) indexing into it, return
2031/// the addressed element of the initializer or null if the index expression is
2032/// invalid.
2033static Constant *
2034GetAddressedElementFromGlobal(GlobalVariable *GV,
2035                              const std::vector<ConstantInt*> &Indices) {
2036  Constant *Init = GV->getInitializer();
2037  for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
2038    uint64_t Idx = Indices[i]->getZExtValue();
2039    if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2040      assert(Idx < CS->getNumOperands() && "Bad struct index!");
2041      Init = cast<Constant>(CS->getOperand(Idx));
2042    } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2043      if (Idx >= CA->getNumOperands()) return 0;  // Bogus program
2044      Init = cast<Constant>(CA->getOperand(Idx));
2045    } else if (isa<ConstantAggregateZero>(Init)) {
2046      if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2047        assert(Idx < STy->getNumElements() && "Bad struct index!");
2048        Init = Constant::getNullValue(STy->getElementType(Idx));
2049      } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
2050        if (Idx >= ATy->getNumElements()) return 0;  // Bogus program
2051        Init = Constant::getNullValue(ATy->getElementType());
2052      } else {
2053        assert(0 && "Unknown constant aggregate type!");
2054      }
2055      return 0;
2056    } else {
2057      return 0; // Unknown initializer type
2058    }
2059  }
2060  return Init;
2061}
2062
2063/// ComputeLoadConstantCompareIterationCount - Given an exit condition of
2064/// 'icmp op load X, cst', try to see if we can compute the trip count.
2065SCEVHandle ScalarEvolutionsImpl::
2066ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS,
2067                                         const Loop *L,
2068                                         ICmpInst::Predicate predicate) {
2069  if (LI->isVolatile()) return UnknownValue;
2070
2071  // Check to see if the loaded pointer is a getelementptr of a global.
2072  GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
2073  if (!GEP) return UnknownValue;
2074
2075  // Make sure that it is really a constant global we are gepping, with an
2076  // initializer, and make sure the first IDX is really 0.
2077  GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
2078  if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
2079      GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
2080      !cast<Constant>(GEP->getOperand(1))->isNullValue())
2081    return UnknownValue;
2082
2083  // Okay, we allow one non-constant index into the GEP instruction.
2084  Value *VarIdx = 0;
2085  std::vector<ConstantInt*> Indexes;
2086  unsigned VarIdxNum = 0;
2087  for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
2088    if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
2089      Indexes.push_back(CI);
2090    } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
2091      if (VarIdx) return UnknownValue;  // Multiple non-constant idx's.
2092      VarIdx = GEP->getOperand(i);
2093      VarIdxNum = i-2;
2094      Indexes.push_back(0);
2095    }
2096
2097  // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
2098  // Check to see if X is a loop variant variable value now.
2099  SCEVHandle Idx = getSCEV(VarIdx);
2100  SCEVHandle Tmp = getSCEVAtScope(Idx, L);
2101  if (!isa<SCEVCouldNotCompute>(Tmp)) Idx = Tmp;
2102
2103  // We can only recognize very limited forms of loop index expressions, in
2104  // particular, only affine AddRec's like {C1,+,C2}.
2105  SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
2106  if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
2107      !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
2108      !isa<SCEVConstant>(IdxExpr->getOperand(1)))
2109    return UnknownValue;
2110
2111  unsigned MaxSteps = MaxBruteForceIterations;
2112  for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
2113    ConstantInt *ItCst =
2114      ConstantInt::get(IdxExpr->getType(), IterationNum);
2115    ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, SE);
2116
2117    // Form the GEP offset.
2118    Indexes[VarIdxNum] = Val;
2119
2120    Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
2121    if (Result == 0) break;  // Cannot compute!
2122
2123    // Evaluate the condition for this iteration.
2124    Result = ConstantExpr::getICmp(predicate, Result, RHS);
2125    if (!isa<ConstantInt>(Result)) break;  // Couldn't decide for sure
2126    if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
2127#if 0
2128      cerr << "\n***\n*** Computed loop count " << *ItCst
2129           << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
2130           << "***\n";
2131#endif
2132      ++NumArrayLenItCounts;
2133      return SE.getConstant(ItCst);   // Found terminating iteration!
2134    }
2135  }
2136  return UnknownValue;
2137}
2138
2139
2140/// CanConstantFold - Return true if we can constant fold an instruction of the
2141/// specified type, assuming that all operands were constants.
