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. We only create one SCEV of a particular shape, so
18 // pointer-comparisons for equality are legal.
20 // One important aspect of the SCEV objects is that they are never cyclic, even
21 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
22 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
23 // recurrence) then we represent it directly as a recurrence node, otherwise we
24 // represent it as a SCEVUnknown node.
26 // In addition to being able to represent expressions of various types, we also
27 // have folders that are used to build the *canonical* representation for a
28 // particular expression. These folders are capable of using a variety of
29 // rewrite rules to simplify the expressions.
31 // Once the folders are defined, we can implement the more interesting
32 // higher-level code, such as the code that recognizes PHI nodes of various
33 // types, computes the execution count of a loop, etc.
35 // TODO: We should use these routines and value representations to implement
36 // dependence analysis!
38 //===----------------------------------------------------------------------===//
40 // There are several good references for the techniques used in this analysis.
42 // Chains of recurrences -- a method to expedite the evaluation
43 // of closed-form functions
44 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
46 // On computational properties of chains of recurrences
49 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
50 // Robert A. van Engelen
52 // Efficient Symbolic Analysis for Optimizing Compilers
53 // Robert A. van Engelen
55 // Using the chains of recurrences algebra for data dependence testing and
56 // induction variable substitution
57 // MS Thesis, Johnie Birch
59 //===----------------------------------------------------------------------===//
61 #define DEBUG_TYPE "scalar-evolution"
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/Constants.h"
64 #include "llvm/DerivedTypes.h"
65 #include "llvm/GlobalVariable.h"
66 #include "llvm/GlobalAlias.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/LLVMContext.h"
69 #include "llvm/Operator.h"
70 #include "llvm/Analysis/ConstantFolding.h"
71 #include "llvm/Analysis/Dominators.h"
72 #include "llvm/Analysis/LoopInfo.h"
73 #include "llvm/Analysis/ValueTracking.h"
74 #include "llvm/Assembly/Writer.h"
75 #include "llvm/Target/TargetData.h"
76 #include "llvm/Support/CommandLine.h"
77 #include "llvm/Support/ConstantRange.h"
78 #include "llvm/Support/ErrorHandling.h"
79 #include "llvm/Support/GetElementPtrTypeIterator.h"
80 #include "llvm/Support/InstIterator.h"
81 #include "llvm/Support/MathExtras.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include "llvm/ADT/Statistic.h"
84 #include "llvm/ADT/STLExtras.h"
85 #include "llvm/ADT/SmallPtrSet.h"
89 STATISTIC(NumArrayLenItCounts,
90 "Number of trip counts computed with array length");
91 STATISTIC(NumTripCountsComputed,
92 "Number of loops with predictable loop counts");
93 STATISTIC(NumTripCountsNotComputed,
94 "Number of loops without predictable loop counts");
95 STATISTIC(NumBruteForceTripCountsComputed,
96 "Number of loops with trip counts computed by force");
98 static cl::opt<unsigned>
99 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
100 cl::desc("Maximum number of iterations SCEV will "
101 "symbolically execute a constant "
105 static RegisterPass<ScalarEvolution>
106 R("scalar-evolution", "Scalar Evolution Analysis", false, true);
107 char ScalarEvolution::ID = 0;
109 //===----------------------------------------------------------------------===//
110 // SCEV class definitions
111 //===----------------------------------------------------------------------===//
113 //===----------------------------------------------------------------------===//
114 // Implementation of the SCEV class.
119 void SCEV::dump() const {
124 bool SCEV::isZero() const {
125 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
126 return SC->getValue()->isZero();
130 bool SCEV::isOne() const {
131 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
132 return SC->getValue()->isOne();
136 bool SCEV::isAllOnesValue() const {
137 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
138 return SC->getValue()->isAllOnesValue();
142 SCEVCouldNotCompute::SCEVCouldNotCompute() :
143 SCEV(FoldingSetNodeID(), scCouldNotCompute) {}
145 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
146 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
150 const Type *SCEVCouldNotCompute::getType() const {
151 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
155 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
156 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
160 bool SCEVCouldNotCompute::hasOperand(const SCEV *) const {
161 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
165 void SCEVCouldNotCompute::print(raw_ostream &OS) const {
166 OS << "***COULDNOTCOMPUTE***";
169 bool SCEVCouldNotCompute::classof(const SCEV *S) {
170 return S->getSCEVType() == scCouldNotCompute;
173 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
175 ID.AddInteger(scConstant);
178 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
179 SCEV *S = SCEVAllocator.Allocate<SCEVConstant>();
180 new (S) SCEVConstant(ID, V);
181 UniqueSCEVs.InsertNode(S, IP);
185 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
186 return getConstant(ConstantInt::get(getContext(), Val));
190 ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
192 ConstantInt::get(cast<IntegerType>(Ty), V, isSigned));
195 const Type *SCEVConstant::getType() const { return V->getType(); }
197 void SCEVConstant::print(raw_ostream &OS) const {
198 WriteAsOperand(OS, V, false);
201 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeID &ID,
202 unsigned SCEVTy, const SCEV *op, const Type *ty)
203 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
205 bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
206 return Op->dominates(BB, DT);
209 bool SCEVCastExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
210 return Op->properlyDominates(BB, DT);
213 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeID &ID,
214 const SCEV *op, const Type *ty)
215 : SCEVCastExpr(ID, scTruncate, op, ty) {
216 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
217 (Ty->isInteger() || isa<PointerType>(Ty)) &&
218 "Cannot truncate non-integer value!");
221 void SCEVTruncateExpr::print(raw_ostream &OS) const {
222 OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
225 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeID &ID,
226 const SCEV *op, const Type *ty)
227 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
228 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
229 (Ty->isInteger() || isa<PointerType>(Ty)) &&
230 "Cannot zero extend non-integer value!");
233 void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
234 OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
237 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeID &ID,
238 const SCEV *op, const Type *ty)
239 : SCEVCastExpr(ID, scSignExtend, op, ty) {
240 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
241 (Ty->isInteger() || isa<PointerType>(Ty)) &&
242 "Cannot sign extend non-integer value!");
245 void SCEVSignExtendExpr::print(raw_ostream &OS) const {
246 OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
249 void SCEVCommutativeExpr::print(raw_ostream &OS) const {
250 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
251 const char *OpStr = getOperationStr();
252 OS << "(" << *Operands[0];
253 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
254 OS << OpStr << *Operands[i];
258 bool SCEVNAryExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
259 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
260 if (!getOperand(i)->dominates(BB, DT))
266 bool SCEVNAryExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
267 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
268 if (!getOperand(i)->properlyDominates(BB, DT))
274 bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
275 return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
278 bool SCEVUDivExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
279 return LHS->properlyDominates(BB, DT) && RHS->properlyDominates(BB, DT);
282 void SCEVUDivExpr::print(raw_ostream &OS) const {
283 OS << "(" << *LHS << " /u " << *RHS << ")";
286 const Type *SCEVUDivExpr::getType() const {
287 // In most cases the types of LHS and RHS will be the same, but in some
288 // crazy cases one or the other may be a pointer. ScalarEvolution doesn't
289 // depend on the type for correctness, but handling types carefully can
290 // avoid extra casts in the SCEVExpander. The LHS is more likely to be
291 // a pointer type than the RHS, so use the RHS' type here.
