1 //===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the implementation of the scalar evolution analysis
11 // engine, which is used primarily to analyze expressions involving induction
12 // variables in loops.
14 // There are several aspects to this library. First is the representation of
15 // scalar expressions, which are represented as subclasses of the SCEV class.
16 // These classes are used to represent certain types of subexpressions that we
17 // can handle. These classes are reference counted, managed by the SCEVHandle
18 // class. We only create one SCEV of a particular shape, so pointer-comparisons
19 // for equality are legal.
21 // One important aspect of the SCEV objects is that they are never cyclic, even
22 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
23 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
24 // recurrence) then we represent it directly as a recurrence node, otherwise we
25 // represent it as a SCEVUnknown node.
27 // In addition to being able to represent expressions of various types, we also
28 // have folders that are used to build the *canonical* representation for a
29 // particular expression. These folders are capable of using a variety of
30 // rewrite rules to simplify the expressions.
32 // Once the folders are defined, we can implement the more interesting
33 // higher-level code, such as the code that recognizes PHI nodes of various
34 // types, computes the execution count of a loop, etc.
36 // TODO: We should use these routines and value representations to implement
37 // dependence analysis!
39 //===----------------------------------------------------------------------===//
41 // There are several good references for the techniques used in this analysis.
43 // Chains of recurrences -- a method to expedite the evaluation
44 // of closed-form functions
45 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
47 // On computational properties of chains of recurrences
50 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
51 // Robert A. van Engelen
53 // Efficient Symbolic Analysis for Optimizing Compilers
54 // Robert A. van Engelen
56 // Using the chains of recurrences algebra for data dependence testing and
57 // induction variable substitution
58 // MS Thesis, Johnie Birch
60 //===----------------------------------------------------------------------===//
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/Constants.h"
64 #include "llvm/DerivedTypes.h"
65 #include "llvm/GlobalVariable.h"
66 #include "llvm/Instructions.h"
67 #include "llvm/Analysis/ConstantFolding.h"
68 #include "llvm/Analysis/LoopInfo.h"
69 #include "llvm/Assembly/Writer.h"
70 #include "llvm/Transforms/Scalar.h"
71 #include "llvm/Support/CFG.h"
72 #include "llvm/Support/CommandLine.h"
73 #include "llvm/Support/Compiler.h"
74 #include "llvm/Support/ConstantRange.h"
75 #include "llvm/Support/InstIterator.h"
76 #include "llvm/Support/ManagedStatic.h"
77 #include "llvm/Support/Streams.h"
78 #include "llvm/ADT/Statistic.h"
85 RegisterPass<ScalarEvolution>
86 R("scalar-evolution", "Scalar Evolution Analysis");
89 NumBruteForceEvaluations("scalar-evolution",
90 "Number of brute force evaluations needed to "
91 "calculate high-order polynomial exit values");
93 NumArrayLenItCounts("scalar-evolution",
94 "Number of trip counts computed with array length");
96 NumTripCountsComputed("scalar-evolution",
97 "Number of loops with predictable loop counts");
99 NumTripCountsNotComputed("scalar-evolution",
100 "Number of loops without predictable loop counts");
102 NumBruteForceTripCountsComputed("scalar-evolution",
103 "Number of loops with trip counts computed by force");
106 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
107 cl::desc("Maximum number of iterations SCEV will "
108 "symbolically execute a constant derived loop"),
112 //===----------------------------------------------------------------------===//
113 // SCEV class definitions
114 //===----------------------------------------------------------------------===//
116 //===----------------------------------------------------------------------===//
117 // Implementation of the SCEV class.
120 void SCEV::dump() const {
124 /// getValueRange - Return the tightest constant bounds that this value is
125 /// known to have. This method is only valid on integer SCEV objects.
126 ConstantRange SCEV::getValueRange() const {
127 const Type *Ty = getType();
128 assert(Ty->isInteger() && "Can't get range for a non-integer SCEV!");
129 Ty = Ty->getUnsignedVersion();
130 // Default to a full range if no better information is available.
131 return ConstantRange(getType());
135 SCEVCouldNotCompute::SCEVCouldNotCompute() : SCEV(scCouldNotCompute) {}
137 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
138 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
142 const Type *SCEVCouldNotCompute::getType() const {
143 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
147 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
148 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
152 SCEVHandle SCEVCouldNotCompute::
153 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
154 const SCEVHandle &Conc) const {
158 void SCEVCouldNotCompute::print(std::ostream &OS) const {
159 OS << "***COULDNOTCOMPUTE***";
162 bool SCEVCouldNotCompute::classof(const SCEV *S) {
163 return S->getSCEVType() == scCouldNotCompute;
167 // SCEVConstants - Only allow the creation of one SCEVConstant for any
168 // particular value. Don't use a SCEVHandle here, or else the object will
170 static ManagedStatic<std::map<ConstantInt*, SCEVConstant*> > SCEVConstants;
173 SCEVConstant::~SCEVConstant() {
174 SCEVConstants->erase(V);
177 SCEVHandle SCEVConstant::get(ConstantInt *V) {
178 // Make sure that SCEVConstant instances are all unsigned.
179 if (V->getType()->isSigned()) {
180 const Type *NewTy = V->getType()->getUnsignedVersion();
181 V = cast<ConstantInt>(ConstantExpr::getCast(V, NewTy));
184 SCEVConstant *&R = (*SCEVConstants)[V];
185 if (R == 0) R = new SCEVConstant(V);
189 ConstantRange SCEVConstant::getValueRange() const {
190 return ConstantRange(V);
193 const Type *SCEVConstant::getType() const { return V->getType(); }
195 void SCEVConstant::print(std::ostream &OS) const {
196 WriteAsOperand(OS, V, false);
199 // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any
200 // particular input. Don't use a SCEVHandle here, or else the object will
202 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
203 SCEVTruncateExpr*> > SCEVTruncates;
205 SCEVTruncateExpr::SCEVTruncateExpr(const SCEVHandle &op, const Type *ty)
206 : SCEV(scTruncate), Op(op), Ty(ty) {
207 assert(Op->getType()->isInteger() && Ty->isInteger() &&
208 "Cannot truncate non-integer value!");
209 assert(Op->getType()->getPrimitiveSize() > Ty->getPrimitiveSize() &&
210 "This is not a truncating conversion!");
213 SCEVTruncateExpr::~SCEVTruncateExpr() {
214 SCEVTruncates->erase(std::make_pair(Op, Ty));
217 ConstantRange SCEVTruncateExpr::getValueRange() const {
218 return getOperand()->getValueRange().truncate(getType());
221 void SCEVTruncateExpr::print(std::ostream &OS) const {
222 OS << "(truncate " << *Op << " to " << *Ty << ")";
225 // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any
226 // particular input. Don't use a SCEVHandle here, or else the object will never
228 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
229 SCEVZeroExtendExpr*> > SCEVZeroExtends;
231 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty)
232 : SCEV(scZeroExtend), Op(op), Ty(ty) {
233 assert(Op->getType()->isInteger() && Ty->isInteger() &&
234 "Cannot zero extend non-integer value!");
235 assert(Op->getType()->getPrimitiveSize() < Ty->getPrimitiveSize() &&
236 "This is not an extending conversion!");
239 SCEVZeroExtendExpr::~SCEVZeroExtendExpr() {
240 SCEVZeroExtends->erase(std::make_pair(Op, Ty));
243 ConstantRange SCEVZeroExtendExpr::getValueRange() const {
244 return getOperand()->getValueRange().zeroExtend(getType());
247 void SCEVZeroExtendExpr::print(std::ostream &OS) const {
248 OS << "(zeroextend " << *Op << " to " << *Ty << ")";
251 // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any
252 // particular input. Don't use a SCEVHandle here, or else the object will never
254 static ManagedStatic<std::map<std::pair<unsigned, std::vector<SCEV*> >,
255 SCEVCommutativeExpr*> > SCEVCommExprs;
257 SCEVCommutativeExpr::~SCEVCommutativeExpr() {
258 SCEVCommExprs->erase(std::make_pair(getSCEVType(),
259 std::vector<SCEV*>(Operands.begin(),
263 void SCEVCommutativeExpr::print(std::ostream &OS) const {
264 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
265 const char *OpStr = getOperationStr();
266 OS << "(" << *Operands[0];
267 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
268 OS << OpStr << *Operands[i];
272 SCEVHandle SCEVCommutativeExpr::
273 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
274 const SCEVHandle &Conc) const {
275 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
276 SCEVHandle H = getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc);
277 if (H != getOperand(i)) {
278 std::vector<SCEVHandle> NewOps;
279 NewOps.reserve(getNumOperands());
280 for (unsigned j = 0; j != i; ++j)
281 NewOps.push_back(getOperand(j));
283 for (++i; i != e; ++i)
284 NewOps.push_back(getOperand(i)->
285 replaceSymbolicValuesWithConcrete(Sym, Conc));
287 if (isa<SCEVAddExpr>(this))
288 return SCEVAddExpr::get(NewOps);
289 else if (isa<SCEVMulExpr>(this))
290 return SCEVMulExpr::get(NewOps);
292 assert(0 && "Unknown commutative expr!");
299 // SCEVSDivs - Only allow the creation of one SCEVSDivExpr for any particular
300 // input. Don't use a SCEVHandle here, or else the object will never be
302 static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>,
303 SCEVSDivExpr*> > SCEVSDivs;
305 SCEVSDivExpr::~SCEVSDivExpr() {
306 SCEVSDivs->erase(std::make_pair(LHS, RHS));
309 void SCEVSDivExpr::print(std::ostream &OS) const {
310 OS << "(" << *LHS << " /s " << *RHS << ")";
313 const Type *SCEVSDivExpr::getType() const {
314 const Type *Ty = LHS->getType();
315 if (Ty->isUnsigned()) Ty = Ty->getSignedVersion();
319 // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
320 // particular input. Don't use a SCEVHandle here, or else the object will never
322 static ManagedStatic<std::map<std::pair<const Loop *, std::vector<SCEV*> >,
323 SCEVAddRecExpr*> > SCEVAddRecExprs;
325 SCEVAddRecExpr::~SCEVAddRecExpr() {
326 SCEVAddRecExprs->erase(std::make_pair(L,
327 std::vector<SCEV*>(Operands.begin(),
331 SCEVHandle SCEVAddRecExpr::
332 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
333 const SCEVHandle &Conc) const {
334 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
335 SCEVHandle H = getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc);
336 if (H != getOperand(i)) {
337 std::vector<SCEVHandle> NewOps;
338 NewOps.reserve(getNumOperands());
339 for (unsigned j = 0; j != i; ++j)
340 NewOps.push_back(getOperand(j));
342 for (++i; i != e; ++i)
343 NewOps.push_back(getOperand(i)->
344 replaceSymbolicValuesWithConcrete(Sym, Conc));
346 return get(NewOps, L);
353 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
354 // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't
355 // contain L and if the start is invariant.
356 return !QueryLoop->contains(L->getHeader()) &&
357 getOperand(0)->isLoopInvariant(QueryLoop);
361 void SCEVAddRecExpr::print(std::ostream &OS) const {
362 OS << "{" << *Operands[0];
363 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
364 OS << ",+," << *Operands[i];
365 OS << "}<" << L->getHeader()->getName() + ">";
368 // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular
369 // value. Don't use a SCEVHandle here, or else the object will never be
371 static ManagedStatic<std::map<Value*, SCEVUnknown*> > SCEVUnknowns;
373 SCEVUnknown::~SCEVUnknown() { SCEVUnknowns->erase(V); }
375 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
376 // All non-instruction values are loop invariant. All instructions are loop
377 // invariant if they are not contained in the specified loop.