2142static bool CanConstantFold(const Instruction *I) {
2143  if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
2144      isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
2145    return true;
2146
2147  if (const CallInst *CI = dyn_cast<CallInst>(I))
2148    if (const Function *F = CI->getCalledFunction())
2149      return canConstantFoldCallTo(F);
2150  return false;
2151}
2152
2153/// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
2154/// in the loop that V is derived from.  We allow arbitrary operations along the
2155/// way, but the operands of an operation must either be constants or a value
2156/// derived from a constant PHI.  If this expression does not fit with these
2157/// constraints, return null.
2158static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
2159  // If this is not an instruction, or if this is an instruction outside of the
2160  // loop, it can't be derived from a loop PHI.
2161  Instruction *I = dyn_cast<Instruction>(V);
2162  if (I == 0 || !L->contains(I->getParent())) return 0;
2163
2164  if (PHINode *PN = dyn_cast<PHINode>(I)) {
2165    if (L->getHeader() == I->getParent())
2166      return PN;
2167    else
2168      // We don't currently keep track of the control flow needed to evaluate
2169      // PHIs, so we cannot handle PHIs inside of loops.
2170      return 0;
2171  }
2172
2173  // If we won't be able to constant fold this expression even if the operands
2174  // are constants, return early.
2175  if (!CanConstantFold(I)) return 0;
2176
2177  // Otherwise, we can evaluate this instruction if all of its operands are
2178  // constant or derived from a PHI node themselves.
2179  PHINode *PHI = 0;
2180  for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
2181    if (!(isa<Constant>(I->getOperand(Op)) ||
2182          isa<GlobalValue>(I->getOperand(Op)))) {
2183      PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
2184      if (P == 0) return 0;  // Not evolving from PHI
2185      if (PHI == 0)
2186        PHI = P;
2187      else if (PHI != P)
2188        return 0;  // Evolving from multiple different PHIs.
2189    }
2190
2191  // This is a expression evolving from a constant PHI!
2192  return PHI;
2193}
2194
2195/// EvaluateExpression - Given an expression that passes the
2196/// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
2197/// in the loop has the value PHIVal.  If we can't fold this expression for some
2198/// reason, return null.
2199static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
2200  if (isa<PHINode>(V)) return PHIVal;
2201  if (Constant *C = dyn_cast<Constant>(V)) return C;
2202  Instruction *I = cast<Instruction>(V);
2203
2204  std::vector<Constant*> Operands;
2205  Operands.resize(I->getNumOperands());
2206
2207  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2208    Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal);
2209    if (Operands[i] == 0) return 0;
2210  }
2211
2212  if (const CmpInst *CI = dyn_cast<CmpInst>(I))
2213    return ConstantFoldCompareInstOperands(CI->getPredicate(),
2214                                           &Operands[0], Operands.size());
2215  else
2216    return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
2217                                    &Operands[0], Operands.size());
2218}
2219
2220/// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
2221/// in the header of its containing loop, we know the loop executes a
2222/// constant number of times, and the PHI node is just a recurrence
2223/// involving constants, fold it.
2224Constant *ScalarEvolutionsImpl::
2225getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& Its, const Loop *L){
2226  std::map<PHINode*, Constant*>::iterator I =
2227    ConstantEvolutionLoopExitValue.find(PN);
2228  if (I != ConstantEvolutionLoopExitValue.end())
2229    return I->second;
2230
2231  if (Its.ugt(APInt(Its.getBitWidth(),MaxBruteForceIterations)))
2232    return ConstantEvolutionLoopExitValue[PN] = 0;  // Not going to evaluate it.
2233
2234  Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
2235
2236  // Since the loop is canonicalized, the PHI node must have two entries.  One
2237  // entry must be a constant (coming in from outside of the loop), and the
2238  // second must be derived from the same PHI.
2239  bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
2240  Constant *StartCST =
2241    dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
2242  if (StartCST == 0)
2243    return RetVal = 0;  // Must be a constant.
2244
2245  Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
2246  PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
2247  if (PN2 != PN)
2248    return RetVal = 0;  // Not derived from same PHI.
2249
2250  // Execute the loop symbolically to determine the exit value.
2251  if (Its.getActiveBits() >= 32)
2252    return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
2253
2254  unsigned NumIterations = Its.getZExtValue(); // must be in range
2255  unsigned IterationNum = 0;
2256  for (Constant *PHIVal = StartCST; ; ++IterationNum) {
2257    if (IterationNum == NumIterations)
2258      return RetVal = PHIVal;  // Got exit value!
2259
2260    // Compute the value of the PHI node for the next iteration.
2261    Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
2262    if (NextPHI == PHIVal)
2263      return RetVal = NextPHI;  // Stopped evolving!
2264    if (NextPHI == 0)
2265      return 0;        // Couldn't evaluate!