292 return RHS->getType();
295 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
296 // Add recurrences are never invariant in the function-body (null loop).
300 // This recurrence is variant w.r.t. QueryLoop if QueryLoop contains L.
301 if (QueryLoop->contains(L))
304 // This recurrence is variant w.r.t. QueryLoop if any of its operands
306 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
307 if (!getOperand(i)->isLoopInvariant(QueryLoop))
310 // Otherwise it's loop-invariant.
314 void SCEVAddRecExpr::print(raw_ostream &OS) const {
315 OS << "{" << *Operands[0];
316 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
317 OS << ",+," << *Operands[i];
318 OS << "}<" << L->getHeader()->getName() + ">";
321 void SCEVFieldOffsetExpr::print(raw_ostream &OS) const {
322 // LLVM struct fields don't have names, so just print the field number.
323 OS << "offsetof(" << *STy << ", " << FieldNo << ")";
326 void SCEVAllocSizeExpr::print(raw_ostream &OS) const {
327 OS << "sizeof(" << *AllocTy << ")";
330 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
331 // All non-instruction values are loop invariant. All instructions are loop
332 // invariant if they are not contained in the specified loop.
333 // Instructions are never considered invariant in the function body
334 // (null loop) because they are defined within the "loop".
335 if (Instruction *I = dyn_cast<Instruction>(V))
336 return L && !L->contains(I);
340 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
341 if (Instruction *I = dyn_cast<Instruction>(getValue()))
342 return DT->dominates(I->getParent(), BB);
346 bool SCEVUnknown::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
347 if (Instruction *I = dyn_cast<Instruction>(getValue()))
348 return DT->properlyDominates(I->getParent(), BB);
352 const Type *SCEVUnknown::getType() const {
356 void SCEVUnknown::print(raw_ostream &OS) const {
357 WriteAsOperand(OS, V, false);
360 //===----------------------------------------------------------------------===//
362 //===----------------------------------------------------------------------===//
364 static bool CompareTypes(const Type *A, const Type *B) {
365 if (A->getTypeID() != B->getTypeID())
366 return A->getTypeID() < B->getTypeID();
367 if (const IntegerType *AI = dyn_cast<IntegerType>(A)) {
368 const IntegerType *BI = cast<IntegerType>(B);
369 return AI->getBitWidth() < BI->getBitWidth();
371 if (const PointerType *AI = dyn_cast<PointerType>(A)) {
372 const PointerType *BI = cast<PointerType>(B);
373 return CompareTypes(AI->getElementType(), BI->getElementType());
375 if (const ArrayType *AI = dyn_cast<ArrayType>(A)) {
376 const ArrayType *BI = cast<ArrayType>(B);
377 if (AI->getNumElements() != BI->getNumElements())
378 return AI->getNumElements() < BI->getNumElements();
379 return CompareTypes(AI->getElementType(), BI->getElementType());
381 if (const VectorType *AI = dyn_cast<VectorType>(A)) {
382 const VectorType *BI = cast<VectorType>(B);
383 if (AI->getNumElements() != BI->getNumElements())
384 return AI->getNumElements() < BI->getNumElements();
385 return CompareTypes(AI->getElementType(), BI->getElementType());
387 if (const StructType *AI = dyn_cast<StructType>(A)) {
388 const StructType *BI = cast<StructType>(B);
389 if (AI->getNumElements() != BI->getNumElements())
390 return AI->getNumElements() < BI->getNumElements();
391 for (unsigned i = 0, e = AI->getNumElements(); i != e; ++i)
392 if (CompareTypes(AI->getElementType(i), BI->getElementType(i)) ||
393 CompareTypes(BI->getElementType(i), AI->getElementType(i)))
394 return CompareTypes(AI->getElementType(i), BI->getElementType(i));
400 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
401 /// than the complexity of the RHS. This comparator is used to canonicalize
403 class SCEVComplexityCompare {
406 explicit SCEVComplexityCompare(LoopInfo *li) : LI(li) {}
408 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
409 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
413 // Primarily, sort the SCEVs by their getSCEVType().
414 if (LHS->getSCEVType() != RHS->getSCEVType())
415 return LHS->getSCEVType() < RHS->getSCEVType();
417 // Aside from the getSCEVType() ordering, the particular ordering
418 // isn't very important except that it's beneficial to be consistent,
419 // so that (a + b) and (b + a) don't end up as different expressions.