378 if (Instruction *I = dyn_cast<Instruction>(V))
379 return !L->contains(I->getParent());
383 const Type *SCEVUnknown::getType() const {
387 void SCEVUnknown::print(std::ostream &OS) const {
388 WriteAsOperand(OS, V, false);
391 //===----------------------------------------------------------------------===//
393 //===----------------------------------------------------------------------===//
396 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
397 /// than the complexity of the RHS. This comparator is used to canonicalize
399 struct VISIBILITY_HIDDEN SCEVComplexityCompare {
400 bool operator()(SCEV *LHS, SCEV *RHS) {
401 return LHS->getSCEVType() < RHS->getSCEVType();
406 /// GroupByComplexity - Given a list of SCEV objects, order them by their
407 /// complexity, and group objects of the same complexity together by value.
408 /// When this routine is finished, we know that any duplicates in the vector are
409 /// consecutive and that complexity is monotonically increasing.
411 /// Note that we go take special precautions to ensure that we get determinstic
412 /// results from this routine. In other words, we don't want the results of
413 /// this to depend on where the addresses of various SCEV objects happened to
416 static void GroupByComplexity(std::vector<SCEVHandle> &Ops) {
417 if (Ops.size() < 2) return; // Noop
418 if (Ops.size() == 2) {
419 // This is the common case, which also happens to be trivially simple.
421 if (Ops[0]->getSCEVType() > Ops[1]->getSCEVType())
422 std::swap(Ops[0], Ops[1]);
426 // Do the rough sort by complexity.
427 std::sort(Ops.begin(), Ops.end(), SCEVComplexityCompare());
429 // Now that we are sorted by complexity, group elements of the same
430 // complexity. Note that this is, at worst, N^2, but the vector is likely to
431 // be extremely short in practice. Note that we take this approach because we
432 // do not want to depend on the addresses of the objects we are grouping.
433 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
435 unsigned Complexity = S->getSCEVType();
437 // If there are any objects of the same complexity and same value as this
439 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
440 if (Ops[j] == S) { // Found a duplicate.
441 // Move it to immediately after i'th element.
442 std::swap(Ops[i+1], Ops[j]);
443 ++i; // no need to rescan it.
444 if (i == e-2) return; // Done!
452 //===----------------------------------------------------------------------===//
453 // Simple SCEV method implementations
454 //===----------------------------------------------------------------------===//
456 /// getIntegerSCEV - Given an integer or FP type, create a constant for the
457 /// specified signed integer value and return a SCEV for the constant.
458 SCEVHandle SCEVUnknown::getIntegerSCEV(int Val, const Type *Ty) {
461 C = Constant::getNullValue(Ty);
462 else if (Ty->isFloatingPoint())
463 C = ConstantFP::get(Ty, Val);
464 else if (Ty->isSigned())
465 C = ConstantInt::get(Ty, Val);
467 C = ConstantInt::get(Ty->getSignedVersion(), Val);
468 C = ConstantExpr::getCast(C, Ty);
470 return SCEVUnknown::get(C);
473 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
474 /// input value to the specified type. If the type must be extended, it is zero
476 static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty) {
477 const Type *SrcTy = V->getType();
478 assert(SrcTy->isInteger() && Ty->isInteger() &&
479 "Cannot truncate or zero extend with non-integer arguments!");
480 if (SrcTy->getPrimitiveSize() == Ty->getPrimitiveSize())
481 return V; // No conversion
482 if (SrcTy->getPrimitiveSize() > Ty->getPrimitiveSize())
483 return SCEVTruncateExpr::get(V, Ty);
484 return SCEVZeroExtendExpr::get(V, Ty);
487 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
489 SCEVHandle SCEV::getNegativeSCEV(const SCEVHandle &V) {
490 if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
491 return SCEVUnknown::get(ConstantExpr::getNeg(VC->getValue()));
493 return SCEVMulExpr::get(V, SCEVUnknown::getIntegerSCEV(-1, V->getType()));
496 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
498 SCEVHandle SCEV::getMinusSCEV(const SCEVHandle &LHS, const SCEVHandle &RHS) {
500 return SCEVAddExpr::get(LHS, SCEV::getNegativeSCEV(RHS));
504 /// PartialFact - Compute V!/(V-NumSteps)!
505 static SCEVHandle PartialFact(SCEVHandle V, unsigned NumSteps) {
506 // Handle this case efficiently, it is common to have constant iteration
507 // counts while computing loop exit values.
508 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
509 uint64_t Val = SC->getValue()->getZExtValue();
511 for (; NumSteps; --NumSteps)
512 Result *= Val-(NumSteps-1);
513 Constant *Res = ConstantInt::get(Type::ULongTy, Result);
514 return SCEVUnknown::get(ConstantExpr::getCast(Res, V->getType()));
517 const Type *Ty = V->getType();
519 return SCEVUnknown::getIntegerSCEV(1, Ty);
521 SCEVHandle Result = V;
522 for (unsigned i = 1; i != NumSteps; ++i)
523 Result = SCEVMulExpr::get(Result, SCEV::getMinusSCEV(V,
524 SCEVUnknown::getIntegerSCEV(i, Ty)));
529 /// evaluateAtIteration - Return the value of this chain of recurrences at
530 /// the specified iteration number. We can evaluate this recurrence by
531 /// multiplying each element in the chain by the binomial coefficient
532 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
534 /// A*choose(It, 0) + B*choose(It, 1) + C*choose(It, 2) + D*choose(It, 3)
536 /// FIXME/VERIFY: I don't trust that this is correct in the face of overflow.
537 /// Is the binomial equation safe using modular arithmetic??
539 SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It) const {
540 SCEVHandle Result = getStart();
542 const Type *Ty = It->getType();
543 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
544 SCEVHandle BC = PartialFact(It, i);
546 SCEVHandle Val = SCEVSDivExpr::get(SCEVMulExpr::get(BC, getOperand(i)),
547 SCEVUnknown::getIntegerSCEV(Divisor,Ty));
548 Result = SCEVAddExpr::get(Result, Val);
554 //===----------------------------------------------------------------------===//
555 // SCEV Expression folder implementations
556 //===----------------------------------------------------------------------===//
558 SCEVHandle SCEVTruncateExpr::get(const SCEVHandle &Op, const Type *Ty) {
559 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
560 return SCEVUnknown::get(ConstantExpr::getCast(SC->getValue(), Ty));
562 // If the input value is a chrec scev made out of constants, truncate
563 // all of the constants.
564 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
565 std::vector<SCEVHandle> Operands;
566 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
567 // FIXME: This should allow truncation of other expression types!
568 if (isa<SCEVConstant>(AddRec->getOperand(i)))
569 Operands.push_back(get(AddRec->getOperand(i), Ty));
572 if (Operands.size() == AddRec->getNumOperands())
573 return SCEVAddRecExpr::get(Operands, AddRec->getLoop());
576 SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)];
577 if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
581 SCEVHandle SCEVZeroExtendExpr::get(const SCEVHandle &Op, const Type *Ty) {
582 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
583 return SCEVUnknown::get(ConstantExpr::getCast(SC->getValue(), Ty));
585 // FIXME: If the input value is a chrec scev, and we can prove that the value
586 // did not overflow the old, smaller, value, we can zero extend all of the
587 // operands (often constants). This would allow analysis of something like
588 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
590 SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)];
591 if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
595 // get - Get a canonical add expression, or something simpler if possible.
596 SCEVHandle SCEVAddExpr::get(std::vector<SCEVHandle> &Ops) {
597 assert(!Ops.empty() && "Cannot get empty add!");
598 if (Ops.size() == 1) return Ops[0];
600 // Sort by complexity, this groups all similar expression types together.
601 GroupByComplexity(Ops);
603 // If there are any constants, fold them together.
605 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
607 assert(Idx < Ops.size());
608 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
609 // We found two constants, fold them together!
610 Constant *Fold = ConstantExpr::getAdd(LHSC->getValue(), RHSC->getValue());
611 if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) {
612 Ops[0] = SCEVConstant::get(CI);
613 Ops.erase(Ops.begin()+1); // Erase the folded element
614 if (Ops.size() == 1) return Ops[0];
615 LHSC = cast<SCEVConstant>(Ops[0]);
617 // If we couldn't fold the expression, move to the next constant. Note
618 // that this is impossible to happen in practice because we always
619 // constant fold constant ints to constant ints.
624 // If we are left with a constant zero being added, strip it off.
625 if (cast<SCEVConstant>(Ops[0])->getValue()->isNullValue()) {
626 Ops.erase(Ops.begin());
631 if (Ops.size() == 1) return Ops[0];
633 // Okay, check to see if the same value occurs in the operand list twice. If
634 // so, merge them together into an multiply expression. Since we sorted the
635 // list, these values are required to be adjacent.
636 const Type *Ty = Ops[0]->getType();
637 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
638 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
639 // Found a match, merge the two values into a multiply, and add any
640 // remaining values to the result.
641 SCEVHandle Two = SCEVUnknown::getIntegerSCEV(2, Ty);
642 SCEVHandle Mul = SCEVMulExpr::get(Ops[i], Two);
645 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
647 return SCEVAddExpr::get(Ops);
650 // Okay, now we know the first non-constant operand. If there are add
651 // operands they would be next.
652 if (Idx < Ops.size()) {
653 bool DeletedAdd = false;
654 while (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
655 // If we have an add, expand the add operands onto the end of the operands
657 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
658 Ops.erase(Ops.begin()+Idx);
662 // If we deleted at least one add, we added operands to the end of the list,
663 // and they are not necessarily sorted. Recurse to resort and resimplify
664 // any operands we just aquired.
669 // Skip over the add expression until we get to a multiply.
670 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
673 // If we are adding something to a multiply expression, make sure the
674 // something is not already an operand of the multiply. If so, merge it into
676 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
677 SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
678 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
679 SCEV *MulOpSCEV = Mul->getOperand(MulOp);
680 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
681 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(MulOpSCEV)) {
682 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
683 SCEVHandle InnerMul = Mul->getOperand(MulOp == 0);
684 if (Mul->getNumOperands() != 2) {
685 // If the multiply has more than two operands, we must get the
687 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
688 MulOps.erase(MulOps.begin()+MulOp);
689 InnerMul = SCEVMulExpr::get(MulOps);
691 SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, Ty);
692 SCEVHandle AddOne = SCEVAddExpr::get(InnerMul, One);
693 SCEVHandle OuterMul = SCEVMulExpr::get(AddOne, Ops[AddOp]);
694 if (Ops.size() == 2) return OuterMul;
696 Ops.erase(Ops.begin()+AddOp);
697 Ops.erase(Ops.begin()+Idx-1);
699 Ops.erase(Ops.begin()+Idx);
700 Ops.erase(Ops.begin()+AddOp-1);
702 Ops.push_back(OuterMul);
703 return SCEVAddExpr::get(Ops);
706 // Check this multiply against other multiplies being added together.