2266    PHIVal = NextPHI;
2267  }
2268}
2269
2270/// ComputeIterationCountExhaustively - If the trip is known to execute a
2271/// constant number of times (the condition evolves only from constants),
2272/// try to evaluate a few iterations of the loop until we get the exit
2273/// condition gets a value of ExitWhen (true or false).  If we cannot
2274/// evaluate the trip count of the loop, return UnknownValue.
2275SCEVHandle ScalarEvolutionsImpl::
2276ComputeIterationCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) {
2277  PHINode *PN = getConstantEvolvingPHI(Cond, L);
2278  if (PN == 0) return UnknownValue;
2279
2280  // Since the loop is canonicalized, the PHI node must have two entries.  One
2281  // entry must be a constant (coming in from outside of the loop), and the
2282  // second must be derived from the same PHI.
2283  bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
2284  Constant *StartCST =
2285    dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
2286  if (StartCST == 0) return UnknownValue;  // Must be a constant.
2287
2288  Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
2289  PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
2290  if (PN2 != PN) return UnknownValue;  // Not derived from same PHI.
2291
2292  // Okay, we find a PHI node that defines the trip count of this loop.  Execute
2293  // the loop symbolically to determine when the condition gets a value of
2294  // "ExitWhen".
2295  unsigned IterationNum = 0;
2296  unsigned MaxIterations = MaxBruteForceIterations;   // Limit analysis.
2297  for (Constant *PHIVal = StartCST;
2298       IterationNum != MaxIterations; ++IterationNum) {
2299    ConstantInt *CondVal =
2300      dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal));
2301
2302    // Couldn't symbolically evaluate.
2303    if (!CondVal) return UnknownValue;
2304
2305    if (CondVal->getValue() == uint64_t(ExitWhen)) {
2306      ConstantEvolutionLoopExitValue[PN] = PHIVal;
2307      ++NumBruteForceTripCountsComputed;
2308      return SE.getConstant(ConstantInt::get(Type::Int32Ty, IterationNum));
2309    }
2310
2311    // Compute the value of the PHI node for the next iteration.
2312    Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
2313    if (NextPHI == 0 || NextPHI == PHIVal)
2314      return UnknownValue;  // Couldn't evaluate or not making progress...
2315    PHIVal = NextPHI;
2316  }
2317
2318  // Too many iterations were needed to evaluate.
2319  return UnknownValue;
2320}
2321
2322/// getSCEVAtScope - Compute the value of the specified expression within the
2323/// indicated loop (which may be null to indicate in no loop).  If the
2324/// expression cannot be evaluated, return UnknownValue.
2325SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) {
2326  // FIXME: this should be turned into a virtual method on SCEV!
2327
2328  if (isa<SCEVConstant>(V)) return V;
2329
2330  // If this instruction is evolved from a constant-evolving PHI, compute the
2331  // exit value from the loop without using SCEVs.
2332  if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
2333    if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
2334      const Loop *LI = this->LI[I->getParent()];
2335      if (LI && LI->getParentLoop() == L)  // Looking for loop exit value.
2336        if (PHINode *PN = dyn_cast<PHINode>(I))
2337          if (PN->getParent() == LI->getHeader()) {
2338            // Okay, there is no closed form solution for the PHI node.  Check
2339            // to see if the loop that contains it has a known iteration count.
2340            // If so, we may be able to force computation of the exit value.
2341            SCEVHandle IterationCount = getIterationCount(LI);
2342            if (SCEVConstant *ICC = dyn_cast<SCEVConstant>(IterationCount)) {
2343              // Okay, we know how many times the containing loop executes.  If
2344              // this is a constant evolving PHI node, get the final value at
2345              // the specified iteration number.
2346              Constant *RV = getConstantEvolutionLoopExitValue(PN,
2347                                                    ICC->getValue()->getValue(),
2348                                                               LI);
2349              if (RV) return SE.getUnknown(RV);
2350            }
2351          }
2352
2353      // Okay, this is an expression that we cannot symbolically evaluate
2354      // into a SCEV.  Check to see if it's possible to symbolically evaluate
2355      // the arguments into constants, and if so, try to constant propagate the
2356      // result.  This is particularly useful for computing loop exit values.
2357      if (CanConstantFold(I)) {
2358        std::vector<Constant*> Operands;
2359        Operands.reserve(I->getNumOperands());
2360        for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2361          Value *Op = I->getOperand(i);
2362          if (Constant *C = dyn_cast<Constant>(Op)) {
2363            Operands.push_back(C);
2364          } else {
2365            // If any of the operands is non-constant and if they are
2366            // non-integer, don't even try to analyze them with scev techniques.