421 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
422 // not as complete as it could be.
423 if (const SCEVUnknown *LU = dyn_cast<SCEVUnknown>(LHS)) {
424 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
426 // Order pointer values after integer values. This helps SCEVExpander
428 if (isa<PointerType>(LU->getType()) && !isa<PointerType>(RU->getType()))
430 if (isa<PointerType>(RU->getType()) && !isa<PointerType>(LU->getType()))
433 // Compare getValueID values.
434 if (LU->getValue()->getValueID() != RU->getValue()->getValueID())
435 return LU->getValue()->getValueID() < RU->getValue()->getValueID();
437 // Sort arguments by their position.
438 if (const Argument *LA = dyn_cast<Argument>(LU->getValue())) {
439 const Argument *RA = cast<Argument>(RU->getValue());
440 return LA->getArgNo() < RA->getArgNo();
443 // For instructions, compare their loop depth, and their opcode.
444 // This is pretty loose.
445 if (Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
446 Instruction *RV = cast<Instruction>(RU->getValue());
448 // Compare loop depths.
449 if (LI->getLoopDepth(LV->getParent()) !=
450 LI->getLoopDepth(RV->getParent()))
451 return LI->getLoopDepth(LV->getParent()) <
452 LI->getLoopDepth(RV->getParent());
455 if (LV->getOpcode() != RV->getOpcode())
456 return LV->getOpcode() < RV->getOpcode();
458 // Compare the number of operands.
459 if (LV->getNumOperands() != RV->getNumOperands())
460 return LV->getNumOperands() < RV->getNumOperands();
466 // Compare constant values.
467 if (const SCEVConstant *LC = dyn_cast<SCEVConstant>(LHS)) {
468 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
469 if (LC->getValue()->getBitWidth() != RC->getValue()->getBitWidth())
470 return LC->getValue()->getBitWidth() < RC->getValue()->getBitWidth();
471 return LC->getValue()->getValue().ult(RC->getValue()->getValue());
474 // Compare addrec loop depths.
475 if (const SCEVAddRecExpr *LA = dyn_cast<SCEVAddRecExpr>(LHS)) {
476 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
477 if (LA->getLoop()->getLoopDepth() != RA->getLoop()->getLoopDepth())
478 return LA->getLoop()->getLoopDepth() < RA->getLoop()->getLoopDepth();
481 // Lexicographically compare n-ary expressions.
482 if (const SCEVNAryExpr *LC = dyn_cast<SCEVNAryExpr>(LHS)) {
483 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
484 for (unsigned i = 0, e = LC->getNumOperands(); i != e; ++i) {
485 if (i >= RC->getNumOperands())
487 if (operator()(LC->getOperand(i), RC->getOperand(i)))
489 if (operator()(RC->getOperand(i), LC->getOperand(i)))
492 return LC->getNumOperands() < RC->getNumOperands();
495 // Lexicographically compare udiv expressions.
496 if (const SCEVUDivExpr *LC = dyn_cast<SCEVUDivExpr>(LHS)) {
497 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
498 if (operator()(LC->getLHS(), RC->getLHS()))
500 if (operator()(RC->getLHS(), LC->getLHS()))
502 if (operator()(LC->getRHS(), RC->getRHS()))
504 if (operator()(RC->getRHS(), LC->getRHS()))
509 // Compare cast expressions by operand.
510 if (const SCEVCastExpr *LC = dyn_cast<SCEVCastExpr>(LHS)) {
511 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
512 return operator()(LC->getOperand(), RC->getOperand());
515 // Compare offsetof expressions.
516 if (const SCEVFieldOffsetExpr *LA = dyn_cast<SCEVFieldOffsetExpr>(LHS)) {
517 const SCEVFieldOffsetExpr *RA = cast<SCEVFieldOffsetExpr>(RHS);
518 if (CompareTypes(LA->getStructType(), RA->getStructType()) ||
519 CompareTypes(RA->getStructType(), LA->getStructType()))
520 return CompareTypes(LA->getStructType(), RA->getStructType());
521 return LA->getFieldNo() < RA->getFieldNo();
524 // Compare sizeof expressions by the allocation type.
525 if (const SCEVAllocSizeExpr *LA = dyn_cast<SCEVAllocSizeExpr>(LHS)) {
526 const SCEVAllocSizeExpr *RA = cast<SCEVAllocSizeExpr>(RHS);
527 return CompareTypes(LA->getAllocType(), RA->getAllocType());
530 llvm_unreachable("Unknown SCEV kind!");
536 /// GroupByComplexity - Given a list of SCEV objects, order them by their
537 /// complexity, and group objects of the same complexity together by value.
538 /// When this routine is finished, we know that any duplicates in the vector are
539 /// consecutive and that complexity is monotonically increasing.
541 /// Note that we go take special precautions to ensure that we get determinstic
542 /// results from this routine. In other words, we don't want the results of
543 /// this to depend on where the addresses of various SCEV objects happened to
546 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
548 if (Ops.size() < 2) return; // Noop
549 if (Ops.size() == 2) {
550 // This is the common case, which also happens to be trivially simple.
552 if (SCEVComplexityCompare(LI)(Ops[1], Ops[0]))
553 std::swap(Ops[0], Ops[1]);
557 // Do the rough sort by complexity.
558 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
560 // Now that we are sorted by complexity, group elements of the same
561 // complexity. Note that this is, at worst, N^2, but the vector is likely to
562 // be extremely short in practice. Note that we take this approach because we
563 // do not want to depend on the addresses of the objects we are grouping.
564 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
565 const SCEV *S = Ops[i];
566 unsigned Complexity = S->getSCEVType();
568 // If there are any objects of the same complexity and same value as this
570 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
571 if (Ops[j] == S) { // Found a duplicate.
572 // Move it to immediately after i'th element.
573 std::swap(Ops[i+1], Ops[j]);
574 ++i; // no need to rescan it.
