1 //===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the implementation of the scalar evolution analysis
11 // engine, which is used primarily to analyze expressions involving induction
12 // variables in loops.
14 // There are several aspects to this library. First is the representation of
15 // scalar expressions, which are represented as subclasses of the SCEV class.
16 // These classes are used to represent certain types of subexpressions that we
17 // can handle. These classes are reference counted, managed by the const SCEV*
18 // class. We only create one SCEV of a particular shape, so pointer-comparisons
19 // for equality are legal.
21 // One important aspect of the SCEV objects is that they are never cyclic, even
22 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
23 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
24 // recurrence) then we represent it directly as a recurrence node, otherwise we
25 // represent it as a SCEVUnknown node.
27 // In addition to being able to represent expressions of various types, we also
28 // have folders that are used to build the *canonical* representation for a
29 // particular expression. These folders are capable of using a variety of
30 // rewrite rules to simplify the expressions.
32 // Once the folders are defined, we can implement the more interesting
33 // higher-level code, such as the code that recognizes PHI nodes of various
34 // types, computes the execution count of a loop, etc.
36 // TODO: We should use these routines and value representations to implement
37 // dependence analysis!
39 //===----------------------------------------------------------------------===//
41 // There are several good references for the techniques used in this analysis.
43 // Chains of recurrences -- a method to expedite the evaluation
44 // of closed-form functions
45 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
47 // On computational properties of chains of recurrences
50 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
51 // Robert A. van Engelen
53 // Efficient Symbolic Analysis for Optimizing Compilers
54 // Robert A. van Engelen
56 // Using the chains of recurrences algebra for data dependence testing and
57 // induction variable substitution
58 // MS Thesis, Johnie Birch
60 //===----------------------------------------------------------------------===//
62 #define DEBUG_TYPE "scalar-evolution"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Constants.h"
65 #include "llvm/DerivedTypes.h"
66 #include "llvm/GlobalVariable.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/Analysis/ConstantFolding.h"
69 #include "llvm/Analysis/Dominators.h"
70 #include "llvm/Analysis/LoopInfo.h"
71 #include "llvm/Analysis/ValueTracking.h"
72 #include "llvm/Assembly/Writer.h"
73 #include "llvm/Target/TargetData.h"
74 #include "llvm/Support/CommandLine.h"
75 #include "llvm/Support/Compiler.h"
76 #include "llvm/Support/ConstantRange.h"
77 #include "llvm/Support/GetElementPtrTypeIterator.h"
78 #include "llvm/Support/InstIterator.h"
79 #include "llvm/Support/MathExtras.h"
80 #include "llvm/Support/raw_ostream.h"
81 #include "llvm/ADT/Statistic.h"
82 #include "llvm/ADT/STLExtras.h"
86 STATISTIC(NumArrayLenItCounts,
87 "Number of trip counts computed with array length");
88 STATISTIC(NumTripCountsComputed,
89 "Number of loops with predictable loop counts");
90 STATISTIC(NumTripCountsNotComputed,
91 "Number of loops without predictable loop counts");
92 STATISTIC(NumBruteForceTripCountsComputed,
93 "Number of loops with trip counts computed by force");
95 static cl::opt<unsigned>
96 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
97 cl::desc("Maximum number of iterations SCEV will "
98 "symbolically execute a constant derived loop"),
101 static RegisterPass<ScalarEvolution>
102 R("scalar-evolution", "Scalar Evolution Analysis", false, true);
103 char ScalarEvolution::ID = 0;
105 //===----------------------------------------------------------------------===//
106 // SCEV class definitions
107 //===----------------------------------------------------------------------===//
109 //===----------------------------------------------------------------------===//
110 // Implementation of the SCEV class.
113 void SCEV::dump() const {
118 void SCEV::print(std::ostream &o) const {
119 raw_os_ostream OS(o);
123 bool SCEV::isZero() const {
124 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
125 return SC->getValue()->isZero();
129 bool SCEV::isOne() const {
130 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
131 return SC->getValue()->isOne();
135 bool SCEV::isAllOnesValue() const {
136 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
137 return SC->getValue()->isAllOnesValue();
141 SCEVCouldNotCompute::SCEVCouldNotCompute() :
142 SCEV(scCouldNotCompute) {}
144 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
145 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
149 const Type *SCEVCouldNotCompute::getType() const {
150 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
154 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
155 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
159 const SCEV* SCEVCouldNotCompute::
160 replaceSymbolicValuesWithConcrete(const SCEV* Sym,
162 ScalarEvolution &SE) const {
166 void SCEVCouldNotCompute::print(raw_ostream &OS) const {
167 OS << "***COULDNOTCOMPUTE***";
170 bool SCEVCouldNotCompute::classof(const SCEV *S) {
171 return S->getSCEVType() == scCouldNotCompute;
175 // SCEVConstants - Only allow the creation of one SCEVConstant for any
176 // particular value. Don't use a const SCEV* here, or else the object will
179 const SCEV* ScalarEvolution::getConstant(ConstantInt *V) {
180 SCEVConstant *&R = SCEVConstants[V];
181 if (R == 0) R = new SCEVConstant(V);
185 const SCEV* ScalarEvolution::getConstant(const APInt& Val) {
186 return getConstant(ConstantInt::get(Val));
190 ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
191 return getConstant(ConstantInt::get(cast<IntegerType>(Ty), V, isSigned));
194 const Type *SCEVConstant::getType() const { return V->getType(); }
196 void SCEVConstant::print(raw_ostream &OS) const {
197 WriteAsOperand(OS, V, false);
200 SCEVCastExpr::SCEVCastExpr(unsigned SCEVTy,
201 const SCEV* op, const Type *ty)
202 : SCEV(SCEVTy), Op(op), Ty(ty) {}
204 bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
205 return Op->dominates(BB, DT);
208 // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any
209 // particular input. Don't use a const SCEV* here, or else the object will
212 SCEVTruncateExpr::SCEVTruncateExpr(const SCEV* op, const Type *ty)
213 : SCEVCastExpr(scTruncate, op, ty) {
214 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
215 (Ty->isInteger() || isa<PointerType>(Ty)) &&
216 "Cannot truncate non-integer value!");
220 void SCEVTruncateExpr::print(raw_ostream &OS) const {
221 OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
224 // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any
225 // particular input. Don't use a const SCEV* here, or else the object will never
228 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEV* op, const Type *ty)
229 : SCEVCastExpr(scZeroExtend, op, ty) {
230 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
231 (Ty->isInteger() || isa<PointerType>(Ty)) &&
232 "Cannot zero extend non-integer value!");
235 void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
236 OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
239 // SCEVSignExtends - Only allow the creation of one SCEVSignExtendExpr for any
240 // particular input. Don't use a const SCEV* here, or else the object will never
243 SCEVSignExtendExpr::SCEVSignExtendExpr(const SCEV* op, const Type *ty)
244 : SCEVCastExpr(scSignExtend, op, ty) {
245 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
246 (Ty->isInteger() || isa<PointerType>(Ty)) &&
247 "Cannot sign extend non-integer value!");
250 void SCEVSignExtendExpr::print(raw_ostream &OS) const {
251 OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
254 // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any
255 // particular input. Don't use a const SCEV* here, or else the object will never
258 void SCEVCommutativeExpr::print(raw_ostream &OS) const {
259 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
260 const char *OpStr = getOperationStr();
261 OS << "(" << *Operands[0];
262 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
263 OS << OpStr << *Operands[i];
267 const SCEV* SCEVCommutativeExpr::
268 replaceSymbolicValuesWithConcrete(const SCEV* Sym,
270 ScalarEvolution &SE) const {
271 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
273 getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
274 if (H != getOperand(i)) {
275 SmallVector<const SCEV*, 8> NewOps;
276 NewOps.reserve(getNumOperands());
277 for (unsigned j = 0; j != i; ++j)
278 NewOps.push_back(getOperand(j));
280 for (++i; i != e; ++i)
281 NewOps.push_back(getOperand(i)->
282 replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
284 if (isa<SCEVAddExpr>(this))
285 return SE.getAddExpr(NewOps);
286 else if (isa<SCEVMulExpr>(this))
287 return SE.getMulExpr(NewOps);
288 else if (isa<SCEVSMaxExpr>(this))
289 return SE.getSMaxExpr(NewOps);
290 else if (isa<SCEVUMaxExpr>(this))
291 return SE.getUMaxExpr(NewOps);
293 assert(0 && "Unknown commutative expr!");
299 bool SCEVNAryExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
300 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
301 if (!getOperand(i)->dominates(BB, DT))
308 // SCEVUDivs - Only allow the creation of one SCEVUDivExpr for any particular
309 // input. Don't use a const SCEV* here, or else the object will never be
312 bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
313 return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
316 void SCEVUDivExpr::print(raw_ostream &OS) const {
317 OS << "(" << *LHS << " /u " << *RHS << ")";
320 const Type *SCEVUDivExpr::getType() const {
321 // In most cases the types of LHS and RHS will be the same, but in some
322 // crazy cases one or the other may be a pointer. ScalarEvolution doesn't
323 // depend on the type for correctness, but handling types carefully can
324 // avoid extra casts in the SCEVExpander. The LHS is more likely to be
325 // a pointer type than the RHS, so use the RHS' type here.
326 return RHS->getType();
329 // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
330 // particular input. Don't use a const SCEV* here, or else the object will never
333 const SCEV* SCEVAddRecExpr::
334 replaceSymbolicValuesWithConcrete(const SCEV* Sym,
336 ScalarEvolution &SE) const {
337 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
339 getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
340 if (H != getOperand(i)) {
341 SmallVector<const SCEV*, 8> NewOps;
342 NewOps.reserve(getNumOperands());
343 for (unsigned j = 0; j != i; ++j)
344 NewOps.push_back(getOperand(j));
346 for (++i; i != e; ++i)
347 NewOps.push_back(getOperand(i)->
348 replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
350 return SE.getAddRecExpr(NewOps, L);
357 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
358 // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't
359 // contain L and if the start is invariant.
360 // Add recurrences are never invariant in the function-body (null loop).
362 !QueryLoop->contains(L->getHeader()) &&
363 getOperand(0)->isLoopInvariant(QueryLoop);
367 void SCEVAddRecExpr::print(raw_ostream &OS) const {
368 OS << "{" << *Operands[0];
369 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
370 OS << ",+," << *Operands[i];
371 OS << "}<" << L->getHeader()->getName() + ">";
374 // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular
375 // value. Don't use a const SCEV* here, or else the object will never be
378 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
379 // All non-instruction values are loop invariant. All instructions are loop
380 // invariant if they are not contained in the specified loop.
381 // Instructions are never considered invariant in the function body
382 // (null loop) because they are defined within the "loop".
383 if (Instruction *I = dyn_cast<Instruction>(V))
384 return L && !L->contains(I->getParent());
388 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
389 if (Instruction *I = dyn_cast<Instruction>(getValue()))
390 return DT->dominates(I->getParent(), BB);
394 const Type *SCEVUnknown::getType() const {
398 void SCEVUnknown::print(raw_ostream &OS) const {
399 WriteAsOperand(OS, V, false);
402 //===----------------------------------------------------------------------===//
404 //===----------------------------------------------------------------------===//
407 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
408 /// than the complexity of the RHS. This comparator is used to canonicalize
410 class VISIBILITY_HIDDEN SCEVComplexityCompare {
413 explicit SCEVComplexityCompare(LoopInfo *li) : LI(li) {}
415 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
416 // Primarily, sort the SCEVs by their getSCEVType().
417 if (LHS->getSCEVType() != RHS->getSCEVType())
418 return LHS->getSCEVType() < RHS->getSCEVType();
420 // Aside from the getSCEVType() ordering, the particular ordering
421 // isn't very important except that it's beneficial to be consistent,
422 // so that (a + b) and (b + a) don't end up as different expressions.
424 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
425 // not as complete as it could be.
426 if (const SCEVUnknown *LU = dyn_cast<SCEVUnknown>(LHS)) {
427 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
429 // Order pointer values after integer values. This helps SCEVExpander
431 if (isa<PointerType>(LU->getType()) && !isa<PointerType>(RU->getType()))
433 if (isa<PointerType>(RU->getType()) && !isa<PointerType>(LU->getType()))
436 // Compare getValueID values.
437 if (LU->getValue()->getValueID() != RU->getValue()->getValueID())
438 return LU->getValue()->getValueID() < RU->getValue()->getValueID();
440 // Sort arguments by their position.
441 if (const Argument *LA = dyn_cast<Argument>(LU->getValue())) {
442 const Argument *RA = cast<Argument>(RU->getValue());
443 return LA->getArgNo() < RA->getArgNo();
446 // For instructions, compare their loop depth, and their opcode.
447 // This is pretty loose.
448 if (Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
449 Instruction *RV = cast<Instruction>(RU->getValue());
451 // Compare loop depths.
452 if (LI->getLoopDepth(LV->getParent()) !=
453 LI->getLoopDepth(RV->getParent()))
454 return LI->getLoopDepth(LV->getParent()) <
455 LI->getLoopDepth(RV->getParent());
458 if (LV->getOpcode() != RV->getOpcode())
459 return LV->getOpcode() < RV->getOpcode();
461 // Compare the number of operands.
462 if (LV->getNumOperands() != RV->getNumOperands())
463 return LV->getNumOperands() < RV->getNumOperands();
469 // Compare constant values.
470 if (const SCEVConstant *LC = dyn_cast<SCEVConstant>(LHS)) {
471 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
472 return LC->getValue()->getValue().ult(RC->getValue()->getValue());
475 // Compare addrec loop depths.
476 if (const SCEVAddRecExpr *LA = dyn_cast<SCEVAddRecExpr>(LHS)) {
477 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
478 if (LA->getLoop()->getLoopDepth() != RA->getLoop()->getLoopDepth())
479 return LA->getLoop()->getLoopDepth() < RA->getLoop()->getLoopDepth();
482 // Lexicographically compare n-ary expressions.
483 if (const SCEVNAryExpr *LC = dyn_cast<SCEVNAryExpr>(LHS)) {
484 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
485 for (unsigned i = 0, e = LC->getNumOperands(); i != e; ++i) {
486 if (i >= RC->getNumOperands())
488 if (operator()(LC->getOperand(i), RC->getOperand(i)))
490 if (operator()(RC->getOperand(i), LC->getOperand(i)))
493 return LC->getNumOperands() < RC->getNumOperands();
496 // Lexicographically compare udiv expressions.
497 if (const SCEVUDivExpr *LC = dyn_cast<SCEVUDivExpr>(LHS)) {
498 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
499 if (operator()(LC->getLHS(), RC->getLHS()))
501 if (operator()(RC->getLHS(), LC->getLHS()))
503 if (operator()(LC->getRHS(), RC->getRHS()))
505 if (operator()(RC->getRHS(), LC->getRHS()))
510 // Compare cast expressions by operand.
511 if (const SCEVCastExpr *LC = dyn_cast<SCEVCastExpr>(LHS)) {
512 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
513 return operator()(LC->getOperand(), RC->getOperand());
516 assert(0 && "Unknown SCEV kind!");
522 /// GroupByComplexity - Given a list of SCEV objects, order them by their
523 /// complexity, and group objects of the same complexity together by value.
524 /// When this routine is finished, we know that any duplicates in the vector are
525 /// consecutive and that complexity is monotonically increasing.
527 /// Note that we go take special precautions to ensure that we get determinstic
528 /// results from this routine. In other words, we don't want the results of
529 /// this to depend on where the addresses of various SCEV objects happened to
532 static void GroupByComplexity(SmallVectorImpl<const SCEV*> &Ops,
534 if (Ops.size() < 2) return; // Noop
535 if (Ops.size() == 2) {
536 // This is the common case, which also happens to be trivially simple.
538 if (SCEVComplexityCompare(LI)(Ops[1], Ops[0]))
539 std::swap(Ops[0], Ops[1]);
543 // Do the rough sort by complexity.
544 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
546 // Now that we are sorted by complexity, group elements of the same
547 // complexity. Note that this is, at worst, N^2, but the vector is likely to
548 // be extremely short in practice. Note that we take this approach because we
549 // do not want to depend on the addresses of the objects we are grouping.
550 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
551 const SCEV *S = Ops[i];
552 unsigned Complexity = S->getSCEVType();
554 // If there are any objects of the same complexity and same value as this
556 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
557 if (Ops[j] == S) { // Found a duplicate.
558 // Move it to immediately after i'th element.
559 std::swap(Ops[i+1], Ops[j]);
560 ++i; // no need to rescan it.
561 if (i == e-2) return; // Done!
569 //===----------------------------------------------------------------------===//
570 // Simple SCEV method implementations
571 //===----------------------------------------------------------------------===//
573 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
575 static const SCEV* BinomialCoefficient(const SCEV* It, unsigned K,
577 const Type* ResultTy) {
578 // Handle the simplest case efficiently.
580 return SE.getTruncateOrZeroExtend(It, ResultTy);
582 // We are using the following formula for BC(It, K):
584 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
586 // Suppose, W is the bitwidth of the return value. We must be prepared for
587 // overflow. Hence, we must assure that the result of our computation is
588 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
589 // safe in modular arithmetic.
591 // However, this code doesn't use exactly that formula; the formula it uses
592 // is something like the following, where T is the number of factors of 2 in
593 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
596 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
598 // This formula is trivially equivalent to the previous formula. However,
599 // this formula can be implemented much more efficiently. The trick is that
600 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
601 // arithmetic. To do exact division in modular arithmetic, all we have
602 // to do is multiply by the inverse. Therefore, this step can be done at
605 // The next issue is how to safely do the division by 2^T. The way this
606 // is done is by doing the multiplication step at a width of at least W + T
607 // bits. This way, the bottom W+T bits of the product are accurate. Then,
608 // when we perform the division by 2^T (which is equivalent to a right shift
609 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
610 // truncated out after the division by 2^T.