2367            if (!isa<IntegerType>(Op->getType()))
2368              return V;
2369
2370            SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L);
2371            if (SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
2372              Operands.push_back(ConstantExpr::getIntegerCast(SC->getValue(),
2373                                                              Op->getType(),
2374                                                              false));
2375            else if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
2376              if (Constant *C = dyn_cast<Constant>(SU->getValue()))
2377                Operands.push_back(ConstantExpr::getIntegerCast(C,
2378                                                                Op->getType(),
2379                                                                false));
2380              else
2381                return V;
2382            } else {
2383              return V;
2384            }
2385          }
2386        }
2387
2388        Constant *C;
2389        if (const CmpInst *CI = dyn_cast<CmpInst>(I))
2390          C = ConstantFoldCompareInstOperands(CI->getPredicate(),
2391                                              &Operands[0], Operands.size());
2392        else
2393          C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
2394                                       &Operands[0], Operands.size());
2395        return SE.getUnknown(C);
2396      }
2397    }
2398
2399    // This is some other type of SCEVUnknown, just return it.
2400    return V;
2401  }
2402
2403  if (SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
2404    // Avoid performing the look-up in the common case where the specified
2405    // expression has no loop-variant portions.
2406    for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
2407      SCEVHandle OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
2408      if (OpAtScope != Comm->getOperand(i)) {
2409        if (OpAtScope == UnknownValue) return UnknownValue;
2410        // Okay, at least one of these operands is loop variant but might be
2411        // foldable.  Build a new instance of the folded commutative expression.
2412        std::vector<SCEVHandle> NewOps(Comm->op_begin(), Comm->op_begin()+i);
2413        NewOps.push_back(OpAtScope);
2414
2415        for (++i; i != e; ++i) {
2416          OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
2417          if (OpAtScope == UnknownValue) return UnknownValue;
2418          NewOps.push_back(OpAtScope);
2419        }
2420        if (isa<SCEVAddExpr>(Comm))
2421          return SE.getAddExpr(NewOps);
2422        if (isa<SCEVMulExpr>(Comm))
2423          return SE.getMulExpr(NewOps);
2424        if (isa<SCEVSMaxExpr>(Comm))
2425          return SE.getSMaxExpr(NewOps);
2426        if (isa<SCEVUMaxExpr>(Comm))
2427          return SE.getUMaxExpr(NewOps);
2428        assert(0 && "Unknown commutative SCEV type!");
2429      }
2430    }
2431    // If we got here, all operands are loop invariant.
2432    return Comm;
2433  }
2434
2435  if (SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
2436    SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L);
2437    if (LHS == UnknownValue) return LHS;
2438    SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L);
2439    if (RHS == UnknownValue) return RHS;
2440    if (LHS == Div->getLHS() && RHS == Div->getRHS())
2441      return Div;   // must be loop invariant
2442    return SE.getUDivExpr(LHS, RHS);
2443  }
2444
2445  // If this is a loop recurrence for a loop that does not contain L, then we
2446  // are dealing with the final value computed by the loop.
2447  if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
2448    if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
2449      // To evaluate this recurrence, we need to know how many times the AddRec
2450      // loop iterates.  Compute this now.
2451      SCEVHandle IterationCount = getIterationCount(AddRec->getLoop());
2452      if (IterationCount == UnknownValue) return UnknownValue;
2453      IterationCount = SE.getTruncateOrZeroExtend(IterationCount,
2454                                                  AddRec->getType());
2455
2456      // If the value is affine, simplify the expression evaluation to just
2457      // Start + Step*IterationCount.
2458      if (AddRec->isAffine())
2459        return SE.getAddExpr(AddRec->getStart(),
2460                             SE.getMulExpr(IterationCount,
2461                                           AddRec->getOperand(1)));
2462
2463      // Otherwise, evaluate it the hard way.
2464      return AddRec->evaluateAtIteration(IterationCount, SE);
2465    }
2466    return UnknownValue;
2467  }
2468
2469  //assert(0 && "Unknown SCEV type!");
2470  return UnknownValue;
2471}
2472
2473
2474/// SolveQuadraticEquation - Find the roots of the quadratic equation for the
2475/// given quadratic chrec {L,+,M,+,N}.  This returns either the two roots (which
2476/// might be the same) or two SCEVCouldNotCompute objects.
2477///
2478static std::pair<SCEVHandle,SCEVHandle>
2479SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
2480  assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
2481  SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
2482  SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
2483  SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
2484
2485  // We currently can only solve this if the coefficients are constants.