575 if (i == e-2) return; // Done!
583 //===----------------------------------------------------------------------===//
584 // Simple SCEV method implementations
585 //===----------------------------------------------------------------------===//
587 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
589 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
591 const Type* ResultTy) {
592 // Handle the simplest case efficiently.
594 return SE.getTruncateOrZeroExtend(It, ResultTy);
596 // We are using the following formula for BC(It, K):
598 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
600 // Suppose, W is the bitwidth of the return value. We must be prepared for
601 // overflow. Hence, we must assure that the result of our computation is
602 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
603 // safe in modular arithmetic.
605 // However, this code doesn't use exactly that formula; the formula it uses
606 // is something like the following, where T is the number of factors of 2 in
607 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
610 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
612 // This formula is trivially equivalent to the previous formula. However,
613 // this formula can be implemented much more efficiently. The trick is that
614 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
615 // arithmetic. To do exact division in modular arithmetic, all we have
616 // to do is multiply by the inverse. Therefore, this step can be done at
619 // The next issue is how to safely do the division by 2^T. The way this
620 // is done is by doing the multiplication step at a width of at least W + T
621 // bits. This way, the bottom W+T bits of the product are accurate. Then,
622 // when we perform the division by 2^T (which is equivalent to a right shift
623 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
624 // truncated out after the division by 2^T.
626 // In comparison to just directly using the first formula, this technique
627 // is much more efficient; using the first formula requires W * K bits,
628 // but this formula less than W + K bits. Also, the first formula requires
629 // a division step, whereas this formula only requires multiplies and shifts.
631 // It doesn't matter whether the subtraction step is done in the calculation
632 // width or the input iteration count's width; if the subtraction overflows,
633 // the result must be zero anyway. We prefer here to do it in the width of
634 // the induction variable because it helps a lot for certain cases; CodeGen
635 // isn't smart enough to ignore the overflow, which leads to much less
636 // efficient code if the width of the subtraction is wider than the native
639 // (It's possible to not widen at all by pulling out factors of 2 before
640 // the multiplication; for example, K=2 can be calculated as
641 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
642 // extra arithmetic, so it's not an obvious win, and it gets
643 // much more complicated for K > 3.)
645 // Protection from insane SCEVs; this bound is conservative,
646 // but it probably doesn't matter.
648 return SE.getCouldNotCompute();
650 unsigned W = SE.getTypeSizeInBits(ResultTy);
652 // Calculate K! / 2^T and T; we divide out the factors of two before
653 // multiplying for calculating K! / 2^T to avoid overflow.
654 // Other overflow doesn't matter because we only care about the bottom
655 // W bits of the result.
656 APInt OddFactorial(W, 1);
658 for (unsigned i = 3; i <= K; ++i) {
660 unsigned TwoFactors = Mult.countTrailingZeros();
662 Mult = Mult.lshr(TwoFactors);
663 OddFactorial *= Mult;
666 // We need at least W + T bits for the multiplication step
667 unsigned CalculationBits = W + T;
669 // Calcuate 2^T, at width T+W.
670 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
672 // Calculate the multiplicative inverse of K! / 2^T;
673 // this multiplication factor will perform the exact division by
675 APInt Mod = APInt::getSignedMinValue(W+1);
676 APInt MultiplyFactor = OddFactorial.zext(W+1);
677 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
678 MultiplyFactor = MultiplyFactor.trunc(W);
680 // Calculate the product, at width T+W
681 const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
683 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
684 for (unsigned i = 1; i != K; ++i) {
685 const SCEV *S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
686 Dividend = SE.getMulExpr(Dividend,
687 SE.getTruncateOrZeroExtend(S, CalculationTy));
691 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
693 // Truncate the result, and divide by K! / 2^T.
695 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
696 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
699 /// evaluateAtIteration - Return the value of this chain of recurrences at
700 /// the specified iteration number. We can evaluate this recurrence by
701 /// multiplying each element in the chain by the binomial coefficient
702 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
704 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
706 /// where BC(It, k) stands for binomial coefficient.
708 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
709 ScalarEvolution &SE) const {
710 const SCEV *Result = getStart();
711 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
712 // The computation is correct in the face of overflow provided that the
713 // multiplication is performed _after_ the evaluation of the binomial
715 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
716 if (isa<SCEVCouldNotCompute>(Coeff))
719 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
724 //===----------------------------------------------------------------------===//
725 // SCEV Expression folder implementations
726 //===----------------------------------------------------------------------===//
728 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
730 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
731 "This is not a truncating conversion!");
732 assert(isSCEVable(Ty) &&
733 "This is not a conversion to a SCEVable type!");
734 Ty = getEffectiveSCEVType(Ty);
737 ID.AddInteger(scTruncate);
741 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
743 // Fold if the operand is constant.
744 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
746 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
748 // trunc(trunc(x)) --> trunc(x)
749 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
750 return getTruncateExpr(ST->getOperand(), Ty);
752 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
753 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
754 return getTruncateOrSignExtend(SS->getOperand(), Ty);
756 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
757 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
758 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
760 // If the input value is a chrec scev, truncate the chrec's operands.
761 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
762 SmallVector<const SCEV *, 4> Operands;
763 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
764 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
765 return getAddRecExpr(Operands, AddRec->getLoop());
768 // The cast wasn't folded; create an explicit cast node.
769 // Recompute the insert position, as it may have been invalidated.
770 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
771 SCEV *S = SCEVAllocator.Allocate<SCEVTruncateExpr>();
772 new (S) SCEVTruncateExpr(ID, Op, Ty);
773 UniqueSCEVs.InsertNode(S, IP);
777 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
779 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
780 "This is not an extending conversion!");
781 assert(isSCEVable(Ty) &&
782 "This is not a conversion to a SCEVable type!");
783 Ty = getEffectiveSCEVType(Ty);
785 // Fold if the operand is constant.
