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 #define DEBUG_TYPE "scalar-evolution"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Constants.h"
65 #include "llvm/DerivedTypes.h"
66 #include "llvm/GlobalVariable.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/Analysis/ConstantFolding.h"
69 #include "llvm/Analysis/LoopInfo.h"
70 #include "llvm/Assembly/Writer.h"
71 #include "llvm/Transforms/Scalar.h"
72 #include "llvm/Support/CFG.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/Compiler.h"
75 #include "llvm/Support/ConstantRange.h"
76 #include "llvm/Support/InstIterator.h"
77 #include "llvm/Support/ManagedStatic.h"
78 #include "llvm/Support/MathExtras.h"
79 #include "llvm/Support/Streams.h"
80 #include "llvm/ADT/Statistic.h"
86 STATISTIC(NumBruteForceEvaluations,
87 "Number of brute force evaluations needed to "
88 "calculate high-order polynomial exit values");
89 STATISTIC(NumArrayLenItCounts,
90 "Number of trip counts computed with array length");
91 STATISTIC(NumTripCountsComputed,
92 "Number of loops with predictable loop counts");
93 STATISTIC(NumTripCountsNotComputed,
94 "Number of loops without predictable loop counts");
95 STATISTIC(NumBruteForceTripCountsComputed,
96 "Number of loops with trip counts computed by force");
99 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
100 cl::desc("Maximum number of iterations SCEV will "
101 "symbolically execute a constant derived loop"),
105 RegisterPass<ScalarEvolution>
106 R("scalar-evolution", "Scalar Evolution Analysis");
109 //===----------------------------------------------------------------------===//
110 // SCEV class definitions
111 //===----------------------------------------------------------------------===//
113 //===----------------------------------------------------------------------===//
114 // Implementation of the SCEV class.
117 void SCEV::dump() const {
121 /// getValueRange - Return the tightest constant bounds that this value is
122 /// known to have. This method is only valid on integer SCEV objects.
123 ConstantRange SCEV::getValueRange() const {
124 const Type *Ty = getType();
125 assert(Ty->isInteger() && "Can't get range for a non-integer SCEV!");
126 // Default to a full range if no better information is available.
127 return ConstantRange(getBitWidth());
130 uint32_t SCEV::getBitWidth() const {
131 if (const IntegerType* ITy = dyn_cast<IntegerType>(getType()))
132 return ITy->getBitWidth();
137 SCEVCouldNotCompute::SCEVCouldNotCompute() : SCEV(scCouldNotCompute) {}
139 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
140 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
144 const Type *SCEVCouldNotCompute::getType() const {
145 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
149 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
150 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
154 SCEVHandle SCEVCouldNotCompute::
155 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
156 const SCEVHandle &Conc) const {
160 void SCEVCouldNotCompute::print(std::ostream &OS) const {
161 OS << "***COULDNOTCOMPUTE***";
164 bool SCEVCouldNotCompute::classof(const SCEV *S) {
165 return S->getSCEVType() == scCouldNotCompute;
169 // SCEVConstants - Only allow the creation of one SCEVConstant for any
170 // particular value. Don't use a SCEVHandle here, or else the object will
172 static ManagedStatic<std::map<ConstantInt*, SCEVConstant*> > SCEVConstants;
175 SCEVConstant::~SCEVConstant() {
176 SCEVConstants->erase(V);
179 SCEVHandle SCEVConstant::get(ConstantInt *V) {
180 SCEVConstant *&R = (*SCEVConstants)[V];
181 if (R == 0) R = new SCEVConstant(V);
185 ConstantRange SCEVConstant::getValueRange() const {
186 return ConstantRange(V->getValue());
189 const Type *SCEVConstant::getType() const { return V->getType(); }
191 void SCEVConstant::print(std::ostream &OS) const {
192 WriteAsOperand(OS, V, false);
195 // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any
196 // particular input. Don't use a SCEVHandle here, or else the object will
198 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
199 SCEVTruncateExpr*> > SCEVTruncates;
201 SCEVTruncateExpr::SCEVTruncateExpr(const SCEVHandle &op, const Type *ty)
202 : SCEV(scTruncate), Op(op), Ty(ty) {
203 assert(Op->getType()->isInteger() && Ty->isInteger() &&
204 "Cannot truncate non-integer value!");
205 assert(Op->getType()->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits()
206 && "This is not a truncating conversion!");
209 SCEVTruncateExpr::~SCEVTruncateExpr() {
210 SCEVTruncates->erase(std::make_pair(Op, Ty));
213 ConstantRange SCEVTruncateExpr::getValueRange() const {
214 return getOperand()->getValueRange().truncate(getBitWidth());
217 void SCEVTruncateExpr::print(std::ostream &OS) const {
218 OS << "(truncate " << *Op << " to " << *Ty << ")";
221 // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any
222 // particular input. Don't use a SCEVHandle here, or else the object will never
224 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>,
225 SCEVZeroExtendExpr*> > SCEVZeroExtends;
227 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty)
228 : SCEV(scZeroExtend), Op(op), Ty(ty) {
229 assert(Op->getType()->isInteger() && Ty->isInteger() &&
230 "Cannot zero extend non-integer value!");
231 assert(Op->getType()->getPrimitiveSizeInBits() < Ty->getPrimitiveSizeInBits()
232 && "This is not an extending conversion!");
235 SCEVZeroExtendExpr::~SCEVZeroExtendExpr() {
236 SCEVZeroExtends->erase(std::make_pair(Op, Ty));
239 ConstantRange SCEVZeroExtendExpr::getValueRange() const {
240 return getOperand()->getValueRange().zeroExtend(getBitWidth());
243 void SCEVZeroExtendExpr::print(std::ostream &OS) const {
244 OS << "(zeroextend " << *Op << " to " << *Ty << ")";
247 // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any
248 // particular input. Don't use a SCEVHandle here, or else the object will never
250 static ManagedStatic<std::map<std::pair<unsigned, std::vector<SCEV*> >,
251 SCEVCommutativeExpr*> > SCEVCommExprs;
253 SCEVCommutativeExpr::~SCEVCommutativeExpr() {
254 SCEVCommExprs->erase(std::make_pair(getSCEVType(),
255 std::vector<SCEV*>(Operands.begin(),
259 void SCEVCommutativeExpr::print(std::ostream &OS) const {
260 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
261 const char *OpStr = getOperationStr();
262 OS << "(" << *Operands[0];
263 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
264 OS << OpStr << *Operands[i];
268 SCEVHandle SCEVCommutativeExpr::
269 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
270 const SCEVHandle &Conc) const {
271 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
272 SCEVHandle H = getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc);
273 if (H != getOperand(i)) {
274 std::vector<SCEVHandle> NewOps;
275 NewOps.reserve(getNumOperands());
276 for (unsigned j = 0; j != i; ++j)
277 NewOps.push_back(getOperand(j));
279 for (++i; i != e; ++i)
280 NewOps.push_back(getOperand(i)->
281 replaceSymbolicValuesWithConcrete(Sym, Conc));
283 if (isa<SCEVAddExpr>(this))
284 return SCEVAddExpr::get(NewOps);
285 else if (isa<SCEVMulExpr>(this))
286 return SCEVMulExpr::get(NewOps);
288 assert(0 && "Unknown commutative expr!");
295 // SCEVSDivs - Only allow the creation of one SCEVSDivExpr for any particular
296 // input. Don't use a SCEVHandle here, or else the object will never be
298 static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>,
299 SCEVSDivExpr*> > SCEVSDivs;
301 SCEVSDivExpr::~SCEVSDivExpr() {
302 SCEVSDivs->erase(std::make_pair(LHS, RHS));
305 void SCEVSDivExpr::print(std::ostream &OS) const {
306 OS << "(" << *LHS << " /s " << *RHS << ")";
309 const Type *SCEVSDivExpr::getType() const {
310 return LHS->getType();
313 // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
314 // particular input. Don't use a SCEVHandle here, or else the object will never
316 static ManagedStatic<std::map<std::pair<const Loop *, std::vector<SCEV*> >,
317 SCEVAddRecExpr*> > SCEVAddRecExprs;
319 SCEVAddRecExpr::~SCEVAddRecExpr() {
320 SCEVAddRecExprs->erase(std::make_pair(L,
321 std::vector<SCEV*>(Operands.begin(),
325 SCEVHandle SCEVAddRecExpr::
326 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
327 const SCEVHandle &Conc) const {
328 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
329 SCEVHandle H = getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc);
330 if (H != getOperand(i)) {
331 std::vector<SCEVHandle> NewOps;
332 NewOps.reserve(getNumOperands());
333 for (unsigned j = 0; j != i; ++j)
334 NewOps.push_back(getOperand(j));
336 for (++i; i != e; ++i)
337 NewOps.push_back(getOperand(i)->
338 replaceSymbolicValuesWithConcrete(Sym, Conc));
340 return get(NewOps, L);
347 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
348 // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't
349 // contain L and if the start is invariant.
350 return !QueryLoop->contains(L->getHeader()) &&
351 getOperand(0)->isLoopInvariant(QueryLoop);
355 void SCEVAddRecExpr::print(std::ostream &OS) const {
356 OS << "{" << *Operands[0];
357 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
358 OS << ",+," << *Operands[i];
359 OS << "}<" << L->getHeader()->getName() + ">";
362 // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular
363 // value. Don't use a SCEVHandle here, or else the object will never be
365 static ManagedStatic<std::map<Value*, SCEVUnknown*> > SCEVUnknowns;
367 SCEVUnknown::~SCEVUnknown() { SCEVUnknowns->erase(V); }
369 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
370 // All non-instruction values are loop invariant. All instructions are loop
371 // invariant if they are not contained in the specified loop.
372 if (Instruction *I = dyn_cast<Instruction>(V))
373 return !L->contains(I->getParent());
377 const Type *SCEVUnknown::getType() const {
381 void SCEVUnknown::print(std::ostream &OS) const {
382 WriteAsOperand(OS, V, false);
385 //===----------------------------------------------------------------------===//
387 //===----------------------------------------------------------------------===//
390 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
391 /// than the complexity of the RHS. This comparator is used to canonicalize
393 struct VISIBILITY_HIDDEN SCEVComplexityCompare {
394 bool operator()(SCEV *LHS, SCEV *RHS) {
395 return LHS->getSCEVType() < RHS->getSCEVType();
400 /// GroupByComplexity - Given a list of SCEV objects, order them by their
401 /// complexity, and group objects of the same complexity together by value.
402 /// When this routine is finished, we know that any duplicates in the vector are
403 /// consecutive and that complexity is monotonically increasing.
405 /// Note that we go take special precautions to ensure that we get determinstic
406 /// results from this routine. In other words, we don't want the results of
407 /// this to depend on where the addresses of various SCEV objects happened to
410 static void GroupByComplexity(std::vector<SCEVHandle> &Ops) {
411 if (Ops.size() < 2) return; // Noop
412 if (Ops.size() == 2) {
413 // This is the common case, which also happens to be trivially simple.