2486  if (!LC || !MC || !NC) {
2487    SCEV *CNC = new SCEVCouldNotCompute();
2488    return std::make_pair(CNC, CNC);
2489  }
2490
2491  uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
2492  const APInt &L = LC->getValue()->getValue();
2493  const APInt &M = MC->getValue()->getValue();
2494  const APInt &N = NC->getValue()->getValue();
2495  APInt Two(BitWidth, 2);
2496  APInt Four(BitWidth, 4);
2497
2498  {
2499    using namespace APIntOps;
2500    const APInt& C = L;
2501    // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
2502    // The B coefficient is M-N/2
2503    APInt B(M);
2504    B -= sdiv(N,Two);
2505
2506    // The A coefficient is N/2
2507    APInt A(N.sdiv(Two));
2508
2509    // Compute the B^2-4ac term.
2510    APInt SqrtTerm(B);
2511    SqrtTerm *= B;
2512    SqrtTerm -= Four * (A * C);
2513
2514    // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
2515    // integer value or else APInt::sqrt() will assert.
2516    APInt SqrtVal(SqrtTerm.sqrt());
2517
2518    // Compute the two solutions for the quadratic formula.
2519    // The divisions must be performed as signed divisions.
2520    APInt NegB(-B);
2521    APInt TwoA( A << 1 );
2522    ConstantInt *Solution1 = ConstantInt::get((NegB + SqrtVal).sdiv(TwoA));
2523    ConstantInt *Solution2 = ConstantInt::get((NegB - SqrtVal).sdiv(TwoA));
2524
2525    return std::make_pair(SE.getConstant(Solution1),
2526                          SE.getConstant(Solution2));
2527    } // end APIntOps namespace
2528}
2529
2530/// HowFarToZero - Return the number of times a backedge comparing the specified
2531/// value to zero will execute.  If not computable, return UnknownValue
2532SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) {
2533  // If the value is a constant
2534  if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2535    // If the value is already zero, the branch will execute zero times.
2536    if (C->getValue()->isZero()) return C;
2537    return UnknownValue;  // Otherwise it will loop infinitely.
2538  }
2539
2540  SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
2541  if (!AddRec || AddRec->getLoop() != L)
2542    return UnknownValue;
2543
2544  if (AddRec->isAffine()) {
2545    // If this is an affine expression the execution count of this branch is
2546    // equal to:
2547    //
2548    //     (0 - Start/Step)    iff   Start % Step == 0
2549    //
2550    // Get the initial value for the loop.
2551    SCEVHandle Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
2552    if (isa<SCEVCouldNotCompute>(Start)) return UnknownValue;
2553    SCEVHandle Step = AddRec->getOperand(1);
2554
2555    Step = getSCEVAtScope(Step, L->getParentLoop());
2556
2557    // Figure out if Start % Step == 0.
2558    // FIXME: We should add DivExpr and RemExpr operations to our AST.
2559    if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
2560      if (StepC->getValue()->equalsInt(1))      // N % 1 == 0
2561        return SE.getNegativeSCEV(Start);  // 0 - Start/1 == -Start
2562      if (StepC->getValue()->isAllOnesValue())  // N % -1 == 0
2563        return Start;                   // 0 - Start/-1 == Start
2564
2565      // Check to see if Start is divisible by SC with no remainder.
2566      if (SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start)) {
2567        ConstantInt *StartCC = StartC->getValue();
2568        Constant *StartNegC = ConstantExpr::getNeg(StartCC);
2569        Constant *Rem = ConstantExpr::getURem(StartNegC, StepC->getValue());
2570        if (Rem->isNullValue()) {
2571          Constant *Result = ConstantExpr::getUDiv(StartNegC,StepC->getValue());
2572          return SE.getUnknown(Result);
2573        }
2574      }
2575    }
2576  } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
2577    // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
2578    // the quadratic equation to solve it.
2579    std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec, SE);
2580    SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2581    SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2582    if (R1) {
2583#if 0
2584      cerr << "HFTZ: " << *V << " - sol#1: " << *R1
2585           << "  sol#2: " << *R2 << "\n";
2586#endif
2587      // Pick the smallest positive root value.
2588      if (ConstantInt *CB =
2589          dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
2590                                   R1->getValue(), R2->getValue()))) {
2591        if (CB->getZExtValue() == false)
2592          std::swap(R1, R2);   // R1 is the minimum root now.
2593
2594        // We can only use this value if the chrec ends up with an exact zero
2595        // value at this index.  When solving for "X*X != 5", for example, we
2596        // should not accept a root of 2.