707 for (unsigned OtherMulIdx = Idx+1;
708 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
710 SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
711 // If MulOp occurs in OtherMul, we can fold the two multiplies
713 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
714 OMulOp != e; ++OMulOp)
715 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
716 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
717 SCEVHandle InnerMul1 = Mul->getOperand(MulOp == 0);
718 if (Mul->getNumOperands() != 2) {
719 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
720 MulOps.erase(MulOps.begin()+MulOp);
721 InnerMul1 = SCEVMulExpr::get(MulOps);
723 SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0);
724 if (OtherMul->getNumOperands() != 2) {
725 std::vector<SCEVHandle> MulOps(OtherMul->op_begin(),
727 MulOps.erase(MulOps.begin()+OMulOp);
728 InnerMul2 = SCEVMulExpr::get(MulOps);
730 SCEVHandle InnerMulSum = SCEVAddExpr::get(InnerMul1,InnerMul2);
731 SCEVHandle OuterMul = SCEVMulExpr::get(MulOpSCEV, InnerMulSum);
732 if (Ops.size() == 2) return OuterMul;
733 Ops.erase(Ops.begin()+Idx);
734 Ops.erase(Ops.begin()+OtherMulIdx-1);
735 Ops.push_back(OuterMul);
736 return SCEVAddExpr::get(Ops);
742 // If there are any add recurrences in the operands list, see if any other
743 // added values are loop invariant. If so, we can fold them into the
745 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
748 // Scan over all recurrences, trying to fold loop invariants into them.
749 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
750 // Scan all of the other operands to this add and add them to the vector if
751 // they are loop invariant w.r.t. the recurrence.
752 std::vector<SCEVHandle> LIOps;
753 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
754 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
755 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
756 LIOps.push_back(Ops[i]);
757 Ops.erase(Ops.begin()+i);
761 // If we found some loop invariants, fold them into the recurrence.
762 if (!LIOps.empty()) {
763 // NLI + LI + { Start,+,Step} --> NLI + { LI+Start,+,Step }
764 LIOps.push_back(AddRec->getStart());
766 std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
767 AddRecOps[0] = SCEVAddExpr::get(LIOps);
769 SCEVHandle NewRec = SCEVAddRecExpr::get(AddRecOps, AddRec->getLoop());
770 // If all of the other operands were loop invariant, we are done.
771 if (Ops.size() == 1) return NewRec;
773 // Otherwise, add the folded AddRec by the non-liv parts.
774 for (unsigned i = 0;; ++i)
775 if (Ops[i] == AddRec) {
779 return SCEVAddExpr::get(Ops);
782 // Okay, if there weren't any loop invariants to be folded, check to see if
783 // there are multiple AddRec's with the same loop induction variable being
784 // added together. If so, we can fold them.
785 for (unsigned OtherIdx = Idx+1;
786 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
787 if (OtherIdx != Idx) {
788 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
789 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
790 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
791 std::vector<SCEVHandle> NewOps(AddRec->op_begin(), AddRec->op_end());
792 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
793 if (i >= NewOps.size()) {
794 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
795 OtherAddRec->op_end());
798 NewOps[i] = SCEVAddExpr::get(NewOps[i], OtherAddRec->getOperand(i));
800 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop());
802 if (Ops.size() == 2) return NewAddRec;
804 Ops.erase(Ops.begin()+Idx);
805 Ops.erase(Ops.begin()+OtherIdx-1);
806 Ops.push_back(NewAddRec);
807 return SCEVAddExpr::get(Ops);
811 // Otherwise couldn't fold anything into this recurrence. Move onto the
815 // Okay, it looks like we really DO need an add expr. Check to see if we
816 // already have one, otherwise create a new one.
817 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
818 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scAddExpr,
820 if (Result == 0) Result = new SCEVAddExpr(Ops);
825 SCEVHandle SCEVMulExpr::get(std::vector<SCEVHandle> &Ops) {
826 assert(!Ops.empty() && "Cannot get empty mul!");
828 // Sort by complexity, this groups all similar expression types together.
829 GroupByComplexity(Ops);
831 // If there are any constants, fold them together.
833 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
835 // C1*(C2+V) -> C1*C2 + C1*V
837 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
838 if (Add->getNumOperands() == 2 &&
839 isa<SCEVConstant>(Add->getOperand(0)))
840 return SCEVAddExpr::get(SCEVMulExpr::get(LHSC, Add->getOperand(0)),
841 SCEVMulExpr::get(LHSC, Add->getOperand(1)));
845 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
846 // We found two constants, fold them together!
847 Constant *Fold = ConstantExpr::getMul(LHSC->getValue(), RHSC->getValue());
848 if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) {
849 Ops[0] = SCEVConstant::get(CI);
850 Ops.erase(Ops.begin()+1); // Erase the folded element
851 if (Ops.size() == 1) return Ops[0];
852 LHSC = cast<SCEVConstant>(Ops[0]);
854 // If we couldn't fold the expression, move to the next constant. Note
855 // that this is impossible to happen in practice because we always
856 // constant fold constant ints to constant ints.
861 // If we are left with a constant one being multiplied, strip it off.
862 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
863 Ops.erase(Ops.begin());
865 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isNullValue()) {
866 // If we have a multiply of zero, it will always be zero.
871 // Skip over the add expression until we get to a multiply.
872 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
878 // If there are mul operands inline them all into this expression.
879 if (Idx < Ops.size()) {
880 bool DeletedMul = false;
881 while (SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
882 // If we have an mul, expand the mul operands onto the end of the operands
884 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
885 Ops.erase(Ops.begin()+Idx);
889 // If we deleted at least one mul, we added operands to the end of the list,
890 // and they are not necessarily sorted. Recurse to resort and resimplify
891 // any operands we just aquired.
896 // If there are any add recurrences in the operands list, see if any other
897 // added values are loop invariant. If so, we can fold them into the
899 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
902 // Scan over all recurrences, trying to fold loop invariants into them.
903 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
904 // Scan all of the other operands to this mul and add them to the vector if
905 // they are loop invariant w.r.t. the recurrence.
906 std::vector<SCEVHandle> LIOps;
907 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
908 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
909 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
910 LIOps.push_back(Ops[i]);
911 Ops.erase(Ops.begin()+i);
915 // If we found some loop invariants, fold them into the recurrence.
916 if (!LIOps.empty()) {
917 // NLI * LI * { Start,+,Step} --> NLI * { LI*Start,+,LI*Step }
918 std::vector<SCEVHandle> NewOps;
919 NewOps.reserve(AddRec->getNumOperands());
920 if (LIOps.size() == 1) {
921 SCEV *Scale = LIOps[0];
922 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
923 NewOps.push_back(SCEVMulExpr::get(Scale, AddRec->getOperand(i)));
925 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
926 std::vector<SCEVHandle> MulOps(LIOps);
927 MulOps.push_back(AddRec->getOperand(i));
928 NewOps.push_back(SCEVMulExpr::get(MulOps));
932 SCEVHandle NewRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop());
934 // If all of the other operands were loop invariant, we are done.
935 if (Ops.size() == 1) return NewRec;
937 // Otherwise, multiply the folded AddRec by the non-liv parts.
938 for (unsigned i = 0;; ++i)
939 if (Ops[i] == AddRec) {
943 return SCEVMulExpr::get(Ops);
946 // Okay, if there weren't any loop invariants to be folded, check to see if
947 // there are multiple AddRec's with the same loop induction variable being
948 // multiplied together. If so, we can fold them.
949 for (unsigned OtherIdx = Idx+1;
950 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
951 if (OtherIdx != Idx) {
952 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
953 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
954 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
955 SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
956 SCEVHandle NewStart = SCEVMulExpr::get(F->getStart(),
958 SCEVHandle B = F->getStepRecurrence();
959 SCEVHandle D = G->getStepRecurrence();
960 SCEVHandle NewStep = SCEVAddExpr::get(SCEVMulExpr::get(F, D),
961 SCEVMulExpr::get(G, B),
962 SCEVMulExpr::get(B, D));
963 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewStart, NewStep,
965 if (Ops.size() == 2) return NewAddRec;
967 Ops.erase(Ops.begin()+Idx);
968 Ops.erase(Ops.begin()+OtherIdx-1);
969 Ops.push_back(NewAddRec);
970 return SCEVMulExpr::get(Ops);
974 // Otherwise couldn't fold anything into this recurrence. Move onto the
978 // Okay, it looks like we really DO need an mul expr. Check to see if we
979 // already have one, otherwise create a new one.
980 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
981 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scMulExpr,
984 Result = new SCEVMulExpr(Ops);
988 SCEVHandle SCEVSDivExpr::get(const SCEVHandle &LHS, const SCEVHandle &RHS) {
989 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
990 if (RHSC->getValue()->equalsInt(1))
991 return LHS; // X sdiv 1 --> x
992 if (RHSC->getValue()->isAllOnesValue())
993 return SCEV::getNegativeSCEV(LHS); // X sdiv -1 --> -x
995 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
996 Constant *LHSCV = LHSC->getValue();
997 Constant *RHSCV = RHSC->getValue();
998 if (LHSCV->getType()->isUnsigned())
999 LHSCV = ConstantExpr::getCast(LHSCV,
1000 LHSCV->getType()->getSignedVersion());
1001 if (RHSCV->getType()->isUnsigned())
1002 RHSCV = ConstantExpr::getCast(RHSCV, LHSCV->getType());
1003 return SCEVUnknown::get(ConstantExpr::getSDiv(LHSCV, RHSCV));
1007 // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow.
1009 SCEVSDivExpr *&Result = (*SCEVSDivs)[std::make_pair(LHS, RHS)];
1010 if (Result == 0) Result = new SCEVSDivExpr(LHS, RHS);
1015 /// SCEVAddRecExpr::get - Get a add recurrence expression for the
1016 /// specified loop. Simplify the expression as much as possible.
1017 SCEVHandle SCEVAddRecExpr::get(const SCEVHandle &Start,
1018 const SCEVHandle &Step, const Loop *L) {
1019 std::vector<SCEVHandle> Operands;
1020 Operands.push_back(Start);
1021 if (SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1022 if (StepChrec->getLoop() == L) {
1023 Operands.insert(Operands.end(), StepChrec->op_begin(),
1024 StepChrec->op_end());
1025 return get(Operands, L);
1028 Operands.push_back(Step);
1029 return get(Operands, L);
1032 /// SCEVAddRecExpr::get - Get a add recurrence expression for the
1033 /// specified loop. Simplify the expression as much as possible.
1034 SCEVHandle SCEVAddRecExpr::get(std::vector<SCEVHandle> &Operands,
1036 if (Operands.size() == 1) return Operands[0];
1038 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Operands.back()))
1039 if (StepC->getValue()->isNullValue()) {
1040 Operands.pop_back();
1041 return get(Operands, L); // { X,+,0 } --> X
1044 SCEVAddRecExpr *&Result =
1045 (*SCEVAddRecExprs)[std::make_pair(L, std::vector<SCEV*>(Operands.begin(),
1047 if (Result == 0) Result = new SCEVAddRecExpr(Operands, L);
1051 SCEVHandle SCEVUnknown::get(Value *V) {
1052 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
1053 return SCEVConstant::get(CI);
1054 SCEVUnknown *&Result = (*SCEVUnknowns)[V];
1055 if (Result == 0) Result = new SCEVUnknown(V);
1060 //===----------------------------------------------------------------------===//
1061 // ScalarEvolutionsImpl Definition and Implementation
1062 //===----------------------------------------------------------------------===//
1064 /// ScalarEvolutionsImpl - This class implements the main driver for the scalar
1068 struct VISIBILITY_HIDDEN ScalarEvolutionsImpl {
1069 /// F - The function we are analyzing.