786 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
787 const Type *IntTy = getEffectiveSCEVType(Ty);
788 Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
789 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
790 return getConstant(cast<ConstantInt>(C));
793 // zext(zext(x)) --> zext(x)
794 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
795 return getZeroExtendExpr(SZ->getOperand(), Ty);
797 // Before doing any expensive analysis, check to see if we've already
798 // computed a SCEV for this Op and Ty.
800 ID.AddInteger(scZeroExtend);
804 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
806 // If the input value is a chrec scev, and we can prove that the value
807 // did not overflow the old, smaller, value, we can zero extend all of the
808 // operands (often constants). This allows analysis of something like
809 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
810 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
811 if (AR->isAffine()) {
812 const SCEV *Start = AR->getStart();
813 const SCEV *Step = AR->getStepRecurrence(*this);
814 unsigned BitWidth = getTypeSizeInBits(AR->getType());
815 const Loop *L = AR->getLoop();
817 // If we have special knowledge that this addrec won't overflow,
818 // we don't need to do any further analysis.
819 if (AR->hasNoUnsignedWrap())
820 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
821 getZeroExtendExpr(Step, Ty),
824 // Check whether the backedge-taken count is SCEVCouldNotCompute.
825 // Note that this serves two purposes: It filters out loops that are
826 // simply not analyzable, and it covers the case where this code is
827 // being called from within backedge-taken count analysis, such that
828 // attempting to ask for the backedge-taken count would likely result
829 // in infinite recursion. In the later case, the analysis code will
830 // cope with a conservative value, and it will take care to purge
831 // that value once it has finished.
832 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
833 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
834 // Manually compute the final value for AR, checking for
837 // Check whether the backedge-taken count can be losslessly casted to
838 // the addrec's type. The count is always unsigned.
839 const SCEV *CastedMaxBECount =
840 getTruncateOrZeroExtend(MaxBECount, Start->getType());
841 const SCEV *RecastedMaxBECount =
842 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
843 if (MaxBECount == RecastedMaxBECount) {
844 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
845 // Check whether Start+Step*MaxBECount has no unsigned overflow.
847 getMulExpr(CastedMaxBECount,
848 getTruncateOrZeroExtend(Step, Start->getType()));
849 const SCEV *Add = getAddExpr(Start, ZMul);
850 const SCEV *OperandExtendedAdd =
851 getAddExpr(getZeroExtendExpr(Start, WideTy),
852 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
853 getZeroExtendExpr(Step, WideTy)));
854 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
855 // Return the expression with the addrec on the outside.
856 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
857 getZeroExtendExpr(Step, Ty),
860 // Similar to above, only this time treat the step value as signed.
861 // This covers loops that count down.
863 getMulExpr(CastedMaxBECount,
864 getTruncateOrSignExtend(Step, Start->getType()));
865 Add = getAddExpr(Start, SMul);
867 getAddExpr(getZeroExtendExpr(Start, WideTy),
868 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
869 getSignExtendExpr(Step, WideTy)));
870 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
871 // Return the expression with the addrec on the outside.
872 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
873 getSignExtendExpr(Step, Ty),
877 // If the backedge is guarded by a comparison with the pre-inc value
878 // the addrec is safe. Also, if the entry is guarded by a comparison
879 // with the start value and the backedge is guarded by a comparison
880 // with the post-inc value, the addrec is safe.
881 if (isKnownPositive(Step)) {
882 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
883 getUnsignedRange(Step).getUnsignedMax());
884 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
885 (isLoopGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
886 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
887 AR->getPostIncExpr(*this), N)))
888 // Return the expression with the addrec on the outside.
889 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
890 getZeroExtendExpr(Step, Ty),
892 } else if (isKnownNegative(Step)) {
893 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
894 getSignedRange(Step).getSignedMin());
895 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) &&
896 (isLoopGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) ||
897 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
898 AR->getPostIncExpr(*this), N)))
899 // Return the expression with the addrec on the outside.
900 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
901 getSignExtendExpr(Step, Ty),
907 // The cast wasn't folded; create an explicit cast node.
908 // Recompute the insert position, as it may have been invalidated.
909 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
910 SCEV *S = SCEVAllocator.Allocate<SCEVZeroExtendExpr>();
911 new (S) SCEVZeroExtendExpr(ID, Op, Ty);
912 UniqueSCEVs.InsertNode(S, IP);
916 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
918 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
919 "This is not an extending conversion!");
920 assert(isSCEVable(Ty) &&
921 "This is not a conversion to a SCEVable type!");
922 Ty = getEffectiveSCEVType(Ty);
924 // Fold if the operand is constant.
925 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
926 const Type *IntTy = getEffectiveSCEVType(Ty);
927 Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
928 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
929 return getConstant(cast<ConstantInt>(C));
932 // sext(sext(x)) --> sext(x)
933 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
934 return getSignExtendExpr(SS->getOperand(), Ty);
936 // Before doing any expensive analysis, check to see if we've already
937 // computed a SCEV for this Op and Ty.
939 ID.AddInteger(scSignExtend);
943 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
945 // If the input value is a chrec scev, and we can prove that the value
946 // did not overflow the old, smaller, value, we can sign extend all of the
947 // operands (often constants). This allows analysis of something like
948 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
949 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
950 if (AR->isAffine()) {
951 const SCEV *Start = AR->getStart();
952 const SCEV *Step = AR->getStepRecurrence(*this);
953 unsigned BitWidth = getTypeSizeInBits(AR->getType());
954 const Loop *L = AR->getLoop();
956 // If we have special knowledge that this addrec won't overflow,
957 // we don't need to do any further analysis.
958 if (AR->hasNoSignedWrap())
959 return getAddRecExpr(getSignExtendExpr(Start, Ty),
960 getSignExtendExpr(Step, Ty),
963 // Check whether the backedge-taken count is SCEVCouldNotCompute.