612 // In comparison to just directly using the first formula, this technique
613 // is much more efficient; using the first formula requires W * K bits,
614 // but this formula less than W + K bits. Also, the first formula requires
615 // a division step, whereas this formula only requires multiplies and shifts.
617 // It doesn't matter whether the subtraction step is done in the calculation
618 // width or the input iteration count's width; if the subtraction overflows,
619 // the result must be zero anyway. We prefer here to do it in the width of
620 // the induction variable because it helps a lot for certain cases; CodeGen
621 // isn't smart enough to ignore the overflow, which leads to much less
622 // efficient code if the width of the subtraction is wider than the native
625 // (It's possible to not widen at all by pulling out factors of 2 before
626 // the multiplication; for example, K=2 can be calculated as
627 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
628 // extra arithmetic, so it's not an obvious win, and it gets
629 // much more complicated for K > 3.)
631 // Protection from insane SCEVs; this bound is conservative,
632 // but it probably doesn't matter.
634 return SE.getCouldNotCompute();
636 unsigned W = SE.getTypeSizeInBits(ResultTy);
638 // Calculate K! / 2^T and T; we divide out the factors of two before
639 // multiplying for calculating K! / 2^T to avoid overflow.
640 // Other overflow doesn't matter because we only care about the bottom
641 // W bits of the result.
642 APInt OddFactorial(W, 1);
644 for (unsigned i = 3; i <= K; ++i) {
646 unsigned TwoFactors = Mult.countTrailingZeros();
648 Mult = Mult.lshr(TwoFactors);
649 OddFactorial *= Mult;
652 // We need at least W + T bits for the multiplication step
653 unsigned CalculationBits = W + T;
655 // Calcuate 2^T, at width T+W.
656 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
658 // Calculate the multiplicative inverse of K! / 2^T;
659 // this multiplication factor will perform the exact division by
661 APInt Mod = APInt::getSignedMinValue(W+1);
662 APInt MultiplyFactor = OddFactorial.zext(W+1);
663 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
664 MultiplyFactor = MultiplyFactor.trunc(W);
666 // Calculate the product, at width T+W
667 const IntegerType *CalculationTy = IntegerType::get(CalculationBits);
668 const SCEV* Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
669 for (unsigned i = 1; i != K; ++i) {
670 const SCEV* S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
671 Dividend = SE.getMulExpr(Dividend,
672 SE.getTruncateOrZeroExtend(S, CalculationTy));
676 const SCEV* DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
678 // Truncate the result, and divide by K! / 2^T.
680 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
681 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
684 /// evaluateAtIteration - Return the value of this chain of recurrences at
685 /// the specified iteration number. We can evaluate this recurrence by
686 /// multiplying each element in the chain by the binomial coefficient
687 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
689 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
691 /// where BC(It, k) stands for binomial coefficient.
693 const SCEV* SCEVAddRecExpr::evaluateAtIteration(const SCEV* It,
694 ScalarEvolution &SE) const {
695 const SCEV* Result = getStart();
696 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
697 // The computation is correct in the face of overflow provided that the
698 // multiplication is performed _after_ the evaluation of the binomial
700 const SCEV* Coeff = BinomialCoefficient(It, i, SE, getType());
701 if (isa<SCEVCouldNotCompute>(Coeff))
704 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
709 //===----------------------------------------------------------------------===//
710 // SCEV Expression folder implementations
711 //===----------------------------------------------------------------------===//
713 const SCEV* ScalarEvolution::getTruncateExpr(const SCEV* Op,
715 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
716 "This is not a truncating conversion!");
717 assert(isSCEVable(Ty) &&
718 "This is not a conversion to a SCEVable type!");
719 Ty = getEffectiveSCEVType(Ty);
721 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
723 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
725 // trunc(trunc(x)) --> trunc(x)
726 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
727 return getTruncateExpr(ST->getOperand(), Ty);
729 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
730 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
731 return getTruncateOrSignExtend(SS->getOperand(), Ty);
733 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
734 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
735 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
737 // If the input value is a chrec scev, truncate the chrec's operands.
738 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
739 SmallVector<const SCEV*, 4> Operands;
740 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
741 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
742 return getAddRecExpr(Operands, AddRec->getLoop());
745 SCEVTruncateExpr *&Result = SCEVTruncates[std::make_pair(Op, Ty)];
746 if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
750 const SCEV* ScalarEvolution::getZeroExtendExpr(const SCEV* Op,
752 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
753 "This is not an extending conversion!");
754 assert(isSCEVable(Ty) &&
755 "This is not a conversion to a SCEVable type!");
756 Ty = getEffectiveSCEVType(Ty);
758 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
759 const Type *IntTy = getEffectiveSCEVType(Ty);
760 Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
761 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
762 return getConstant(cast<ConstantInt>(C));
765 // zext(zext(x)) --> zext(x)
766 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
767 return getZeroExtendExpr(SZ->getOperand(), Ty);
769 // If the input value is a chrec scev, and we can prove that the value
770 // did not overflow the old, smaller, value, we can zero extend all of the
771 // operands (often constants). This allows analysis of something like
772 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
773 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
774 if (AR->isAffine()) {
775 // Check whether the backedge-taken count is SCEVCouldNotCompute.
776 // Note that this serves two purposes: It filters out loops that are
777 // simply not analyzable, and it covers the case where this code is
778 // being called from within backedge-taken count analysis, such that
779 // attempting to ask for the backedge-taken count would likely result
780 // in infinite recursion. In the later case, the analysis code will
781 // cope with a conservative value, and it will take care to purge
782 // that value once it has finished.
783 const SCEV* MaxBECount = getMaxBackedgeTakenCount(AR->getLoop());
784 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
785 // Manually compute the final value for AR, checking for
787 const SCEV* Start = AR->getStart();
788 const SCEV* Step = AR->getStepRecurrence(*this);
790 // Check whether the backedge-taken count can be losslessly casted to
791 // the addrec's type. The count is always unsigned.
792 const SCEV* CastedMaxBECount =
793 getTruncateOrZeroExtend(MaxBECount, Start->getType());
794 const SCEV* RecastedMaxBECount =
795 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
796 if (MaxBECount == RecastedMaxBECount) {
798 IntegerType::get(getTypeSizeInBits(Start->getType()) * 2);
799 // Check whether Start+Step*MaxBECount has no unsigned overflow.
801 getMulExpr(CastedMaxBECount,
802 getTruncateOrZeroExtend(Step, Start->getType()));
803 const SCEV* Add = getAddExpr(Start, ZMul);
804 const SCEV* OperandExtendedAdd =
805 getAddExpr(getZeroExtendExpr(Start, WideTy),
806 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
807 getZeroExtendExpr(Step, WideTy)));
808 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
809 // Return the expression with the addrec on the outside.
810 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
811 getZeroExtendExpr(Step, Ty),
814 // Similar to above, only this time treat the step value as signed.
815 // This covers loops that count down.
817 getMulExpr(CastedMaxBECount,
818 getTruncateOrSignExtend(Step, Start->getType()));
819 Add = getAddExpr(Start, SMul);
821 getAddExpr(getZeroExtendExpr(Start, WideTy),
822 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
823 getSignExtendExpr(Step, WideTy)));
824 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
825 // Return the expression with the addrec on the outside.
826 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
827 getSignExtendExpr(Step, Ty),
833 SCEVZeroExtendExpr *&Result = SCEVZeroExtends[std::make_pair(Op, Ty)];
834 if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
838 const SCEV* ScalarEvolution::getSignExtendExpr(const SCEV* Op,
840 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
841 "This is not an extending conversion!");
842 assert(isSCEVable(Ty) &&
843 "This is not a conversion to a SCEVable type!");
844 Ty = getEffectiveSCEVType(Ty);
846 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
847 const Type *IntTy = getEffectiveSCEVType(Ty);
848 Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
849 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
850 return getConstant(cast<ConstantInt>(C));
853 // sext(sext(x)) --> sext(x)
854 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
855 return getSignExtendExpr(SS->getOperand(), Ty);
857 // If the input value is a chrec scev, and we can prove that the value
858 // did not overflow the old, smaller, value, we can sign extend all of the
859 // operands (often constants). This allows analysis of something like
860 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
861 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
862 if (AR->isAffine()) {
863 // Check whether the backedge-taken count is SCEVCouldNotCompute.
864 // Note that this serves two purposes: It filters out loops that are
865 // simply not analyzable, and it covers the case where this code is
866 // being called from within backedge-taken count analysis, such that
867 // attempting to ask for the backedge-taken count would likely result
868 // in infinite recursion. In the later case, the analysis code will
869 // cope with a conservative value, and it will take care to purge
870 // that value once it has finished.
871 const SCEV* MaxBECount = getMaxBackedgeTakenCount(AR->getLoop());
872 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
873 // Manually compute the final value for AR, checking for
875 const SCEV* Start = AR->getStart();
876 const SCEV* Step = AR->getStepRecurrence(*this);
878 // Check whether the backedge-taken count can be losslessly casted to
879 // the addrec's type. The count is always unsigned.
880 const SCEV* CastedMaxBECount =
881 getTruncateOrZeroExtend(MaxBECount, Start->getType());
882 const SCEV* RecastedMaxBECount =
883 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
884 if (MaxBECount == RecastedMaxBECount) {
886 IntegerType::get(getTypeSizeInBits(Start->getType()) * 2);
887 // Check whether Start+Step*MaxBECount has no signed overflow.
889 getMulExpr(CastedMaxBECount,
890 getTruncateOrSignExtend(Step, Start->getType()));
891 const SCEV* Add = getAddExpr(Start, SMul);
892 const SCEV* OperandExtendedAdd =
893 getAddExpr(getSignExtendExpr(Start, WideTy),
894 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
895 getSignExtendExpr(Step, WideTy)));
896 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
897 // Return the expression with the addrec on the outside.
898 return getAddRecExpr(getSignExtendExpr(Start, Ty),
899 getSignExtendExpr(Step, Ty),
905 SCEVSignExtendExpr *&Result = SCEVSignExtends[std::make_pair(Op, Ty)];
906 if (Result == 0) Result = new SCEVSignExtendExpr(Op, Ty);
910 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
911 /// unspecified bits out to the given type.
913 const SCEV* ScalarEvolution::getAnyExtendExpr(const SCEV* Op,
915 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
916 "This is not an extending conversion!");
917 assert(isSCEVable(Ty) &&
918 "This is not a conversion to a SCEVable type!");
919 Ty = getEffectiveSCEVType(Ty);
921 // Sign-extend negative constants.
922 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
923 if (SC->getValue()->getValue().isNegative())
924 return getSignExtendExpr(Op, Ty);
926 // Peel off a truncate cast.
927 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
928 const SCEV* NewOp = T->getOperand();
929 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
930 return getAnyExtendExpr(NewOp, Ty);
931 return getTruncateOrNoop(NewOp, Ty);
934 // Next try a zext cast. If the cast is folded, use it.
935 const SCEV* ZExt = getZeroExtendExpr(Op, Ty);
936 if (!isa<SCEVZeroExtendExpr>(ZExt))
939 // Next try a sext cast. If the cast is folded, use it.
940 const SCEV* SExt = getSignExtendExpr(Op, Ty);
941 if (!isa<SCEVSignExtendExpr>(SExt))
944 // If the expression is obviously signed, use the sext cast value.
945 if (isa<SCEVSMaxExpr>(Op))
948 // Absent any other information, use the zext cast value.
952 /// CollectAddOperandsWithScales - Process the given Ops list, which is
953 /// a list of operands to be added under the given scale, update the given
954 /// map. This is a helper function for getAddRecExpr. As an example of
955 /// what it does, given a sequence of operands that would form an add
956 /// expression like this:
958 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
960 /// where A and B are constants, update the map with these values:
962 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
964 /// and add 13 + A*B*29 to AccumulatedConstant.
965 /// This will allow getAddRecExpr to produce this:
967 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
969 /// This form often exposes folding opportunities that are hidden in
970 /// the original operand list.
972 /// Return true iff it appears that any interesting folding opportunities
973 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
974 /// the common case where no interesting opportunities are present, and
975 /// is also used as a check to avoid infinite recursion.
978 CollectAddOperandsWithScales(DenseMap<const SCEV*, APInt> &M,
979 SmallVector<const SCEV*, 8> &NewOps,
980 APInt &AccumulatedConstant,
981 const SmallVectorImpl<const SCEV*> &Ops,
983 ScalarEvolution &SE) {
984 bool Interesting = false;
986 // Iterate over the add operands.
987 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
988 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
989 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
991 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
992 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
993 // A multiplication of a constant with another add; recurse.
995 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
996 cast<SCEVAddExpr>(Mul->getOperand(1))
1000 // A multiplication of a constant with some other value. Update
1002 SmallVector<const SCEV*, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1003 const SCEV* Key = SE.getMulExpr(MulOps);
1004 std::pair<DenseMap<const SCEV*, APInt>::iterator, bool> Pair =
1005 M.insert(std::make_pair(Key, APInt()));
1007 Pair.first->second = NewScale;
1008 NewOps.push_back(Pair.first->first);
1010 Pair.first->second += NewScale;
1011 // The map already had an entry for this value, which may indicate
1012 // a folding opportunity.
1016 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1017 // Pull a buried constant out to the outside.
1018 if (Scale != 1 || AccumulatedConstant != 0 || C->isZero())
1020 AccumulatedConstant += Scale * C->getValue()->getValue();
1022 // An ordinary operand. Update the map.
1023 std::pair<DenseMap<const SCEV*, APInt>::iterator, bool> Pair =
1024 M.insert(std::make_pair(Ops[i], APInt()));
1026 Pair.first->second = Scale;
1027 NewOps.push_back(Pair.first->first);
1029 Pair.first->second += Scale;
1030 // The map already had an entry for this value, which may indicate
1031 // a folding opportunity.
1041 struct APIntCompare {
1042 bool operator()(const APInt &LHS, const APInt &RHS) const {
1043 return LHS.ult(RHS);
1048 /// getAddExpr - Get a canonical add expression, or something simpler if
1050 const SCEV* ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV*> &Ops) {
1051 assert(!Ops.empty() && "Cannot get empty add!");
1052 if (Ops.size() == 1) return Ops[0];
1054 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1055 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1056 getEffectiveSCEVType(Ops[0]->getType()) &&
1057 "SCEVAddExpr operand types don't match!");
1060 // Sort by complexity, this groups all similar expression types together.
1061 GroupByComplexity(Ops, LI);
1063 // If there are any constants, fold them together.
1065 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1067 assert(Idx < Ops.size());
1068 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1069 // We found two constants, fold them together!
1070 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1071 RHSC->getValue()->getValue());
1072 if (Ops.size() == 2) return Ops[0];
1073 Ops.erase(Ops.begin()+1); // Erase the folded element
1074 LHSC = cast<SCEVConstant>(Ops[0]);
1077 // If we are left with a constant zero being added, strip it off.
1078 if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1079 Ops.erase(Ops.begin());
1084 if (Ops.size() == 1) return Ops[0];
1086 // Okay, check to see if the same value occurs in the operand list twice. If
1087 // so, merge them together into an multiply expression. Since we sorted the
1088 // list, these values are required to be adjacent.
1089 const Type *Ty = Ops[0]->getType();
1090 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1091 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1092 // Found a match, merge the two values into a multiply, and add any
1093 // remaining values to the result.
1094 const SCEV* Two = getIntegerSCEV(2, Ty);
1095 const SCEV* Mul = getMulExpr(Ops[i], Two);
1096 if (Ops.size() == 2)
1098 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1100 return getAddExpr(Ops);
1103 // Check for truncates. If all the operands are truncated from the same
1104 // type, see if factoring out the truncate would permit the result to be
1105 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1106 // if the contents of the resulting outer trunc fold to something simple.
1107 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1108 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1109 const Type *DstType = Trunc->getType();
1110 const Type *SrcType = Trunc->getOperand()->getType();
1111 SmallVector<const SCEV*, 8> LargeOps;
1113 // Check all the operands to see if they can be represented in the
1114 // source type of the truncate.
1115 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1116 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1117 if (T->getOperand()->getType() != SrcType) {
1121 LargeOps.push_back(T->getOperand());
1122 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1123 // This could be either sign or zero extension, but sign extension
1124 // is much more likely to be foldable here.
1125 LargeOps.push_back(getSignExtendExpr(C, SrcType));
1126 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1127 SmallVector<const SCEV*, 8> LargeMulOps;
1128 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1129 if (const SCEVTruncateExpr *T =
1130 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1131 if (T->getOperand()->getType() != SrcType) {
1135 LargeMulOps.push_back(T->getOperand());
1136 } else if (const SCEVConstant *C =
1137 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1138 // This could be either sign or zero extension, but sign extension
1139 // is much more likely to be foldable here.
1140 LargeMulOps.push_back(getSignExtendExpr(C, SrcType));
1147 LargeOps.push_back(getMulExpr(LargeMulOps));
1154 // Evaluate the expression in the larger type.
1155 const SCEV* Fold = getAddExpr(LargeOps);
1156 // If it folds to something simple, use it. Otherwise, don't.
1157 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1158 return getTruncateExpr(Fold, DstType);
1162 // Skip past any other cast SCEVs.
1163 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1166 // If there are add operands they would be next.