415 if (Ops[0]->getSCEVType() > Ops[1]->getSCEVType())
416 std::swap(Ops[0], Ops[1]);
420 // Do the rough sort by complexity.
421 std::sort(Ops.begin(), Ops.end(), SCEVComplexityCompare());
423 // Now that we are sorted by complexity, group elements of the same
424 // complexity. Note that this is, at worst, N^2, but the vector is likely to
425 // be extremely short in practice. Note that we take this approach because we
426 // do not want to depend on the addresses of the objects we are grouping.
427 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
429 unsigned Complexity = S->getSCEVType();
431 // If there are any objects of the same complexity and same value as this
433 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
434 if (Ops[j] == S) { // Found a duplicate.
435 // Move it to immediately after i'th element.
436 std::swap(Ops[i+1], Ops[j]);
437 ++i; // no need to rescan it.
438 if (i == e-2) return; // Done!
446 //===----------------------------------------------------------------------===//
447 // Simple SCEV method implementations
448 //===----------------------------------------------------------------------===//
450 /// getIntegerSCEV - Given an integer or FP type, create a constant for the
451 /// specified signed integer value and return a SCEV for the constant.
452 SCEVHandle SCEVUnknown::getIntegerSCEV(int Val, const Type *Ty) {
455 C = Constant::getNullValue(Ty);
456 else if (Ty->isFloatingPoint())
457 C = ConstantFP::get(Ty, Val);
459 C = ConstantInt::get(Ty, Val);
460 return SCEVUnknown::get(C);
463 SCEVHandle SCEVUnknown::getIntegerSCEV(const APInt& Val) {
464 return SCEVUnknown::get(ConstantInt::get(Val));
467 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
468 /// input value to the specified type. If the type must be extended, it is zero
470 static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty) {
471 const Type *SrcTy = V->getType();
472 assert(SrcTy->isInteger() && Ty->isInteger() &&
473 "Cannot truncate or zero extend with non-integer arguments!");
474 if (SrcTy->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
475 return V; // No conversion
476 if (SrcTy->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits())
477 return SCEVTruncateExpr::get(V, Ty);
478 return SCEVZeroExtendExpr::get(V, Ty);
481 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
483 SCEVHandle SCEV::getNegativeSCEV(const SCEVHandle &V) {
484 if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
485 return SCEVUnknown::get(ConstantExpr::getNeg(VC->getValue()));
487 return SCEVMulExpr::get(V, SCEVUnknown::getIntegerSCEV(-1, V->getType()));
490 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
492 SCEVHandle SCEV::getMinusSCEV(const SCEVHandle &LHS, const SCEVHandle &RHS) {
494 return SCEVAddExpr::get(LHS, SCEV::getNegativeSCEV(RHS));
498 /// PartialFact - Compute V!/(V-NumSteps)!
499 static SCEVHandle PartialFact(SCEVHandle V, unsigned NumSteps) {
500 // Handle this case efficiently, it is common to have constant iteration
501 // counts while computing loop exit values.
502 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
503 APInt Val = SC->getValue()->getValue();
504 APInt Result(Val.getBitWidth(), 1);
505 for (; NumSteps; --NumSteps)
506 Result *= Val-(NumSteps-1);
507 return SCEVUnknown::get(ConstantInt::get(Result));
510 const Type *Ty = V->getType();
512 return SCEVUnknown::getIntegerSCEV(1, Ty);
514 SCEVHandle Result = V;
515 for (unsigned i = 1; i != NumSteps; ++i)
516 Result = SCEVMulExpr::get(Result, SCEV::getMinusSCEV(V,
517 SCEVUnknown::getIntegerSCEV(i, Ty)));
522 /// evaluateAtIteration - Return the value of this chain of recurrences at
523 /// the specified iteration number. We can evaluate this recurrence by
524 /// multiplying each element in the chain by the binomial coefficient
525 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
527 /// A*choose(It, 0) + B*choose(It, 1) + C*choose(It, 2) + D*choose(It, 3)
529 /// FIXME/VERIFY: I don't trust that this is correct in the face of overflow.
530 /// Is the binomial equation safe using modular arithmetic??
532 SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It) const {
533 SCEVHandle Result = getStart();
535 const Type *Ty = It->getType();
536 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
537 SCEVHandle BC = PartialFact(It, i);
539 SCEVHandle Val = SCEVSDivExpr::get(SCEVMulExpr::get(BC, getOperand(i)),
540 SCEVUnknown::getIntegerSCEV(Divisor,Ty));
541 Result = SCEVAddExpr::get(Result, Val);
547 //===----------------------------------------------------------------------===//
548 // SCEV Expression folder implementations
549 //===----------------------------------------------------------------------===//
551 SCEVHandle SCEVTruncateExpr::get(const SCEVHandle &Op, const Type *Ty) {
552 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
553 return SCEVUnknown::get(
554 ConstantExpr::getTrunc(SC->getValue(), Ty));
556 // If the input value is a chrec scev made out of constants, truncate
557 // all of the constants.
558 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
559 std::vector<SCEVHandle> Operands;
560 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
561 // FIXME: This should allow truncation of other expression types!
562 if (isa<SCEVConstant>(AddRec->getOperand(i)))
563 Operands.push_back(get(AddRec->getOperand(i), Ty));
566 if (Operands.size() == AddRec->getNumOperands())
567 return SCEVAddRecExpr::get(Operands, AddRec->getLoop());
570 SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)];
571 if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
575 SCEVHandle SCEVZeroExtendExpr::get(const SCEVHandle &Op, const Type *Ty) {
576 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
577 return SCEVUnknown::get(
578 ConstantExpr::getZExt(SC->getValue(), Ty));
580 // FIXME: If the input value is a chrec scev, and we can prove that the value
581 // did not overflow the old, smaller, value, we can zero extend all of the
582 // operands (often constants). This would allow analysis of something like
583 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
585 SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)];
586 if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
590 // get - Get a canonical add expression, or something simpler if possible.
591 SCEVHandle SCEVAddExpr::get(std::vector<SCEVHandle> &Ops) {
592 assert(!Ops.empty() && "Cannot get empty add!");
593 if (Ops.size() == 1) return Ops[0];
595 // Sort by complexity, this groups all similar expression types together.
596 GroupByComplexity(Ops);
598 // If there are any constants, fold them together.
600 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
602 assert(Idx < Ops.size());
603 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
604 // We found two constants, fold them together!
605 Constant *Fold = ConstantExpr::getAdd(LHSC->getValue(), RHSC->getValue());
606 if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) {
607 Ops[0] = SCEVConstant::get(CI);
608 Ops.erase(Ops.begin()+1); // Erase the folded element
609 if (Ops.size() == 1) return Ops[0];
610 LHSC = cast<SCEVConstant>(Ops[0]);
612 // If we couldn't fold the expression, move to the next constant. Note
613 // that this is impossible to happen in practice because we always
614 // constant fold constant ints to constant ints.
619 // If we are left with a constant zero being added, strip it off.
620 if (cast<SCEVConstant>(Ops[0])->getValue()->isNullValue()) {
621 Ops.erase(Ops.begin());
626 if (Ops.size() == 1) return Ops[0];
628 // Okay, check to see if the same value occurs in the operand list twice. If
629 // so, merge them together into an multiply expression. Since we sorted the
630 // list, these values are required to be adjacent.
631 const Type *Ty = Ops[0]->getType();
632 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
633 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
634 // Found a match, merge the two values into a multiply, and add any
635 // remaining values to the result.
636 SCEVHandle Two = SCEVUnknown::getIntegerSCEV(2, Ty);
637 SCEVHandle Mul = SCEVMulExpr::get(Ops[i], Two);
640 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
642 return SCEVAddExpr::get(Ops);
645 // Okay, now we know the first non-constant operand. If there are add
646 // operands they would be next.
647 if (Idx < Ops.size()) {
648 bool DeletedAdd = false;
649 while (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
650 // If we have an add, expand the add operands onto the end of the operands
652 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
653 Ops.erase(Ops.begin()+Idx);
657 // If we deleted at least one add, we added operands to the end of the list,
658 // and they are not necessarily sorted. Recurse to resort and resimplify
659 // any operands we just aquired.
664 // Skip over the add expression until we get to a multiply.
665 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
668 // If we are adding something to a multiply expression, make sure the
669 // something is not already an operand of the multiply. If so, merge it into
671 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
672 SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
673 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
674 SCEV *MulOpSCEV = Mul->getOperand(MulOp);
675 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
676 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(MulOpSCEV)) {
677 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
678 SCEVHandle InnerMul = Mul->getOperand(MulOp == 0);
679 if (Mul->getNumOperands() != 2) {
680 // If the multiply has more than two operands, we must get the
682 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
683 MulOps.erase(MulOps.begin()+MulOp);
684 InnerMul = SCEVMulExpr::get(MulOps);
686 SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, Ty);
687 SCEVHandle AddOne = SCEVAddExpr::get(InnerMul, One);
688 SCEVHandle OuterMul = SCEVMulExpr::get(AddOne, Ops[AddOp]);
689 if (Ops.size() == 2) return OuterMul;
691 Ops.erase(Ops.begin()+AddOp);
692 Ops.erase(Ops.begin()+Idx-1);
694 Ops.erase(Ops.begin()+Idx);
695 Ops.erase(Ops.begin()+AddOp-1);
697 Ops.push_back(OuterMul);
698 return SCEVAddExpr::get(Ops);
701 // Check this multiply against other multiplies being added together.
702 for (unsigned OtherMulIdx = Idx+1;
703 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
705 SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
706 // If MulOp occurs in OtherMul, we can fold the two multiplies
708 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
709 OMulOp != e; ++OMulOp)
710 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
711 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
712 SCEVHandle InnerMul1 = Mul->getOperand(MulOp == 0);
713 if (Mul->getNumOperands() != 2) {
714 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
715 MulOps.erase(MulOps.begin()+MulOp);
716 InnerMul1 = SCEVMulExpr::get(MulOps);
718 SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0);
719 if (OtherMul->getNumOperands() != 2) {
720 std::vector<SCEVHandle> MulOps(OtherMul->op_begin(),
722 MulOps.erase(MulOps.begin()+OMulOp);
723 InnerMul2 = SCEVMulExpr::get(MulOps);
725 SCEVHandle InnerMulSum = SCEVAddExpr::get(InnerMul1,InnerMul2);
726 SCEVHandle OuterMul = SCEVMulExpr::get(MulOpSCEV, InnerMulSum);
727 if (Ops.size() == 2) return OuterMul;
728 Ops.erase(Ops.begin()+Idx);
729 Ops.erase(Ops.begin()+OtherMulIdx-1);
730 Ops.push_back(OuterMul);
731 return SCEVAddExpr::get(Ops);
737 // If there are any add recurrences in the operands list, see if any other
738 // added values are loop invariant. If so, we can fold them into the
740 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
743 // Scan over all recurrences, trying to fold loop invariants into them.