1073 /// LI - The loop information for the function we are currently analyzing.
1077 /// UnknownValue - This SCEV is used to represent unknown trip counts and
1079 SCEVHandle UnknownValue;
1081 /// Scalars - This is a cache of the scalars we have analyzed so far.
1083 std::map<Value*, SCEVHandle> Scalars;
1085 /// IterationCounts - Cache the iteration count of the loops for this
1086 /// function as they are computed.
1087 std::map<const Loop*, SCEVHandle> IterationCounts;
1089 /// ConstantEvolutionLoopExitValue - This map contains entries for all of
1090 /// the PHI instructions that we attempt to compute constant evolutions for.
1091 /// This allows us to avoid potentially expensive recomputation of these
1092 /// properties. An instruction maps to null if we are unable to compute its
1094 std::map<PHINode*, Constant*> ConstantEvolutionLoopExitValue;
1097 ScalarEvolutionsImpl(Function &f, LoopInfo &li)
1098 : F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {}
1100 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1101 /// expression and create a new one.
1102 SCEVHandle getSCEV(Value *V);
1104 /// hasSCEV - Return true if the SCEV for this value has already been
1106 bool hasSCEV(Value *V) const {
1107 return Scalars.count(V);
1110 /// setSCEV - Insert the specified SCEV into the map of current SCEVs for
1111 /// the specified value.
1112 void setSCEV(Value *V, const SCEVHandle &H) {
1113 bool isNew = Scalars.insert(std::make_pair(V, H)).second;
1114 assert(isNew && "This entry already existed!");
1118 /// getSCEVAtScope - Compute the value of the specified expression within
1119 /// the indicated loop (which may be null to indicate in no loop). If the
1120 /// expression cannot be evaluated, return UnknownValue itself.
1121 SCEVHandle getSCEVAtScope(SCEV *V, const Loop *L);
1124 /// hasLoopInvariantIterationCount - Return true if the specified loop has
1125 /// an analyzable loop-invariant iteration count.
1126 bool hasLoopInvariantIterationCount(const Loop *L);
1128 /// getIterationCount - If the specified loop has a predictable iteration
1129 /// count, return it. Note that it is not valid to call this method on a
1130 /// loop without a loop-invariant iteration count.
1131 SCEVHandle getIterationCount(const Loop *L);
1133 /// deleteInstructionFromRecords - This method should be called by the
1134 /// client before it removes an instruction from the program, to make sure
1135 /// that no dangling references are left around.
1136 void deleteInstructionFromRecords(Instruction *I);
1139 /// createSCEV - We know that there is no SCEV for the specified value.
1140 /// Analyze the expression.
1141 SCEVHandle createSCEV(Value *V);
1143 /// createNodeForPHI - Provide the special handling we need to analyze PHI
1145 SCEVHandle createNodeForPHI(PHINode *PN);
1147 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value
1148 /// for the specified instruction and replaces any references to the
1149 /// symbolic value SymName with the specified value. This is used during
1151 void ReplaceSymbolicValueWithConcrete(Instruction *I,
1152 const SCEVHandle &SymName,
1153 const SCEVHandle &NewVal);
1155 /// ComputeIterationCount - Compute the number of times the specified loop
1157 SCEVHandle ComputeIterationCount(const Loop *L);
1159 /// ComputeLoadConstantCompareIterationCount - Given an exit condition of
1160 /// 'setcc load X, cst', try to se if we can compute the trip count.
1161 SCEVHandle ComputeLoadConstantCompareIterationCount(LoadInst *LI,
1164 unsigned SetCCOpcode);
1166 /// ComputeIterationCountExhaustively - If the trip is known to execute a
1167 /// constant number of times (the condition evolves only from constants),
1168 /// try to evaluate a few iterations of the loop until we get the exit
1169 /// condition gets a value of ExitWhen (true or false). If we cannot
1170 /// evaluate the trip count of the loop, return UnknownValue.
1171 SCEVHandle ComputeIterationCountExhaustively(const Loop *L, Value *Cond,
1174 /// HowFarToZero - Return the number of times a backedge comparing the
1175 /// specified value to zero will execute. If not computable, return
1177 SCEVHandle HowFarToZero(SCEV *V, const Loop *L);
1179 /// HowFarToNonZero - Return the number of times a backedge checking the
1180 /// specified value for nonzero will execute. If not computable, return
1182 SCEVHandle HowFarToNonZero(SCEV *V, const Loop *L);
1184 /// HowManyLessThans - Return the number of times a backedge containing the
1185 /// specified less-than comparison will execute. If not computable, return
1187 SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L);
1189 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
1190 /// in the header of its containing loop, we know the loop executes a
1191 /// constant number of times, and the PHI node is just a recurrence
1192 /// involving constants, fold it.
1193 Constant *getConstantEvolutionLoopExitValue(PHINode *PN, uint64_t Its,
1198 //===----------------------------------------------------------------------===//
1199 // Basic SCEV Analysis and PHI Idiom Recognition Code
1202 /// deleteInstructionFromRecords - This method should be called by the
1203 /// client before it removes an instruction from the program, to make sure
1204 /// that no dangling references are left around.
1205 void ScalarEvolutionsImpl::deleteInstructionFromRecords(Instruction *I) {
1207 if (PHINode *PN = dyn_cast<PHINode>(I))
1208 ConstantEvolutionLoopExitValue.erase(PN);
1212 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1213 /// expression and create a new one.
1214 SCEVHandle ScalarEvolutionsImpl::getSCEV(Value *V) {
1215 assert(V->getType() != Type::VoidTy && "Can't analyze void expressions!");
1217 std::map<Value*, SCEVHandle>::iterator I = Scalars.find(V);
1218 if (I != Scalars.end()) return I->second;
1219 SCEVHandle S = createSCEV(V);
1220 Scalars.insert(std::make_pair(V, S));
1224 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value for
1225 /// the specified instruction and replaces any references to the symbolic value
1226 /// SymName with the specified value. This is used during PHI resolution.
1227 void ScalarEvolutionsImpl::
1228 ReplaceSymbolicValueWithConcrete(Instruction *I, const SCEVHandle &SymName,
1229 const SCEVHandle &NewVal) {
1230 std::map<Value*, SCEVHandle>::iterator SI = Scalars.find(I);
1231 if (SI == Scalars.end()) return;
1234 SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal);
1235 if (NV == SI->second) return; // No change.
1237 SI->second = NV; // Update the scalars map!
1239 // Any instruction values that use this instruction might also need to be
1241 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1243 ReplaceSymbolicValueWithConcrete(cast<Instruction>(*UI), SymName, NewVal);
1246 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
1247 /// a loop header, making it a potential recurrence, or it doesn't.
1249 SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) {
1250 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
1251 if (const Loop *L = LI.getLoopFor(PN->getParent()))
1252 if (L->getHeader() == PN->getParent()) {
1253 // If it lives in the loop header, it has two incoming values, one
1254 // from outside the loop, and one from inside.
1255 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1256 unsigned BackEdge = IncomingEdge^1;
1258 // While we are analyzing this PHI node, handle its value symbolically.
1259 SCEVHandle SymbolicName = SCEVUnknown::get(PN);
1260 assert(Scalars.find(PN) == Scalars.end() &&
1261 "PHI node already processed?");
1262 Scalars.insert(std::make_pair(PN, SymbolicName));
1264 // Using this symbolic name for the PHI, analyze the value coming around
1266 SCEVHandle BEValue = getSCEV(PN->getIncomingValue(BackEdge));
1268 // NOTE: If BEValue is loop invariant, we know that the PHI node just
1269 // has a special value for the first iteration of the loop.
1271 // If the value coming around the backedge is an add with the symbolic
1272 // value we just inserted, then we found a simple induction variable!
1273 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
1274 // If there is a single occurrence of the symbolic value, replace it
1275 // with a recurrence.
1276 unsigned FoundIndex = Add->getNumOperands();
1277 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1278 if (Add->getOperand(i) == SymbolicName)
1279 if (FoundIndex == e) {
1284 if (FoundIndex != Add->getNumOperands()) {
1285 // Create an add with everything but the specified operand.
1286 std::vector<SCEVHandle> Ops;
1287 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1288 if (i != FoundIndex)
1289 Ops.push_back(Add->getOperand(i));
1290 SCEVHandle Accum = SCEVAddExpr::get(Ops);
1292 // This is not a valid addrec if the step amount is varying each
1293 // loop iteration, but is not itself an addrec in this loop.
1294 if (Accum->isLoopInvariant(L) ||
1295 (isa<SCEVAddRecExpr>(Accum) &&
1296 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
1297 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1298 SCEVHandle PHISCEV = SCEVAddRecExpr::get(StartVal, Accum, L);
1300 // Okay, for the entire analysis of this edge we assumed the PHI
1301 // to be symbolic. We now need to go back and update all of the
1302 // entries for the scalars that use the PHI (except for the PHI
1303 // itself) to use the new analyzed value instead of the "symbolic"
1305 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1309 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(BEValue)) {
1310 // Otherwise, this could be a loop like this:
1311 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
1312 // In this case, j = {1,+,1} and BEValue is j.
1313 // Because the other in-value of i (0) fits the evolution of BEValue
1314 // i really is an addrec evolution.
1315 if (AddRec->getLoop() == L && AddRec->isAffine()) {
1316 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1318 // If StartVal = j.start - j.stride, we can use StartVal as the
1319 // initial step of the addrec evolution.
1320 if (StartVal == SCEV::getMinusSCEV(AddRec->getOperand(0),
1321 AddRec->getOperand(1))) {
1322 SCEVHandle PHISCEV =
1323 SCEVAddRecExpr::get(StartVal, AddRec->getOperand(1), L);
1325 // Okay, for the entire analysis of this edge we assumed the PHI
1326 // to be symbolic. We now need to go back and update all of the
1327 // entries for the scalars that use the PHI (except for the PHI
1328 // itself) to use the new analyzed value instead of the "symbolic"
1330 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1336 return SymbolicName;
1339 // If it's not a loop phi, we can't handle it yet.
1340 return SCEVUnknown::get(PN);
1344 /// createSCEV - We know that there is no SCEV for the specified value.
1345 /// Analyze the expression.
1347 SCEVHandle ScalarEvolutionsImpl::createSCEV(Value *V) {
1348 if (Instruction *I = dyn_cast<Instruction>(V)) {
1349 switch (I->getOpcode()) {
1350 case Instruction::Add:
1351 return SCEVAddExpr::get(getSCEV(I->getOperand(0)),
1352 getSCEV(I->getOperand(1)));
1353 case Instruction::Mul:
1354 return SCEVMulExpr::get(getSCEV(I->getOperand(0)),
1355 getSCEV(I->getOperand(1)));
1356 case Instruction::SDiv:
1357 return SCEVSDivExpr::get(getSCEV(I->getOperand(0)),
1358 getSCEV(I->getOperand(1)));
1361 case Instruction::Sub:
1362 return SCEV::getMinusSCEV(getSCEV(I->getOperand(0)),
1363 getSCEV(I->getOperand(1)));
1365 case Instruction::Shl:
1366 // Turn shift left of a constant amount into a multiply.