964 // Note that this serves two purposes: It filters out loops that are
965 // simply not analyzable, and it covers the case where this code is
966 // being called from within backedge-taken count analysis, such that
967 // attempting to ask for the backedge-taken count would likely result
968 // in infinite recursion. In the later case, the analysis code will
969 // cope with a conservative value, and it will take care to purge
970 // that value once it has finished.
971 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
972 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
973 // Manually compute the final value for AR, checking for
976 // Check whether the backedge-taken count can be losslessly casted to
977 // the addrec's type. The count is always unsigned.
978 const SCEV *CastedMaxBECount =
979 getTruncateOrZeroExtend(MaxBECount, Start->getType());
980 const SCEV *RecastedMaxBECount =
981 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
982 if (MaxBECount == RecastedMaxBECount) {
983 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
984 // Check whether Start+Step*MaxBECount has no signed overflow.
986 getMulExpr(CastedMaxBECount,
987 getTruncateOrSignExtend(Step, Start->getType()));
988 const SCEV *Add = getAddExpr(Start, SMul);
989 const SCEV *OperandExtendedAdd =
990 getAddExpr(getSignExtendExpr(Start, WideTy),
991 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
992 getSignExtendExpr(Step, WideTy)));
993 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
994 // Return the expression with the addrec on the outside.
995 return getAddRecExpr(getSignExtendExpr(Start, Ty),
996 getSignExtendExpr(Step, Ty),
999 // Similar to above, only this time treat the step value as unsigned.
1000 // This covers loops that count up with an unsigned step.
1002 getMulExpr(CastedMaxBECount,
1003 getTruncateOrZeroExtend(Step, Start->getType()));
1004 Add = getAddExpr(Start, UMul);
1005 OperandExtendedAdd =
1006 getAddExpr(getSignExtendExpr(Start, WideTy),
1007 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1008 getZeroExtendExpr(Step, WideTy)));
1009 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1010 // Return the expression with the addrec on the outside.
1011 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1012 getZeroExtendExpr(Step, Ty),
1016 // If the backedge is guarded by a comparison with the pre-inc value
1017 // the addrec is safe. Also, if the entry is guarded by a comparison
1018 // with the start value and the backedge is guarded by a comparison
1019 // with the post-inc value, the addrec is safe.
1020 if (isKnownPositive(Step)) {
1021 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1022 getSignedRange(Step).getSignedMax());
1023 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1024 (isLoopGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1025 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1026 AR->getPostIncExpr(*this), N)))
1027 // Return the expression with the addrec on the outside.
1028 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1029 getSignExtendExpr(Step, Ty),
1031 } else if (isKnownNegative(Step)) {
1032 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1033 getSignedRange(Step).getSignedMin());
1034 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1035 (isLoopGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1036 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1037 AR->getPostIncExpr(*this), N)))
1038 // Return the expression with the addrec on the outside.
1039 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1040 getSignExtendExpr(Step, Ty),
1046 // The cast wasn't folded; create an explicit cast node.
1047 // Recompute the insert position, as it may have been invalidated.
1048 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1049 SCEV *S = SCEVAllocator.Allocate<SCEVSignExtendExpr>();
1050 new (S) SCEVSignExtendExpr(ID, Op, Ty);
1051 UniqueSCEVs.InsertNode(S, IP);
1055 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1056 /// unspecified bits out to the given type.
1058 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1060 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1061 "This is not an extending conversion!");
1062 assert(isSCEVable(Ty) &&
1063 "This is not a conversion to a SCEVable type!");
1064 Ty = getEffectiveSCEVType(Ty);
1066 // Sign-extend negative constants.
1067 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1068 if (SC->getValue()->getValue().isNegative())
1069 return getSignExtendExpr(Op, Ty);
1071 // Peel off a truncate cast.
1072 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1073 const SCEV *NewOp = T->getOperand();
1074 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1075 return getAnyExtendExpr(NewOp, Ty);
1076 return getTruncateOrNoop(NewOp, Ty);
1079 // Next try a zext cast. If the cast is folded, use it.
1080 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1081 if (!isa<SCEVZeroExtendExpr>(ZExt))
1084 // Next try a sext cast. If the cast is folded, use it.
1085 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1086 if (!isa<SCEVSignExtendExpr>(SExt))
1089 // If the expression is obviously signed, use the sext cast value.
1090 if (isa<SCEVSMaxExpr>(Op))
1093 // Absent any other information, use the zext cast value.
1097 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1098 /// a list of operands to be added under the given scale, update the given
1099 /// map. This is a helper function for getAddRecExpr. As an example of
1100 /// what it does, given a sequence of operands that would form an add
1101 /// expression like this:
1103 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1105 /// where A and B are constants, update the map with these values:
1107 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1109 /// and add 13 + A*B*29 to AccumulatedConstant.
1110 /// This will allow getAddRecExpr to produce this:
1112 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1114 /// This form often exposes folding opportunities that are hidden in
1115 /// the original operand list.
1117 /// Return true iff it appears that any interesting folding opportunities
1118 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1119 /// the common case where no interesting opportunities are present, and
1120 /// is also used as a check to avoid infinite recursion.
1123 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1124 SmallVector<const SCEV *, 8> &NewOps,
1125 APInt &AccumulatedConstant,
1126 const SmallVectorImpl<const SCEV *> &Ops,
1128 ScalarEvolution &SE) {
1129 bool Interesting = false;
1131 // Iterate over the add operands.
1132 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1133 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1134 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1136 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1137 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1138 // A multiplication of a constant with another add; recurse.
1140 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1141 cast<SCEVAddExpr>(Mul->getOperand(1))
1145 // A multiplication of a constant with some other value. Update
1147 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1148 const SCEV *Key = SE.getMulExpr(MulOps);
1149 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1150 M.insert(std::make_pair(Key, NewScale));
1152 NewOps.push_back(Pair.first->first);
1154 Pair.first->second += NewScale;
1155 // The map already had an entry for this value, which may indicate
1156 // a folding opportunity.