744 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
745 // Scan all of the other operands to this add and add them to the vector if
746 // they are loop invariant w.r.t. the recurrence.
747 std::vector<SCEVHandle> LIOps;
748 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
749 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
750 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
751 LIOps.push_back(Ops[i]);
752 Ops.erase(Ops.begin()+i);
756 // If we found some loop invariants, fold them into the recurrence.
757 if (!LIOps.empty()) {
758 // NLI + LI + { Start,+,Step} --> NLI + { LI+Start,+,Step }
759 LIOps.push_back(AddRec->getStart());
761 std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
762 AddRecOps[0] = SCEVAddExpr::get(LIOps);
764 SCEVHandle NewRec = SCEVAddRecExpr::get(AddRecOps, AddRec->getLoop());
765 // If all of the other operands were loop invariant, we are done.
766 if (Ops.size() == 1) return NewRec;
768 // Otherwise, add the folded AddRec by the non-liv parts.
769 for (unsigned i = 0;; ++i)
770 if (Ops[i] == AddRec) {
774 return SCEVAddExpr::get(Ops);
777 // Okay, if there weren't any loop invariants to be folded, check to see if
778 // there are multiple AddRec's with the same loop induction variable being
779 // added together. If so, we can fold them.
780 for (unsigned OtherIdx = Idx+1;
781 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
782 if (OtherIdx != Idx) {
783 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
784 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
785 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
786 std::vector<SCEVHandle> NewOps(AddRec->op_begin(), AddRec->op_end());
787 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
788 if (i >= NewOps.size()) {
789 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
790 OtherAddRec->op_end());
793 NewOps[i] = SCEVAddExpr::get(NewOps[i], OtherAddRec->getOperand(i));
795 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop());
797 if (Ops.size() == 2) return NewAddRec;
799 Ops.erase(Ops.begin()+Idx);
800 Ops.erase(Ops.begin()+OtherIdx-1);
801 Ops.push_back(NewAddRec);
802 return SCEVAddExpr::get(Ops);
806 // Otherwise couldn't fold anything into this recurrence. Move onto the
810 // Okay, it looks like we really DO need an add expr. Check to see if we
811 // already have one, otherwise create a new one.
812 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
813 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scAddExpr,
815 if (Result == 0) Result = new SCEVAddExpr(Ops);
820 SCEVHandle SCEVMulExpr::get(std::vector<SCEVHandle> &Ops) {
821 assert(!Ops.empty() && "Cannot get empty mul!");
823 // Sort by complexity, this groups all similar expression types together.
824 GroupByComplexity(Ops);
826 // If there are any constants, fold them together.
828 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
830 // C1*(C2+V) -> C1*C2 + C1*V
832 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
833 if (Add->getNumOperands() == 2 &&
834 isa<SCEVConstant>(Add->getOperand(0)))
835 return SCEVAddExpr::get(SCEVMulExpr::get(LHSC, Add->getOperand(0)),
836 SCEVMulExpr::get(LHSC, Add->getOperand(1)));
840 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
841 // We found two constants, fold them together!
842 Constant *Fold = ConstantExpr::getMul(LHSC->getValue(), RHSC->getValue());
843 if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) {
844 Ops[0] = SCEVConstant::get(CI);
845 Ops.erase(Ops.begin()+1); // Erase the folded element
846 if (Ops.size() == 1) return Ops[0];
847 LHSC = cast<SCEVConstant>(Ops[0]);
849 // If we couldn't fold the expression, move to the next constant. Note
850 // that this is impossible to happen in practice because we always
851 // constant fold constant ints to constant ints.
856 // If we are left with a constant one being multiplied, strip it off.
857 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
858 Ops.erase(Ops.begin());
860 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isNullValue()) {
861 // If we have a multiply of zero, it will always be zero.
866 // Skip over the add expression until we get to a multiply.
867 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
873 // If there are mul operands inline them all into this expression.
874 if (Idx < Ops.size()) {
875 bool DeletedMul = false;
876 while (SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
877 // If we have an mul, expand the mul operands onto the end of the operands
879 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
880 Ops.erase(Ops.begin()+Idx);
884 // If we deleted at least one mul, we added operands to the end of the list,
885 // and they are not necessarily sorted. Recurse to resort and resimplify
886 // any operands we just aquired.
891 // If there are any add recurrences in the operands list, see if any other
892 // added values are loop invariant. If so, we can fold them into the
894 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
897 // Scan over all recurrences, trying to fold loop invariants into them.
898 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
899 // Scan all of the other operands to this mul and add them to the vector if
900 // they are loop invariant w.r.t. the recurrence.
901 std::vector<SCEVHandle> LIOps;
902 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
903 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
904 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
905 LIOps.push_back(Ops[i]);
906 Ops.erase(Ops.begin()+i);
910 // If we found some loop invariants, fold them into the recurrence.
911 if (!LIOps.empty()) {
912 // NLI * LI * { Start,+,Step} --> NLI * { LI*Start,+,LI*Step }
913 std::vector<SCEVHandle> NewOps;
914 NewOps.reserve(AddRec->getNumOperands());
915 if (LIOps.size() == 1) {
916 SCEV *Scale = LIOps[0];
917 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
918 NewOps.push_back(SCEVMulExpr::get(Scale, AddRec->getOperand(i)));
920 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
921 std::vector<SCEVHandle> MulOps(LIOps);
922 MulOps.push_back(AddRec->getOperand(i));
923 NewOps.push_back(SCEVMulExpr::get(MulOps));
927 SCEVHandle NewRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop());
929 // If all of the other operands were loop invariant, we are done.
930 if (Ops.size() == 1) return NewRec;
932 // Otherwise, multiply the folded AddRec by the non-liv parts.
933 for (unsigned i = 0;; ++i)
934 if (Ops[i] == AddRec) {
938 return SCEVMulExpr::get(Ops);
941 // Okay, if there weren't any loop invariants to be folded, check to see if
942 // there are multiple AddRec's with the same loop induction variable being
943 // multiplied together. If so, we can fold them.
944 for (unsigned OtherIdx = Idx+1;
945 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
946 if (OtherIdx != Idx) {
947 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
948 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
949 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
950 SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
951 SCEVHandle NewStart = SCEVMulExpr::get(F->getStart(),
953 SCEVHandle B = F->getStepRecurrence();
954 SCEVHandle D = G->getStepRecurrence();
955 SCEVHandle NewStep = SCEVAddExpr::get(SCEVMulExpr::get(F, D),
956 SCEVMulExpr::get(G, B),
957 SCEVMulExpr::get(B, D));
958 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewStart, NewStep,
960 if (Ops.size() == 2) return NewAddRec;
962 Ops.erase(Ops.begin()+Idx);
963 Ops.erase(Ops.begin()+OtherIdx-1);
964 Ops.push_back(NewAddRec);
965 return SCEVMulExpr::get(Ops);
969 // Otherwise couldn't fold anything into this recurrence. Move onto the
973 // Okay, it looks like we really DO need an mul expr. Check to see if we
974 // already have one, otherwise create a new one.
975 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
976 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scMulExpr,
979 Result = new SCEVMulExpr(Ops);
983 SCEVHandle SCEVSDivExpr::get(const SCEVHandle &LHS, const SCEVHandle &RHS) {
984 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
985 if (RHSC->getValue()->equalsInt(1))
986 return LHS; // X sdiv 1 --> x
987 if (RHSC->getValue()->isAllOnesValue())
988 return SCEV::getNegativeSCEV(LHS); // X sdiv -1 --> -x
990 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
991 Constant *LHSCV = LHSC->getValue();
992 Constant *RHSCV = RHSC->getValue();
993 return SCEVUnknown::get(ConstantExpr::getSDiv(LHSCV, RHSCV));
997 // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow.
999 SCEVSDivExpr *&Result = (*SCEVSDivs)[std::make_pair(LHS, RHS)];
1000 if (Result == 0) Result = new SCEVSDivExpr(LHS, RHS);
1005 /// SCEVAddRecExpr::get - Get a add recurrence expression for the
1006 /// specified loop. Simplify the expression as much as possible.
1007 SCEVHandle SCEVAddRecExpr::get(const SCEVHandle &Start,
1008 const SCEVHandle &Step, const Loop *L) {
1009 std::vector<SCEVHandle> Operands;
1010 Operands.push_back(Start);
1011 if (SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1012 if (StepChrec->getLoop() == L) {
1013 Operands.insert(Operands.end(), StepChrec->op_begin(),
1014 StepChrec->op_end());
1015 return get(Operands, L);
1018 Operands.push_back(Step);
1019 return get(Operands, L);
1022 /// SCEVAddRecExpr::get - Get a add recurrence expression for the
1023 /// specified loop. Simplify the expression as much as possible.
1024 SCEVHandle SCEVAddRecExpr::get(std::vector<SCEVHandle> &Operands,
1026 if (Operands.size() == 1) return Operands[0];
1028 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Operands.back()))
1029 if (StepC->getValue()->isNullValue()) {
1030 Operands.pop_back();
1031 return get(Operands, L); // { X,+,0 } --> X
1034 SCEVAddRecExpr *&Result =
1035 (*SCEVAddRecExprs)[std::make_pair(L, std::vector<SCEV*>(Operands.begin(),
1037 if (Result == 0) Result = new SCEVAddRecExpr(Operands, L);
1041 SCEVHandle SCEVUnknown::get(Value *V) {
1042 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
1043 return SCEVConstant::get(CI);
1044 SCEVUnknown *&Result = (*SCEVUnknowns)[V];
1045 if (Result == 0) Result = new SCEVUnknown(V);
1050 //===----------------------------------------------------------------------===//
1051 // ScalarEvolutionsImpl Definition and Implementation
1052 //===----------------------------------------------------------------------===//
1054 /// ScalarEvolutionsImpl - This class implements the main driver for the scalar
1058 struct VISIBILITY_HIDDEN ScalarEvolutionsImpl {
1059 /// F - The function we are analyzing.
1063 /// LI - The loop information for the function we are currently analyzing.
1067 /// UnknownValue - This SCEV is used to represent unknown trip counts and
1069 SCEVHandle UnknownValue;
1071 /// Scalars - This is a cache of the scalars we have analyzed so far.
1073 std::map<Value*, SCEVHandle> Scalars;
1075 /// IterationCounts - Cache the iteration count of the loops for this
1076 /// function as they are computed.
1077 std::map<const Loop*, SCEVHandle> IterationCounts;
1079 /// ConstantEvolutionLoopExitValue - This map contains entries for all of
1080 /// the PHI instructions that we attempt to compute constant evolutions for.
1081 /// This allows us to avoid potentially expensive recomputation of these
1082 /// properties. An instruction maps to null if we are unable to compute its
1084 std::map<PHINode*, Constant*> ConstantEvolutionLoopExitValue;
1087 ScalarEvolutionsImpl(Function &f, LoopInfo &li)
1088 : F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {}
1090 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1091 /// expression and create a new one.