1367 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1368 Constant *X = ConstantInt::get(V->getType(), 1);
1369 X = ConstantExpr::getShl(X, SA);
1370 return SCEVMulExpr::get(getSCEV(I->getOperand(0)), getSCEV(X));
1374 case Instruction::Trunc:
1375 if (I->getType()->isInteger() && I->getOperand(0)->getType()->isInteger())
1376 return SCEVTruncateExpr::get(getSCEV(I->getOperand(0)),
1377 I->getType()->getUnsignedVersion());
1380 case Instruction::ZExt:
1381 if (I->getType()->isInteger() && I->getOperand(0)->getType()->isInteger())
1382 return SCEVZeroExtendExpr::get(getSCEV(I->getOperand(0)),
1383 I->getType()->getUnsignedVersion());
1386 case Instruction::BitCast:
1387 // BitCasts are no-op casts so we just eliminate the cast.
1388 return getSCEV(I->getOperand(0));
1390 case Instruction::PHI:
1391 return createNodeForPHI(cast<PHINode>(I));
1393 default: // We cannot analyze this expression.
1398 return SCEVUnknown::get(V);
1403 //===----------------------------------------------------------------------===//
1404 // Iteration Count Computation Code
1407 /// getIterationCount - If the specified loop has a predictable iteration
1408 /// count, return it. Note that it is not valid to call this method on a
1409 /// loop without a loop-invariant iteration count.
1410 SCEVHandle ScalarEvolutionsImpl::getIterationCount(const Loop *L) {
1411 std::map<const Loop*, SCEVHandle>::iterator I = IterationCounts.find(L);
1412 if (I == IterationCounts.end()) {
1413 SCEVHandle ItCount = ComputeIterationCount(L);
1414 I = IterationCounts.insert(std::make_pair(L, ItCount)).first;
1415 if (ItCount != UnknownValue) {
1416 assert(ItCount->isLoopInvariant(L) &&
1417 "Computed trip count isn't loop invariant for loop!");
1418 ++NumTripCountsComputed;
1419 } else if (isa<PHINode>(L->getHeader()->begin())) {
1420 // Only count loops that have phi nodes as not being computable.
1421 ++NumTripCountsNotComputed;
1427 /// ComputeIterationCount - Compute the number of times the specified loop
1429 SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) {
1430 // If the loop has a non-one exit block count, we can't analyze it.
1431 std::vector<BasicBlock*> ExitBlocks;
1432 L->getExitBlocks(ExitBlocks);
1433 if (ExitBlocks.size() != 1) return UnknownValue;
1435 // Okay, there is one exit block. Try to find the condition that causes the
1436 // loop to be exited.
1437 BasicBlock *ExitBlock = ExitBlocks[0];
1439 BasicBlock *ExitingBlock = 0;
1440 for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock);
1442 if (L->contains(*PI)) {
1443 if (ExitingBlock == 0)
1446 return UnknownValue; // More than one block exiting!
1448 assert(ExitingBlock && "No exits from loop, something is broken!");
1450 // Okay, we've computed the exiting block. See what condition causes us to
1453 // FIXME: we should be able to handle switch instructions (with a single exit)
1454 // FIXME: We should handle cast of int to bool as well
1455 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1456 if (ExitBr == 0) return UnknownValue;
1457 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
1458 SetCondInst *ExitCond = dyn_cast<SetCondInst>(ExitBr->getCondition());
1459 if (ExitCond == 0) // Not a setcc
1460 return ComputeIterationCountExhaustively(L, ExitBr->getCondition(),
1461 ExitBr->getSuccessor(0) == ExitBlock);
1463 // If the condition was exit on true, convert the condition to exit on false.
1464 Instruction::BinaryOps Cond;
1465 if (ExitBr->getSuccessor(1) == ExitBlock)
1466 Cond = ExitCond->getOpcode();
1468 Cond = ExitCond->getInverseCondition();
1470 // Handle common loops like: for (X = "string"; *X; ++X)
1471 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
1472 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
1474 ComputeLoadConstantCompareIterationCount(LI, RHS, L, Cond);
1475 if (!isa<SCEVCouldNotCompute>(ItCnt)) return ItCnt;
1478 SCEVHandle LHS = getSCEV(ExitCond->getOperand(0));
1479 SCEVHandle RHS = getSCEV(ExitCond->getOperand(1));
1481 // Try to evaluate any dependencies out of the loop.
1482 SCEVHandle Tmp = getSCEVAtScope(LHS, L);
1483 if (!isa<SCEVCouldNotCompute>(Tmp)) LHS = Tmp;
1484 Tmp = getSCEVAtScope(RHS, L);
1485 if (!isa<SCEVCouldNotCompute>(Tmp)) RHS = Tmp;
1487 // At this point, we would like to compute how many iterations of the loop the
1488 // predicate will return true for these inputs.
1489 if (isa<SCEVConstant>(LHS) && !isa<SCEVConstant>(RHS)) {
1490 // If there is a constant, force it into the RHS.
1491 std::swap(LHS, RHS);
1492 Cond = SetCondInst::getSwappedCondition(Cond);
1495 // FIXME: think about handling pointer comparisons! i.e.:
1496 // while (P != P+100) ++P;
1498 // If we have a comparison of a chrec against a constant, try to use value
1499 // ranges to answer this query.
1500 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
1501 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
1502 if (AddRec->getLoop() == L) {
1503 // Form the comparison range using the constant of the correct type so
1504 // that the ConstantRange class knows to do a signed or unsigned
1506 ConstantInt *CompVal = RHSC->getValue();
1507 const Type *RealTy = ExitCond->getOperand(0)->getType();
1508 CompVal = dyn_cast<ConstantInt>(ConstantExpr::getCast(CompVal, RealTy));
1510 // Form the constant range.
1511 ConstantRange CompRange(Cond, CompVal);
1513 // Now that we have it, if it's signed, convert it to an unsigned
1515 if (CompRange.getLower()->getType()->isSigned()) {
1516 const Type *NewTy = RHSC->getValue()->getType();
1517 Constant *NewL = ConstantExpr::getCast(CompRange.getLower(), NewTy);
1518 Constant *NewU = ConstantExpr::getCast(CompRange.getUpper(), NewTy);
1519 CompRange = ConstantRange(NewL, NewU);
1522 SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange);
1523 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
1528 case Instruction::SetNE: // while (X != Y)
1529 // Convert to: while (X-Y != 0)
1530 if (LHS->getType()->isInteger()) {
1531 SCEVHandle TC = HowFarToZero(SCEV::getMinusSCEV(LHS, RHS), L);
1532 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1535 case Instruction::SetEQ:
1536 // Convert to: while (X-Y == 0) // while (X == Y)
1537 if (LHS->getType()->isInteger()) {
1538 SCEVHandle TC = HowFarToNonZero(SCEV::getMinusSCEV(LHS, RHS), L);
1539 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1542 case Instruction::SetLT:
1543 if (LHS->getType()->isInteger() &&
1544 ExitCond->getOperand(0)->getType()->isSigned()) {
1545 SCEVHandle TC = HowManyLessThans(LHS, RHS, L);
1546 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1549 case Instruction::SetGT:
1550 if (LHS->getType()->isInteger() &&
1551 ExitCond->getOperand(0)->getType()->isSigned()) {
1552 SCEVHandle TC = HowManyLessThans(RHS, LHS, L);
1553 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1558 llvm_cerr << "ComputeIterationCount ";
1559 if (ExitCond->getOperand(0)->getType()->isUnsigned())
1560 llvm_cerr << "[unsigned] ";
1561 llvm_cerr << *LHS << " "
1562 << Instruction::getOpcodeName(Cond) << " " << *RHS << "\n";
1567 return ComputeIterationCountExhaustively(L, ExitCond,
1568 ExitBr->getSuccessor(0) == ExitBlock);
1571 static ConstantInt *
1572 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, Constant *C) {
1573 SCEVHandle InVal = SCEVConstant::get(cast<ConstantInt>(C));
1574 SCEVHandle Val = AddRec->evaluateAtIteration(InVal);
1575 assert(isa<SCEVConstant>(Val) &&
1576 "Evaluation of SCEV at constant didn't fold correctly?");
1577 return cast<SCEVConstant>(Val)->getValue();
1580 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
1581 /// and a GEP expression (missing the pointer index) indexing into it, return
1582 /// the addressed element of the initializer or null if the index expression is
1585 GetAddressedElementFromGlobal(GlobalVariable *GV,
1586 const std::vector<ConstantInt*> &Indices) {
1587 Constant *Init = GV->getInitializer();
1588 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1589 uint64_t Idx = Indices[i]->getZExtValue();
1590 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1591 assert(Idx < CS->getNumOperands() && "Bad struct index!");
1592 Init = cast<Constant>(CS->getOperand(Idx));
1593 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1594 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
1595 Init = cast<Constant>(CA->getOperand(Idx));
1596 } else if (isa<ConstantAggregateZero>(Init)) {
1597 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1598 assert(Idx < STy->getNumElements() && "Bad struct index!");
1599 Init = Constant::getNullValue(STy->getElementType(Idx));
1600 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
1601 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
1602 Init = Constant::getNullValue(ATy->getElementType());
1604 assert(0 && "Unknown constant aggregate type!");
1608 return 0; // Unknown initializer type
1614 /// ComputeLoadConstantCompareIterationCount - Given an exit condition of
1615 /// 'setcc load X, cst', try to se if we can compute the trip count.
1616 SCEVHandle ScalarEvolutionsImpl::
1617 ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS,
1618 const Loop *L, unsigned SetCCOpcode) {
1619 if (LI->isVolatile()) return UnknownValue;
1621 // Check to see if the loaded pointer is a getelementptr of a global.
1622 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
1623 if (!GEP) return UnknownValue;
1625 // Make sure that it is really a constant global we are gepping, with an
1626 // initializer, and make sure the first IDX is really 0.
1627 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1628 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1629 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
1630 !cast<Constant>(GEP->getOperand(1))->isNullValue())
1631 return UnknownValue;
1633 // Okay, we allow one non-constant index into the GEP instruction.
1635 std::vector<ConstantInt*> Indexes;
1636 unsigned VarIdxNum = 0;
1637 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
1638 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
1639 Indexes.push_back(CI);
1640 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
1641 if (VarIdx) return UnknownValue; // Multiple non-constant idx's.
1642 VarIdx = GEP->getOperand(i);
1644 Indexes.push_back(0);
1647 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
1648 // Check to see if X is a loop variant variable value now.
1649 SCEVHandle Idx = getSCEV(VarIdx);
1650 SCEVHandle Tmp = getSCEVAtScope(Idx, L);
1651 if (!isa<SCEVCouldNotCompute>(Tmp)) Idx = Tmp;
1653 // We can only recognize very limited forms of loop index expressions, in
1654 // particular, only affine AddRec's like {C1,+,C2}.
1655 SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
1656 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
1657 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
1658 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
1659 return UnknownValue;
1661 unsigned MaxSteps = MaxBruteForceIterations;
1662 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
1663 ConstantInt *ItCst =
1664 ConstantInt::get(IdxExpr->getType()->getUnsignedVersion(), IterationNum);
1665 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst);
1667 // Form the GEP offset.
1668 Indexes[VarIdxNum] = Val;
1670 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
1671 if (Result == 0) break; // Cannot compute!