1160 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1161 // Pull a buried constant out to the outside.
1162 if (Scale != 1 || AccumulatedConstant != 0 || C->isZero())
1164 AccumulatedConstant += Scale * C->getValue()->getValue();
1166 // An ordinary operand. Update the map.
1167 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1168 M.insert(std::make_pair(Ops[i], Scale));
1170 NewOps.push_back(Pair.first->first);
1172 Pair.first->second += Scale;
1173 // The map already had an entry for this value, which may indicate
1174 // a folding opportunity.
1184 struct APIntCompare {
1185 bool operator()(const APInt &LHS, const APInt &RHS) const {
1186 return LHS.ult(RHS);
1191 /// getAddExpr - Get a canonical add expression, or something simpler if
1193 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1194 bool HasNUW, bool HasNSW) {
1195 assert(!Ops.empty() && "Cannot get empty add!");
1196 if (Ops.size() == 1) return Ops[0];
1198 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1199 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1200 getEffectiveSCEVType(Ops[0]->getType()) &&
1201 "SCEVAddExpr operand types don't match!");
1204 // Sort by complexity, this groups all similar expression types together.
1205 GroupByComplexity(Ops, LI);
1207 // If there are any constants, fold them together.
1209 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1211 assert(Idx < Ops.size());
1212 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1213 // We found two constants, fold them together!
1214 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1215 RHSC->getValue()->getValue());
1216 if (Ops.size() == 2) return Ops[0];
1217 Ops.erase(Ops.begin()+1); // Erase the folded element
1218 LHSC = cast<SCEVConstant>(Ops[0]);
1221 // If we are left with a constant zero being added, strip it off.
1222 if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1223 Ops.erase(Ops.begin());
1228 if (Ops.size() == 1) return Ops[0];
1230 // Okay, check to see if the same value occurs in the operand list twice. If
1231 // so, merge them together into an multiply expression. Since we sorted the
1232 // list, these values are required to be adjacent.
1233 const Type *Ty = Ops[0]->getType();
1234 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1235 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1236 // Found a match, merge the two values into a multiply, and add any
1237 // remaining values to the result.
1238 const SCEV *Two = getIntegerSCEV(2, Ty);
1239 const SCEV *Mul = getMulExpr(Ops[i], Two);
1240 if (Ops.size() == 2)
1242 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1244 return getAddExpr(Ops, HasNUW, HasNSW);
1247 // Check for truncates. If all the operands are truncated from the same
1248 // type, see if factoring out the truncate would permit the result to be
1249 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1250 // if the contents of the resulting outer trunc fold to something simple.
1251 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1252 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1253 const Type *DstType = Trunc->getType();
1254 const Type *SrcType = Trunc->getOperand()->getType();
1255 SmallVector<const SCEV *, 8> LargeOps;
1257 // Check all the operands to see if they can be represented in the
1258 // source type of the truncate.
1259 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1260 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1261 if (T->getOperand()->getType() != SrcType) {
1265 LargeOps.push_back(T->getOperand());
1266 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1267 // This could be either sign or zero extension, but sign extension
1268 // is much more likely to be foldable here.
1269 LargeOps.push_back(getSignExtendExpr(C, SrcType));
1270 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1271 SmallVector<const SCEV *, 8> LargeMulOps;
1272 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1273 if (const SCEVTruncateExpr *T =
1274 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1275 if (T->getOperand()->getType() != SrcType) {
1279 LargeMulOps.push_back(T->getOperand());
1280 } else if (const SCEVConstant *C =
1281 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1282 // This could be either sign or zero extension, but sign extension
1283 // is much more likely to be foldable here.
1284 LargeMulOps.push_back(getSignExtendExpr(C, SrcType));
1291 LargeOps.push_back(getMulExpr(LargeMulOps));
1298 // Evaluate the expression in the larger type.
1299 const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
1300 // If it folds to something simple, use it. Otherwise, don't.
1301 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1302 return getTruncateExpr(Fold, DstType);
1306 // Skip past any other cast SCEVs.
1307 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1310 // If there are add operands they would be next.
1311 if (Idx < Ops.size()) {
1312 bool DeletedAdd = false;
1313 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1314 // If we have an add, expand the add operands onto the end of the operands
1316 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
1317 Ops.erase(Ops.begin()+Idx);
1321 // If we deleted at least one add, we added operands to the end of the list,
1322 // and they are not necessarily sorted. Recurse to resort and resimplify
1323 // any operands we just aquired.
1325 return getAddExpr(Ops);
1328 // Skip over the add expression until we get to a multiply.
1329 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1332 // Check to see if there are any folding opportunities present with
1333 // operands multiplied by constant values.
1334 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1335 uint64_t BitWidth = getTypeSizeInBits(Ty);
1336 DenseMap<const SCEV *, APInt> M;
1337 SmallVector<const SCEV *, 8> NewOps;
1338 APInt AccumulatedConstant(BitWidth, 0);
1339 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1340 Ops, APInt(BitWidth, 1), *this)) {
1341 // Some interesting folding opportunity is present, so its worthwhile to
1342 // re-generate the operands list. Group the operands by constant scale,
1343 // to avoid multiplying by the same constant scale multiple times.
1344 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1345 for (SmallVector<const SCEV *, 8>::iterator I = NewOps.begin(),
1346 E = NewOps.end(); I != E; ++I)
1347 MulOpLists[M.find(*I)->second].push_back(*I);
1348 // Re-generate the operands list.