2597        SCEVHandle Val = AddRec->evaluateAtIteration(R1, SE);
2598        if (Val->isZero())
2599          return R1;  // We found a quadratic root!
2600      }
2601    }
2602  }
2603
2604  return UnknownValue;
2605}
2606
2607/// HowFarToNonZero - Return the number of times a backedge checking the
2608/// specified value for nonzero will execute.  If not computable, return
2609/// UnknownValue
2610SCEVHandle ScalarEvolutionsImpl::HowFarToNonZero(SCEV *V, const Loop *L) {
2611  // Loops that look like: while (X == 0) are very strange indeed.  We don't
2612  // handle them yet except for the trivial case.  This could be expanded in the
2613  // future as needed.
2614
2615  // If the value is a constant, check to see if it is known to be non-zero
2616  // already.  If so, the backedge will execute zero times.
2617  if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2618    if (!C->getValue()->isNullValue())
2619      return SE.getIntegerSCEV(0, C->getType());
2620    return UnknownValue;  // Otherwise it will loop infinitely.
2621  }
2622
2623  // We could implement others, but I really doubt anyone writes loops like
2624  // this, and if they did, they would already be constant folded.
2625  return UnknownValue;
2626}
2627
2628/// HowManyLessThans - Return the number of times a backedge containing the
2629/// specified less-than comparison will execute.  If not computable, return
2630/// UnknownValue.
2631SCEVHandle ScalarEvolutionsImpl::
2632HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L, bool isSigned) {
2633  // Only handle:  "ADDREC < LoopInvariant".
2634  if (!RHS->isLoopInvariant(L)) return UnknownValue;
2635
2636  SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
2637  if (!AddRec || AddRec->getLoop() != L)
2638    return UnknownValue;
2639
2640  if (AddRec->isAffine()) {
2641    // FORNOW: We only support unit strides.
2642    SCEVHandle One = SE.getIntegerSCEV(1, RHS->getType());
2643    if (AddRec->getOperand(1) != One)
2644      return UnknownValue;
2645
2646    // We know the LHS is of the form {n,+,1} and the RHS is some loop-invariant
2647    // m.  So, we count the number of iterations in which {n,+,1} < m is true.
2648    // Note that we cannot simply return max(m-n,0) because it's not safe to
2649    // treat m-n as signed nor unsigned due to overflow possibility.
2650
2651    // First, we get the value of the LHS in the first iteration: n
2652    SCEVHandle Start = AddRec->getOperand(0);
2653
2654    // Then, we get the value of the LHS in the first iteration in which the
2655    // above condition doesn't hold.  This equals to max(m,n).
2656    SCEVHandle End = isSigned ? SE.getSMaxExpr(RHS, Start)
2657                              : SE.getUMaxExpr(RHS, Start);
2658
2659    // Finally, we subtract these two values to get the number of times the
2660    // backedge is executed: max(m,n)-n.
2661    return SE.getMinusSCEV(End, Start);
2662  }
2663
2664  return UnknownValue;
2665}
2666
2667/// getNumIterationsInRange - Return the number of iterations of this loop that
2668/// produce values in the specified constant range.  Another way of looking at
2669/// this is that it returns the first iteration number where the value is not in
2670/// the condition, thus computing the exit count. If the iteration count can't
2671/// be computed, an instance of SCEVCouldNotCompute is returned.
2672SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
2673                                                   ScalarEvolution &SE) const {
2674  if (Range.isFullSet())  // Infinite loop.
2675    return new SCEVCouldNotCompute();
2676
2677  // If the start is a non-zero constant, shift the range to simplify things.
2678  if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
2679    if (!SC->getValue()->isZero()) {
2680      std::vector<SCEVHandle> Operands(op_begin(), op_end());
2681      Operands[0] = SE.getIntegerSCEV(0, SC->getType());
2682      SCEVHandle Shifted = SE.getAddRecExpr(Operands, getLoop());
2683      if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast<SCEVAddRecExpr>(Shifted))
2684        return ShiftedAddRec->getNumIterationsInRange(
2685                           Range.subtract(SC->getValue()->getValue()), SE);
2686      // This is strange and shouldn't happen.
2687      return new SCEVCouldNotCompute();
2688    }
2689
2690  // The only time we can solve this is when we have all constant indices.
2691  // Otherwise, we cannot determine the overflow conditions.
2692  for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
2693    if (!isa<SCEVConstant>(getOperand(i)))
2694      return new SCEVCouldNotCompute();
2695
2696
2697  // Okay at this point we know that all elements of the chrec are constants and
2698  // that the start element is zero.