1167 if (Idx < Ops.size()) {
1168 bool DeletedAdd = false;
1169 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1170 // If we have an add, expand the add operands onto the end of the operands
1172 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
1173 Ops.erase(Ops.begin()+Idx);
1177 // If we deleted at least one add, we added operands to the end of the list,
1178 // and they are not necessarily sorted. Recurse to resort and resimplify
1179 // any operands we just aquired.
1181 return getAddExpr(Ops);
1184 // Skip over the add expression until we get to a multiply.
1185 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1188 // Check to see if there are any folding opportunities present with
1189 // operands multiplied by constant values.
1190 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1191 uint64_t BitWidth = getTypeSizeInBits(Ty);
1192 DenseMap<const SCEV*, APInt> M;
1193 SmallVector<const SCEV*, 8> NewOps;
1194 APInt AccumulatedConstant(BitWidth, 0);
1195 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1196 Ops, APInt(BitWidth, 1), *this)) {
1197 // Some interesting folding opportunity is present, so its worthwhile to
1198 // re-generate the operands list. Group the operands by constant scale,
1199 // to avoid multiplying by the same constant scale multiple times.
1200 std::map<APInt, SmallVector<const SCEV*, 4>, APIntCompare> MulOpLists;
1201 for (SmallVector<const SCEV*, 8>::iterator I = NewOps.begin(),
1202 E = NewOps.end(); I != E; ++I)
1203 MulOpLists[M.find(*I)->second].push_back(*I);
1204 // Re-generate the operands list.
1206 if (AccumulatedConstant != 0)
1207 Ops.push_back(getConstant(AccumulatedConstant));
1208 for (std::map<APInt, SmallVector<const SCEV*, 4>, APIntCompare>::iterator I =
1209 MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1211 Ops.push_back(getMulExpr(getConstant(I->first), getAddExpr(I->second)));
1213 return getIntegerSCEV(0, Ty);
1214 if (Ops.size() == 1)
1216 return getAddExpr(Ops);
1220 // If we are adding something to a multiply expression, make sure the
1221 // something is not already an operand of the multiply. If so, merge it into
1223 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1224 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1225 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1226 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1227 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1228 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(Ops[AddOp])) {
1229 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1230 const SCEV* InnerMul = Mul->getOperand(MulOp == 0);
1231 if (Mul->getNumOperands() != 2) {
1232 // If the multiply has more than two operands, we must get the
1234 SmallVector<const SCEV*, 4> MulOps(Mul->op_begin(), Mul->op_end());
1235 MulOps.erase(MulOps.begin()+MulOp);
1236 InnerMul = getMulExpr(MulOps);
1238 const SCEV* One = getIntegerSCEV(1, Ty);
1239 const SCEV* AddOne = getAddExpr(InnerMul, One);
1240 const SCEV* OuterMul = getMulExpr(AddOne, Ops[AddOp]);
1241 if (Ops.size() == 2) return OuterMul;
1243 Ops.erase(Ops.begin()+AddOp);
1244 Ops.erase(Ops.begin()+Idx-1);
1246 Ops.erase(Ops.begin()+Idx);
1247 Ops.erase(Ops.begin()+AddOp-1);
1249 Ops.push_back(OuterMul);
1250 return getAddExpr(Ops);
1253 // Check this multiply against other multiplies being added together.
1254 for (unsigned OtherMulIdx = Idx+1;
1255 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1257 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1258 // If MulOp occurs in OtherMul, we can fold the two multiplies
1260 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1261 OMulOp != e; ++OMulOp)
1262 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1263 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1264 const SCEV* InnerMul1 = Mul->getOperand(MulOp == 0);
1265 if (Mul->getNumOperands() != 2) {
1266 SmallVector<const SCEV*, 4> MulOps(Mul->op_begin(), Mul->op_end());
1267 MulOps.erase(MulOps.begin()+MulOp);
1268 InnerMul1 = getMulExpr(MulOps);
1270 const SCEV* InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1271 if (OtherMul->getNumOperands() != 2) {
1272 SmallVector<const SCEV*, 4> MulOps(OtherMul->op_begin(),
1273 OtherMul->op_end());
1274 MulOps.erase(MulOps.begin()+OMulOp);
1275 InnerMul2 = getMulExpr(MulOps);
1277 const SCEV* InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1278 const SCEV* OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1279 if (Ops.size() == 2) return OuterMul;
1280 Ops.erase(Ops.begin()+Idx);
1281 Ops.erase(Ops.begin()+OtherMulIdx-1);
1282 Ops.push_back(OuterMul);
1283 return getAddExpr(Ops);
1289 // If there are any add recurrences in the operands list, see if any other
1290 // added values are loop invariant. If so, we can fold them into the
1292 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1295 // Scan over all recurrences, trying to fold loop invariants into them.
1296 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1297 // Scan all of the other operands to this add and add them to the vector if
1298 // they are loop invariant w.r.t. the recurrence.
1299 SmallVector<const SCEV*, 8> LIOps;
1300 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1301 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1302 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1303 LIOps.push_back(Ops[i]);
1304 Ops.erase(Ops.begin()+i);
1308 // If we found some loop invariants, fold them into the recurrence.
1309 if (!LIOps.empty()) {
1310 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1311 LIOps.push_back(AddRec->getStart());
1313 SmallVector<const SCEV*, 4> AddRecOps(AddRec->op_begin(),
1315 AddRecOps[0] = getAddExpr(LIOps);
1317 const SCEV* NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
1318 // If all of the other operands were loop invariant, we are done.
1319 if (Ops.size() == 1) return NewRec;
1321 // Otherwise, add the folded AddRec by the non-liv parts.
1322 for (unsigned i = 0;; ++i)
1323 if (Ops[i] == AddRec) {
1327 return getAddExpr(Ops);
1330 // Okay, if there weren't any loop invariants to be folded, check to see if
1331 // there are multiple AddRec's with the same loop induction variable being
1332 // added together. If so, we can fold them.
1333 for (unsigned OtherIdx = Idx+1;
1334 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1335 if (OtherIdx != Idx) {
1336 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1337 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1338 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
1339 SmallVector<const SCEV*, 4> NewOps(AddRec->op_begin(), AddRec->op_end());
1340 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
1341 if (i >= NewOps.size()) {
1342 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
1343 OtherAddRec->op_end());
1346 NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
1348 const SCEV* NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
1350 if (Ops.size() == 2) return NewAddRec;
1352 Ops.erase(Ops.begin()+Idx);
1353 Ops.erase(Ops.begin()+OtherIdx-1);
1354 Ops.push_back(NewAddRec);
1355 return getAddExpr(Ops);
1359 // Otherwise couldn't fold anything into this recurrence. Move onto the
1363 // Okay, it looks like we really DO need an add expr. Check to see if we
1364 // already have one, otherwise create a new one.
1365 std::vector<const SCEV*> SCEVOps(Ops.begin(), Ops.end());
1366 SCEVCommutativeExpr *&Result = SCEVCommExprs[std::make_pair(scAddExpr,
1368 if (Result == 0) Result = new SCEVAddExpr(Ops);
1373 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1375 const SCEV* ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV*> &Ops) {
1376 assert(!Ops.empty() && "Cannot get empty mul!");
1378 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1379 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1380 getEffectiveSCEVType(Ops[0]->getType()) &&
1381 "SCEVMulExpr operand types don't match!");
1384 // Sort by complexity, this groups all similar expression types together.
1385 GroupByComplexity(Ops, LI);
1387 // If there are any constants, fold them together.
1389 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1391 // C1*(C2+V) -> C1*C2 + C1*V
1392 if (Ops.size() == 2)
1393 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1394 if (Add->getNumOperands() == 2 &&
1395 isa<SCEVConstant>(Add->getOperand(0)))
1396 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1397 getMulExpr(LHSC, Add->getOperand(1)));
1401 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1402 // We found two constants, fold them together!
1403 ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() *
1404 RHSC->getValue()->getValue());
1405 Ops[0] = getConstant(Fold);
1406 Ops.erase(Ops.begin()+1); // Erase the folded element
1407 if (Ops.size() == 1) return Ops[0];
1408 LHSC = cast<SCEVConstant>(Ops[0]);
1411 // If we are left with a constant one being multiplied, strip it off.
1412 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1413 Ops.erase(Ops.begin());
1415 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1416 // If we have a multiply of zero, it will always be zero.
1421 // Skip over the add expression until we get to a multiply.
1422 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1425 if (Ops.size() == 1)
1428 // If there are mul operands inline them all into this expression.
1429 if (Idx < Ops.size()) {
1430 bool DeletedMul = false;
1431 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1432 // If we have an mul, expand the mul operands onto the end of the operands
1434 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1435 Ops.erase(Ops.begin()+Idx);
1439 // If we deleted at least one mul, we added operands to the end of the list,
1440 // and they are not necessarily sorted. Recurse to resort and resimplify
1441 // any operands we just aquired.
1443 return getMulExpr(Ops);
1446 // If there are any add recurrences in the operands list, see if any other
1447 // added values are loop invariant. If so, we can fold them into the
1449 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1452 // Scan over all recurrences, trying to fold loop invariants into them.
1453 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1454 // Scan all of the other operands to this mul and add them to the vector if
1455 // they are loop invariant w.r.t. the recurrence.
1456 SmallVector<const SCEV*, 8> LIOps;
1457 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1458 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1459 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1460 LIOps.push_back(Ops[i]);
1461 Ops.erase(Ops.begin()+i);
1465 // If we found some loop invariants, fold them into the recurrence.
1466 if (!LIOps.empty()) {
1467 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1468 SmallVector<const SCEV*, 4> NewOps;
1469 NewOps.reserve(AddRec->getNumOperands());
1470 if (LIOps.size() == 1) {
1471 const SCEV *Scale = LIOps[0];
1472 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1473 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1475 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1476 SmallVector<const SCEV*, 4> MulOps(LIOps.begin(), LIOps.end());
1477 MulOps.push_back(AddRec->getOperand(i));
1478 NewOps.push_back(getMulExpr(MulOps));
1482 const SCEV* NewRec = getAddRecExpr(NewOps, AddRec->getLoop());
1484 // If all of the other operands were loop invariant, we are done.
1485 if (Ops.size() == 1) return NewRec;
1487 // Otherwise, multiply the folded AddRec by the non-liv parts.
1488 for (unsigned i = 0;; ++i)
1489 if (Ops[i] == AddRec) {
1493 return getMulExpr(Ops);
1496 // Okay, if there weren't any loop invariants to be folded, check to see if
1497 // there are multiple AddRec's with the same loop induction variable being
1498 // multiplied together. If so, we can fold them.
1499 for (unsigned OtherIdx = Idx+1;
1500 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1501 if (OtherIdx != Idx) {
1502 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1503 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1504 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1505 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1506 const SCEV* NewStart = getMulExpr(F->getStart(),
1508 const SCEV* B = F->getStepRecurrence(*this);
1509 const SCEV* D = G->getStepRecurrence(*this);
1510 const SCEV* NewStep = getAddExpr(getMulExpr(F, D),
1513 const SCEV* NewAddRec = getAddRecExpr(NewStart, NewStep,
1515 if (Ops.size() == 2) return NewAddRec;
1517 Ops.erase(Ops.begin()+Idx);
1518 Ops.erase(Ops.begin()+OtherIdx-1);
1519 Ops.push_back(NewAddRec);
1520 return getMulExpr(Ops);
1524 // Otherwise couldn't fold anything into this recurrence. Move onto the
1528 // Okay, it looks like we really DO need an mul expr. Check to see if we
1529 // already have one, otherwise create a new one.
1530 std::vector<const SCEV*> SCEVOps(Ops.begin(), Ops.end());
1531 SCEVCommutativeExpr *&Result = SCEVCommExprs[std::make_pair(scMulExpr,
1534 Result = new SCEVMulExpr(Ops);
1538 /// getUDivExpr - Get a canonical multiply expression, or something simpler if
1540 const SCEV* ScalarEvolution::getUDivExpr(const SCEV* LHS,
1542 assert(getEffectiveSCEVType(LHS->getType()) ==
1543 getEffectiveSCEVType(RHS->getType()) &&
1544 "SCEVUDivExpr operand types don't match!");
1546 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1547 if (RHSC->getValue()->equalsInt(1))
1548 return LHS; // X udiv 1 --> x
1550 return getIntegerSCEV(0, LHS->getType()); // value is undefined
1552 // Determine if the division can be folded into the operands of
1554 // TODO: Generalize this to non-constants by using known-bits information.
1555 const Type *Ty = LHS->getType();
1556 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1557 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ;
1558 // For non-power-of-two values, effectively round the value up to the
1559 // nearest power of two.
1560 if (!RHSC->getValue()->getValue().isPowerOf2())
1562 const IntegerType *ExtTy =
1563 IntegerType::get(getTypeSizeInBits(Ty) + MaxShiftAmt);
1564 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1565 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1566 if (const SCEVConstant *Step =
1567 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1568 if (!Step->getValue()->getValue()
1569 .urem(RHSC->getValue()->getValue()) &&
1570 getZeroExtendExpr(AR, ExtTy) ==
1571 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1572 getZeroExtendExpr(Step, ExtTy),
1574 SmallVector<const SCEV*, 4> Operands;
1575 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1576 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1577 return getAddRecExpr(Operands, AR->getLoop());
1579 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1580 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1581 SmallVector<const SCEV*, 4> Operands;
1582 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1583 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1584 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1585 // Find an operand that's safely divisible.
1586 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1587 const SCEV* Op = M->getOperand(i);
1588 const SCEV* Div = getUDivExpr(Op, RHSC);
1589 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1590 const SmallVectorImpl<const SCEV*> &MOperands = M->getOperands();
1591 Operands = SmallVector<const SCEV*, 4>(MOperands.begin(),
1594 return getMulExpr(Operands);
1598 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1599 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1600 SmallVector<const SCEV*, 4> Operands;
1601 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1602 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1603 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1605 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1606 const SCEV* Op = getUDivExpr(A->getOperand(i), RHS);
1607 if (isa<SCEVUDivExpr>(Op) || getMulExpr(Op, RHS) != A->getOperand(i))
1609 Operands.push_back(Op);
1611 if (Operands.size() == A->getNumOperands())
1612 return getAddExpr(Operands);
1616 // Fold if both operands are constant.
1617 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1618 Constant *LHSCV = LHSC->getValue();
1619 Constant *RHSCV = RHSC->getValue();
1620 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
1625 SCEVUDivExpr *&Result = SCEVUDivs[std::make_pair(LHS, RHS)];
1626 if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
1631 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1632 /// Simplify the expression as much as possible.
1633 const SCEV* ScalarEvolution::getAddRecExpr(const SCEV* Start,
1634 const SCEV* Step, const Loop *L) {
1635 SmallVector<const SCEV*, 4> Operands;
1636 Operands.push_back(Start);
1637 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1638 if (StepChrec->getLoop() == L) {
1639 Operands.insert(Operands.end(), StepChrec->op_begin(),
1640 StepChrec->op_end());
1641 return getAddRecExpr(Operands, L);
1644 Operands.push_back(Step);
1645 return getAddRecExpr(Operands, L);
1648 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1649 /// Simplify the expression as much as possible.
1650 const SCEV* ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV*> &Operands,
1652 if (Operands.size() == 1) return Operands[0];
1654 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
1655 assert(getEffectiveSCEVType(Operands[i]->getType()) ==
1656 getEffectiveSCEVType(Operands[0]->getType()) &&
1657 "SCEVAddRecExpr operand types don't match!");
1660 if (Operands.back()->isZero()) {
1661 Operands.pop_back();
1662 return getAddRecExpr(Operands, L); // {X,+,0} --> X
1665 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
1666 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
1667 const Loop* NestedLoop = NestedAR->getLoop();
1668 if (L->getLoopDepth() < NestedLoop->getLoopDepth()) {
1669 SmallVector<const SCEV*, 4> NestedOperands(NestedAR->op_begin(),
1670 NestedAR->op_end());
1671 Operands[0] = NestedAR->getStart();
1672 NestedOperands[0] = getAddRecExpr(Operands, L);
1673 return getAddRecExpr(NestedOperands, NestedLoop);
1677 std::vector<const SCEV*> SCEVOps(Operands.begin(), Operands.end());
1678 SCEVAddRecExpr *&Result = SCEVAddRecExprs[std::make_pair(L, SCEVOps)];
1679 if (Result == 0) Result = new SCEVAddRecExpr(Operands, L);
1683 const SCEV* ScalarEvolution::getSMaxExpr(const SCEV* LHS,
1685 SmallVector<const SCEV*, 2> Ops;
1688 return getSMaxExpr(Ops);
1692 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV*> &Ops) {
1693 assert(!Ops.empty() && "Cannot get empty smax!");
1694 if (Ops.size() == 1) return Ops[0];
1696 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1697 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1698 getEffectiveSCEVType(Ops[0]->getType()) &&
1699 "SCEVSMaxExpr operand types don't match!");
1702 // Sort by complexity, this groups all similar expression types together.
1703 GroupByComplexity(Ops, LI);
1705 // If there are any constants, fold them together.
1707 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1709 assert(Idx < Ops.size());
1710 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1711 // We found two constants, fold them together!
1712 ConstantInt *Fold = ConstantInt::get(
1713 APIntOps::smax(LHSC->getValue()->getValue(),
1714 RHSC->getValue()->getValue()));
1715 Ops[0] = getConstant(Fold);
1716 Ops.erase(Ops.begin()+1); // Erase the folded element
1717 if (Ops.size() == 1) return Ops[0];
1718 LHSC = cast<SCEVConstant>(Ops[0]);
1721 // If we are left with a constant -inf, strip it off.
1722 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
1723 Ops.erase(Ops.begin());
1728 if (Ops.size() == 1) return Ops[0];
1730 // Find the first SMax
1731 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
1734 // Check to see if one of the operands is an SMax. If so, expand its operands
1735 // onto our operand list, and recurse to simplify.