1092 SCEVHandle getSCEV(Value *V);
1094 /// hasSCEV - Return true if the SCEV for this value has already been
1096 bool hasSCEV(Value *V) const {
1097 return Scalars.count(V);
1100 /// setSCEV - Insert the specified SCEV into the map of current SCEVs for
1101 /// the specified value.
1102 void setSCEV(Value *V, const SCEVHandle &H) {
1103 bool isNew = Scalars.insert(std::make_pair(V, H)).second;
1104 assert(isNew && "This entry already existed!");
1108 /// getSCEVAtScope - Compute the value of the specified expression within
1109 /// the indicated loop (which may be null to indicate in no loop). If the
1110 /// expression cannot be evaluated, return UnknownValue itself.
1111 SCEVHandle getSCEVAtScope(SCEV *V, const Loop *L);
1114 /// hasLoopInvariantIterationCount - Return true if the specified loop has
1115 /// an analyzable loop-invariant iteration count.
1116 bool hasLoopInvariantIterationCount(const Loop *L);
1118 /// getIterationCount - If the specified loop has a predictable iteration
1119 /// count, return it. Note that it is not valid to call this method on a
1120 /// loop without a loop-invariant iteration count.
1121 SCEVHandle getIterationCount(const Loop *L);
1123 /// deleteInstructionFromRecords - This method should be called by the
1124 /// client before it removes an instruction from the program, to make sure
1125 /// that no dangling references are left around.
1126 void deleteInstructionFromRecords(Instruction *I);
1129 /// createSCEV - We know that there is no SCEV for the specified value.
1130 /// Analyze the expression.
1131 SCEVHandle createSCEV(Value *V);
1133 /// createNodeForPHI - Provide the special handling we need to analyze PHI
1135 SCEVHandle createNodeForPHI(PHINode *PN);
1137 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value
1138 /// for the specified instruction and replaces any references to the
1139 /// symbolic value SymName with the specified value. This is used during
1141 void ReplaceSymbolicValueWithConcrete(Instruction *I,
1142 const SCEVHandle &SymName,
1143 const SCEVHandle &NewVal);
1145 /// ComputeIterationCount - Compute the number of times the specified loop
1147 SCEVHandle ComputeIterationCount(const Loop *L);
1149 /// ComputeLoadConstantCompareIterationCount - Given an exit condition of
1150 /// 'setcc load X, cst', try to se if we can compute the trip count.
1151 SCEVHandle ComputeLoadConstantCompareIterationCount(LoadInst *LI,
1154 ICmpInst::Predicate p);
1156 /// ComputeIterationCountExhaustively - If the trip is known to execute a
1157 /// constant number of times (the condition evolves only from constants),
1158 /// try to evaluate a few iterations of the loop until we get the exit
1159 /// condition gets a value of ExitWhen (true or false). If we cannot
1160 /// evaluate the trip count of the loop, return UnknownValue.
1161 SCEVHandle ComputeIterationCountExhaustively(const Loop *L, Value *Cond,
1164 /// HowFarToZero - Return the number of times a backedge comparing the
1165 /// specified value to zero will execute. If not computable, return
1167 SCEVHandle HowFarToZero(SCEV *V, const Loop *L);
1169 /// HowFarToNonZero - Return the number of times a backedge checking the
1170 /// specified value for nonzero will execute. If not computable, return
1172 SCEVHandle HowFarToNonZero(SCEV *V, const Loop *L);
1174 /// HowManyLessThans - Return the number of times a backedge containing the
1175 /// specified less-than comparison will execute. If not computable, return
1177 SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L);
1179 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
1180 /// in the header of its containing loop, we know the loop executes a
1181 /// constant number of times, and the PHI node is just a recurrence
1182 /// involving constants, fold it.
1183 Constant *getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& Its,
1188 //===----------------------------------------------------------------------===//
1189 // Basic SCEV Analysis and PHI Idiom Recognition Code
1192 /// deleteInstructionFromRecords - This method should be called by the
1193 /// client before it removes an instruction from the program, to make sure
1194 /// that no dangling references are left around.
1195 void ScalarEvolutionsImpl::deleteInstructionFromRecords(Instruction *I) {
1197 if (PHINode *PN = dyn_cast<PHINode>(I))
1198 ConstantEvolutionLoopExitValue.erase(PN);
1202 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1203 /// expression and create a new one.
1204 SCEVHandle ScalarEvolutionsImpl::getSCEV(Value *V) {
1205 assert(V->getType() != Type::VoidTy && "Can't analyze void expressions!");
1207 std::map<Value*, SCEVHandle>::iterator I = Scalars.find(V);
1208 if (I != Scalars.end()) return I->second;
1209 SCEVHandle S = createSCEV(V);
1210 Scalars.insert(std::make_pair(V, S));
1214 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value for
1215 /// the specified instruction and replaces any references to the symbolic value
1216 /// SymName with the specified value. This is used during PHI resolution.
1217 void ScalarEvolutionsImpl::
1218 ReplaceSymbolicValueWithConcrete(Instruction *I, const SCEVHandle &SymName,
1219 const SCEVHandle &NewVal) {
1220 std::map<Value*, SCEVHandle>::iterator SI = Scalars.find(I);
1221 if (SI == Scalars.end()) return;
1224 SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal);
1225 if (NV == SI->second) return; // No change.
1227 SI->second = NV; // Update the scalars map!
1229 // Any instruction values that use this instruction might also need to be
1231 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1233 ReplaceSymbolicValueWithConcrete(cast<Instruction>(*UI), SymName, NewVal);
1236 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
1237 /// a loop header, making it a potential recurrence, or it doesn't.
1239 SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) {
1240 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
1241 if (const Loop *L = LI.getLoopFor(PN->getParent()))
1242 if (L->getHeader() == PN->getParent()) {
1243 // If it lives in the loop header, it has two incoming values, one
1244 // from outside the loop, and one from inside.
1245 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1246 unsigned BackEdge = IncomingEdge^1;
1248 // While we are analyzing this PHI node, handle its value symbolically.
1249 SCEVHandle SymbolicName = SCEVUnknown::get(PN);
1250 assert(Scalars.find(PN) == Scalars.end() &&
1251 "PHI node already processed?");
1252 Scalars.insert(std::make_pair(PN, SymbolicName));
1254 // Using this symbolic name for the PHI, analyze the value coming around
1256 SCEVHandle BEValue = getSCEV(PN->getIncomingValue(BackEdge));
1258 // NOTE: If BEValue is loop invariant, we know that the PHI node just
1259 // has a special value for the first iteration of the loop.
1261 // If the value coming around the backedge is an add with the symbolic
1262 // value we just inserted, then we found a simple induction variable!
1263 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
1264 // If there is a single occurrence of the symbolic value, replace it
1265 // with a recurrence.
1266 unsigned FoundIndex = Add->getNumOperands();
1267 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1268 if (Add->getOperand(i) == SymbolicName)
1269 if (FoundIndex == e) {
1274 if (FoundIndex != Add->getNumOperands()) {
1275 // Create an add with everything but the specified operand.
1276 std::vector<SCEVHandle> Ops;
1277 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1278 if (i != FoundIndex)
1279 Ops.push_back(Add->getOperand(i));
1280 SCEVHandle Accum = SCEVAddExpr::get(Ops);
1282 // This is not a valid addrec if the step amount is varying each
1283 // loop iteration, but is not itself an addrec in this loop.
1284 if (Accum->isLoopInvariant(L) ||
1285 (isa<SCEVAddRecExpr>(Accum) &&
1286 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
1287 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1288 SCEVHandle PHISCEV = SCEVAddRecExpr::get(StartVal, Accum, L);
1290 // Okay, for the entire analysis of this edge we assumed the PHI
1291 // to be symbolic. We now need to go back and update all of the
1292 // entries for the scalars that use the PHI (except for the PHI
1293 // itself) to use the new analyzed value instead of the "symbolic"
1295 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1299 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(BEValue)) {
1300 // Otherwise, this could be a loop like this:
1301 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
1302 // In this case, j = {1,+,1} and BEValue is j.
1303 // Because the other in-value of i (0) fits the evolution of BEValue
1304 // i really is an addrec evolution.
1305 if (AddRec->getLoop() == L && AddRec->isAffine()) {
1306 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1308 // If StartVal = j.start - j.stride, we can use StartVal as the
1309 // initial step of the addrec evolution.
1310 if (StartVal == SCEV::getMinusSCEV(AddRec->getOperand(0),
1311 AddRec->getOperand(1))) {
1312 SCEVHandle PHISCEV =
1313 SCEVAddRecExpr::get(StartVal, AddRec->getOperand(1), L);
1315 // Okay, for the entire analysis of this edge we assumed the PHI
1316 // to be symbolic. We now need to go back and update all of the
1317 // entries for the scalars that use the PHI (except for the PHI
1318 // itself) to use the new analyzed value instead of the "symbolic"
1320 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
1326 return SymbolicName;
1329 // If it's not a loop phi, we can't handle it yet.
1330 return SCEVUnknown::get(PN);
1333 /// GetConstantFactor - Determine the largest constant factor that S has. For
1334 /// example, turn {4,+,8} -> 4. (S umod result) should always equal zero.
1335 static APInt GetConstantFactor(SCEVHandle S) {
1336 if (SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
1337 APInt V = C->getValue()->getValue();
1338 if (!V.isMinValue())
1340 else // Zero is a multiple of everything.
1341 return APInt(C->getBitWidth(), 1).shl(C->getBitWidth()-1);
1344 if (SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
1345 return GetConstantFactor(T->getOperand()) &
1346 cast<IntegerType>(T->getType())->getMask();
1347 if (SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S))
1348 return GetConstantFactor(E->getOperand());
1350 if (SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
1351 // The result is the min of all operands.
1352 APInt Res = GetConstantFactor(A->getOperand(0));
1353 for (unsigned i = 1, e = A->getNumOperands();
1354 i != e && Res.ugt(APInt(Res.getBitWidth(),1)); ++i)
1355 Res = APIntOps::umin(Res, GetConstantFactor(A->getOperand(i)));
1359 if (SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
1360 // The result is the product of all the operands.
1361 APInt Res = GetConstantFactor(M->getOperand(0));
1362 for (unsigned i = 1, e = M->getNumOperands(); i != e; ++i)
1363 Res *= GetConstantFactor(M->getOperand(i));
1367 if (SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
1368 // For now, we just handle linear expressions.
1369 if (A->getNumOperands() == 2) {
1370 // We want the GCD between the start and the stride value.
1371 APInt Start = GetConstantFactor(A->getOperand(0));
1373 return APInt(A->getBitWidth(),1);
1374 APInt Stride = GetConstantFactor(A->getOperand(1));
1375 return APIntOps::GreatestCommonDivisor(Start, Stride);
1379 // SCEVSDivExpr, SCEVUnknown.
1380 return APInt(S->getBitWidth(), 1);
1383 /// createSCEV - We know that there is no SCEV for the specified value.
1384 /// Analyze the expression.