1673 // Evaluate the condition for this iteration.
1674 Result = ConstantExpr::get(SetCCOpcode, Result, RHS);
1675 if (!isa<ConstantBool>(Result)) break; // Couldn't decide for sure
1676 if (cast<ConstantBool>(Result)->getValue() == false) {
1678 llvm_cerr << "\n***\n*** Computed loop count " << *ItCst
1679 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
1682 ++NumArrayLenItCounts;
1683 return SCEVConstant::get(ItCst); // Found terminating iteration!
1686 return UnknownValue;
1690 /// CanConstantFold - Return true if we can constant fold an instruction of the
1691 /// specified type, assuming that all operands were constants.
1692 static bool CanConstantFold(const Instruction *I) {
1693 if (isa<BinaryOperator>(I) || isa<ShiftInst>(I) ||
1694 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
1697 if (const CallInst *CI = dyn_cast<CallInst>(I))
1698 if (const Function *F = CI->getCalledFunction())
1699 return canConstantFoldCallTo((Function*)F); // FIXME: elim cast
1703 /// ConstantFold - Constant fold an instruction of the specified type with the
1704 /// specified constant operands. This function may modify the operands vector.
1705 static Constant *ConstantFold(const Instruction *I,
1706 std::vector<Constant*> &Operands) {
1707 if (isa<BinaryOperator>(I) || isa<ShiftInst>(I))
1708 return ConstantExpr::get(I->getOpcode(), Operands[0], Operands[1]);
1710 if (isa<CastInst>(I))
1711 return ConstantExpr::getCast(I->getOpcode(), Operands[0], I->getType());
1713 switch (I->getOpcode()) {
1714 case Instruction::Select:
1715 return ConstantExpr::getSelect(Operands[0], Operands[1], Operands[2]);
1716 case Instruction::Call:
1717 if (Function *GV = dyn_cast<Function>(Operands[0])) {
1718 Operands.erase(Operands.begin());
1719 return ConstantFoldCall(cast<Function>(GV), Operands);
1722 case Instruction::GetElementPtr:
1723 Constant *Base = Operands[0];
1724 Operands.erase(Operands.begin());
1725 return ConstantExpr::getGetElementPtr(Base, Operands);
1731 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
1732 /// in the loop that V is derived from. We allow arbitrary operations along the
1733 /// way, but the operands of an operation must either be constants or a value
1734 /// derived from a constant PHI. If this expression does not fit with these
1735 /// constraints, return null.
1736 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
1737 // If this is not an instruction, or if this is an instruction outside of the
1738 // loop, it can't be derived from a loop PHI.
1739 Instruction *I = dyn_cast<Instruction>(V);
1740 if (I == 0 || !L->contains(I->getParent())) return 0;
1742 if (PHINode *PN = dyn_cast<PHINode>(I))
1743 if (L->getHeader() == I->getParent())
1746 // We don't currently keep track of the control flow needed to evaluate
1747 // PHIs, so we cannot handle PHIs inside of loops.
1750 // If we won't be able to constant fold this expression even if the operands
1751 // are constants, return early.
1752 if (!CanConstantFold(I)) return 0;
1754 // Otherwise, we can evaluate this instruction if all of its operands are
1755 // constant or derived from a PHI node themselves.
1757 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
1758 if (!(isa<Constant>(I->getOperand(Op)) ||
1759 isa<GlobalValue>(I->getOperand(Op)))) {
1760 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
1761 if (P == 0) return 0; // Not evolving from PHI
1765 return 0; // Evolving from multiple different PHIs.
1768 // This is a expression evolving from a constant PHI!
1772 /// EvaluateExpression - Given an expression that passes the
1773 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
1774 /// in the loop has the value PHIVal. If we can't fold this expression for some
1775 /// reason, return null.
1776 static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
1777 if (isa<PHINode>(V)) return PHIVal;
1778 if (GlobalValue *GV = dyn_cast<GlobalValue>(V))
1780 if (Constant *C = dyn_cast<Constant>(V)) return C;
1781 Instruction *I = cast<Instruction>(V);
1783 std::vector<Constant*> Operands;
1784 Operands.resize(I->getNumOperands());
1786 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
1787 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal);
1788 if (Operands[i] == 0) return 0;
1791 return ConstantFold(I, Operands);
1794 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
1795 /// in the header of its containing loop, we know the loop executes a
1796 /// constant number of times, and the PHI node is just a recurrence
1797 /// involving constants, fold it.
1798 Constant *ScalarEvolutionsImpl::
1799 getConstantEvolutionLoopExitValue(PHINode *PN, uint64_t Its, const Loop *L) {
1800 std::map<PHINode*, Constant*>::iterator I =
1801 ConstantEvolutionLoopExitValue.find(PN);
1802 if (I != ConstantEvolutionLoopExitValue.end())
1805 if (Its > MaxBruteForceIterations)
1806 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
1808 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
1810 // Since the loop is canonicalized, the PHI node must have two entries. One
1811 // entry must be a constant (coming in from outside of the loop), and the
1812 // second must be derived from the same PHI.
1813 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
1814 Constant *StartCST =
1815 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
1817 return RetVal = 0; // Must be a constant.
1819 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
1820 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
1822 return RetVal = 0; // Not derived from same PHI.
1824 // Execute the loop symbolically to determine the exit value.
1825 unsigned IterationNum = 0;
1826 unsigned NumIterations = Its;
1827 if (NumIterations != Its)
1828 return RetVal = 0; // More than 2^32 iterations??
1830 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
1831 if (IterationNum == NumIterations)
1832 return RetVal = PHIVal; // Got exit value!
1834 // Compute the value of the PHI node for the next iteration.
1835 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
1836 if (NextPHI == PHIVal)
1837 return RetVal = NextPHI; // Stopped evolving!
1839 return 0; // Couldn't evaluate!
1844 /// ComputeIterationCountExhaustively - If the trip is known to execute a
1845 /// constant number of times (the condition evolves only from constants),
1846 /// try to evaluate a few iterations of the loop until we get the exit
1847 /// condition gets a value of ExitWhen (true or false). If we cannot
1848 /// evaluate the trip count of the loop, return UnknownValue.
1849 SCEVHandle ScalarEvolutionsImpl::
1850 ComputeIterationCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) {
1851 PHINode *PN = getConstantEvolvingPHI(Cond, L);
1852 if (PN == 0) return UnknownValue;
1854 // Since the loop is canonicalized, the PHI node must have two entries. One
1855 // entry must be a constant (coming in from outside of the loop), and the
1856 // second must be derived from the same PHI.
1857 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
1858 Constant *StartCST =
1859 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
1860 if (StartCST == 0) return UnknownValue; // Must be a constant.
1862 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
1863 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
1864 if (PN2 != PN) return UnknownValue; // Not derived from same PHI.
1866 // Okay, we find a PHI node that defines the trip count of this loop. Execute
1867 // the loop symbolically to determine when the condition gets a value of
1869 unsigned IterationNum = 0;
1870 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
1871 for (Constant *PHIVal = StartCST;
1872 IterationNum != MaxIterations; ++IterationNum) {
1873 ConstantBool *CondVal =
1874 dyn_cast_or_null<ConstantBool>(EvaluateExpression(Cond, PHIVal));
1875 if (!CondVal) return UnknownValue; // Couldn't symbolically evaluate.
1877 if (CondVal->getValue() == ExitWhen) {
1878 ConstantEvolutionLoopExitValue[PN] = PHIVal;
1879 ++NumBruteForceTripCountsComputed;
1880 return SCEVConstant::get(ConstantInt::get(Type::UIntTy, IterationNum));
1883 // Compute the value of the PHI node for the next iteration.
1884 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
1885 if (NextPHI == 0 || NextPHI == PHIVal)
1886 return UnknownValue; // Couldn't evaluate or not making progress...
1890 // Too many iterations were needed to evaluate.
1891 return UnknownValue;
1894 /// getSCEVAtScope - Compute the value of the specified expression within the
1895 /// indicated loop (which may be null to indicate in no loop). If the
1896 /// expression cannot be evaluated, return UnknownValue.
1897 SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) {
1898 // FIXME: this should be turned into a virtual method on SCEV!
1900 if (isa<SCEVConstant>(V)) return V;
1902 // If this instruction is evolves from a constant-evolving PHI, compute the
1903 // exit value from the loop without using SCEVs.
1904 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
1905 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
1906 const Loop *LI = this->LI[I->getParent()];
1907 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
1908 if (PHINode *PN = dyn_cast<PHINode>(I))
1909 if (PN->getParent() == LI->getHeader()) {
1910 // Okay, there is no closed form solution for the PHI node. Check
1911 // to see if the loop that contains it has a known iteration count.
1912 // If so, we may be able to force computation of the exit value.
1913 SCEVHandle IterationCount = getIterationCount(LI);
1914 if (SCEVConstant *ICC = dyn_cast<SCEVConstant>(IterationCount)) {
1915 // Okay, we know how many times the containing loop executes. If
1916 // this is a constant evolving PHI node, get the final value at
1917 // the specified iteration number.
1918 Constant *RV = getConstantEvolutionLoopExitValue(PN,
1919 ICC->getValue()->getZExtValue(),
1921 if (RV) return SCEVUnknown::get(RV);
1925 // Okay, this is a some expression that we cannot symbolically evaluate
1926 // into a SCEV. Check to see if it's possible to symbolically evaluate
1927 // the arguments into constants, and if see, try to constant propagate the
1928 // result. This is particularly useful for computing loop exit values.
1929 if (CanConstantFold(I)) {
1930 std::vector<Constant*> Operands;
1931 Operands.reserve(I->getNumOperands());
1932 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
1933 Value *Op = I->getOperand(i);
1934 if (Constant *C = dyn_cast<Constant>(Op)) {
1935 Operands.push_back(C);
1937 SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L);
1938 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
1939 Operands.push_back(ConstantExpr::getCast(SC->getValue(),
1941 else if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
1942 if (Constant *C = dyn_cast<Constant>(SU->getValue()))
1943 Operands.push_back(ConstantExpr::getCast(C, Op->getType()));
1951 return SCEVUnknown::get(ConstantFold(I, Operands));
1955 // This is some other type of SCEVUnknown, just return it.
1959 if (SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
1960 // Avoid performing the look-up in the common case where the specified
1961 // expression has no loop-variant portions.
1962 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
1963 SCEVHandle OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
1964 if (OpAtScope != Comm->getOperand(i)) {
1965 if (OpAtScope == UnknownValue) return UnknownValue;
1966 // Okay, at least one of these operands is loop variant but might be
1967 // foldable. Build a new instance of the folded commutative expression.
1968 std::vector<SCEVHandle> NewOps(Comm->op_begin(), Comm->op_begin()+i);
1969 NewOps.push_back(OpAtScope);
1971 for (++i; i != e; ++i) {
1972 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
1973 if (OpAtScope == UnknownValue) return UnknownValue;
1974 NewOps.push_back(OpAtScope);
1976 if (isa<SCEVAddExpr>(Comm))
1977 return SCEVAddExpr::get(NewOps);
1978 assert(isa<SCEVMulExpr>(Comm) && "Only know about add and mul!");
1979 return SCEVMulExpr::get(NewOps);
1982 // If we got here, all operands are loop invariant.