1736 if (Idx < Ops.size()) {
1737 bool DeletedSMax = false;
1738 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
1739 Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
1740 Ops.erase(Ops.begin()+Idx);
1745 return getSMaxExpr(Ops);
1748 // Okay, check to see if the same value occurs in the operand list twice. If
1749 // so, delete one. Since we sorted the list, these values are required to
1751 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1752 if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
1753 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1757 if (Ops.size() == 1) return Ops[0];
1759 assert(!Ops.empty() && "Reduced smax down to nothing!");
1761 // Okay, it looks like we really DO need an smax expr. Check to see if we
1762 // already have one, otherwise create a new one.
1763 std::vector<const SCEV*> SCEVOps(Ops.begin(), Ops.end());
1764 SCEVCommutativeExpr *&Result = SCEVCommExprs[std::make_pair(scSMaxExpr,
1766 if (Result == 0) Result = new SCEVSMaxExpr(Ops);
1770 const SCEV* ScalarEvolution::getUMaxExpr(const SCEV* LHS,
1772 SmallVector<const SCEV*, 2> Ops;
1775 return getUMaxExpr(Ops);
1779 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV*> &Ops) {
1780 assert(!Ops.empty() && "Cannot get empty umax!");
1781 if (Ops.size() == 1) return Ops[0];
1783 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1784 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1785 getEffectiveSCEVType(Ops[0]->getType()) &&
1786 "SCEVUMaxExpr operand types don't match!");
1789 // Sort by complexity, this groups all similar expression types together.
1790 GroupByComplexity(Ops, LI);
1792 // If there are any constants, fold them together.
1794 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1796 assert(Idx < Ops.size());
1797 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1798 // We found two constants, fold them together!
1799 ConstantInt *Fold = ConstantInt::get(
1800 APIntOps::umax(LHSC->getValue()->getValue(),
1801 RHSC->getValue()->getValue()));
1802 Ops[0] = getConstant(Fold);
1803 Ops.erase(Ops.begin()+1); // Erase the folded element
1804 if (Ops.size() == 1) return Ops[0];
1805 LHSC = cast<SCEVConstant>(Ops[0]);
1808 // If we are left with a constant zero, strip it off.
1809 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
1810 Ops.erase(Ops.begin());
1815 if (Ops.size() == 1) return Ops[0];
1817 // Find the first UMax
1818 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
1821 // Check to see if one of the operands is a UMax. If so, expand its operands
1822 // onto our operand list, and recurse to simplify.
1823 if (Idx < Ops.size()) {
1824 bool DeletedUMax = false;
1825 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
1826 Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
1827 Ops.erase(Ops.begin()+Idx);
1832 return getUMaxExpr(Ops);
1835 // Okay, check to see if the same value occurs in the operand list twice. If
1836 // so, delete one. Since we sorted the list, these values are required to
1838 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1839 if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
1840 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1844 if (Ops.size() == 1) return Ops[0];
1846 assert(!Ops.empty() && "Reduced umax down to nothing!");
1848 // Okay, it looks like we really DO need a umax expr. Check to see if we
1849 // already have one, otherwise create a new one.
1850 std::vector<const SCEV*> SCEVOps(Ops.begin(), Ops.end());
1851 SCEVCommutativeExpr *&Result = SCEVCommExprs[std::make_pair(scUMaxExpr,
1853 if (Result == 0) Result = new SCEVUMaxExpr(Ops);
1857 const SCEV* ScalarEvolution::getSMinExpr(const SCEV* LHS,
1859 // ~smax(~x, ~y) == smin(x, y).
1860 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
1863 const SCEV* ScalarEvolution::getUMinExpr(const SCEV* LHS,
1865 // ~umax(~x, ~y) == umin(x, y)
1866 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
1869 const SCEV* ScalarEvolution::getUnknown(Value *V) {
1870 // Don't attempt to do anything other than create a SCEVUnknown object
1871 // here. createSCEV only calls getUnknown after checking for all other
1872 // interesting possibilities, and any other code that calls getUnknown
1873 // is doing so in order to hide a value from SCEV canonicalization.
1875 SCEVUnknown *&Result = SCEVUnknowns[V];
1876 if (Result == 0) Result = new SCEVUnknown(V);
1880 //===----------------------------------------------------------------------===//
1881 // Basic SCEV Analysis and PHI Idiom Recognition Code
1884 /// isSCEVable - Test if values of the given type are analyzable within
1885 /// the SCEV framework. This primarily includes integer types, and it
1886 /// can optionally include pointer types if the ScalarEvolution class
1887 /// has access to target-specific information.
1888 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
1889 // Integers are always SCEVable.
1890 if (Ty->isInteger())
1893 // Pointers are SCEVable if TargetData information is available
1894 // to provide pointer size information.
1895 if (isa<PointerType>(Ty))
1898 // Otherwise it's not SCEVable.
1902 /// getTypeSizeInBits - Return the size in bits of the specified type,
1903 /// for which isSCEVable must return true.
1904 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
1905 assert(isSCEVable(Ty) && "Type is not SCEVable!");
1907 // If we have a TargetData, use it!
1909 return TD->getTypeSizeInBits(Ty);
1911 // Otherwise, we support only integer types.
1912 assert(Ty->isInteger() && "isSCEVable permitted a non-SCEVable type!");
1913 return Ty->getPrimitiveSizeInBits();
1916 /// getEffectiveSCEVType - Return a type with the same bitwidth as
1917 /// the given type and which represents how SCEV will treat the given
1918 /// type, for which isSCEVable must return true. For pointer types,
1919 /// this is the pointer-sized integer type.
1920 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
1921 assert(isSCEVable(Ty) && "Type is not SCEVable!");
1923 if (Ty->isInteger())
1926 assert(isa<PointerType>(Ty) && "Unexpected non-pointer non-integer type!");
1927 return TD->getIntPtrType();
1930 const SCEV* ScalarEvolution::getCouldNotCompute() {
1931 return CouldNotCompute;
1934 /// hasSCEV - Return true if the SCEV for this value has already been
1936 bool ScalarEvolution::hasSCEV(Value *V) const {
1937 return Scalars.count(V);
1940 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1941 /// expression and create a new one.
1942 const SCEV* ScalarEvolution::getSCEV(Value *V) {
1943 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
1945 std::map<SCEVCallbackVH, const SCEV*>::iterator I = Scalars.find(V);
1946 if (I != Scalars.end()) return I->second;
1947 const SCEV* S = createSCEV(V);
1948 Scalars.insert(std::make_pair(SCEVCallbackVH(V, this), S));
1952 /// getIntegerSCEV - Given a SCEVable type, create a constant for the
1953 /// specified signed integer value and return a SCEV for the constant.
1954 const SCEV* ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
1955 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
1956 return getConstant(ConstantInt::get(ITy, Val));
1959 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
1961 const SCEV* ScalarEvolution::getNegativeSCEV(const SCEV* V) {
1962 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
1963 return getConstant(cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
1965 const Type *Ty = V->getType();
1966 Ty = getEffectiveSCEVType(Ty);
1967 return getMulExpr(V, getConstant(ConstantInt::getAllOnesValue(Ty)));
1970 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
1971 const SCEV* ScalarEvolution::getNotSCEV(const SCEV* V) {
1972 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
1973 return getConstant(cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
1975 const Type *Ty = V->getType();
1976 Ty = getEffectiveSCEVType(Ty);
1977 const SCEV* AllOnes = getConstant(ConstantInt::getAllOnesValue(Ty));
1978 return getMinusSCEV(AllOnes, V);
1981 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
1983 const SCEV* ScalarEvolution::getMinusSCEV(const SCEV* LHS,
1986 return getAddExpr(LHS, getNegativeSCEV(RHS));
1989 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
1990 /// input value to the specified type. If the type must be extended, it is zero
1993 ScalarEvolution::getTruncateOrZeroExtend(const SCEV* V,
1995 const Type *SrcTy = V->getType();
1996 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
1997 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
1998 "Cannot truncate or zero extend with non-integer arguments!");
1999 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2000 return V; // No conversion
2001 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2002 return getTruncateExpr(V, Ty);
2003 return getZeroExtendExpr(V, Ty);
2006 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2007 /// input value to the specified type. If the type must be extended, it is sign
2010 ScalarEvolution::getTruncateOrSignExtend(const SCEV* V,
2012 const Type *SrcTy = V->getType();
2013 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
2014 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
2015 "Cannot truncate or zero extend with non-integer arguments!");
2016 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2017 return V; // No conversion
2018 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2019 return getTruncateExpr(V, Ty);
2020 return getSignExtendExpr(V, Ty);
2023 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2024 /// input value to the specified type. If the type must be extended, it is zero
2025 /// extended. The conversion must not be narrowing.
2027 ScalarEvolution::getNoopOrZeroExtend(const SCEV* V, const Type *Ty) {
2028 const Type *SrcTy = V->getType();
2029 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
2030 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
2031 "Cannot noop or zero extend with non-integer arguments!");
2032 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2033 "getNoopOrZeroExtend cannot truncate!");
2034 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2035 return V; // No conversion
2036 return getZeroExtendExpr(V, Ty);
2039 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2040 /// input value to the specified type. If the type must be extended, it is sign
2041 /// extended. The conversion must not be narrowing.
2043 ScalarEvolution::getNoopOrSignExtend(const SCEV* V, const Type *Ty) {
2044 const Type *SrcTy = V->getType();
2045 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
2046 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
2047 "Cannot noop or sign extend with non-integer arguments!");
2048 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2049 "getNoopOrSignExtend cannot truncate!");
2050 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2051 return V; // No conversion
2052 return getSignExtendExpr(V, Ty);
2055 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2056 /// the input value to the specified type. If the type must be extended,
2057 /// it is extended with unspecified bits. The conversion must not be
2060 ScalarEvolution::getNoopOrAnyExtend(const SCEV* V, const Type *Ty) {
2061 const Type *SrcTy = V->getType();
2062 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
2063 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
2064 "Cannot noop or any extend with non-integer arguments!");
2065 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2066 "getNoopOrAnyExtend cannot truncate!");
2067 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2068 return V; // No conversion
2069 return getAnyExtendExpr(V, Ty);
2072 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2073 /// input value to the specified type. The conversion must not be widening.
2075 ScalarEvolution::getTruncateOrNoop(const SCEV* V, const Type *Ty) {
2076 const Type *SrcTy = V->getType();
2077 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
2078 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
2079 "Cannot truncate or noop with non-integer arguments!");
2080 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2081 "getTruncateOrNoop cannot extend!");
2082 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2083 return V; // No conversion
2084 return getTruncateExpr(V, Ty);
2087 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2088 /// the types using zero-extension, and then perform a umax operation
2090 const SCEV* ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV* LHS,
2092 const SCEV* PromotedLHS = LHS;
2093 const SCEV* PromotedRHS = RHS;
2095 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2096 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2098 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2100 return getUMaxExpr(PromotedLHS, PromotedRHS);
2103 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2104 /// the types using zero-extension, and then perform a umin operation
2106 const SCEV* ScalarEvolution::getUMinFromMismatchedTypes(const SCEV* LHS,
2108 const SCEV* PromotedLHS = LHS;
2109 const SCEV* PromotedRHS = RHS;
2111 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2112 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2114 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2116 return getUMinExpr(PromotedLHS, PromotedRHS);
2119 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value for
2120 /// the specified instruction and replaces any references to the symbolic value
2121 /// SymName with the specified value. This is used during PHI resolution.
2122 void ScalarEvolution::
2123 ReplaceSymbolicValueWithConcrete(Instruction *I, const SCEV* SymName,
2124 const SCEV* NewVal) {
2125 std::map<SCEVCallbackVH, const SCEV*>::iterator SI =
2126 Scalars.find(SCEVCallbackVH(I, this));
2127 if (SI == Scalars.end()) return;
2130 SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal, *this);
2131 if (NV == SI->second) return; // No change.
2133 SI->second = NV; // Update the scalars map!
2135 // Any instruction values that use this instruction might also need to be
2137 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
2139 ReplaceSymbolicValueWithConcrete(cast<Instruction>(*UI), SymName, NewVal);
2142 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2143 /// a loop header, making it a potential recurrence, or it doesn't.
2145 const SCEV* ScalarEvolution::createNodeForPHI(PHINode *PN) {
2146 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
2147 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2148 if (L->getHeader() == PN->getParent()) {
2149 // If it lives in the loop header, it has two incoming values, one
2150 // from outside the loop, and one from inside.
2151 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
2152 unsigned BackEdge = IncomingEdge^1;
2154 // While we are analyzing this PHI node, handle its value symbolically.
2155 const SCEV* SymbolicName = getUnknown(PN);
2156 assert(Scalars.find(PN) == Scalars.end() &&
2157 "PHI node already processed?");
2158 Scalars.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2160 // Using this symbolic name for the PHI, analyze the value coming around
2162 const SCEV* BEValue = getSCEV(PN->getIncomingValue(BackEdge));
2164 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2165 // has a special value for the first iteration of the loop.
2167 // If the value coming around the backedge is an add with the symbolic
2168 // value we just inserted, then we found a simple induction variable!
2169 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2170 // If there is a single occurrence of the symbolic value, replace it
2171 // with a recurrence.
2172 unsigned FoundIndex = Add->getNumOperands();
2173 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2174 if (Add->getOperand(i) == SymbolicName)
2175 if (FoundIndex == e) {
2180 if (FoundIndex != Add->getNumOperands()) {
2181 // Create an add with everything but the specified operand.
2182 SmallVector<const SCEV*, 8> Ops;
2183 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2184 if (i != FoundIndex)
2185 Ops.push_back(Add->getOperand(i));
2186 const SCEV* Accum = getAddExpr(Ops);
2188 // This is not a valid addrec if the step amount is varying each
2189 // loop iteration, but is not itself an addrec in this loop.
2190 if (Accum->isLoopInvariant(L) ||
2191 (isa<SCEVAddRecExpr>(Accum) &&
2192 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2193 const SCEV* StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
2194 const SCEV* PHISCEV = getAddRecExpr(StartVal, Accum, L);
2196 // Okay, for the entire analysis of this edge we assumed the PHI
2197 // to be symbolic. We now need to go back and update all of the
2198 // entries for the scalars that use the PHI (except for the PHI
2199 // itself) to use the new analyzed value instead of the "symbolic"
2201 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
2205 } else if (const SCEVAddRecExpr *AddRec =
2206 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2207 // Otherwise, this could be a loop like this:
2208 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2209 // In this case, j = {1,+,1} and BEValue is j.
2210 // Because the other in-value of i (0) fits the evolution of BEValue
2211 // i really is an addrec evolution.
2212 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2213 const SCEV* StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
2215 // If StartVal = j.start - j.stride, we can use StartVal as the
2216 // initial step of the addrec evolution.
2217 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2218 AddRec->getOperand(1))) {
2219 const SCEV* PHISCEV =
2220 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2222 // Okay, for the entire analysis of this edge we assumed the PHI
2223 // to be symbolic. We now need to go back and update all of the
2224 // entries for the scalars that use the PHI (except for the PHI
2225 // itself) to use the new analyzed value instead of the "symbolic"
2227 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
2233 return SymbolicName;
2236 // If it's not a loop phi, we can't handle it yet.
2237 return getUnknown(PN);
2240 /// createNodeForGEP - Expand GEP instructions into add and multiply
2241 /// operations. This allows them to be analyzed by regular SCEV code.
2243 const SCEV* ScalarEvolution::createNodeForGEP(User *GEP) {
2245 const Type *IntPtrTy = TD->getIntPtrType();
2246 Value *Base = GEP->getOperand(0);
2247 // Don't attempt to analyze GEPs over unsized objects.
2248 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2249 return getUnknown(GEP);
2250 const SCEV* TotalOffset = getIntegerSCEV(0, IntPtrTy);
2251 gep_type_iterator GTI = gep_type_begin(GEP);
2252 for (GetElementPtrInst::op_iterator I = next(GEP->op_begin()),
2256 // Compute the (potentially symbolic) offset in bytes for this index.
2257 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2258 // For a struct, add the member offset.
2259 const StructLayout &SL = *TD->getStructLayout(STy);
2260 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2261 uint64_t Offset = SL.getElementOffset(FieldNo);
2262 TotalOffset = getAddExpr(TotalOffset,
2263 getIntegerSCEV(Offset, IntPtrTy));
2265 // For an array, add the element offset, explicitly scaled.
2266 const SCEV* LocalOffset = getSCEV(Index);
2267 if (!isa<PointerType>(LocalOffset->getType()))
2268 // Getelementptr indicies are signed.
2269 LocalOffset = getTruncateOrSignExtend(LocalOffset,
2272 getMulExpr(LocalOffset,
2273 getIntegerSCEV(TD->getTypeAllocSize(*GTI),
2275 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2278 return getAddExpr(getSCEV(Base), TotalOffset);
2281 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2282 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2283 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2284 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2286 ScalarEvolution::GetMinTrailingZeros(const SCEV* S) {
2287 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2288 return C->getValue()->getValue().countTrailingZeros();
2290 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2291 return std::min(GetMinTrailingZeros(T->getOperand()),
2292 (uint32_t)getTypeSizeInBits(T->getType()));
2294 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2295 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2296 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2297 getTypeSizeInBits(E->getType()) : OpRes;
2300 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2301 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2302 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2303 getTypeSizeInBits(E->getType()) : OpRes;
2306 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2307 // The result is the min of all operands results.
2308 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2309 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2310 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2314 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2315 // The result is the sum of all operands results.