1386 SCEVHandle ScalarEvolutionsImpl::createSCEV(Value *V) {
1387 if (Instruction *I = dyn_cast<Instruction>(V)) {
1388 switch (I->getOpcode()) {
1389 case Instruction::Add:
1390 return SCEVAddExpr::get(getSCEV(I->getOperand(0)),
1391 getSCEV(I->getOperand(1)));
1392 case Instruction::Mul:
1393 return SCEVMulExpr::get(getSCEV(I->getOperand(0)),
1394 getSCEV(I->getOperand(1)));
1395 case Instruction::SDiv:
1396 return SCEVSDivExpr::get(getSCEV(I->getOperand(0)),
1397 getSCEV(I->getOperand(1)));
1400 case Instruction::Sub:
1401 return SCEV::getMinusSCEV(getSCEV(I->getOperand(0)),
1402 getSCEV(I->getOperand(1)));
1403 case Instruction::Or:
1404 // If the RHS of the Or is a constant, we may have something like:
1405 // X*4+1 which got turned into X*4|1. Handle this as an add so loop
1406 // optimizations will transparently handle this case.
1407 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
1408 SCEVHandle LHS = getSCEV(I->getOperand(0));
1409 APInt CommonFact = GetConstantFactor(LHS);
1410 assert(!CommonFact.isMinValue() &&
1411 "Common factor should at least be 1!");
1412 CommonFact.zextOrTrunc(CI->getValue().getBitWidth());
1413 if (CommonFact.ugt(CI->getValue())) {
1414 // If the LHS is a multiple that is larger than the RHS, use +.
1415 return SCEVAddExpr::get(LHS,
1416 getSCEV(I->getOperand(1)));
1421 case Instruction::Shl:
1422 // Turn shift left of a constant amount into a multiply.
1423 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1424 Constant *X = ConstantInt::get(V->getType(), 1);
1425 X = ConstantExpr::getShl(X, SA);
1426 return SCEVMulExpr::get(getSCEV(I->getOperand(0)), getSCEV(X));
1430 case Instruction::Trunc:
1431 return SCEVTruncateExpr::get(getSCEV(I->getOperand(0)), I->getType());
1433 case Instruction::ZExt:
1434 return SCEVZeroExtendExpr::get(getSCEV(I->getOperand(0)), I->getType());
1436 case Instruction::BitCast:
1437 // BitCasts are no-op casts so we just eliminate the cast.
1438 if (I->getType()->isInteger() &&
1439 I->getOperand(0)->getType()->isInteger())
1440 return getSCEV(I->getOperand(0));
1443 case Instruction::PHI:
1444 return createNodeForPHI(cast<PHINode>(I));
1446 default: // We cannot analyze this expression.
1451 return SCEVUnknown::get(V);
1456 //===----------------------------------------------------------------------===//
1457 // Iteration Count Computation Code
1460 /// getIterationCount - If the specified loop has a predictable iteration
1461 /// count, return it. Note that it is not valid to call this method on a
1462 /// loop without a loop-invariant iteration count.
1463 SCEVHandle ScalarEvolutionsImpl::getIterationCount(const Loop *L) {
1464 std::map<const Loop*, SCEVHandle>::iterator I = IterationCounts.find(L);
1465 if (I == IterationCounts.end()) {
1466 SCEVHandle ItCount = ComputeIterationCount(L);
1467 I = IterationCounts.insert(std::make_pair(L, ItCount)).first;
1468 if (ItCount != UnknownValue) {
1469 assert(ItCount->isLoopInvariant(L) &&
1470 "Computed trip count isn't loop invariant for loop!");
1471 ++NumTripCountsComputed;
1472 } else if (isa<PHINode>(L->getHeader()->begin())) {
1473 // Only count loops that have phi nodes as not being computable.
1474 ++NumTripCountsNotComputed;
1480 /// ComputeIterationCount - Compute the number of times the specified loop
1482 SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) {
1483 // If the loop has a non-one exit block count, we can't analyze it.
1484 std::vector<BasicBlock*> ExitBlocks;
1485 L->getExitBlocks(ExitBlocks);
1486 if (ExitBlocks.size() != 1) return UnknownValue;
1488 // Okay, there is one exit block. Try to find the condition that causes the
1489 // loop to be exited.
1490 BasicBlock *ExitBlock = ExitBlocks[0];
1492 BasicBlock *ExitingBlock = 0;
1493 for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock);
1495 if (L->contains(*PI)) {
1496 if (ExitingBlock == 0)
1499 return UnknownValue; // More than one block exiting!
1501 assert(ExitingBlock && "No exits from loop, something is broken!");
1503 // Okay, we've computed the exiting block. See what condition causes us to
1506 // FIXME: we should be able to handle switch instructions (with a single exit)
1507 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1508 if (ExitBr == 0) return UnknownValue;
1509 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
1511 // At this point, we know we have a conditional branch that determines whether
1512 // the loop is exited. However, we don't know if the branch is executed each
1513 // time through the loop. If not, then the execution count of the branch will
1514 // not be equal to the trip count of the loop.
1516 // Currently we check for this by checking to see if the Exit branch goes to
1517 // the loop header. If so, we know it will always execute the same number of
1518 // times as the loop. We also handle the case where the exit block *is* the
1519 // loop header. This is common for un-rotated loops. More extensive analysis
1520 // could be done to handle more cases here.
1521 if (ExitBr->getSuccessor(0) != L->getHeader() &&
1522 ExitBr->getSuccessor(1) != L->getHeader() &&
1523 ExitBr->getParent() != L->getHeader())
1524 return UnknownValue;
1526 ICmpInst *ExitCond = dyn_cast<ICmpInst>(ExitBr->getCondition());
1528 // If its not an integer comparison then compute it the hard way.
1529 // Note that ICmpInst deals with pointer comparisons too so we must check
1530 // the type of the operand.
1531 if (ExitCond == 0 || isa<PointerType>(ExitCond->getOperand(0)->getType()))
1532 return ComputeIterationCountExhaustively(L, ExitBr->getCondition(),
1533 ExitBr->getSuccessor(0) == ExitBlock);
1535 // If the condition was exit on true, convert the condition to exit on false
1536 ICmpInst::Predicate Cond;
1537 if (ExitBr->getSuccessor(1) == ExitBlock)
1538 Cond = ExitCond->getPredicate();
1540 Cond = ExitCond->getInversePredicate();
1542 // Handle common loops like: for (X = "string"; *X; ++X)
1543 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
1544 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
1546 ComputeLoadConstantCompareIterationCount(LI, RHS, L, Cond);
1547 if (!isa<SCEVCouldNotCompute>(ItCnt)) return ItCnt;
1550 SCEVHandle LHS = getSCEV(ExitCond->getOperand(0));
1551 SCEVHandle RHS = getSCEV(ExitCond->getOperand(1));
1553 // Try to evaluate any dependencies out of the loop.
1554 SCEVHandle Tmp = getSCEVAtScope(LHS, L);
1555 if (!isa<SCEVCouldNotCompute>(Tmp)) LHS = Tmp;
1556 Tmp = getSCEVAtScope(RHS, L);
1557 if (!isa<SCEVCouldNotCompute>(Tmp)) RHS = Tmp;
1559 // At this point, we would like to compute how many iterations of the
1560 // loop the predicate will return true for these inputs.
1561 if (isa<SCEVConstant>(LHS) && !isa<SCEVConstant>(RHS)) {
1562 // If there is a constant, force it into the RHS.
1563 std::swap(LHS, RHS);
1564 Cond = ICmpInst::getSwappedPredicate(Cond);
1567 // FIXME: think about handling pointer comparisons! i.e.:
1568 // while (P != P+100) ++P;
1570 // If we have a comparison of a chrec against a constant, try to use value
1571 // ranges to answer this query.
1572 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
1573 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
1574 if (AddRec->getLoop() == L) {
1575 // Form the comparison range using the constant of the correct type so
1576 // that the ConstantRange class knows to do a signed or unsigned
1578 ConstantInt *CompVal = RHSC->getValue();
1579 const Type *RealTy = ExitCond->getOperand(0)->getType();
1580 CompVal = dyn_cast<ConstantInt>(
1581 ConstantExpr::getBitCast(CompVal, RealTy));
1583 // Form the constant range.
1584 ConstantRange CompRange(
1585 ICmpInst::makeConstantRange(Cond, CompVal->getValue()));
1587 SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange,
1588 false /*Always treat as unsigned range*/);
1589 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
1594 case ICmpInst::ICMP_NE: { // while (X != Y)
1595 // Convert to: while (X-Y != 0)
1596 SCEVHandle TC = HowFarToZero(SCEV::getMinusSCEV(LHS, RHS), L);
1597 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1600 case ICmpInst::ICMP_EQ: {
1601 // Convert to: while (X-Y == 0) // while (X == Y)
1602 SCEVHandle TC = HowFarToNonZero(SCEV::getMinusSCEV(LHS, RHS), L);
1603 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1606 case ICmpInst::ICMP_SLT: {
1607 SCEVHandle TC = HowManyLessThans(LHS, RHS, L);
1608 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1611 case ICmpInst::ICMP_SGT: {
1612 SCEVHandle TC = HowManyLessThans(RHS, LHS, L);
1613 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
1618 cerr << "ComputeIterationCount ";
1619 if (ExitCond->getOperand(0)->getType()->isUnsigned())
1620 cerr << "[unsigned] ";
1622 << Instruction::getOpcodeName(Instruction::ICmp)
1623 << " " << *RHS << "\n";
1627 return ComputeIterationCountExhaustively(L, ExitCond,
1628 ExitBr->getSuccessor(0) == ExitBlock);
1631 static ConstantInt *
1632 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, Constant *C) {
1633 SCEVHandle InVal = SCEVConstant::get(cast<ConstantInt>(C));
1634 SCEVHandle Val = AddRec->evaluateAtIteration(InVal);
1635 assert(isa<SCEVConstant>(Val) &&
1636 "Evaluation of SCEV at constant didn't fold correctly?");
1637 return cast<SCEVConstant>(Val)->getValue();
1640 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
1641 /// and a GEP expression (missing the pointer index) indexing into it, return
1642 /// the addressed element of the initializer or null if the index expression is
1645 GetAddressedElementFromGlobal(GlobalVariable *GV,
1646 const std::vector<ConstantInt*> &Indices) {
1647 Constant *Init = GV->getInitializer();
1648 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
1649 uint64_t Idx = Indices[i]->getZExtValue();
1650 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1651 assert(Idx < CS->getNumOperands() && "Bad struct index!");
1652 Init = cast<Constant>(CS->getOperand(Idx));
1653 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1654 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
1655 Init = cast<Constant>(CA->getOperand(Idx));
1656 } else if (isa<ConstantAggregateZero>(Init)) {
1657 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1658 assert(Idx < STy->getNumElements() && "Bad struct index!");
1659 Init = Constant::getNullValue(STy->getElementType(Idx));
1660 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
1661 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
1662 Init = Constant::getNullValue(ATy->getElementType());
1664 assert(0 && "Unknown constant aggregate type!");
1668 return 0; // Unknown initializer type
1674 /// ComputeLoadConstantCompareIterationCount - Given an exit condition of
1675 /// 'setcc load X, cst', try to se if we can compute the trip count.