1986 if (SCEVSDivExpr *Div = dyn_cast<SCEVSDivExpr>(V)) {
1987 SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L);
1988 if (LHS == UnknownValue) return LHS;
1989 SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L);
1990 if (RHS == UnknownValue) return RHS;
1991 if (LHS == Div->getLHS() && RHS == Div->getRHS())
1992 return Div; // must be loop invariant
1993 return SCEVSDivExpr::get(LHS, RHS);
1996 // If this is a loop recurrence for a loop that does not contain L, then we
1997 // are dealing with the final value computed by the loop.
1998 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
1999 if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
2000 // To evaluate this recurrence, we need to know how many times the AddRec
2001 // loop iterates. Compute this now.
2002 SCEVHandle IterationCount = getIterationCount(AddRec->getLoop());
2003 if (IterationCount == UnknownValue) return UnknownValue;
2004 IterationCount = getTruncateOrZeroExtend(IterationCount,
2007 // If the value is affine, simplify the expression evaluation to just
2008 // Start + Step*IterationCount.
2009 if (AddRec->isAffine())
2010 return SCEVAddExpr::get(AddRec->getStart(),
2011 SCEVMulExpr::get(IterationCount,
2012 AddRec->getOperand(1)));
2014 // Otherwise, evaluate it the hard way.
2015 return AddRec->evaluateAtIteration(IterationCount);
2017 return UnknownValue;
2020 //assert(0 && "Unknown SCEV type!");
2021 return UnknownValue;
2025 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
2026 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
2027 /// might be the same) or two SCEVCouldNotCompute objects.
2029 static std::pair<SCEVHandle,SCEVHandle>
2030 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec) {
2031 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
2032 SCEVConstant *L = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
2033 SCEVConstant *M = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
2034 SCEVConstant *N = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
2036 // We currently can only solve this if the coefficients are constants.
2037 if (!L || !M || !N) {
2038 SCEV *CNC = new SCEVCouldNotCompute();
2039 return std::make_pair(CNC, CNC);
2042 Constant *C = L->getValue();
2043 Constant *Two = ConstantInt::get(C->getType(), 2);
2045 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
2046 // The B coefficient is M-N/2
2047 Constant *B = ConstantExpr::getSub(M->getValue(),
2048 ConstantExpr::getSDiv(N->getValue(),
2050 // The A coefficient is N/2
2051 Constant *A = ConstantExpr::getSDiv(N->getValue(), Two);
2053 // Compute the B^2-4ac term.
2054 Constant *SqrtTerm =
2055 ConstantExpr::getMul(ConstantInt::get(C->getType(), 4),
2056 ConstantExpr::getMul(A, C));
2057 SqrtTerm = ConstantExpr::getSub(ConstantExpr::getMul(B, B), SqrtTerm);
2059 // Compute floor(sqrt(B^2-4ac))
2060 ConstantInt *SqrtVal =
2061 cast<ConstantInt>(ConstantExpr::getCast(SqrtTerm,
2062 SqrtTerm->getType()->getUnsignedVersion()));
2063 uint64_t SqrtValV = SqrtVal->getZExtValue();
2064 uint64_t SqrtValV2 = (uint64_t)sqrt((double)SqrtValV);
2065 // The square root might not be precise for arbitrary 64-bit integer
2066 // values. Do some sanity checks to ensure it's correct.
2067 if (SqrtValV2*SqrtValV2 > SqrtValV ||
2068 (SqrtValV2+1)*(SqrtValV2+1) <= SqrtValV) {
2069 SCEV *CNC = new SCEVCouldNotCompute();
2070 return std::make_pair(CNC, CNC);
2073 SqrtVal = ConstantInt::get(Type::ULongTy, SqrtValV2);
2074 SqrtTerm = ConstantExpr::getCast(SqrtVal, SqrtTerm->getType());
2076 Constant *NegB = ConstantExpr::getNeg(B);
2077 Constant *TwoA = ConstantExpr::getMul(A, Two);
2079 // The divisions must be performed as signed divisions.
2080 const Type *SignedTy = NegB->getType()->getSignedVersion();
2081 NegB = ConstantExpr::getCast(NegB, SignedTy);
2082 TwoA = ConstantExpr::getCast(TwoA, SignedTy);
2083 SqrtTerm = ConstantExpr::getCast(SqrtTerm, SignedTy);
2085 Constant *Solution1 =
2086 ConstantExpr::getSDiv(ConstantExpr::getAdd(NegB, SqrtTerm), TwoA);
2087 Constant *Solution2 =
2088 ConstantExpr::getSDiv(ConstantExpr::getSub(NegB, SqrtTerm), TwoA);
2089 return std::make_pair(SCEVUnknown::get(Solution1),
2090 SCEVUnknown::get(Solution2));
2093 /// HowFarToZero - Return the number of times a backedge comparing the specified
2094 /// value to zero will execute. If not computable, return UnknownValue
2095 SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) {
2096 // If the value is a constant
2097 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2098 // If the value is already zero, the branch will execute zero times.
2099 if (C->getValue()->isNullValue()) return C;
2100 return UnknownValue; // Otherwise it will loop infinitely.
2103 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
2104 if (!AddRec || AddRec->getLoop() != L)
2105 return UnknownValue;
2107 if (AddRec->isAffine()) {
2108 // If this is an affine expression the execution count of this branch is
2111 // (0 - Start/Step) iff Start % Step == 0
2113 // Get the initial value for the loop.
2114 SCEVHandle Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
2115 if (isa<SCEVCouldNotCompute>(Start)) return UnknownValue;
2116 SCEVHandle Step = AddRec->getOperand(1);
2118 Step = getSCEVAtScope(Step, L->getParentLoop());
2120 // Figure out if Start % Step == 0.
2121 // FIXME: We should add DivExpr and RemExpr operations to our AST.
2122 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
2123 if (StepC->getValue()->equalsInt(1)) // N % 1 == 0
2124 return SCEV::getNegativeSCEV(Start); // 0 - Start/1 == -Start
2125 if (StepC->getValue()->isAllOnesValue()) // N % -1 == 0
2126 return Start; // 0 - Start/-1 == Start
2128 // Check to see if Start is divisible by SC with no remainder.
2129 if (SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start)) {
2130 ConstantInt *StartCC = StartC->getValue();
2131 Constant *StartNegC = ConstantExpr::getNeg(StartCC);
2132 Constant *Rem = ConstantExpr::getSRem(StartNegC, StepC->getValue());
2133 if (Rem->isNullValue()) {
2134 Constant *Result =ConstantExpr::getSDiv(StartNegC,StepC->getValue());
2135 return SCEVUnknown::get(Result);
2139 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
2140 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
2141 // the quadratic equation to solve it.
2142 std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec);
2143 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2144 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2147 llvm_cerr << "HFTZ: " << *V << " - sol#1: " << *R1
2148 << " sol#2: " << *R2 << "\n";
2150 // Pick the smallest positive root value.
2151 assert(R1->getType()->isUnsigned()&&"Didn't canonicalize to unsigned?");
2152 if (ConstantBool *CB =
2153 dyn_cast<ConstantBool>(ConstantExpr::getSetLT(R1->getValue(),
2155 if (CB->getValue() == false)
2156 std::swap(R1, R2); // R1 is the minimum root now.
2158 // We can only use this value if the chrec ends up with an exact zero
2159 // value at this index. When solving for "X*X != 5", for example, we
2160 // should not accept a root of 2.
2161 SCEVHandle Val = AddRec->evaluateAtIteration(R1);
2162 if (SCEVConstant *EvalVal = dyn_cast<SCEVConstant>(Val))
2163 if (EvalVal->getValue()->isNullValue())
2164 return R1; // We found a quadratic root!
2169 return UnknownValue;
2172 /// HowFarToNonZero - Return the number of times a backedge checking the
2173 /// specified value for nonzero will execute. If not computable, return
2175 SCEVHandle ScalarEvolutionsImpl::HowFarToNonZero(SCEV *V, const Loop *L) {
2176 // Loops that look like: while (X == 0) are very strange indeed. We don't
2177 // handle them yet except for the trivial case. This could be expanded in the
2178 // future as needed.
2180 // If the value is a constant, check to see if it is known to be non-zero
2181 // already. If so, the backedge will execute zero times.
2182 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2183 Constant *Zero = Constant::getNullValue(C->getValue()->getType());
2184 Constant *NonZero = ConstantExpr::getSetNE(C->getValue(), Zero);
2185 if (NonZero == ConstantBool::getTrue())
2186 return getSCEV(Zero);
2187 return UnknownValue; // Otherwise it will loop infinitely.
2190 // We could implement others, but I really doubt anyone writes loops like
2191 // this, and if they did, they would already be constant folded.
2192 return UnknownValue;
2195 /// HowManyLessThans - Return the number of times a backedge containing the
2196 /// specified less-than comparison will execute. If not computable, return
2198 SCEVHandle ScalarEvolutionsImpl::
2199 HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L) {
2200 // Only handle: "ADDREC < LoopInvariant".
2201 if (!RHS->isLoopInvariant(L)) return UnknownValue;
2203 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
2204 if (!AddRec || AddRec->getLoop() != L)
2205 return UnknownValue;
2207 if (AddRec->isAffine()) {
2208 // FORNOW: We only support unit strides.
2209 SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, RHS->getType());
2210 if (AddRec->getOperand(1) != One)
2211 return UnknownValue;
2213 // The number of iterations for "[n,+,1] < m", is m-n. However, we don't
2214 // know that m is >= n on input to the loop. If it is, the condition return
2215 // true zero times. What we really should return, for full generality, is
2216 // SMAX(0, m-n). Since we cannot check this, we will instead check for a
2217 // canonical loop form: most do-loops will have a check that dominates the
2218 // loop, that only enters the loop if [n-1]<m. If we can find this check,
2219 // we know that the SMAX will evaluate to m-n, because we know that m >= n.
2221 // Search for the check.
2222 BasicBlock *Preheader = L->getLoopPreheader();
2223 BasicBlock *PreheaderDest = L->getHeader();
2224 if (Preheader == 0) return UnknownValue;
2226 BranchInst *LoopEntryPredicate =
2227 dyn_cast<BranchInst>(Preheader->getTerminator());
2228 if (!LoopEntryPredicate) return UnknownValue;
2230 // This might be a critical edge broken out. If the loop preheader ends in
2231 // an unconditional branch to the loop, check to see if the preheader has a
2232 // single predecessor, and if so, look for its terminator.
2233 while (LoopEntryPredicate->isUnconditional()) {
2234 PreheaderDest = Preheader;
2235 Preheader = Preheader->getSinglePredecessor();
2236 if (!Preheader) return UnknownValue; // Multiple preds.
2238 LoopEntryPredicate =
2239 dyn_cast<BranchInst>(Preheader->getTerminator());
2240 if (!LoopEntryPredicate) return UnknownValue;
2243 // Now that we found a conditional branch that dominates the loop, check to
2244 // see if it is the comparison we are looking for.