2316 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2317 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2318 for (unsigned i = 1, e = M->getNumOperands();
2319 SumOpRes != BitWidth && i != e; ++i)
2320 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2325 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2326 // The result is the min of all operands results.
2327 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2328 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2329 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2333 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2334 // The result is the min of all operands results.
2335 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2336 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2337 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2341 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2342 // The result is the min of all operands results.
2343 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2344 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2345 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2349 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2350 // For a SCEVUnknown, ask ValueTracking.
2351 unsigned BitWidth = getTypeSizeInBits(U->getType());
2352 APInt Mask = APInt::getAllOnesValue(BitWidth);
2353 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2354 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
2355 return Zeros.countTrailingOnes();
2363 ScalarEvolution::GetMinLeadingZeros(const SCEV* S) {
2364 // TODO: Handle other SCEV expression types here.
2366 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2367 return C->getValue()->getValue().countLeadingZeros();
2369 if (const SCEVZeroExtendExpr *C = dyn_cast<SCEVZeroExtendExpr>(S)) {
2370 // A zero-extension cast adds zero bits.
2371 return GetMinLeadingZeros(C->getOperand()) +
2372 (getTypeSizeInBits(C->getType()) -
2373 getTypeSizeInBits(C->getOperand()->getType()));
2376 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2377 // For a SCEVUnknown, ask ValueTracking.
2378 unsigned BitWidth = getTypeSizeInBits(U->getType());
2379 APInt Mask = APInt::getAllOnesValue(BitWidth);
2380 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2381 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
2382 return Zeros.countLeadingOnes();
2389 ScalarEvolution::GetMinSignBits(const SCEV* S) {
2390 // TODO: Handle other SCEV expression types here.
2392 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
2393 const APInt &A = C->getValue()->getValue();
2394 return A.isNegative() ? A.countLeadingOnes() :
2395 A.countLeadingZeros();
2398 if (const SCEVSignExtendExpr *C = dyn_cast<SCEVSignExtendExpr>(S)) {
2399 // A sign-extension cast adds sign bits.
2400 return GetMinSignBits(C->getOperand()) +
2401 (getTypeSizeInBits(C->getType()) -
2402 getTypeSizeInBits(C->getOperand()->getType()));
2405 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2406 unsigned BitWidth = getTypeSizeInBits(A->getType());
2408 // Special case decrementing a value (ADD X, -1):
2409 if (const SCEVConstant *CRHS = dyn_cast<SCEVConstant>(A->getOperand(0)))
2410 if (CRHS->isAllOnesValue()) {
2411 SmallVector<const SCEV *, 4> OtherOps(A->op_begin() + 1, A->op_end());
2412 const SCEV *OtherOpsAdd = getAddExpr(OtherOps);
2413 unsigned LZ = GetMinLeadingZeros(OtherOpsAdd);
2415 // If the input is known to be 0 or 1, the output is 0/-1, which is all
2417 if (LZ == BitWidth - 1)
2420 // If we are subtracting one from a positive number, there is no carry
2421 // out of the result.
2423 return GetMinSignBits(OtherOpsAdd);
2426 // Add can have at most one carry bit. Thus we know that the output
2427 // is, at worst, one more bit than the inputs.
2428 unsigned Min = BitWidth;
2429 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2430 unsigned N = GetMinSignBits(A->getOperand(i));
2431 Min = std::min(Min, N) - 1;
2432 if (Min == 0) return 1;
2437 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2438 // For a SCEVUnknown, ask ValueTracking.
2439 return ComputeNumSignBits(U->getValue(), TD);
2445 /// createSCEV - We know that there is no SCEV for the specified value.
2446 /// Analyze the expression.
2448 const SCEV* ScalarEvolution::createSCEV(Value *V) {
2449 if (!isSCEVable(V->getType()))
2450 return getUnknown(V);
2452 unsigned Opcode = Instruction::UserOp1;
2453 if (Instruction *I = dyn_cast<Instruction>(V))
2454 Opcode = I->getOpcode();
2455 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2456 Opcode = CE->getOpcode();
2457 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2458 return getConstant(CI);
2459 else if (isa<ConstantPointerNull>(V))
2460 return getIntegerSCEV(0, V->getType());
2461 else if (isa<UndefValue>(V))
2462 return getIntegerSCEV(0, V->getType());
2464 return getUnknown(V);
2466 User *U = cast<User>(V);
2468 case Instruction::Add:
2469 return getAddExpr(getSCEV(U->getOperand(0)),
2470 getSCEV(U->getOperand(1)));
2471 case Instruction::Mul:
2472 return getMulExpr(getSCEV(U->getOperand(0)),
2473 getSCEV(U->getOperand(1)));
2474 case Instruction::UDiv:
2475 return getUDivExpr(getSCEV(U->getOperand(0)),
2476 getSCEV(U->getOperand(1)));
2477 case Instruction::Sub:
2478 return getMinusSCEV(getSCEV(U->getOperand(0)),
2479 getSCEV(U->getOperand(1)));
2480 case Instruction::And:
2481 // For an expression like x&255 that merely masks off the high bits,
2482 // use zext(trunc(x)) as the SCEV expression.
2483 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2484 if (CI->isNullValue())
2485 return getSCEV(U->getOperand(1));
2486 if (CI->isAllOnesValue())
2487 return getSCEV(U->getOperand(0));
2488 const APInt &A = CI->getValue();
2490 // Instcombine's ShrinkDemandedConstant may strip bits out of
2491 // constants, obscuring what would otherwise be a low-bits mask.
2492 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
2493 // knew about to reconstruct a low-bits mask value.
2494 unsigned LZ = A.countLeadingZeros();
2495 unsigned BitWidth = A.getBitWidth();
2496 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
2497 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2498 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
2500 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
2502 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
2504 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
2505 IntegerType::get(BitWidth - LZ)),
2510 case Instruction::Or:
2511 // If the RHS of the Or is a constant, we may have something like:
2512 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
2513 // optimizations will transparently handle this case.
2515 // In order for this transformation to be safe, the LHS must be of the
2516 // form X*(2^n) and the Or constant must be less than 2^n.
2517 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2518 const SCEV* LHS = getSCEV(U->getOperand(0));
2519 const APInt &CIVal = CI->getValue();
2520 if (GetMinTrailingZeros(LHS) >=
2521 (CIVal.getBitWidth() - CIVal.countLeadingZeros()))
2522 return getAddExpr(LHS, getSCEV(U->getOperand(1)));
2525 case Instruction::Xor:
2526 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2527 // If the RHS of the xor is a signbit, then this is just an add.
2528 // Instcombine turns add of signbit into xor as a strength reduction step.
2529 if (CI->getValue().isSignBit())
2530 return getAddExpr(getSCEV(U->getOperand(0)),
2531 getSCEV(U->getOperand(1)));
2533 // If the RHS of xor is -1, then this is a not operation.
2534 if (CI->isAllOnesValue())
2535 return getNotSCEV(getSCEV(U->getOperand(0)));
2537 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
2538 // This is a variant of the check for xor with -1, and it handles
2539 // the case where instcombine has trimmed non-demanded bits out
2540 // of an xor with -1.
2541 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
2542 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
2543 if (BO->getOpcode() == Instruction::And &&
2544 LCI->getValue() == CI->getValue())
2545 if (const SCEVZeroExtendExpr *Z =
2546 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
2547 const Type *UTy = U->getType();
2548 const SCEV* Z0 = Z->getOperand();
2549 const Type *Z0Ty = Z0->getType();
2550 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
2552 // If C is a low-bits mask, the zero extend is zerving to
2553 // mask off the high bits. Complement the operand and
2554 // re-apply the zext.
2555 if (APIntOps::isMask(Z0TySize, CI->getValue()))
2556 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
2558 // If C is a single bit, it may be in the sign-bit position
2559 // before the zero-extend. In this case, represent the xor
2560 // using an add, which is equivalent, and re-apply the zext.
2561 APInt Trunc = APInt(CI->getValue()).trunc(Z0TySize);
2562 if (APInt(Trunc).zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
2564 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
2570 case Instruction::Shl:
2571 // Turn shift left of a constant amount into a multiply.
2572 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
2573 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
2574 Constant *X = ConstantInt::get(
2575 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
2576 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
2580 case Instruction::LShr:
2581 // Turn logical shift right of a constant into a unsigned divide.
2582 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
2583 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
2584 Constant *X = ConstantInt::get(
2585 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
2586 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
2590 case Instruction::AShr:
2591 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
2592 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
2593 if (Instruction *L = dyn_cast<Instruction>(U->getOperand(0)))
2594 if (L->getOpcode() == Instruction::Shl &&
2595 L->getOperand(1) == U->getOperand(1)) {
2596 unsigned BitWidth = getTypeSizeInBits(U->getType());
2597 uint64_t Amt = BitWidth - CI->getZExtValue();
2598 if (Amt == BitWidth)
2599 return getSCEV(L->getOperand(0)); // shift by zero --> noop
2601 return getIntegerSCEV(0, U->getType()); // value is undefined
2603 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
2604 IntegerType::get(Amt)),
2609 case Instruction::Trunc:
2610 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
2612 case Instruction::ZExt:
2613 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
2615 case Instruction::SExt:
2616 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
2618 case Instruction::BitCast:
2619 // BitCasts are no-op casts so we just eliminate the cast.
2620 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
2621 return getSCEV(U->getOperand(0));
2624 case Instruction::IntToPtr:
2625 if (!TD) break; // Without TD we can't analyze pointers.
2626 return getTruncateOrZeroExtend(getSCEV(U->getOperand(0)),
2627 TD->getIntPtrType());
2629 case Instruction::PtrToInt:
2630 if (!TD) break; // Without TD we can't analyze pointers.
2631 return getTruncateOrZeroExtend(getSCEV(U->getOperand(0)),
2634 case Instruction::GetElementPtr:
2635 if (!TD) break; // Without TD we can't analyze pointers.
2636 return createNodeForGEP(U);
2638 case Instruction::PHI:
2639 return createNodeForPHI(cast<PHINode>(U));
2641 case Instruction::Select:
2642 // This could be a smax or umax that was lowered earlier.
2643 // Try to recover it.
2644 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
2645 Value *LHS = ICI->getOperand(0);
2646 Value *RHS = ICI->getOperand(1);
2647 switch (ICI->getPredicate()) {
2648 case ICmpInst::ICMP_SLT:
2649 case ICmpInst::ICMP_SLE:
2650 std::swap(LHS, RHS);
2652 case ICmpInst::ICMP_SGT:
2653 case ICmpInst::ICMP_SGE:
2654 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
2655 return getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
2656 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
2657 return getSMinExpr(getSCEV(LHS), getSCEV(RHS));
2659 case ICmpInst::ICMP_ULT:
2660 case ICmpInst::ICMP_ULE:
2661 std::swap(LHS, RHS);
2663 case ICmpInst::ICMP_UGT:
2664 case ICmpInst::ICMP_UGE:
2665 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
2666 return getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
2667 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
2668 return getUMinExpr(getSCEV(LHS), getSCEV(RHS));
2670 case ICmpInst::ICMP_NE:
2671 // n != 0 ? n : 1 -> umax(n, 1)
2672 if (LHS == U->getOperand(1) &&
2673 isa<ConstantInt>(U->getOperand(2)) &&
2674 cast<ConstantInt>(U->getOperand(2))->isOne() &&
2675 isa<ConstantInt>(RHS) &&
2676 cast<ConstantInt>(RHS)->isZero())
2677 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(2)));
2679 case ICmpInst::ICMP_EQ:
2680 // n == 0 ? 1 : n -> umax(n, 1)
2681 if (LHS == U->getOperand(2) &&
2682 isa<ConstantInt>(U->getOperand(1)) &&
2683 cast<ConstantInt>(U->getOperand(1))->isOne() &&
2684 isa<ConstantInt>(RHS) &&
2685 cast<ConstantInt>(RHS)->isZero())
2686 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(1)));
2693 default: // We cannot analyze this expression.
2697 return getUnknown(V);
2702 //===----------------------------------------------------------------------===//
2703 // Iteration Count Computation Code
2706 /// getBackedgeTakenCount - If the specified loop has a predictable
2707 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
2708 /// object. The backedge-taken count is the number of times the loop header
2709 /// will be branched to from within the loop. This is one less than the
2710 /// trip count of the loop, since it doesn't count the first iteration,
2711 /// when the header is branched to from outside the loop.
2713 /// Note that it is not valid to call this method on a loop without a
2714 /// loop-invariant backedge-taken count (see
2715 /// hasLoopInvariantBackedgeTakenCount).
2717 const SCEV* ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
2718 return getBackedgeTakenInfo(L).Exact;
2721 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
2722 /// return the least SCEV value that is known never to be less than the
2723 /// actual backedge taken count.
2724 const SCEV* ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
2725 return getBackedgeTakenInfo(L).Max;
2728 const ScalarEvolution::BackedgeTakenInfo &
2729 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
2730 // Initially insert a CouldNotCompute for this loop. If the insertion
2731 // succeeds, procede to actually compute a backedge-taken count and
2732 // update the value. The temporary CouldNotCompute value tells SCEV
2733 // code elsewhere that it shouldn't attempt to request a new
2734 // backedge-taken count, which could result in infinite recursion.
2735 std::pair<std::map<const Loop*, BackedgeTakenInfo>::iterator, bool> Pair =
2736 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
2738 BackedgeTakenInfo ItCount = ComputeBackedgeTakenCount(L);
2739 if (ItCount.Exact != CouldNotCompute) {
2740 assert(ItCount.Exact->isLoopInvariant(L) &&
2741 ItCount.Max->isLoopInvariant(L) &&
2742 "Computed trip count isn't loop invariant for loop!");
2743 ++NumTripCountsComputed;
2745 // Update the value in the map.
2746 Pair.first->second = ItCount;
2748 if (ItCount.Max != CouldNotCompute)
2749 // Update the value in the map.
2750 Pair.first->second = ItCount;
2751 if (isa<PHINode>(L->getHeader()->begin()))
2752 // Only count loops that have phi nodes as not being computable.
2753 ++NumTripCountsNotComputed;
2756 // Now that we know more about the trip count for this loop, forget any
2757 // existing SCEV values for PHI nodes in this loop since they are only
2758 // conservative estimates made without the benefit
2759 // of trip count information.
2760 if (ItCount.hasAnyInfo())
2763 return Pair.first->second;
2766 /// forgetLoopBackedgeTakenCount - This method should be called by the
2767 /// client when it has changed a loop in a way that may effect
2768 /// ScalarEvolution's ability to compute a trip count, or if the loop
2770 void ScalarEvolution::forgetLoopBackedgeTakenCount(const Loop *L) {
2771 BackedgeTakenCounts.erase(L);
2775 /// forgetLoopPHIs - Delete the memoized SCEVs associated with the
2776 /// PHI nodes in the given loop. This is used when the trip count of
2777 /// the loop may have changed.
2778 void ScalarEvolution::forgetLoopPHIs(const Loop *L) {
2779 BasicBlock *Header = L->getHeader();
2781 // Push all Loop-header PHIs onto the Worklist stack, except those
2782 // that are presently represented via a SCEVUnknown. SCEVUnknown for
2783 // a PHI either means that it has an unrecognized structure, or it's
2784 // a PHI that's in the progress of being computed by createNodeForPHI.
2785 // In the former case, additional loop trip count information isn't
2786 // going to change anything. In the later case, createNodeForPHI will
2787 // perform the necessary updates on its own when it gets to that point.
2788 SmallVector<Instruction *, 16> Worklist;
2789 for (BasicBlock::iterator I = Header->begin();
2790 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2791 std::map<SCEVCallbackVH, const SCEV*>::iterator It = Scalars.find((Value*)I);
2792 if (It != Scalars.end() && !isa<SCEVUnknown>(It->second))
2793 Worklist.push_back(PN);
2796 while (!Worklist.empty()) {
2797 Instruction *I = Worklist.pop_back_val();
2798 if (Scalars.erase(I))
2799 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2801 Worklist.push_back(cast<Instruction>(UI));
2805 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
2806 /// of the specified loop will execute.
2807 ScalarEvolution::BackedgeTakenInfo
2808 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
2809 SmallVector<BasicBlock*, 8> ExitingBlocks;
2810 L->getExitingBlocks(ExitingBlocks);
2812 // Examine all exits and pick the most conservative values.
2813 const SCEV* BECount = CouldNotCompute;
2814 const SCEV* MaxBECount = CouldNotCompute;
2815 bool CouldNotComputeBECount = false;
2816 bool CouldNotComputeMaxBECount = false;
2817 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2818 BackedgeTakenInfo NewBTI =
2819 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
2821 if (NewBTI.Exact == CouldNotCompute) {
2822 // We couldn't compute an exact value for this exit, so
2823 // we won't be able to compute an exact value for the loop.
2824 CouldNotComputeBECount = true;
2825 BECount = CouldNotCompute;
2826 } else if (!CouldNotComputeBECount) {
2827 if (BECount == CouldNotCompute)
2828 BECount = NewBTI.Exact;
2830 // TODO: More analysis could be done here. For example, a
2831 // loop with a short-circuiting && operator has an exact count
2832 // of the min of both sides.
2833 CouldNotComputeBECount = true;
2834 BECount = CouldNotCompute;
2837 if (NewBTI.Max == CouldNotCompute) {
2838 // We couldn't compute an maximum value for this exit, so
2839 // we won't be able to compute an maximum value for the loop.
2840 CouldNotComputeMaxBECount = true;
2841 MaxBECount = CouldNotCompute;
2842 } else if (!CouldNotComputeMaxBECount) {
2843 if (MaxBECount == CouldNotCompute)
2844 MaxBECount = NewBTI.Max;
2846 MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, NewBTI.Max);
2850 return BackedgeTakenInfo(BECount, MaxBECount);
2853 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
2854 /// of the specified loop will execute if it exits via the specified block.