1676 SCEVHandle ScalarEvolutionsImpl::
1677 ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS,
1679 ICmpInst::Predicate predicate) {
1680 if (LI->isVolatile()) return UnknownValue;
1682 // Check to see if the loaded pointer is a getelementptr of a global.
1683 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
1684 if (!GEP) return UnknownValue;
1686 // Make sure that it is really a constant global we are gepping, with an
1687 // initializer, and make sure the first IDX is really 0.
1688 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
1689 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
1690 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
1691 !cast<Constant>(GEP->getOperand(1))->isNullValue())
1692 return UnknownValue;
1694 // Okay, we allow one non-constant index into the GEP instruction.
1696 std::vector<ConstantInt*> Indexes;
1697 unsigned VarIdxNum = 0;
1698 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
1699 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
1700 Indexes.push_back(CI);
1701 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
1702 if (VarIdx) return UnknownValue; // Multiple non-constant idx's.
1703 VarIdx = GEP->getOperand(i);
1705 Indexes.push_back(0);
1708 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
1709 // Check to see if X is a loop variant variable value now.
1710 SCEVHandle Idx = getSCEV(VarIdx);
1711 SCEVHandle Tmp = getSCEVAtScope(Idx, L);
1712 if (!isa<SCEVCouldNotCompute>(Tmp)) Idx = Tmp;
1714 // We can only recognize very limited forms of loop index expressions, in
1715 // particular, only affine AddRec's like {C1,+,C2}.
1716 SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
1717 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
1718 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
1719 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
1720 return UnknownValue;
1722 unsigned MaxSteps = MaxBruteForceIterations;
1723 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
1724 ConstantInt *ItCst =
1725 ConstantInt::get(IdxExpr->getType(), IterationNum);
1726 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst);
1728 // Form the GEP offset.
1729 Indexes[VarIdxNum] = Val;
1731 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
1732 if (Result == 0) break; // Cannot compute!
1734 // Evaluate the condition for this iteration.
1735 Result = ConstantExpr::getICmp(predicate, Result, RHS);
1736 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
1737 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
1739 cerr << "\n***\n*** Computed loop count " << *ItCst
1740 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
1743 ++NumArrayLenItCounts;
1744 return SCEVConstant::get(ItCst); // Found terminating iteration!
1747 return UnknownValue;
1751 /// CanConstantFold - Return true if we can constant fold an instruction of the
1752 /// specified type, assuming that all operands were constants.
1753 static bool CanConstantFold(const Instruction *I) {
1754 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
1755 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
1758 if (const CallInst *CI = dyn_cast<CallInst>(I))
1759 if (const Function *F = CI->getCalledFunction())
1760 return canConstantFoldCallTo((Function*)F); // FIXME: elim cast
1764 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
1765 /// in the loop that V is derived from. We allow arbitrary operations along the
1766 /// way, but the operands of an operation must either be constants or a value
1767 /// derived from a constant PHI. If this expression does not fit with these
1768 /// constraints, return null.
1769 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
1770 // If this is not an instruction, or if this is an instruction outside of the
1771 // loop, it can't be derived from a loop PHI.
1772 Instruction *I = dyn_cast<Instruction>(V);
1773 if (I == 0 || !L->contains(I->getParent())) return 0;
1775 if (PHINode *PN = dyn_cast<PHINode>(I))
1776 if (L->getHeader() == I->getParent())
1779 // We don't currently keep track of the control flow needed to evaluate
1780 // PHIs, so we cannot handle PHIs inside of loops.
1783 // If we won't be able to constant fold this expression even if the operands
1784 // are constants, return early.
1785 if (!CanConstantFold(I)) return 0;
1787 // Otherwise, we can evaluate this instruction if all of its operands are
1788 // constant or derived from a PHI node themselves.
1790 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
1791 if (!(isa<Constant>(I->getOperand(Op)) ||
1792 isa<GlobalValue>(I->getOperand(Op)))) {
1793 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
1794 if (P == 0) return 0; // Not evolving from PHI
1798 return 0; // Evolving from multiple different PHIs.
1801 // This is a expression evolving from a constant PHI!
1805 /// EvaluateExpression - Given an expression that passes the
1806 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
1807 /// in the loop has the value PHIVal. If we can't fold this expression for some
1808 /// reason, return null.
1809 static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
1810 if (isa<PHINode>(V)) return PHIVal;
1811 if (GlobalValue *GV = dyn_cast<GlobalValue>(V))
1813 if (Constant *C = dyn_cast<Constant>(V)) return C;
1814 Instruction *I = cast<Instruction>(V);
1816 std::vector<Constant*> Operands;
1817 Operands.resize(I->getNumOperands());
1819 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
1820 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal);
1821 if (Operands[i] == 0) return 0;
1824 return ConstantFoldInstOperands(I, &Operands[0], Operands.size());
1827 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
1828 /// in the header of its containing loop, we know the loop executes a
1829 /// constant number of times, and the PHI node is just a recurrence
1830 /// involving constants, fold it.
1831 Constant *ScalarEvolutionsImpl::
1832 getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& Its, const Loop *L){
1833 std::map<PHINode*, Constant*>::iterator I =
1834 ConstantEvolutionLoopExitValue.find(PN);
1835 if (I != ConstantEvolutionLoopExitValue.end())
1838 if (Its.ugt(APInt(Its.getBitWidth(),MaxBruteForceIterations)))
1839 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
1841 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
1843 // Since the loop is canonicalized, the PHI node must have two entries. One
1844 // entry must be a constant (coming in from outside of the loop), and the
1845 // second must be derived from the same PHI.
1846 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
1847 Constant *StartCST =
1848 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
1850 return RetVal = 0; // Must be a constant.
1852 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
1853 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
1855 return RetVal = 0; // Not derived from same PHI.
1857 // Execute the loop symbolically to determine the exit value.
1858 if (Its.getActiveBits() >= 32)
1859 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
1861 unsigned NumIterations = Its.getZExtValue(); // must be in range
1862 unsigned IterationNum = 0;
1863 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
1864 if (IterationNum == NumIterations)
1865 return RetVal = PHIVal; // Got exit value!
1867 // Compute the value of the PHI node for the next iteration.
1868 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
1869 if (NextPHI == PHIVal)
1870 return RetVal = NextPHI; // Stopped evolving!
1872 return 0; // Couldn't evaluate!
1877 /// ComputeIterationCountExhaustively - If the trip is known to execute a
1878 /// constant number of times (the condition evolves only from constants),
1879 /// try to evaluate a few iterations of the loop until we get the exit
1880 /// condition gets a value of ExitWhen (true or false). If we cannot
1881 /// evaluate the trip count of the loop, return UnknownValue.
1882 SCEVHandle ScalarEvolutionsImpl::
1883 ComputeIterationCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) {
1884 PHINode *PN = getConstantEvolvingPHI(Cond, L);
1885 if (PN == 0) return UnknownValue;
1887 // Since the loop is canonicalized, the PHI node must have two entries. One
1888 // entry must be a constant (coming in from outside of the loop), and the
1889 // second must be derived from the same PHI.
1890 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
1891 Constant *StartCST =
1892 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
1893 if (StartCST == 0) return UnknownValue; // Must be a constant.
1895 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
1896 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
1897 if (PN2 != PN) return UnknownValue; // Not derived from same PHI.
1899 // Okay, we find a PHI node that defines the trip count of this loop. Execute
1900 // the loop symbolically to determine when the condition gets a value of
1902 unsigned IterationNum = 0;
1903 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
1904 for (Constant *PHIVal = StartCST;
1905 IterationNum != MaxIterations; ++IterationNum) {
1906 ConstantInt *CondVal =
1907 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal));
1909 // Couldn't symbolically evaluate.
1910 if (!CondVal) return UnknownValue;
1912 if (CondVal->getValue() == uint64_t(ExitWhen)) {
1913 ConstantEvolutionLoopExitValue[PN] = PHIVal;
1914 ++NumBruteForceTripCountsComputed;
1915 return SCEVConstant::get(ConstantInt::get(Type::Int32Ty, IterationNum));
1918 // Compute the value of the PHI node for the next iteration.
1919 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
1920 if (NextPHI == 0 || NextPHI == PHIVal)
1921 return UnknownValue; // Couldn't evaluate or not making progress...
1925 // Too many iterations were needed to evaluate.
1926 return UnknownValue;
1929 /// getSCEVAtScope - Compute the value of the specified expression within the
1930 /// indicated loop (which may be null to indicate in no loop). If the
1931 /// expression cannot be evaluated, return UnknownValue.
1932 SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) {
1933 // FIXME: this should be turned into a virtual method on SCEV!
1935 if (isa<SCEVConstant>(V)) return V;
1937 // If this instruction is evolves from a constant-evolving PHI, compute the
1938 // exit value from the loop without using SCEVs.
1939 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
1940 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
1941 const Loop *LI = this->LI[I->getParent()];
1942 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
1943 if (PHINode *PN = dyn_cast<PHINode>(I))
1944 if (PN->getParent() == LI->getHeader()) {
1945 // Okay, there is no closed form solution for the PHI node. Check
1946 // to see if the loop that contains it has a known iteration count.
1947 // If so, we may be able to force computation of the exit value.
1948 SCEVHandle IterationCount = getIterationCount(LI);
1949 if (SCEVConstant *ICC = dyn_cast<SCEVConstant>(IterationCount)) {
1950 // Okay, we know how many times the containing loop executes. If
1951 // this is a constant evolving PHI node, get the final value at
1952 // the specified iteration number.
1953 Constant *RV = getConstantEvolutionLoopExitValue(PN,
1954 ICC->getValue()->getValue(),
1956 if (RV) return SCEVUnknown::get(RV);
1960 // Okay, this is an expression that we cannot symbolically evaluate
1961 // into a SCEV. Check to see if it's possible to symbolically evaluate
1962 // the arguments into constants, and if so, try to constant propagate the
1963 // result. This is particularly useful for computing loop exit values.
1964 if (CanConstantFold(I)) {
1965 std::vector<Constant*> Operands;
1966 Operands.reserve(I->getNumOperands());
1967 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
1968 Value *Op = I->getOperand(i);
1969 if (Constant *C = dyn_cast<Constant>(Op)) {
1970 Operands.push_back(C);
1972 SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L);
1973 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
1974 Operands.push_back(ConstantExpr::getIntegerCast(SC->getValue(),
1977 else if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
1978 if (Constant *C = dyn_cast<Constant>(SU->getValue()))
1979 Operands.push_back(ConstantExpr::getIntegerCast(C,
1989 Constant *C =ConstantFoldInstOperands(I, &Operands[0], Operands.size());
1990 return SCEVUnknown::get(C);
1994 // This is some other type of SCEVUnknown, just return it.