2699
2700  // First check to see if the range contains zero.  If not, the first
2701  // iteration exits.
2702  if (!Range.contains(APInt(getBitWidth(),0)))
2703    return SE.getConstant(ConstantInt::get(getType(),0));
2704
2705  if (isAffine()) {
2706    // If this is an affine expression then we have this situation:
2707    //   Solve {0,+,A} in Range  ===  Ax in Range
2708
2709    // We know that zero is in the range.  If A is positive then we know that
2710    // the upper value of the range must be the first possible exit value.
2711    // If A is negative then the lower of the range is the last possible loop
2712    // value.  Also note that we already checked for a full range.
2713    APInt One(getBitWidth(),1);
2714    APInt A     = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
2715    APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
2716
2717    // The exit value should be (End+A)/A.
2718    APInt ExitVal = (End + A).udiv(A);
2719    ConstantInt *ExitValue = ConstantInt::get(ExitVal);
2720
2721    // Evaluate at the exit value.  If we really did fall out of the valid
2722    // range, then we computed our trip count, otherwise wrap around or other
2723    // things must have happened.
2724    ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
2725    if (Range.contains(Val->getValue()))
2726      return new SCEVCouldNotCompute();  // Something strange happened
2727
2728    // Ensure that the previous value is in the range.  This is a sanity check.
2729    assert(Range.contains(
2730           EvaluateConstantChrecAtConstant(this,
2731           ConstantInt::get(ExitVal - One), SE)->getValue()) &&
2732           "Linear scev computation is off in a bad way!");
2733    return SE.getConstant(ExitValue);
2734  } else if (isQuadratic()) {
2735    // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
2736    // quadratic equation to solve it.  To do this, we must frame our problem in
2737    // terms of figuring out when zero is crossed, instead of when
2738    // Range.getUpper() is crossed.
2739    std::vector<SCEVHandle> NewOps(op_begin(), op_end());
2740    NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
2741    SCEVHandle NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
2742
2743    // Next, solve the constructed addrec
2744    std::pair<SCEVHandle,SCEVHandle> Roots =
2745      SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
2746    SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2747    SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2748    if (R1) {
2749      // Pick the smallest positive root value.
2750      if (ConstantInt *CB =
2751          dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
2752                                   R1->getValue(), R2->getValue()))) {
2753        if (CB->getZExtValue() == false)
2754          std::swap(R1, R2);   // R1 is the minimum root now.
2755
2756        // Make sure the root is not off by one.  The returned iteration should
2757        // not be in the range, but the previous one should be.  When solving
2758        // for "X*X < 5", for example, we should not return a root of 2.
2759        ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
2760                                                             R1->getValue(),
2761                                                             SE);
2762        if (Range.contains(R1Val->getValue())) {
2763          // The next iteration must be out of the range...
2764          ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()+1);
2765
2766          R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
2767          if (!Range.contains(R1Val->getValue()))
2768            return SE.getConstant(NextVal);
2769          return new SCEVCouldNotCompute();  // Something strange happened
2770        }
2771
2772        // If R1 was not in the range, then it is a good return value.  Make
2773        // sure that R1-1 WAS in the range though, just in case.
2774        ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()-1);
2775        R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
2776        if (Range.contains(R1Val->getValue()))
2777          return R1;
2778        return new SCEVCouldNotCompute();  // Something strange happened
2779      }
2780    }
2781  }
2782
2783  // Fallback, if this is a general polynomial, figure out the progression
2784  // through brute force: evaluate until we find an iteration that fails the
2785  // test.  This is likely to be slow, but getting an accurate trip count is
2786  // incredibly important, we will be able to simplify the exit test a lot, and
2787  // we are almost guaranteed to get a trip count in this case.
2788  ConstantInt *TestVal = ConstantInt::get(getType(), 0);
2789  ConstantInt *EndVal  = TestVal;  // Stop when we wrap around.
2790  do {
2791    ++NumBruteForceEvaluations;
2792    SCEVHandle Val = evaluateAtIteration(SE.getConstant(TestVal), SE);
2793    if (!isa<SCEVConstant>(Val))  // This shouldn't happen.
2794      return new SCEVCouldNotCompute();
2795
2796    // Check to see if we found the value!
2797    if (!Range.contains(cast<SCEVConstant>(Val)->getValue()->getValue()))
2798      return SE.getConstant(TestVal);
2799
2800    // Increment to test the next index.