2855 ScalarEvolution::BackedgeTakenInfo
2856 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
2857 BasicBlock *ExitingBlock) {
2859 // Okay, we've chosen an exiting block. See what condition causes us to
2860 // exit at this block.
2862 // FIXME: we should be able to handle switch instructions (with a single exit)
2863 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2864 if (ExitBr == 0) return CouldNotCompute;
2865 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
2867 // At this point, we know we have a conditional branch that determines whether
2868 // the loop is exited. However, we don't know if the branch is executed each
2869 // time through the loop. If not, then the execution count of the branch will
2870 // not be equal to the trip count of the loop.
2872 // Currently we check for this by checking to see if the Exit branch goes to
2873 // the loop header. If so, we know it will always execute the same number of
2874 // times as the loop. We also handle the case where the exit block *is* the
2875 // loop header. This is common for un-rotated loops.
2877 // If both of those tests fail, walk up the unique predecessor chain to the
2878 // header, stopping if there is an edge that doesn't exit the loop. If the
2879 // header is reached, the execution count of the branch will be equal to the
2880 // trip count of the loop.
2882 // More extensive analysis could be done to handle more cases here.
2884 if (ExitBr->getSuccessor(0) != L->getHeader() &&
2885 ExitBr->getSuccessor(1) != L->getHeader() &&
2886 ExitBr->getParent() != L->getHeader()) {
2887 // The simple checks failed, try climbing the unique predecessor chain
2888 // up to the header.
2890 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
2891 BasicBlock *Pred = BB->getUniquePredecessor();
2893 return CouldNotCompute;
2894 TerminatorInst *PredTerm = Pred->getTerminator();
2895 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
2896 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
2899 // If the predecessor has a successor that isn't BB and isn't
2900 // outside the loop, assume the worst.
2901 if (L->contains(PredSucc))
2902 return CouldNotCompute;
2904 if (Pred == L->getHeader()) {
2911 return CouldNotCompute;
2914 // Procede to the next level to examine the exit condition expression.
2915 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
2916 ExitBr->getSuccessor(0),
2917 ExitBr->getSuccessor(1));
2920 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
2921 /// backedge of the specified loop will execute if its exit condition
2922 /// were a conditional branch of ExitCond, TBB, and FBB.
2923 ScalarEvolution::BackedgeTakenInfo
2924 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
2928 // Check if the controlling expression for this loop is an and or or. In
2929 // such cases, an exact backedge-taken count may be infeasible, but a
2930 // maximum count may still be feasible.
2931 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
2932 if (BO->getOpcode() == Instruction::And) {
2933 // Recurse on the operands of the and.
2934 BackedgeTakenInfo BTI0 =
2935 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
2936 BackedgeTakenInfo BTI1 =
2937 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
2938 const SCEV* BECount = CouldNotCompute;
2939 const SCEV* MaxBECount = CouldNotCompute;
2940 if (L->contains(TBB)) {
2941 // Both conditions must be true for the loop to continue executing.
2942 // Choose the less conservative count.
2943 if (BTI0.Exact == CouldNotCompute || BTI1.Exact == CouldNotCompute)
2944 BECount = CouldNotCompute;
2946 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
2947 if (BTI0.Max == CouldNotCompute)
2948 MaxBECount = BTI1.Max;
2949 else if (BTI1.Max == CouldNotCompute)
2950 MaxBECount = BTI0.Max;
2952 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
2954 // Both conditions must be true for the loop to exit.
2955 assert(L->contains(FBB) && "Loop block has no successor in loop!");
2956 if (BTI0.Exact != CouldNotCompute && BTI1.Exact != CouldNotCompute)
2957 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
2958 if (BTI0.Max != CouldNotCompute && BTI1.Max != CouldNotCompute)
2959 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
2962 return BackedgeTakenInfo(BECount, MaxBECount);
2964 if (BO->getOpcode() == Instruction::Or) {
2965 // Recurse on the operands of the or.
2966 BackedgeTakenInfo BTI0 =
2967 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
2968 BackedgeTakenInfo BTI1 =
2969 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
2970 const SCEV* BECount = CouldNotCompute;
2971 const SCEV* MaxBECount = CouldNotCompute;
2972 if (L->contains(FBB)) {
2973 // Both conditions must be false for the loop to continue executing.
2974 // Choose the less conservative count.
2975 if (BTI0.Exact == CouldNotCompute || BTI1.Exact == CouldNotCompute)
2976 BECount = CouldNotCompute;
2978 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
2979 if (BTI0.Max == CouldNotCompute)
2980 MaxBECount = BTI1.Max;
2981 else if (BTI1.Max == CouldNotCompute)
2982 MaxBECount = BTI0.Max;
2984 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
2986 // Both conditions must be false for the loop to exit.
2987 assert(L->contains(TBB) && "Loop block has no successor in loop!");
2988 if (BTI0.Exact != CouldNotCompute && BTI1.Exact != CouldNotCompute)
2989 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
2990 if (BTI0.Max != CouldNotCompute && BTI1.Max != CouldNotCompute)
2991 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
2994 return BackedgeTakenInfo(BECount, MaxBECount);
2998 // With an icmp, it may be feasible to compute an exact backedge-taken count.
2999 // Procede to the next level to examine the icmp.
3000 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
3001 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
3003 // If it's not an integer or pointer comparison then compute it the hard way.
3004 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3007 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
3008 /// backedge of the specified loop will execute if its exit condition
3009 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
3010 ScalarEvolution::BackedgeTakenInfo
3011 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
3016 // If the condition was exit on true, convert the condition to exit on false
3017 ICmpInst::Predicate Cond;
3018 if (!L->contains(FBB))
3019 Cond = ExitCond->getPredicate();
3021 Cond = ExitCond->getInversePredicate();
3023 // Handle common loops like: for (X = "string"; *X; ++X)
3024 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
3025 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
3027 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
3028 if (!isa<SCEVCouldNotCompute>(ItCnt)) {
3029 unsigned BitWidth = getTypeSizeInBits(ItCnt->getType());
3030 return BackedgeTakenInfo(ItCnt,
3031 isa<SCEVConstant>(ItCnt) ? ItCnt :
3032 getConstant(APInt::getMaxValue(BitWidth)-1));
3036 const SCEV* LHS = getSCEV(ExitCond->getOperand(0));
3037 const SCEV* RHS = getSCEV(ExitCond->getOperand(1));
3039 // Try to evaluate any dependencies out of the loop.
3040 LHS = getSCEVAtScope(LHS, L);
3041 RHS = getSCEVAtScope(RHS, L);
3043 // At this point, we would like to compute how many iterations of the
3044 // loop the predicate will return true for these inputs.
3045 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
3046 // If there is a loop-invariant, force it into the RHS.
3047 std::swap(LHS, RHS);
3048 Cond = ICmpInst::getSwappedPredicate(Cond);
3051 // If we have a comparison of a chrec against a constant, try to use value
3052 // ranges to answer this query.
3053 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
3054 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
3055 if (AddRec->getLoop() == L) {
3056 // Form the constant range.
3057 ConstantRange CompRange(
3058 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
3060 const SCEV* Ret = AddRec->getNumIterationsInRange(CompRange, *this);
3061 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
3065 case ICmpInst::ICMP_NE: { // while (X != Y)
3066 // Convert to: while (X-Y != 0)
3067 const SCEV* TC = HowFarToZero(getMinusSCEV(LHS, RHS), L);
3068 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3071 case ICmpInst::ICMP_EQ: {
3072 // Convert to: while (X-Y == 0) // while (X == Y)
3073 const SCEV* TC = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
3074 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3077 case ICmpInst::ICMP_SLT: {
3078 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
3079 if (BTI.hasAnyInfo()) return BTI;
3082 case ICmpInst::ICMP_SGT: {
3083 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3084 getNotSCEV(RHS), L, true);
3085 if (BTI.hasAnyInfo()) return BTI;
3088 case ICmpInst::ICMP_ULT: {
3089 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
3090 if (BTI.hasAnyInfo()) return BTI;
3093 case ICmpInst::ICMP_UGT: {
3094 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3095 getNotSCEV(RHS), L, false);
3096 if (BTI.hasAnyInfo()) return BTI;
3101 errs() << "ComputeBackedgeTakenCount ";
3102 if (ExitCond->getOperand(0)->getType()->isUnsigned())
3103 errs() << "[unsigned] ";
3104 errs() << *LHS << " "
3105 << Instruction::getOpcodeName(Instruction::ICmp)
3106 << " " << *RHS << "\n";
3111 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3114 static ConstantInt *
3115 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
3116 ScalarEvolution &SE) {
3117 const SCEV* InVal = SE.getConstant(C);
3118 const SCEV* Val = AddRec->evaluateAtIteration(InVal, SE);
3119 assert(isa<SCEVConstant>(Val) &&
3120 "Evaluation of SCEV at constant didn't fold correctly?");
3121 return cast<SCEVConstant>(Val)->getValue();
3124 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
3125 /// and a GEP expression (missing the pointer index) indexing into it, return
3126 /// the addressed element of the initializer or null if the index expression is
3129 GetAddressedElementFromGlobal(GlobalVariable *GV,
3130 const std::vector<ConstantInt*> &Indices) {
3131 Constant *Init = GV->getInitializer();
3132 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
3133 uint64_t Idx = Indices[i]->getZExtValue();
3134 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
3135 assert(Idx < CS->getNumOperands() && "Bad struct index!");
3136 Init = cast<Constant>(CS->getOperand(Idx));
3137 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
3138 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
3139 Init = cast<Constant>(CA->getOperand(Idx));
3140 } else if (isa<ConstantAggregateZero>(Init)) {
3141 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
3142 assert(Idx < STy->getNumElements() && "Bad struct index!");
3143 Init = Constant::getNullValue(STy->getElementType(Idx));
3144 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
3145 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
3146 Init = Constant::getNullValue(ATy->getElementType());
3148 assert(0 && "Unknown constant aggregate type!");
3152 return 0; // Unknown initializer type
3158 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
3159 /// 'icmp op load X, cst', try to see if we can compute the backedge
3160 /// execution count.
3161 const SCEV* ScalarEvolution::
3162 ComputeLoadConstantCompareBackedgeTakenCount(LoadInst *LI, Constant *RHS,
3164 ICmpInst::Predicate predicate) {
3165 if (LI->isVolatile()) return CouldNotCompute;
3167 // Check to see if the loaded pointer is a getelementptr of a global.
3168 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
3169 if (!GEP) return CouldNotCompute;
3171 // Make sure that it is really a constant global we are gepping, with an
3172 // initializer, and make sure the first IDX is really 0.
3173 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
3174 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
3175 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
3176 !cast<Constant>(GEP->getOperand(1))->isNullValue())
3177 return CouldNotCompute;
3179 // Okay, we allow one non-constant index into the GEP instruction.
3181 std::vector<ConstantInt*> Indexes;
3182 unsigned VarIdxNum = 0;
3183 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
3184 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
3185 Indexes.push_back(CI);
3186 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
3187 if (VarIdx) return CouldNotCompute; // Multiple non-constant idx's.
3188 VarIdx = GEP->getOperand(i);
3190 Indexes.push_back(0);
3193 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
3194 // Check to see if X is a loop variant variable value now.
3195 const SCEV* Idx = getSCEV(VarIdx);
3196 Idx = getSCEVAtScope(Idx, L);
3198 // We can only recognize very limited forms of loop index expressions, in
3199 // particular, only affine AddRec's like {C1,+,C2}.
3200 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
3201 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
3202 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
3203 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
3204 return CouldNotCompute;
3206 unsigned MaxSteps = MaxBruteForceIterations;
3207 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
3208 ConstantInt *ItCst =
3209 ConstantInt::get(cast<IntegerType>(IdxExpr->getType()), IterationNum);
3210 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
3212 // Form the GEP offset.
3213 Indexes[VarIdxNum] = Val;
3215 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
3216 if (Result == 0) break; // Cannot compute!
3218 // Evaluate the condition for this iteration.
3219 Result = ConstantExpr::getICmp(predicate, Result, RHS);
3220 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
3221 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
3223 errs() << "\n***\n*** Computed loop count " << *ItCst
3224 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
3227 ++NumArrayLenItCounts;
3228 return getConstant(ItCst); // Found terminating iteration!
3231 return CouldNotCompute;
3235 /// CanConstantFold - Return true if we can constant fold an instruction of the
3236 /// specified type, assuming that all operands were constants.
3237 static bool CanConstantFold(const Instruction *I) {
3238 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
3239 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
3242 if (const CallInst *CI = dyn_cast<CallInst>(I))
3243 if (const Function *F = CI->getCalledFunction())
3244 return canConstantFoldCallTo(F);
3248 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
3249 /// in the loop that V is derived from. We allow arbitrary operations along the
3250 /// way, but the operands of an operation must either be constants or a value
3251 /// derived from a constant PHI. If this expression does not fit with these
3252 /// constraints, return null.
3253 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
3254 // If this is not an instruction, or if this is an instruction outside of the
3255 // loop, it can't be derived from a loop PHI.
3256 Instruction *I = dyn_cast<Instruction>(V);
3257 if (I == 0 || !L->contains(I->getParent())) return 0;
3259 if (PHINode *PN = dyn_cast<PHINode>(I)) {
3260 if (L->getHeader() == I->getParent())
3263 // We don't currently keep track of the control flow needed to evaluate
3264 // PHIs, so we cannot handle PHIs inside of loops.
3268 // If we won't be able to constant fold this expression even if the operands
3269 // are constants, return early.
3270 if (!CanConstantFold(I)) return 0;
3272 // Otherwise, we can evaluate this instruction if all of its operands are
3273 // constant or derived from a PHI node themselves.
3275 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
3276 if (!(isa<Constant>(I->getOperand(Op)) ||
3277 isa<GlobalValue>(I->getOperand(Op)))) {
3278 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
3279 if (P == 0) return 0; // Not evolving from PHI
3283 return 0; // Evolving from multiple different PHIs.
3286 // This is a expression evolving from a constant PHI!
3290 /// EvaluateExpression - Given an expression that passes the
3291 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
3292 /// in the loop has the value PHIVal. If we can't fold this expression for some
3293 /// reason, return null.
3294 static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
3295 if (isa<PHINode>(V)) return PHIVal;
3296 if (Constant *C = dyn_cast<Constant>(V)) return C;
3297 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
3298 Instruction *I = cast<Instruction>(V);
3300 std::vector<Constant*> Operands;
3301 Operands.resize(I->getNumOperands());
3303 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3304 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal);
3305 if (Operands[i] == 0) return 0;
3308 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
3309 return ConstantFoldCompareInstOperands(CI->getPredicate(),
3310 &Operands[0], Operands.size());
3312 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
3313 &Operands[0], Operands.size());
3316 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
3317 /// in the header of its containing loop, we know the loop executes a
3318 /// constant number of times, and the PHI node is just a recurrence
3319 /// involving constants, fold it.
3320 Constant *ScalarEvolution::
3321 getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& BEs, const Loop *L){
3322 std::map<PHINode*, Constant*>::iterator I =
3323 ConstantEvolutionLoopExitValue.find(PN);
3324 if (I != ConstantEvolutionLoopExitValue.end())
3327 if (BEs.ugt(APInt(BEs.getBitWidth(),MaxBruteForceIterations)))
3328 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
3330 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
3332 // Since the loop is canonicalized, the PHI node must have two entries. One
3333 // entry must be a constant (coming in from outside of the loop), and the
3334 // second must be derived from the same PHI.
3335 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
3336 Constant *StartCST =
3337 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
3339 return RetVal = 0; // Must be a constant.
3341 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
3342 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
3344 return RetVal = 0; // Not derived from same PHI.
3346 // Execute the loop symbolically to determine the exit value.
3347 if (BEs.getActiveBits() >= 32)
3348 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
3350 unsigned NumIterations = BEs.getZExtValue(); // must be in range
3351 unsigned IterationNum = 0;
3352 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
3353 if (IterationNum == NumIterations)
3354 return RetVal = PHIVal; // Got exit value!
3356 // Compute the value of the PHI node for the next iteration.
3357 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
3358 if (NextPHI == PHIVal)
3359 return RetVal = NextPHI; // Stopped evolving!
3361 return 0; // Couldn't evaluate!
3366 /// ComputeBackedgeTakenCountExhaustively - If the trip is known to execute a
3367 /// constant number of times (the condition evolves only from constants),
3368 /// try to evaluate a few iterations of the loop until we get the exit
3369 /// condition gets a value of ExitWhen (true or false). If we cannot
3370 /// evaluate the trip count of the loop, return CouldNotCompute.
3371 const SCEV* ScalarEvolution::
3372 ComputeBackedgeTakenCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) {
3373 PHINode *PN = getConstantEvolvingPHI(Cond, L);
3374 if (PN == 0) return CouldNotCompute;
3376 // Since the loop is canonicalized, the PHI node must have two entries. One
3377 // entry must be a constant (coming in from outside of the loop), and the
3378 // second must be derived from the same PHI.
3379 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
3380 Constant *StartCST =
3381 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
3382 if (StartCST == 0) return CouldNotCompute; // Must be a constant.
3384 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
3385 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
3386 if (PN2 != PN) return CouldNotCompute; // Not derived from same PHI.
3388 // Okay, we find a PHI node that defines the trip count of this loop. Execute
3389 // the loop symbolically to determine when the condition gets a value of
3391 unsigned IterationNum = 0;
3392 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
3393 for (Constant *PHIVal = StartCST;
3394 IterationNum != MaxIterations; ++IterationNum) {
3395 ConstantInt *CondVal =
3396 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal));
3398 // Couldn't symbolically evaluate.