1998 if (SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
1999 // Avoid performing the look-up in the common case where the specified
2000 // expression has no loop-variant portions.
2001 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
2002 SCEVHandle OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
2003 if (OpAtScope != Comm->getOperand(i)) {
2004 if (OpAtScope == UnknownValue) return UnknownValue;
2005 // Okay, at least one of these operands is loop variant but might be
2006 // foldable. Build a new instance of the folded commutative expression.
2007 std::vector<SCEVHandle> NewOps(Comm->op_begin(), Comm->op_begin()+i);
2008 NewOps.push_back(OpAtScope);
2010 for (++i; i != e; ++i) {
2011 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
2012 if (OpAtScope == UnknownValue) return UnknownValue;
2013 NewOps.push_back(OpAtScope);
2015 if (isa<SCEVAddExpr>(Comm))
2016 return SCEVAddExpr::get(NewOps);
2017 assert(isa<SCEVMulExpr>(Comm) && "Only know about add and mul!");
2018 return SCEVMulExpr::get(NewOps);
2021 // If we got here, all operands are loop invariant.
2025 if (SCEVSDivExpr *Div = dyn_cast<SCEVSDivExpr>(V)) {
2026 SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L);
2027 if (LHS == UnknownValue) return LHS;
2028 SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L);
2029 if (RHS == UnknownValue) return RHS;
2030 if (LHS == Div->getLHS() && RHS == Div->getRHS())
2031 return Div; // must be loop invariant
2032 return SCEVSDivExpr::get(LHS, RHS);
2035 // If this is a loop recurrence for a loop that does not contain L, then we
2036 // are dealing with the final value computed by the loop.
2037 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
2038 if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
2039 // To evaluate this recurrence, we need to know how many times the AddRec
2040 // loop iterates. Compute this now.
2041 SCEVHandle IterationCount = getIterationCount(AddRec->getLoop());
2042 if (IterationCount == UnknownValue) return UnknownValue;
2043 IterationCount = getTruncateOrZeroExtend(IterationCount,
2046 // If the value is affine, simplify the expression evaluation to just
2047 // Start + Step*IterationCount.
2048 if (AddRec->isAffine())
2049 return SCEVAddExpr::get(AddRec->getStart(),
2050 SCEVMulExpr::get(IterationCount,
2051 AddRec->getOperand(1)));
2053 // Otherwise, evaluate it the hard way.
2054 return AddRec->evaluateAtIteration(IterationCount);
2056 return UnknownValue;
2059 //assert(0 && "Unknown SCEV type!");
2060 return UnknownValue;
2064 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
2065 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
2066 /// might be the same) or two SCEVCouldNotCompute objects.
2068 static std::pair<SCEVHandle,SCEVHandle>
2069 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec) {
2070 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
2071 SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
2072 SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
2073 SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
2075 // We currently can only solve this if the coefficients are constants.
2076 if (!LC || !MC || !NC) {
2077 SCEV *CNC = new SCEVCouldNotCompute();
2078 return std::make_pair(CNC, CNC);
2081 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
2082 APInt L(LC->getValue()->getValue());
2083 APInt M(MC->getValue()->getValue());
2084 APInt N(MC->getValue()->getValue());
2085 APInt Two(BitWidth, 2);
2086 APInt Four(BitWidth, 4);
2089 using namespace APIntOps;
2091 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
2092 // The B coefficient is M-N/2
2096 // The A coefficient is N/2
2100 // Compute the B^2-4ac term.
2103 SqrtTerm -= Four * (A * C);
2105 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
2106 // integer value or else APInt::sqrt() will assert.
2107 APInt SqrtVal(SqrtTerm.sqrt());
2109 // Compute the two solutions for the quadratic formula.
2110 // The divisions must be performed as signed divisions.
2112 APInt TwoA( A * Two );
2113 ConstantInt *Solution1 = ConstantInt::get((NegB + SqrtVal).sdiv(TwoA));
2114 ConstantInt *Solution2 = ConstantInt::get((NegB - SqrtVal).sdiv(TwoA));
2116 return std::make_pair(SCEVUnknown::get(Solution1),
2117 SCEVUnknown::get(Solution2));
2118 } // end APIntOps namespace
2121 /// HowFarToZero - Return the number of times a backedge comparing the specified
2122 /// value to zero will execute. If not computable, return UnknownValue
2123 SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) {
2124 // If the value is a constant
2125 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2126 // If the value is already zero, the branch will execute zero times.
2127 if (C->getValue()->isNullValue()) return C;
2128 return UnknownValue; // Otherwise it will loop infinitely.
2131 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
2132 if (!AddRec || AddRec->getLoop() != L)
2133 return UnknownValue;
2135 if (AddRec->isAffine()) {
2136 // If this is an affine expression the execution count of this branch is
2139 // (0 - Start/Step) iff Start % Step == 0
2141 // Get the initial value for the loop.
2142 SCEVHandle Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
2143 if (isa<SCEVCouldNotCompute>(Start)) return UnknownValue;
2144 SCEVHandle Step = AddRec->getOperand(1);
2146 Step = getSCEVAtScope(Step, L->getParentLoop());
2148 // Figure out if Start % Step == 0.
2149 // FIXME: We should add DivExpr and RemExpr operations to our AST.
2150 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
2151 if (StepC->getValue()->equalsInt(1)) // N % 1 == 0
2152 return SCEV::getNegativeSCEV(Start); // 0 - Start/1 == -Start
2153 if (StepC->getValue()->isAllOnesValue()) // N % -1 == 0
2154 return Start; // 0 - Start/-1 == Start
2156 // Check to see if Start is divisible by SC with no remainder.
2157 if (SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start)) {
2158 ConstantInt *StartCC = StartC->getValue();
2159 Constant *StartNegC = ConstantExpr::getNeg(StartCC);
2160 Constant *Rem = ConstantExpr::getSRem(StartNegC, StepC->getValue());
2161 if (Rem->isNullValue()) {
2162 Constant *Result =ConstantExpr::getSDiv(StartNegC,StepC->getValue());
2163 return SCEVUnknown::get(Result);
2167 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
2168 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
2169 // the quadratic equation to solve it.
2170 std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec);
2171 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2172 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2175 cerr << "HFTZ: " << *V << " - sol#1: " << *R1
2176 << " sol#2: " << *R2 << "\n";
2178 // Pick the smallest positive root value.
2179 if (ConstantInt *CB =
2180 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
2181 R1->getValue(), R2->getValue()))) {
2182 if (CB->getZExtValue() == false)
2183 std::swap(R1, R2); // R1 is the minimum root now.
2185 // We can only use this value if the chrec ends up with an exact zero
2186 // value at this index. When solving for "X*X != 5", for example, we
2187 // should not accept a root of 2.
2188 SCEVHandle Val = AddRec->evaluateAtIteration(R1);
2189 if (SCEVConstant *EvalVal = dyn_cast<SCEVConstant>(Val))
2190 if (EvalVal->getValue()->isNullValue())
2191 return R1; // We found a quadratic root!
2196 return UnknownValue;
2199 /// HowFarToNonZero - Return the number of times a backedge checking the
2200 /// specified value for nonzero will execute. If not computable, return
2202 SCEVHandle ScalarEvolutionsImpl::HowFarToNonZero(SCEV *V, const Loop *L) {
2203 // Loops that look like: while (X == 0) are very strange indeed. We don't
2204 // handle them yet except for the trivial case. This could be expanded in the
2205 // future as needed.
2207 // If the value is a constant, check to see if it is known to be non-zero
2208 // already. If so, the backedge will execute zero times.
2209 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2210 Constant *Zero = Constant::getNullValue(C->getValue()->getType());
2212 ConstantExpr::getICmp(ICmpInst::ICMP_NE, C->getValue(), Zero);
2213 if (NonZero == ConstantInt::getTrue())
2214 return getSCEV(Zero);
2215 return UnknownValue; // Otherwise it will loop infinitely.
2218 // We could implement others, but I really doubt anyone writes loops like
2219 // this, and if they did, they would already be constant folded.
2220 return UnknownValue;
2223 /// HowManyLessThans - Return the number of times a backedge containing the
2224 /// specified less-than comparison will execute. If not computable, return
2226 SCEVHandle ScalarEvolutionsImpl::
2227 HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L) {
2228 // Only handle: "ADDREC < LoopInvariant".
2229 if (!RHS->isLoopInvariant(L)) return UnknownValue;
2231 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
2232 if (!AddRec || AddRec->getLoop() != L)
2233 return UnknownValue;
2235 if (AddRec->isAffine()) {
2236 // FORNOW: We only support unit strides.
2237 SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, RHS->getType());
2238 if (AddRec->getOperand(1) != One)
2239 return UnknownValue;
2241 // The number of iterations for "[n,+,1] < m", is m-n. However, we don't
2242 // know that m is >= n on input to the loop. If it is, the condition return
2243 // true zero times. What we really should return, for full generality, is
2244 // SMAX(0, m-n). Since we cannot check this, we will instead check for a
2245 // canonical loop form: most do-loops will have a check that dominates the
2246 // loop, that only enters the loop if [n-1]<m. If we can find this check,
2247 // we know that the SMAX will evaluate to m-n, because we know that m >= n.
2249 // Search for the check.
2250 BasicBlock *Preheader = L->getLoopPreheader();
2251 BasicBlock *PreheaderDest = L->getHeader();
2252 if (Preheader == 0) return UnknownValue;
2254 BranchInst *LoopEntryPredicate =
2255 dyn_cast<BranchInst>(Preheader->getTerminator());
2256 if (!LoopEntryPredicate) return UnknownValue;
2258 // This might be a critical edge broken out. If the loop preheader ends in
2259 // an unconditional branch to the loop, check to see if the preheader has a
2260 // single predecessor, and if so, look for its terminator.
2261 while (LoopEntryPredicate->isUnconditional()) {
2262 PreheaderDest = Preheader;
2263 Preheader = Preheader->getSinglePredecessor();
2264 if (!Preheader) return UnknownValue; // Multiple preds.
2266 LoopEntryPredicate =
2267 dyn_cast<BranchInst>(Preheader->getTerminator());
2268 if (!LoopEntryPredicate) return UnknownValue;
2271 // Now that we found a conditional branch that dominates the loop, check to
2272 // see if it is the comparison we are looking for.
2273 if (ICmpInst *ICI = dyn_cast<ICmpInst>(LoopEntryPredicate->getCondition())){
2274 Value *PreCondLHS = ICI->getOperand(0);
2275 Value *PreCondRHS = ICI->getOperand(1);
2276 ICmpInst::Predicate Cond;
2277 if (LoopEntryPredicate->getSuccessor(0) == PreheaderDest)
2278 Cond = ICI->getPredicate();
2280 Cond = ICI->getInversePredicate();
2283 case ICmpInst::ICMP_UGT:
2284 std::swap(PreCondLHS, PreCondRHS);
2285 Cond = ICmpInst::ICMP_ULT;
2287 case ICmpInst::ICMP_SGT:
2288 std::swap(PreCondLHS, PreCondRHS);
2289 Cond = ICmpInst::ICMP_SLT;
2294 if (Cond == ICmpInst::ICMP_SLT) {
2295 if (PreCondLHS->getType()->isInteger()) {
2296 if (RHS != getSCEV(PreCondRHS))
2297 return UnknownValue; // Not a comparison against 'm'.