2801    TestVal = ConstantInt::get(TestVal->getValue()+1);
2802  } while (TestVal != EndVal);
2803
2804  return new SCEVCouldNotCompute();
2805}
2806
2807
2808
2809//===----------------------------------------------------------------------===//
2810//                   ScalarEvolution Class Implementation
2811//===----------------------------------------------------------------------===//
2812
2813bool ScalarEvolution::runOnFunction(Function &F) {
2814  Impl = new ScalarEvolutionsImpl(*this, F, getAnalysis<LoopInfo>());
2815  return false;
2816}
2817
2818void ScalarEvolution::releaseMemory() {
2819  delete (ScalarEvolutionsImpl*)Impl;
2820  Impl = 0;
2821}
2822
2823void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
2824  AU.setPreservesAll();
2825  AU.addRequiredTransitive<LoopInfo>();
2826}
2827
2828SCEVHandle ScalarEvolution::getSCEV(Value *V) const {
2829  return ((ScalarEvolutionsImpl*)Impl)->getSCEV(V);
2830}
2831
2832/// hasSCEV - Return true if the SCEV for this value has already been
2833/// computed.
2834bool ScalarEvolution::hasSCEV(Value *V) const {
2835  return ((ScalarEvolutionsImpl*)Impl)->hasSCEV(V);
2836}
2837
2838
2839/// setSCEV - Insert the specified SCEV into the map of current SCEVs for
2840/// the specified value.
2841void ScalarEvolution::setSCEV(Value *V, const SCEVHandle &H) {
2842  ((ScalarEvolutionsImpl*)Impl)->setSCEV(V, H);
2843}
2844
2845
2846SCEVHandle ScalarEvolution::getIterationCount(const Loop *L) const {
2847  return ((ScalarEvolutionsImpl*)Impl)->getIterationCount(L);
2848}
2849
2850bool ScalarEvolution::hasLoopInvariantIterationCount(const Loop *L) const {
2851  return !isa<SCEVCouldNotCompute>(getIterationCount(L));
2852}
2853
2854SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) const {
2855  return ((ScalarEvolutionsImpl*)Impl)->getSCEVAtScope(getSCEV(V), L);
2856}
2857
2858void ScalarEvolution::deleteValueFromRecords(Value *V) const {
2859  return ((ScalarEvolutionsImpl*)Impl)->deleteValueFromRecords(V);
2860}
2861
2862static void PrintLoopInfo(std::ostream &OS, const ScalarEvolution *SE,
2863                          const Loop *L) {
2864  // Print all inner loops first
2865  for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
2866    PrintLoopInfo(OS, SE, *I);
2867
2868  OS << "Loop " << L->getHeader()->getName() << ": ";
2869
2870  SmallVector<BasicBlock*, 8> ExitBlocks;
2871  L->getExitBlocks(ExitBlocks);
2872  if (ExitBlocks.size() != 1)
2873    OS << "<multiple exits> ";
2874
2875  if (SE->hasLoopInvariantIterationCount(L)) {
2876    OS << *SE->getIterationCount(L) << " iterations! ";
2877  } else {
2878    OS << "Unpredictable iteration count. ";
2879  }
2880
2881  OS << "\n";
2882}
2883
2884void ScalarEvolution::print(std::ostream &OS, const Module* ) const {
2885  Function &F = ((ScalarEvolutionsImpl*)Impl)->F;
2886  LoopInfo &LI = ((ScalarEvolutionsImpl*)Impl)->LI;
2887
2888  OS << "Classifying expressions for: " << F.getName() << "\n";
2889  for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
2890    if (I->getType()->isInteger()) {
2891      OS << *I;
2892      OS << "  --> ";
2893      SCEVHandle SV = getSCEV(&*I);
2894      SV->print(OS);
2895      OS << "\t\t";
2896
2897      if ((*I).getType()->isInteger()) {
2898        ConstantRange Bounds = SV->getValueRange();
2899        if (!Bounds.isFullSet())
2900          OS << "Bounds: " << Bounds << " ";
2901      }
2902
2903      if (const Loop *L = LI.getLoopFor((*I).getParent())) {
2904        OS << "Exits: ";
2905        SCEVHandle ExitValue = getSCEVAtScope(&*I, L->getParentLoop());
2906        if (isa<SCEVCouldNotCompute>(ExitValue)) {
2907          OS << "<<Unknown>>";
2908        } else {
2909          OS << *ExitValue;
2910        }
2911      }
2912
2913
2914      OS << "\n";
2915    }
2916
2917  OS << "Determining loop execution counts for: " << F.getName() << "\n";
2918  for (LoopInfo::iterator I = LI.begin(), E = LI.end(); I != E; ++I)
2919    PrintLoopInfo(OS, this, *I);
2920}
2921