3399 if (!CondVal) return CouldNotCompute;
3401 if (CondVal->getValue() == uint64_t(ExitWhen)) {
3402 ConstantEvolutionLoopExitValue[PN] = PHIVal;
3403 ++NumBruteForceTripCountsComputed;
3404 return getConstant(Type::Int32Ty, IterationNum);
3407 // Compute the value of the PHI node for the next iteration.
3408 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
3409 if (NextPHI == 0 || NextPHI == PHIVal)
3410 return CouldNotCompute; // Couldn't evaluate or not making progress...
3414 // Too many iterations were needed to evaluate.
3415 return CouldNotCompute;
3418 /// getSCEVAtScope - Return a SCEV expression handle for the specified value
3419 /// at the specified scope in the program. The L value specifies a loop
3420 /// nest to evaluate the expression at, where null is the top-level or a
3421 /// specified loop is immediately inside of the loop.
3423 /// This method can be used to compute the exit value for a variable defined
3424 /// in a loop by querying what the value will hold in the parent loop.
3426 /// In the case that a relevant loop exit value cannot be computed, the
3427 /// original value V is returned.
3428 const SCEV* ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
3429 // FIXME: this should be turned into a virtual method on SCEV!
3431 if (isa<SCEVConstant>(V)) return V;
3433 // If this instruction is evolved from a constant-evolving PHI, compute the
3434 // exit value from the loop without using SCEVs.
3435 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
3436 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
3437 const Loop *LI = (*this->LI)[I->getParent()];
3438 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
3439 if (PHINode *PN = dyn_cast<PHINode>(I))
3440 if (PN->getParent() == LI->getHeader()) {
3441 // Okay, there is no closed form solution for the PHI node. Check
3442 // to see if the loop that contains it has a known backedge-taken
3443 // count. If so, we may be able to force computation of the exit
3445 const SCEV* BackedgeTakenCount = getBackedgeTakenCount(LI);
3446 if (const SCEVConstant *BTCC =
3447 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
3448 // Okay, we know how many times the containing loop executes. If
3449 // this is a constant evolving PHI node, get the final value at
3450 // the specified iteration number.
3451 Constant *RV = getConstantEvolutionLoopExitValue(PN,
3452 BTCC->getValue()->getValue(),
3454 if (RV) return getUnknown(RV);
3458 // Okay, this is an expression that we cannot symbolically evaluate
3459 // into a SCEV. Check to see if it's possible to symbolically evaluate
3460 // the arguments into constants, and if so, try to constant propagate the
3461 // result. This is particularly useful for computing loop exit values.
3462 if (CanConstantFold(I)) {
3463 // Check to see if we've folded this instruction at this loop before.
3464 std::map<const Loop *, Constant *> &Values = ValuesAtScopes[I];
3465 std::pair<std::map<const Loop *, Constant *>::iterator, bool> Pair =
3466 Values.insert(std::make_pair(L, static_cast<Constant *>(0)));
3468 return Pair.first->second ? &*getUnknown(Pair.first->second) : V;
3470 std::vector<Constant*> Operands;
3471 Operands.reserve(I->getNumOperands());
3472 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3473 Value *Op = I->getOperand(i);
3474 if (Constant *C = dyn_cast<Constant>(Op)) {
3475 Operands.push_back(C);
3477 // If any of the operands is non-constant and if they are
3478 // non-integer and non-pointer, don't even try to analyze them
3479 // with scev techniques.
3480 if (!isSCEVable(Op->getType()))
3483 const SCEV* OpV = getSCEVAtScope(getSCEV(Op), L);
3484 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
3485 Constant *C = SC->getValue();
3486 if (C->getType() != Op->getType())
3487 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3491 Operands.push_back(C);
3492 } else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
3493 if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
3494 if (C->getType() != Op->getType())
3496 ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3500 Operands.push_back(C);
3510 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
3511 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
3512 &Operands[0], Operands.size());
3514 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
3515 &Operands[0], Operands.size());
3516 Pair.first->second = C;
3517 return getUnknown(C);
3521 // This is some other type of SCEVUnknown, just return it.
3525 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
3526 // Avoid performing the look-up in the common case where the specified
3527 // expression has no loop-variant portions.
3528 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
3529 const SCEV* OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
3530 if (OpAtScope != Comm->getOperand(i)) {
3531 // Okay, at least one of these operands is loop variant but might be
3532 // foldable. Build a new instance of the folded commutative expression.
3533 SmallVector<const SCEV*, 8> NewOps(Comm->op_begin(), Comm->op_begin()+i);
3534 NewOps.push_back(OpAtScope);
3536 for (++i; i != e; ++i) {
3537 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
3538 NewOps.push_back(OpAtScope);
3540 if (isa<SCEVAddExpr>(Comm))
3541 return getAddExpr(NewOps);
3542 if (isa<SCEVMulExpr>(Comm))
3543 return getMulExpr(NewOps);
3544 if (isa<SCEVSMaxExpr>(Comm))
3545 return getSMaxExpr(NewOps);
3546 if (isa<SCEVUMaxExpr>(Comm))
3547 return getUMaxExpr(NewOps);
3548 assert(0 && "Unknown commutative SCEV type!");
3551 // If we got here, all operands are loop invariant.
3555 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
3556 const SCEV* LHS = getSCEVAtScope(Div->getLHS(), L);
3557 const SCEV* RHS = getSCEVAtScope(Div->getRHS(), L);
3558 if (LHS == Div->getLHS() && RHS == Div->getRHS())
3559 return Div; // must be loop invariant
3560 return getUDivExpr(LHS, RHS);
3563 // If this is a loop recurrence for a loop that does not contain L, then we
3564 // are dealing with the final value computed by the loop.
3565 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
3566 if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
3567 // To evaluate this recurrence, we need to know how many times the AddRec
3568 // loop iterates. Compute this now.
3569 const SCEV* BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
3570 if (BackedgeTakenCount == CouldNotCompute) return AddRec;
3572 // Then, evaluate the AddRec.
3573 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
3578 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
3579 const SCEV* Op = getSCEVAtScope(Cast->getOperand(), L);
3580 if (Op == Cast->getOperand())
3581 return Cast; // must be loop invariant
3582 return getZeroExtendExpr(Op, Cast->getType());
3585 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
3586 const SCEV* Op = getSCEVAtScope(Cast->getOperand(), L);
3587 if (Op == Cast->getOperand())
3588 return Cast; // must be loop invariant
3589 return getSignExtendExpr(Op, Cast->getType());
3592 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
3593 const SCEV* Op = getSCEVAtScope(Cast->getOperand(), L);
3594 if (Op == Cast->getOperand())
3595 return Cast; // must be loop invariant
3596 return getTruncateExpr(Op, Cast->getType());
3599 assert(0 && "Unknown SCEV type!");
3603 /// getSCEVAtScope - This is a convenience function which does
3604 /// getSCEVAtScope(getSCEV(V), L).
3605 const SCEV* ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
3606 return getSCEVAtScope(getSCEV(V), L);
3609 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
3610 /// following equation:
3612 /// A * X = B (mod N)
3614 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
3615 /// A and B isn't important.
3617 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
3618 static const SCEV* SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
3619 ScalarEvolution &SE) {
3620 uint32_t BW = A.getBitWidth();
3621 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
3622 assert(A != 0 && "A must be non-zero.");
3626 // The gcd of A and N may have only one prime factor: 2. The number of
3627 // trailing zeros in A is its multiplicity
3628 uint32_t Mult2 = A.countTrailingZeros();
3631 // 2. Check if B is divisible by D.
3633 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
3634 // is not less than multiplicity of this prime factor for D.
3635 if (B.countTrailingZeros() < Mult2)
3636 return SE.getCouldNotCompute();
3638 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
3641 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
3642 // bit width during computations.
3643 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
3644 APInt Mod(BW + 1, 0);
3645 Mod.set(BW - Mult2); // Mod = N / D
3646 APInt I = AD.multiplicativeInverse(Mod);
3648 // 4. Compute the minimum unsigned root of the equation:
3649 // I * (B / D) mod (N / D)
3650 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
3652 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
3654 return SE.getConstant(Result.trunc(BW));
3657 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
3658 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
3659 /// might be the same) or two SCEVCouldNotCompute objects.
3661 static std::pair<const SCEV*,const SCEV*>
3662 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
3663 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
3664 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
3665 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
3666 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
3668 // We currently can only solve this if the coefficients are constants.
3669 if (!LC || !MC || !NC) {
3670 const SCEV *CNC = SE.getCouldNotCompute();
3671 return std::make_pair(CNC, CNC);
3674 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
3675 const APInt &L = LC->getValue()->getValue();
3676 const APInt &M = MC->getValue()->getValue();
3677 const APInt &N = NC->getValue()->getValue();
3678 APInt Two(BitWidth, 2);
3679 APInt Four(BitWidth, 4);
3682 using namespace APIntOps;
3684 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
3685 // The B coefficient is M-N/2
3689 // The A coefficient is N/2
3690 APInt A(N.sdiv(Two));
3692 // Compute the B^2-4ac term.
3695 SqrtTerm -= Four * (A * C);
3697 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
3698 // integer value or else APInt::sqrt() will assert.
3699 APInt SqrtVal(SqrtTerm.sqrt());
3701 // Compute the two solutions for the quadratic formula.
3702 // The divisions must be performed as signed divisions.
3704 APInt TwoA( A << 1 );
3705 if (TwoA.isMinValue()) {
3706 const SCEV *CNC = SE.getCouldNotCompute();
3707 return std::make_pair(CNC, CNC);
3710 ConstantInt *Solution1 = ConstantInt::get((NegB + SqrtVal).sdiv(TwoA));
3711 ConstantInt *Solution2 = ConstantInt::get((NegB - SqrtVal).sdiv(TwoA));
3713 return std::make_pair(SE.getConstant(Solution1),
3714 SE.getConstant(Solution2));
3715 } // end APIntOps namespace
3718 /// HowFarToZero - Return the number of times a backedge comparing the specified
3719 /// value to zero will execute. If not computable, return CouldNotCompute.
3720 const SCEV* ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
3721 // If the value is a constant
3722 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
3723 // If the value is already zero, the branch will execute zero times.
3724 if (C->getValue()->isZero()) return C;
3725 return CouldNotCompute; // Otherwise it will loop infinitely.
3728 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
3729 if (!AddRec || AddRec->getLoop() != L)
3730 return CouldNotCompute;
3732 if (AddRec->isAffine()) {
3733 // If this is an affine expression, the execution count of this branch is
3734 // the minimum unsigned root of the following equation:
3736 // Start + Step*N = 0 (mod 2^BW)
3740 // Step*N = -Start (mod 2^BW)
3742 // where BW is the common bit width of Start and Step.
3744 // Get the initial value for the loop.
3745 const SCEV* Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
3746 const SCEV* Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
3748 if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
3749 // For now we handle only constant steps.
3751 // First, handle unitary steps.
3752 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
3753 return getNegativeSCEV(Start); // N = -Start (as unsigned)
3754 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
3755 return Start; // N = Start (as unsigned)
3757 // Then, try to solve the above equation provided that Start is constant.
3758 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
3759 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
3760 -StartC->getValue()->getValue(),
3763 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
3764 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
3765 // the quadratic equation to solve it.
3766 std::pair<const SCEV*,const SCEV*> Roots = SolveQuadraticEquation(AddRec,
3768 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
3769 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
3772 errs() << "HFTZ: " << *V << " - sol#1: " << *R1
3773 << " sol#2: " << *R2 << "\n";
3775 // Pick the smallest positive root value.
3776 if (ConstantInt *CB =
3777 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
3778 R1->getValue(), R2->getValue()))) {
3779 if (CB->getZExtValue() == false)
3780 std::swap(R1, R2); // R1 is the minimum root now.
3782 // We can only use this value if the chrec ends up with an exact zero
3783 // value at this index. When solving for "X*X != 5", for example, we
3784 // should not accept a root of 2.
3785 const SCEV* Val = AddRec->evaluateAtIteration(R1, *this);
3787 return R1; // We found a quadratic root!
3792 return CouldNotCompute;
3795 /// HowFarToNonZero - Return the number of times a backedge checking the
3796 /// specified value for nonzero will execute. If not computable, return
3798 const SCEV* ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
3799 // Loops that look like: while (X == 0) are very strange indeed. We don't
3800 // handle them yet except for the trivial case. This could be expanded in the
3801 // future as needed.
3803 // If the value is a constant, check to see if it is known to be non-zero
3804 // already. If so, the backedge will execute zero times.
3805 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
3806 if (!C->getValue()->isNullValue())
3807 return getIntegerSCEV(0, C->getType());
3808 return CouldNotCompute; // Otherwise it will loop infinitely.
3811 // We could implement others, but I really doubt anyone writes loops like
3812 // this, and if they did, they would already be constant folded.
3813 return CouldNotCompute;
3816 /// getLoopPredecessor - If the given loop's header has exactly one unique
3817 /// predecessor outside the loop, return it. Otherwise return null.
3819 BasicBlock *ScalarEvolution::getLoopPredecessor(const Loop *L) {
3820 BasicBlock *Header = L->getHeader();
3821 BasicBlock *Pred = 0;
3822 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header);
3824 if (!L->contains(*PI)) {
3825 if (Pred && Pred != *PI) return 0; // Multiple predecessors.
3831 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
3832 /// (which may not be an immediate predecessor) which has exactly one
3833 /// successor from which BB is reachable, or null if no such block is
3837 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
3838 // If the block has a unique predecessor, then there is no path from the
3839 // predecessor to the block that does not go through the direct edge
3840 // from the predecessor to the block.
3841 if (BasicBlock *Pred = BB->getSinglePredecessor())
3844 // A loop's header is defined to be a block that dominates the loop.
3845 // If the header has a unique predecessor outside the loop, it must be
3846 // a block that has exactly one successor that can reach the loop.
3847 if (Loop *L = LI->getLoopFor(BB))
3848 return getLoopPredecessor(L);
3853 /// HasSameValue - SCEV structural equivalence is usually sufficient for
3854 /// testing whether two expressions are equal, however for the purposes of
3855 /// looking for a condition guarding a loop, it can be useful to be a little
3856 /// more general, since a front-end may have replicated the controlling
3859 static bool HasSameValue(const SCEV* A, const SCEV* B) {
3860 // Quick check to see if they are the same SCEV.
3861 if (A == B) return true;
3863 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
3864 // two different instructions with the same value. Check for this case.
3865 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
3866 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
3867 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
3868 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
3869 if (AI->isIdenticalTo(BI))
3872 // Otherwise assume they may have a different value.
3876 /// isLoopGuardedByCond - Test whether entry to the loop is protected by
3877 /// a conditional between LHS and RHS. This is used to help avoid max
3878 /// expressions in loop trip counts.
3879 bool ScalarEvolution::isLoopGuardedByCond(const Loop *L,
3880 ICmpInst::Predicate Pred,
3881 const SCEV *LHS, const SCEV *RHS) {
3882 // Interpret a null as meaning no loop, where there is obviously no guard
3883 // (interprocedural conditions notwithstanding).
3884 if (!L) return false;
3886 BasicBlock *Predecessor = getLoopPredecessor(L);
3887 BasicBlock *PredecessorDest = L->getHeader();
3889 // Starting at the loop predecessor, climb up the predecessor chain, as long
3890 // as there are predecessors that can be found that have unique successors
3891 // leading to the original header.
3893 PredecessorDest = Predecessor,
3894 Predecessor = getPredecessorWithUniqueSuccessorForBB(Predecessor)) {
3896 BranchInst *LoopEntryPredicate =
3897 dyn_cast<BranchInst>(Predecessor->getTerminator());
3898 if (!LoopEntryPredicate ||
3899 LoopEntryPredicate->isUnconditional())
3902 ICmpInst *ICI = dyn_cast<ICmpInst>(LoopEntryPredicate->getCondition());
3905 // Now that we found a conditional branch that dominates the loop, check to
3906 // see if it is the comparison we are looking for.
3907 Value *PreCondLHS = ICI->getOperand(0);
3908 Value *PreCondRHS = ICI->getOperand(1);
3909 ICmpInst::Predicate Cond;
3910 if (LoopEntryPredicate->getSuccessor(0) == PredecessorDest)
3911 Cond = ICI->getPredicate();
3913 Cond = ICI->getInversePredicate();
3916 ; // An exact match.
3917 else if (!ICmpInst::isTrueWhenEqual(Cond) && Pred == ICmpInst::ICMP_NE)
3918 ; // The actual condition is beyond sufficient.
3920 // Check a few special cases.
3922 case ICmpInst::ICMP_UGT:
3923 if (Pred == ICmpInst::ICMP_ULT) {
3924 std::swap(PreCondLHS, PreCondRHS);
3925 Cond = ICmpInst::ICMP_ULT;
3929 case ICmpInst::ICMP_SGT:
3930 if (Pred == ICmpInst::ICMP_SLT) {
3931 std::swap(PreCondLHS, PreCondRHS);
3932 Cond = ICmpInst::ICMP_SLT;
3936 case ICmpInst::ICMP_NE:
3937 // Expressions like (x >u 0) are often canonicalized to (x != 0),
3938 // so check for this case by checking if the NE is comparing against
3939 // a minimum or maximum constant.