2245 SetCondInst *SCI =dyn_cast<SetCondInst>(LoopEntryPredicate->getCondition());
2246 if (!SCI) return UnknownValue;
2247 Value *PreCondLHS = SCI->getOperand(0);
2248 Value *PreCondRHS = SCI->getOperand(1);
2249 Instruction::BinaryOps Cond;
2250 if (LoopEntryPredicate->getSuccessor(0) == PreheaderDest)
2251 Cond = SCI->getOpcode();
2253 Cond = SCI->getInverseCondition();
2256 case Instruction::SetGT:
2257 std::swap(PreCondLHS, PreCondRHS);
2258 Cond = Instruction::SetLT;
2260 case Instruction::SetLT:
2261 if (PreCondLHS->getType()->isInteger() &&
2262 PreCondLHS->getType()->isSigned()) {
2263 if (RHS != getSCEV(PreCondRHS))
2264 return UnknownValue; // Not a comparison against 'm'.
2266 if (SCEV::getMinusSCEV(AddRec->getOperand(0), One)
2267 != getSCEV(PreCondLHS))
2268 return UnknownValue; // Not a comparison against 'n-1'.
2271 return UnknownValue;
2276 //llvm_cerr << "Computed Loop Trip Count as: " <<
2277 // *SCEV::getMinusSCEV(RHS, AddRec->getOperand(0)) << "\n";
2278 return SCEV::getMinusSCEV(RHS, AddRec->getOperand(0));
2281 return UnknownValue;
2284 /// getNumIterationsInRange - Return the number of iterations of this loop that
2285 /// produce values in the specified constant range. Another way of looking at
2286 /// this is that it returns the first iteration number where the value is not in
2287 /// the condition, thus computing the exit count. If the iteration count can't
2288 /// be computed, an instance of SCEVCouldNotCompute is returned.
2289 SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range) const {
2290 if (Range.isFullSet()) // Infinite loop.
2291 return new SCEVCouldNotCompute();
2293 // If the start is a non-zero constant, shift the range to simplify things.
2294 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
2295 if (!SC->getValue()->isNullValue()) {
2296 std::vector<SCEVHandle> Operands(op_begin(), op_end());
2297 Operands[0] = SCEVUnknown::getIntegerSCEV(0, SC->getType());
2298 SCEVHandle Shifted = SCEVAddRecExpr::get(Operands, getLoop());
2299 if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast<SCEVAddRecExpr>(Shifted))
2300 return ShiftedAddRec->getNumIterationsInRange(
2301 Range.subtract(SC->getValue()));
2302 // This is strange and shouldn't happen.
2303 return new SCEVCouldNotCompute();
2306 // The only time we can solve this is when we have all constant indices.
2307 // Otherwise, we cannot determine the overflow conditions.
2308 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
2309 if (!isa<SCEVConstant>(getOperand(i)))
2310 return new SCEVCouldNotCompute();
2313 // Okay at this point we know that all elements of the chrec are constants and
2314 // that the start element is zero.
2316 // First check to see if the range contains zero. If not, the first
2318 ConstantInt *Zero = ConstantInt::get(getType(), 0);
2319 if (!Range.contains(Zero)) return SCEVConstant::get(Zero);
2322 // If this is an affine expression then we have this situation:
2323 // Solve {0,+,A} in Range === Ax in Range
2325 // Since we know that zero is in the range, we know that the upper value of
2326 // the range must be the first possible exit value. Also note that we
2327 // already checked for a full range.
2328 ConstantInt *Upper = cast<ConstantInt>(Range.getUpper());
2329 ConstantInt *A = cast<SCEVConstant>(getOperand(1))->getValue();
2330 ConstantInt *One = ConstantInt::get(getType(), 1);
2332 // The exit value should be (Upper+A-1)/A.
2333 Constant *ExitValue = Upper;
2335 ExitValue = ConstantExpr::getSub(ConstantExpr::getAdd(Upper, A), One);
2336 ExitValue = ConstantExpr::getSDiv(ExitValue, A);
2338 assert(isa<ConstantInt>(ExitValue) &&
2339 "Constant folding of integers not implemented?");
2341 // Evaluate at the exit value. If we really did fall out of the valid
2342 // range, then we computed our trip count, otherwise wrap around or other
2343 // things must have happened.
2344 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue);
2345 if (Range.contains(Val))
2346 return new SCEVCouldNotCompute(); // Something strange happened
2348 // Ensure that the previous value is in the range. This is a sanity check.
2349 assert(Range.contains(EvaluateConstantChrecAtConstant(this,
2350 ConstantExpr::getSub(ExitValue, One))) &&
2351 "Linear scev computation is off in a bad way!");
2352 return SCEVConstant::get(cast<ConstantInt>(ExitValue));
2353 } else if (isQuadratic()) {
2354 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
2355 // quadratic equation to solve it. To do this, we must frame our problem in
2356 // terms of figuring out when zero is crossed, instead of when
2357 // Range.getUpper() is crossed.
2358 std::vector<SCEVHandle> NewOps(op_begin(), op_end());
2359 NewOps[0] = SCEV::getNegativeSCEV(SCEVUnknown::get(Range.getUpper()));
2360 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, getLoop());
2362 // Next, solve the constructed addrec
2363 std::pair<SCEVHandle,SCEVHandle> Roots =
2364 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec));
2365 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2366 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2368 // Pick the smallest positive root value.
2369 assert(R1->getType()->isUnsigned() && "Didn't canonicalize to unsigned?");
2370 if (ConstantBool *CB =
2371 dyn_cast<ConstantBool>(ConstantExpr::getSetLT(R1->getValue(),
2373 if (CB->getValue() == false)
2374 std::swap(R1, R2); // R1 is the minimum root now.
2376 // Make sure the root is not off by one. The returned iteration should
2377 // not be in the range, but the previous one should be. When solving
2378 // for "X*X < 5", for example, we should not return a root of 2.
2379 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
2381 if (Range.contains(R1Val)) {
2382 // The next iteration must be out of the range...
2384 ConstantExpr::getAdd(R1->getValue(),
2385 ConstantInt::get(R1->getType(), 1));
2387 R1Val = EvaluateConstantChrecAtConstant(this, NextVal);
2388 if (!Range.contains(R1Val))
2389 return SCEVUnknown::get(NextVal);
2390 return new SCEVCouldNotCompute(); // Something strange happened
2393 // If R1 was not in the range, then it is a good return value. Make
2394 // sure that R1-1 WAS in the range though, just in case.
2396 ConstantExpr::getSub(R1->getValue(),
2397 ConstantInt::get(R1->getType(), 1));
2398 R1Val = EvaluateConstantChrecAtConstant(this, NextVal);
2399 if (Range.contains(R1Val))
2401 return new SCEVCouldNotCompute(); // Something strange happened
2406 // Fallback, if this is a general polynomial, figure out the progression
2407 // through brute force: evaluate until we find an iteration that fails the
2408 // test. This is likely to be slow, but getting an accurate trip count is
2409 // incredibly important, we will be able to simplify the exit test a lot, and
2410 // we are almost guaranteed to get a trip count in this case.
2411 ConstantInt *TestVal = ConstantInt::get(getType(), 0);
2412 ConstantInt *One = ConstantInt::get(getType(), 1);
2413 ConstantInt *EndVal = TestVal; // Stop when we wrap around.
2415 ++NumBruteForceEvaluations;
2416 SCEVHandle Val = evaluateAtIteration(SCEVConstant::get(TestVal));
2417 if (!isa<SCEVConstant>(Val)) // This shouldn't happen.
2418 return new SCEVCouldNotCompute();
2420 // Check to see if we found the value!
2421 if (!Range.contains(cast<SCEVConstant>(Val)->getValue()))
2422 return SCEVConstant::get(TestVal);
2424 // Increment to test the next index.
2425 TestVal = cast<ConstantInt>(ConstantExpr::getAdd(TestVal, One));
2426 } while (TestVal != EndVal);
2428 return new SCEVCouldNotCompute();
2433 //===----------------------------------------------------------------------===//
2434 // ScalarEvolution Class Implementation
2435 //===----------------------------------------------------------------------===//
2437 bool ScalarEvolution::runOnFunction(Function &F) {
2438 Impl = new ScalarEvolutionsImpl(F, getAnalysis<LoopInfo>());
2442 void ScalarEvolution::releaseMemory() {
2443 delete (ScalarEvolutionsImpl*)Impl;
2447 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
2448 AU.setPreservesAll();
2449 AU.addRequiredTransitive<LoopInfo>();
2452 SCEVHandle ScalarEvolution::getSCEV(Value *V) const {
2453 return ((ScalarEvolutionsImpl*)Impl)->getSCEV(V);
2456 /// hasSCEV - Return true if the SCEV for this value has already been
2458 bool ScalarEvolution::hasSCEV(Value *V) const {
2459 return ((ScalarEvolutionsImpl*)Impl)->hasSCEV(V);
2463 /// setSCEV - Insert the specified SCEV into the map of current SCEVs for
2464 /// the specified value.
2465 void ScalarEvolution::setSCEV(Value *V, const SCEVHandle &H) {
2466 ((ScalarEvolutionsImpl*)Impl)->setSCEV(V, H);
2470 SCEVHandle ScalarEvolution::getIterationCount(const Loop *L) const {
2471 return ((ScalarEvolutionsImpl*)Impl)->getIterationCount(L);
2474 bool ScalarEvolution::hasLoopInvariantIterationCount(const Loop *L) const {
2475 return !isa<SCEVCouldNotCompute>(getIterationCount(L));
2478 SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) const {
2479 return ((ScalarEvolutionsImpl*)Impl)->getSCEVAtScope(getSCEV(V), L);
2482 void ScalarEvolution::deleteInstructionFromRecords(Instruction *I) const {
2483 return ((ScalarEvolutionsImpl*)Impl)->deleteInstructionFromRecords(I);
2486 static void PrintLoopInfo(std::ostream &OS, const ScalarEvolution *SE,
2488 // Print all inner loops first
2489 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
2490 PrintLoopInfo(OS, SE, *I);
2492 llvm_cerr << "Loop " << L->getHeader()->getName() << ": ";
2494 std::vector<BasicBlock*> ExitBlocks;
2495 L->getExitBlocks(ExitBlocks);
2496 if (ExitBlocks.size() != 1)
2497 llvm_cerr << "<multiple exits> ";
2499 if (SE->hasLoopInvariantIterationCount(L)) {
2500 llvm_cerr << *SE->getIterationCount(L) << " iterations! ";
2502 llvm_cerr << "Unpredictable iteration count. ";
2508 void ScalarEvolution::print(std::ostream &OS, const Module* ) const {
2509 Function &F = ((ScalarEvolutionsImpl*)Impl)->F;
2510 LoopInfo &LI = ((ScalarEvolutionsImpl*)Impl)->LI;
2512 OS << "Classifying expressions for: " << F.getName() << "\n";
2513 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
2514 if (I->getType()->isInteger()) {
2517 SCEVHandle SV = getSCEV(&*I);
2521 if ((*I).getType()->isIntegral()) {
2522 ConstantRange Bounds = SV->getValueRange();
2523 if (!Bounds.isFullSet())
2524 OS << "Bounds: " << Bounds << " ";
2527 if (const Loop *L = LI.getLoopFor((*I).getParent())) {
2529 SCEVHandle ExitValue = getSCEVAtScope(&*I, L->getParentLoop());
2530 if (isa<SCEVCouldNotCompute>(ExitValue)) {
2531 OS << "<<Unknown>>";
2541 OS << "Determining loop execution counts for: " << F.getName() << "\n";
2542 for (LoopInfo::iterator I = LI.begin(), E = LI.end(); I != E; ++I)
2543 PrintLoopInfo(OS, this, *I);