2299 if (SCEV::getMinusSCEV(AddRec->getOperand(0), One)
2300 != getSCEV(PreCondLHS))
2301 return UnknownValue; // Not a comparison against 'n-1'.
2303 else return UnknownValue;
2304 } else if (Cond == ICmpInst::ICMP_ULT)
2305 return UnknownValue;
2307 // cerr << "Computed Loop Trip Count as: "
2308 // << // *SCEV::getMinusSCEV(RHS, AddRec->getOperand(0)) << "\n";
2309 return SCEV::getMinusSCEV(RHS, AddRec->getOperand(0));
2312 return UnknownValue;
2315 return UnknownValue;
2318 /// getNumIterationsInRange - Return the number of iterations of this loop that
2319 /// produce values in the specified constant range. Another way of looking at
2320 /// this is that it returns the first iteration number where the value is not in
2321 /// the condition, thus computing the exit count. If the iteration count can't
2322 /// be computed, an instance of SCEVCouldNotCompute is returned.
2323 SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
2324 bool isSigned) const {
2325 if (Range.isFullSet()) // Infinite loop.
2326 return new SCEVCouldNotCompute();
2328 // If the start is a non-zero constant, shift the range to simplify things.
2329 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
2330 if (!SC->getValue()->isNullValue()) {
2331 std::vector<SCEVHandle> Operands(op_begin(), op_end());
2332 Operands[0] = SCEVUnknown::getIntegerSCEV(0, SC->getType());
2333 SCEVHandle Shifted = SCEVAddRecExpr::get(Operands, getLoop());
2334 if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast<SCEVAddRecExpr>(Shifted))
2335 return ShiftedAddRec->getNumIterationsInRange(
2336 Range.subtract(SC->getValue()->getValue()),isSigned);
2337 // This is strange and shouldn't happen.
2338 return new SCEVCouldNotCompute();
2341 // The only time we can solve this is when we have all constant indices.
2342 // Otherwise, we cannot determine the overflow conditions.
2343 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
2344 if (!isa<SCEVConstant>(getOperand(i)))
2345 return new SCEVCouldNotCompute();
2348 // Okay at this point we know that all elements of the chrec are constants and
2349 // that the start element is zero.
2351 // First check to see if the range contains zero. If not, the first
2353 if (!Range.contains(APInt(getBitWidth(),0)))
2354 return SCEVConstant::get(ConstantInt::get(getType(),0));
2357 // If this is an affine expression then we have this situation:
2358 // Solve {0,+,A} in Range === Ax in Range
2360 // Since we know that zero is in the range, we know that the upper value of
2361 // the range must be the first possible exit value. Also note that we
2362 // already checked for a full range.
2363 const APInt &Upper = Range.getUpper();
2364 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
2365 APInt One(getBitWidth(),1);
2367 // The exit value should be (Upper+A-1)/A.
2368 APInt ExitVal(Upper);
2370 ExitVal = (Upper + A - One).sdiv(A);
2371 ConstantInt *ExitValue = ConstantInt::get(ExitVal);
2373 // Evaluate at the exit value. If we really did fall out of the valid
2374 // range, then we computed our trip count, otherwise wrap around or other
2375 // things must have happened.
2376 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue);
2377 if (Range.contains(Val->getValue()))
2378 return new SCEVCouldNotCompute(); // Something strange happened
2380 // Ensure that the previous value is in the range. This is a sanity check.
2381 assert(Range.contains(
2382 EvaluateConstantChrecAtConstant(this,
2383 ConstantInt::get(ExitVal - One))->getValue()) &&
2384 "Linear scev computation is off in a bad way!");
2385 return SCEVConstant::get(cast<ConstantInt>(ExitValue));
2386 } else if (isQuadratic()) {
2387 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
2388 // quadratic equation to solve it. To do this, we must frame our problem in
2389 // terms of figuring out when zero is crossed, instead of when
2390 // Range.getUpper() is crossed.
2391 std::vector<SCEVHandle> NewOps(op_begin(), op_end());
2392 NewOps[0] = SCEV::getNegativeSCEV(SCEVUnknown::get(
2393 ConstantInt::get(Range.getUpper())));
2394 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, getLoop());
2396 // Next, solve the constructed addrec
2397 std::pair<SCEVHandle,SCEVHandle> Roots =
2398 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec));
2399 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2400 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2402 // Pick the smallest positive root value.
2403 if (ConstantInt *CB =
2404 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
2405 R1->getValue(), R2->getValue()))) {
2406 if (CB->getZExtValue() == false)
2407 std::swap(R1, R2); // R1 is the minimum root now.
2409 // Make sure the root is not off by one. The returned iteration should
2410 // not be in the range, but the previous one should be. When solving
2411 // for "X*X < 5", for example, we should not return a root of 2.
2412 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
2414 if (Range.contains(R1Val->getValue())) {
2415 // The next iteration must be out of the range...
2417 ConstantExpr::getAdd(R1->getValue(),
2418 ConstantInt::get(R1->getType(), 1));
2420 R1Val = EvaluateConstantChrecAtConstant(this, NextVal);
2421 if (!Range.contains(R1Val->getValue()))
2422 return SCEVUnknown::get(NextVal);
2423 return new SCEVCouldNotCompute(); // Something strange happened
2426 // If R1 was not in the range, then it is a good return value. Make
2427 // sure that R1-1 WAS in the range though, just in case.
2429 ConstantExpr::getSub(R1->getValue(),
2430 ConstantInt::get(R1->getType(), 1));
2431 R1Val = EvaluateConstantChrecAtConstant(this, NextVal);
2432 if (Range.contains(R1Val->getValue()))
2434 return new SCEVCouldNotCompute(); // Something strange happened
2439 // Fallback, if this is a general polynomial, figure out the progression
2440 // through brute force: evaluate until we find an iteration that fails the
2441 // test. This is likely to be slow, but getting an accurate trip count is
2442 // incredibly important, we will be able to simplify the exit test a lot, and
2443 // we are almost guaranteed to get a trip count in this case.
2444 ConstantInt *TestVal = ConstantInt::get(getType(), 0);
2445 ConstantInt *One = ConstantInt::get(getType(), 1);
2446 ConstantInt *EndVal = TestVal; // Stop when we wrap around.
2448 ++NumBruteForceEvaluations;
2449 SCEVHandle Val = evaluateAtIteration(SCEVConstant::get(TestVal));
2450 if (!isa<SCEVConstant>(Val)) // This shouldn't happen.
2451 return new SCEVCouldNotCompute();
2453 // Check to see if we found the value!
2454 if (!Range.contains(cast<SCEVConstant>(Val)->getValue()->getValue()))
2455 return SCEVConstant::get(TestVal);
2457 // Increment to test the next index.
2458 TestVal = cast<ConstantInt>(ConstantExpr::getAdd(TestVal, One));
2459 } while (TestVal != EndVal);
2461 return new SCEVCouldNotCompute();
2466 //===----------------------------------------------------------------------===//
2467 // ScalarEvolution Class Implementation
2468 //===----------------------------------------------------------------------===//
2470 bool ScalarEvolution::runOnFunction(Function &F) {
2471 Impl = new ScalarEvolutionsImpl(F, getAnalysis<LoopInfo>());
2475 void ScalarEvolution::releaseMemory() {
2476 delete (ScalarEvolutionsImpl*)Impl;
2480 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
2481 AU.setPreservesAll();
2482 AU.addRequiredTransitive<LoopInfo>();
2485 SCEVHandle ScalarEvolution::getSCEV(Value *V) const {
2486 return ((ScalarEvolutionsImpl*)Impl)->getSCEV(V);
2489 /// hasSCEV - Return true if the SCEV for this value has already been
2491 bool ScalarEvolution::hasSCEV(Value *V) const {
2492 return ((ScalarEvolutionsImpl*)Impl)->hasSCEV(V);
2496 /// setSCEV - Insert the specified SCEV into the map of current SCEVs for
2497 /// the specified value.
2498 void ScalarEvolution::setSCEV(Value *V, const SCEVHandle &H) {
2499 ((ScalarEvolutionsImpl*)Impl)->setSCEV(V, H);
2503 SCEVHandle ScalarEvolution::getIterationCount(const Loop *L) const {
2504 return ((ScalarEvolutionsImpl*)Impl)->getIterationCount(L);
2507 bool ScalarEvolution::hasLoopInvariantIterationCount(const Loop *L) const {
2508 return !isa<SCEVCouldNotCompute>(getIterationCount(L));
2511 SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) const {
2512 return ((ScalarEvolutionsImpl*)Impl)->getSCEVAtScope(getSCEV(V), L);
2515 void ScalarEvolution::deleteInstructionFromRecords(Instruction *I) const {
2516 return ((ScalarEvolutionsImpl*)Impl)->deleteInstructionFromRecords(I);
2519 static void PrintLoopInfo(std::ostream &OS, const ScalarEvolution *SE,
2521 // Print all inner loops first
2522 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
2523 PrintLoopInfo(OS, SE, *I);
2525 cerr << "Loop " << L->getHeader()->getName() << ": ";
2527 std::vector<BasicBlock*> ExitBlocks;
2528 L->getExitBlocks(ExitBlocks);
2529 if (ExitBlocks.size() != 1)
2530 cerr << "<multiple exits> ";
2532 if (SE->hasLoopInvariantIterationCount(L)) {
2533 cerr << *SE->getIterationCount(L) << " iterations! ";
2535 cerr << "Unpredictable iteration count. ";
2541 void ScalarEvolution::print(std::ostream &OS, const Module* ) const {
2542 Function &F = ((ScalarEvolutionsImpl*)Impl)->F;
2543 LoopInfo &LI = ((ScalarEvolutionsImpl*)Impl)->LI;
2545 OS << "Classifying expressions for: " << F.getName() << "\n";
2546 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
2547 if (I->getType()->isInteger()) {
2550 SCEVHandle SV = getSCEV(&*I);
2554 if ((*I).getType()->isInteger()) {
2555 ConstantRange Bounds = SV->getValueRange();
2556 if (!Bounds.isFullSet())
2557 OS << "Bounds: " << Bounds << " ";
2560 if (const Loop *L = LI.getLoopFor((*I).getParent())) {
2562 SCEVHandle ExitValue = getSCEVAtScope(&*I, L->getParentLoop());
2563 if (isa<SCEVCouldNotCompute>(ExitValue)) {
2564 OS << "<<Unknown>>";
2574 OS << "Determining loop execution counts for: " << F.getName() << "\n";
2575 for (LoopInfo::iterator I = LI.begin(), E = LI.end(); I != E; ++I)
2576 PrintLoopInfo(OS, this, *I);