3940 if (!ICmpInst::isTrueWhenEqual(Pred))
3941 if (ConstantInt *CI = dyn_cast<ConstantInt>(PreCondRHS)) {
3942 const APInt &A = CI->getValue();
3944 case ICmpInst::ICMP_SLT:
3945 if (A.isMaxSignedValue()) break;
3947 case ICmpInst::ICMP_SGT:
3948 if (A.isMinSignedValue()) break;
3950 case ICmpInst::ICMP_ULT:
3951 if (A.isMaxValue()) break;
3953 case ICmpInst::ICMP_UGT:
3954 if (A.isMinValue()) break;
3959 Cond = ICmpInst::ICMP_NE;
3960 // NE is symmetric but the original comparison may not be. Swap
3961 // the operands if necessary so that they match below.
3962 if (isa<SCEVConstant>(LHS))
3963 std::swap(PreCondLHS, PreCondRHS);
3968 // We weren't able to reconcile the condition.
3972 if (!PreCondLHS->getType()->isInteger()) continue;
3974 const SCEV* PreCondLHSSCEV = getSCEV(PreCondLHS);
3975 const SCEV* PreCondRHSSCEV = getSCEV(PreCondRHS);
3976 if ((HasSameValue(LHS, PreCondLHSSCEV) &&
3977 HasSameValue(RHS, PreCondRHSSCEV)) ||
3978 (HasSameValue(LHS, getNotSCEV(PreCondRHSSCEV)) &&
3979 HasSameValue(RHS, getNotSCEV(PreCondLHSSCEV))))
3986 /// getBECount - Subtract the end and start values and divide by the step,
3987 /// rounding up, to get the number of times the backedge is executed. Return
3988 /// CouldNotCompute if an intermediate computation overflows.
3989 const SCEV* ScalarEvolution::getBECount(const SCEV* Start,
3992 const Type *Ty = Start->getType();
3993 const SCEV* NegOne = getIntegerSCEV(-1, Ty);
3994 const SCEV* Diff = getMinusSCEV(End, Start);
3995 const SCEV* RoundUp = getAddExpr(Step, NegOne);
3997 // Add an adjustment to the difference between End and Start so that
3998 // the division will effectively round up.
3999 const SCEV* Add = getAddExpr(Diff, RoundUp);
4001 // Check Add for unsigned overflow.
4002 // TODO: More sophisticated things could be done here.
4003 const Type *WideTy = IntegerType::get(getTypeSizeInBits(Ty) + 1);
4004 const SCEV* OperandExtendedAdd =
4005 getAddExpr(getZeroExtendExpr(Diff, WideTy),
4006 getZeroExtendExpr(RoundUp, WideTy));
4007 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
4008 return CouldNotCompute;
4010 return getUDivExpr(Add, Step);
4013 /// HowManyLessThans - Return the number of times a backedge containing the
4014 /// specified less-than comparison will execute. If not computable, return
4015 /// CouldNotCompute.
4016 ScalarEvolution::BackedgeTakenInfo ScalarEvolution::
4017 HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
4018 const Loop *L, bool isSigned) {
4019 // Only handle: "ADDREC < LoopInvariant".
4020 if (!RHS->isLoopInvariant(L)) return CouldNotCompute;
4022 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
4023 if (!AddRec || AddRec->getLoop() != L)
4024 return CouldNotCompute;
4026 if (AddRec->isAffine()) {
4027 // FORNOW: We only support unit strides.
4028 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
4029 const SCEV* Step = AddRec->getStepRecurrence(*this);
4031 // TODO: handle non-constant strides.
4032 const SCEVConstant *CStep = dyn_cast<SCEVConstant>(Step);
4033 if (!CStep || CStep->isZero())
4034 return CouldNotCompute;
4035 if (CStep->isOne()) {
4036 // With unit stride, the iteration never steps past the limit value.
4037 } else if (CStep->getValue()->getValue().isStrictlyPositive()) {
4038 if (const SCEVConstant *CLimit = dyn_cast<SCEVConstant>(RHS)) {
4039 // Test whether a positive iteration iteration can step past the limit
4040 // value and past the maximum value for its type in a single step.
4042 APInt Max = APInt::getSignedMaxValue(BitWidth);
4043 if ((Max - CStep->getValue()->getValue())
4044 .slt(CLimit->getValue()->getValue()))
4045 return CouldNotCompute;
4047 APInt Max = APInt::getMaxValue(BitWidth);
4048 if ((Max - CStep->getValue()->getValue())
4049 .ult(CLimit->getValue()->getValue()))
4050 return CouldNotCompute;
4053 // TODO: handle non-constant limit values below.
4054 return CouldNotCompute;
4056 // TODO: handle negative strides below.
4057 return CouldNotCompute;
4059 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
4060 // m. So, we count the number of iterations in which {n,+,s} < m is true.
4061 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
4062 // treat m-n as signed nor unsigned due to overflow possibility.
4064 // First, we get the value of the LHS in the first iteration: n
4065 const SCEV* Start = AddRec->getOperand(0);
4067 // Determine the minimum constant start value.
4068 const SCEV* MinStart = isa<SCEVConstant>(Start) ? Start :
4069 getConstant(isSigned ? APInt::getSignedMinValue(BitWidth) :
4070 APInt::getMinValue(BitWidth));
4072 // If we know that the condition is true in order to enter the loop,
4073 // then we know that it will run exactly (m-n)/s times. Otherwise, we
4074 // only know that it will execute (max(m,n)-n)/s times. In both cases,
4075 // the division must round up.
4076 const SCEV* End = RHS;
4077 if (!isLoopGuardedByCond(L,
4078 isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
4079 getMinusSCEV(Start, Step), RHS))
4080 End = isSigned ? getSMaxExpr(RHS, Start)
4081 : getUMaxExpr(RHS, Start);
4083 // Determine the maximum constant end value.
4084 const SCEV* MaxEnd =
4085 isa<SCEVConstant>(End) ? End :
4086 getConstant(isSigned ? APInt::getSignedMaxValue(BitWidth)
4087 .ashr(GetMinSignBits(End) - 1) :
4088 APInt::getMaxValue(BitWidth)
4089 .lshr(GetMinLeadingZeros(End)));
4091 // Finally, we subtract these two values and divide, rounding up, to get
4092 // the number of times the backedge is executed.
4093 const SCEV* BECount = getBECount(Start, End, Step);
4095 // The maximum backedge count is similar, except using the minimum start
4096 // value and the maximum end value.
4097 const SCEV* MaxBECount = getBECount(MinStart, MaxEnd, Step);;
4099 return BackedgeTakenInfo(BECount, MaxBECount);
4102 return CouldNotCompute;
4105 /// getNumIterationsInRange - Return the number of iterations of this loop that
4106 /// produce values in the specified constant range. Another way of looking at
4107 /// this is that it returns the first iteration number where the value is not in
4108 /// the condition, thus computing the exit count. If the iteration count can't
4109 /// be computed, an instance of SCEVCouldNotCompute is returned.
4110 const SCEV* SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
4111 ScalarEvolution &SE) const {
4112 if (Range.isFullSet()) // Infinite loop.
4113 return SE.getCouldNotCompute();
4115 // If the start is a non-zero constant, shift the range to simplify things.
4116 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
4117 if (!SC->getValue()->isZero()) {
4118 SmallVector<const SCEV*, 4> Operands(op_begin(), op_end());
4119 Operands[0] = SE.getIntegerSCEV(0, SC->getType());
4120 const SCEV* Shifted = SE.getAddRecExpr(Operands, getLoop());
4121 if (const SCEVAddRecExpr *ShiftedAddRec =
4122 dyn_cast<SCEVAddRecExpr>(Shifted))
4123 return ShiftedAddRec->getNumIterationsInRange(
4124 Range.subtract(SC->getValue()->getValue()), SE);
4125 // This is strange and shouldn't happen.
4126 return SE.getCouldNotCompute();
4129 // The only time we can solve this is when we have all constant indices.
4130 // Otherwise, we cannot determine the overflow conditions.
4131 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
4132 if (!isa<SCEVConstant>(getOperand(i)))
4133 return SE.getCouldNotCompute();
4136 // Okay at this point we know that all elements of the chrec are constants and
4137 // that the start element is zero.
4139 // First check to see if the range contains zero. If not, the first
4141 unsigned BitWidth = SE.getTypeSizeInBits(getType());
4142 if (!Range.contains(APInt(BitWidth, 0)))
4143 return SE.getIntegerSCEV(0, getType());
4146 // If this is an affine expression then we have this situation:
4147 // Solve {0,+,A} in Range === Ax in Range
4149 // We know that zero is in the range. If A is positive then we know that
4150 // the upper value of the range must be the first possible exit value.
4151 // If A is negative then the lower of the range is the last possible loop
4152 // value. Also note that we already checked for a full range.
4153 APInt One(BitWidth,1);
4154 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
4155 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
4157 // The exit value should be (End+A)/A.
4158 APInt ExitVal = (End + A).udiv(A);
4159 ConstantInt *ExitValue = ConstantInt::get(ExitVal);
4161 // Evaluate at the exit value. If we really did fall out of the valid
4162 // range, then we computed our trip count, otherwise wrap around or other
4163 // things must have happened.
4164 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
4165 if (Range.contains(Val->getValue()))
4166 return SE.getCouldNotCompute(); // Something strange happened
4168 // Ensure that the previous value is in the range. This is a sanity check.
4169 assert(Range.contains(
4170 EvaluateConstantChrecAtConstant(this,
4171 ConstantInt::get(ExitVal - One), SE)->getValue()) &&
4172 "Linear scev computation is off in a bad way!");
4173 return SE.getConstant(ExitValue);
4174 } else if (isQuadratic()) {
4175 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
4176 // quadratic equation to solve it. To do this, we must frame our problem in
4177 // terms of figuring out when zero is crossed, instead of when
4178 // Range.getUpper() is crossed.
4179 SmallVector<const SCEV*, 4> NewOps(op_begin(), op_end());
4180 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
4181 const SCEV* NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
4183 // Next, solve the constructed addrec
4184 std::pair<const SCEV*,const SCEV*> Roots =
4185 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
4186 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4187 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4189 // Pick the smallest positive root value.
4190 if (ConstantInt *CB =
4191 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
4192 R1->getValue(), R2->getValue()))) {
4193 if (CB->getZExtValue() == false)
4194 std::swap(R1, R2); // R1 is the minimum root now.
4196 // Make sure the root is not off by one. The returned iteration should
4197 // not be in the range, but the previous one should be. When solving
4198 // for "X*X < 5", for example, we should not return a root of 2.
4199 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
4202 if (Range.contains(R1Val->getValue())) {
4203 // The next iteration must be out of the range...
4204 ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()+1);
4206 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
4207 if (!Range.contains(R1Val->getValue()))
4208 return SE.getConstant(NextVal);
4209 return SE.getCouldNotCompute(); // Something strange happened
4212 // If R1 was not in the range, then it is a good return value. Make
4213 // sure that R1-1 WAS in the range though, just in case.
4214 ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()-1);
4215 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
4216 if (Range.contains(R1Val->getValue()))
4218 return SE.getCouldNotCompute(); // Something strange happened
4223 return SE.getCouldNotCompute();
4228 //===----------------------------------------------------------------------===//
4229 // SCEVCallbackVH Class Implementation
4230 //===----------------------------------------------------------------------===//
4232 void ScalarEvolution::SCEVCallbackVH::deleted() {
4233 assert(SE && "SCEVCallbackVH called with a non-null ScalarEvolution!");
4234 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
4235 SE->ConstantEvolutionLoopExitValue.erase(PN);
4236 if (Instruction *I = dyn_cast<Instruction>(getValPtr()))
4237 SE->ValuesAtScopes.erase(I);
4238 SE->Scalars.erase(getValPtr());
4239 // this now dangles!
4242 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *) {
4243 assert(SE && "SCEVCallbackVH called with a non-null ScalarEvolution!");
4245 // Forget all the expressions associated with users of the old value,
4246 // so that future queries will recompute the expressions using the new
4248 SmallVector<User *, 16> Worklist;
4249 Value *Old = getValPtr();
4250 bool DeleteOld = false;
4251 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
4253 Worklist.push_back(*UI);
4254 while (!Worklist.empty()) {
4255 User *U = Worklist.pop_back_val();
4256 // Deleting the Old value will cause this to dangle. Postpone
4257 // that until everything else is done.
4262 if (PHINode *PN = dyn_cast<PHINode>(U))
4263 SE->ConstantEvolutionLoopExitValue.erase(PN);
4264 if (Instruction *I = dyn_cast<Instruction>(U))
4265 SE->ValuesAtScopes.erase(I);
4266 if (SE->Scalars.erase(U))
4267 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
4269 Worklist.push_back(*UI);
4272 if (PHINode *PN = dyn_cast<PHINode>(Old))
4273 SE->ConstantEvolutionLoopExitValue.erase(PN);
4274 if (Instruction *I = dyn_cast<Instruction>(Old))
4275 SE->ValuesAtScopes.erase(I);
4276 SE->Scalars.erase(Old);
4277 // this now dangles!
4282 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
4283 : CallbackVH(V), SE(se) {}
4285 //===----------------------------------------------------------------------===//
4286 // ScalarEvolution Class Implementation
4287 //===----------------------------------------------------------------------===//
4289 ScalarEvolution::ScalarEvolution()
4290 : FunctionPass(&ID), CouldNotCompute(new SCEVCouldNotCompute()) {
4293 bool ScalarEvolution::runOnFunction(Function &F) {
4295 LI = &getAnalysis<LoopInfo>();
4296 TD = getAnalysisIfAvailable<TargetData>();
4300 void ScalarEvolution::releaseMemory() {
4302 BackedgeTakenCounts.clear();
4303 ConstantEvolutionLoopExitValue.clear();
4304 ValuesAtScopes.clear();
4306 for (std::map<ConstantInt*, SCEVConstant*>::iterator
4307 I = SCEVConstants.begin(), E = SCEVConstants.end(); I != E; ++I)
4309 for (std::map<std::pair<const SCEV*, const Type*>,
4310 SCEVTruncateExpr*>::iterator I = SCEVTruncates.begin(),
4311 E = SCEVTruncates.end(); I != E; ++I)
4313 for (std::map<std::pair<const SCEV*, const Type*>,
4314 SCEVZeroExtendExpr*>::iterator I = SCEVZeroExtends.begin(),
4315 E = SCEVZeroExtends.end(); I != E; ++I)
4317 for (std::map<std::pair<unsigned, std::vector<const SCEV*> >,
4318 SCEVCommutativeExpr*>::iterator I = SCEVCommExprs.begin(),
4319 E = SCEVCommExprs.end(); I != E; ++I)
4321 for (std::map<std::pair<const SCEV*, const SCEV*>, SCEVUDivExpr*>::iterator
4322 I = SCEVUDivs.begin(), E = SCEVUDivs.end(); I != E; ++I)
4324 for (std::map<std::pair<const SCEV*, const Type*>,
4325 SCEVSignExtendExpr*>::iterator I = SCEVSignExtends.begin(),
4326 E = SCEVSignExtends.end(); I != E; ++I)
4328 for (std::map<std::pair<const Loop *, std::vector<const SCEV*> >,
4329 SCEVAddRecExpr*>::iterator I = SCEVAddRecExprs.begin(),
4330 E = SCEVAddRecExprs.end(); I != E; ++I)
4332 for (std::map<Value*, SCEVUnknown*>::iterator I = SCEVUnknowns.begin(),
4333 E = SCEVUnknowns.end(); I != E; ++I)
4336 SCEVConstants.clear();
4337 SCEVTruncates.clear();
4338 SCEVZeroExtends.clear();
4339 SCEVCommExprs.clear();
4341 SCEVSignExtends.clear();
4342 SCEVAddRecExprs.clear();
4343 SCEVUnknowns.clear();
4346 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
4347 AU.setPreservesAll();
4348 AU.addRequiredTransitive<LoopInfo>();
4351 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
4352 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
4355 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
4357 // Print all inner loops first
4358 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
4359 PrintLoopInfo(OS, SE, *I);
4361 OS << "Loop " << L->getHeader()->getName() << ": ";
4363 SmallVector<BasicBlock*, 8> ExitBlocks;
4364 L->getExitBlocks(ExitBlocks);
4365 if (ExitBlocks.size() != 1)
4366 OS << "<multiple exits> ";
4368 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
4369 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
4371 OS << "Unpredictable backedge-taken count. ";
4375 OS << "Loop " << L->getHeader()->getName() << ": ";
4377 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
4378 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
4380 OS << "Unpredictable max backedge-taken count. ";
4386 void ScalarEvolution::print(raw_ostream &OS, const Module* ) const {
4387 // ScalarEvolution's implementaiton of the print method is to print
4388 // out SCEV values of all instructions that are interesting. Doing
4389 // this potentially causes it to create new SCEV objects though,
4390 // which technically conflicts with the const qualifier. This isn't
4391 // observable from outside the class though (the hasSCEV function
4392 // notwithstanding), so casting away the const isn't dangerous.
4393 ScalarEvolution &SE = *const_cast<ScalarEvolution*>(this);
4395 OS << "Classifying expressions for: " << F->getName() << "\n";
4396 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
4397 if (isSCEVable(I->getType())) {
4400 const SCEV* SV = SE.getSCEV(&*I);
4403 const Loop *L = LI->getLoopFor((*I).getParent());
4405 const SCEV* AtUse = SE.getSCEVAtScope(SV, L);
4412 OS << "\t\t" "Exits: ";
4413 const SCEV* ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
4414 if (!ExitValue->isLoopInvariant(L)) {
4415 OS << "<<Unknown>>";
4424 OS << "Determining loop execution counts for: " << F->getName() << "\n";
4425 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
4426 PrintLoopInfo(OS, &SE, *I);
4429 void ScalarEvolution::print(std::ostream &o, const Module *M) const {
4430 raw_os_ostream OS(o);