1350 if (AccumulatedConstant != 0)
1351 Ops.push_back(getConstant(AccumulatedConstant));
1352 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1353 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1355 Ops.push_back(getMulExpr(getConstant(I->first),
1356 getAddExpr(I->second)));
1358 return getIntegerSCEV(0, Ty);
1359 if (Ops.size() == 1)
1361 return getAddExpr(Ops);
1365 // If we are adding something to a multiply expression, make sure the
1366 // something is not already an operand of the multiply. If so, merge it into
1368 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1369 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1370 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1371 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1372 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1373 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(Ops[AddOp])) {
1374 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1375 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1376 if (Mul->getNumOperands() != 2) {
1377 // If the multiply has more than two operands, we must get the
1379 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(), Mul->op_end());
1380 MulOps.erase(MulOps.begin()+MulOp);
1381 InnerMul = getMulExpr(MulOps);
1383 const SCEV *One = getIntegerSCEV(1, Ty);
1384 const SCEV *AddOne = getAddExpr(InnerMul, One);
1385 const SCEV *OuterMul = getMulExpr(AddOne, Ops[AddOp]);
1386 if (Ops.size() == 2) return OuterMul;
1388 Ops.erase(Ops.begin()+AddOp);
1389 Ops.erase(Ops.begin()+Idx-1);
1391 Ops.erase(Ops.begin()+Idx);
1392 Ops.erase(Ops.begin()+AddOp-1);
1394 Ops.push_back(OuterMul);
1395 return getAddExpr(Ops);
1398 // Check this multiply against other multiplies being added together.
1399 for (unsigned OtherMulIdx = Idx+1;
1400 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1402 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1403 // If MulOp occurs in OtherMul, we can fold the two multiplies
1405 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1406 OMulOp != e; ++OMulOp)
1407 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1408 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1409 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1410 if (Mul->getNumOperands() != 2) {
1411 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1413 MulOps.erase(MulOps.begin()+MulOp);
1414 InnerMul1 = getMulExpr(MulOps);
1416 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1417 if (OtherMul->getNumOperands() != 2) {
1418 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1419 OtherMul->op_end());
1420 MulOps.erase(MulOps.begin()+OMulOp);
1421 InnerMul2 = getMulExpr(MulOps);
1423 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1424 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1425 if (Ops.size() == 2) return OuterMul;
1426 Ops.erase(Ops.begin()+Idx);
1427 Ops.erase(Ops.begin()+OtherMulIdx-1);
1428 Ops.push_back(OuterMul);
1429 return getAddExpr(Ops);
1435 // If there are any add recurrences in the operands list, see if any other
1436 // added values are loop invariant. If so, we can fold them into the
1438 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1441 // Scan over all recurrences, trying to fold loop invariants into them.
1442 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1443 // Scan all of the other operands to this add and add them to the vector if
1444 // they are loop invariant w.r.t. the recurrence.
1445 SmallVector<const SCEV *, 8> LIOps;
1446 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1447 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1448 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1449 LIOps.push_back(Ops[i]);
1450 Ops.erase(Ops.begin()+i);
1454 // If we found some loop invariants, fold them into the recurrence.
1455 if (!LIOps.empty()) {
1456 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1457 LIOps.push_back(AddRec->getStart());
1459 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1461 AddRecOps[0] = getAddExpr(LIOps);
1463 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
1464 // If all of the other operands were loop invariant, we are done.
1465 if (Ops.size() == 1) return NewRec;
1467 // Otherwise, add the folded AddRec by the non-liv parts.
1468 for (unsigned i = 0;; ++i)
1469 if (Ops[i] == AddRec) {
1473 return getAddExpr(Ops);
1476 // Okay, if there weren't any loop invariants to be folded, check to see if
1477 // there are multiple AddRec's with the same loop induction variable being
1478 // added together. If so, we can fold them.
1479 for (unsigned OtherIdx = Idx+1;
1480 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1481 if (OtherIdx != Idx) {
1482 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1483 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1484 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
1485 SmallVector<const SCEV *, 4> NewOps(AddRec->op_begin(),
1487 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
1488 if (i >= NewOps.size()) {
1489 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
1490 OtherAddRec->op_end());
1493 NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
1495 const SCEV *NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
1497 if (Ops.size() == 2) return NewAddRec;
1499 Ops.erase(Ops.begin()+Idx);
1500 Ops.erase(Ops.begin()+OtherIdx-1);
1501 Ops.push_back(NewAddRec);
1502 return getAddExpr(Ops);
1506 // Otherwise couldn't fold anything into this recurrence. Move onto the
1510 // Okay, it looks like we really DO need an add expr. Check to see if we
1511 // already have one, otherwise create a new one.
1512 FoldingSetNodeID ID;
1513 ID.AddInteger(scAddExpr);
1514 ID.AddInteger(Ops.size());
1515 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1516 ID.AddPointer(Ops[i]);
1518 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1519 SCEVAddExpr *S = SCEVAllocator.Allocate<SCEVAddExpr>();
1520 new (S) SCEVAddExpr(ID, Ops);
1521 UniqueSCEVs.InsertNode(S, IP);
1522 if (HasNUW) S->setHasNoUnsignedWrap(true);
1523 if (HasNSW) S->setHasNoSignedWrap(true);
1528 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1530 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1531 bool HasNUW, bool HasNSW) {
1532 assert(!Ops.empty() && "Cannot get empty mul!");
1534 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1535 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1536 getEffectiveSCEVType(Ops[0]->getType()) &&
1537 "SCEVMulExpr operand types don't match!");
1540 // Sort by complexity, this groups all similar expression types together.
1541 GroupByComplexity(Ops, LI);
1543 // If there are any constants, fold them together.
1545 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1547 // C1*(C2+V) -> C1*C2 + C1*V
1548 if (Ops.size() == 2)
1549 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1550 if (Add->getNumOperands() == 2 &&
1551 isa<SCEVConstant>(Add->getOperand(0)))
1552 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1553 getMulExpr(LHSC, Add->getOperand(1)));
1557 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1558 // We found two constants, fold them together!
1559 ConstantInt *Fold = ConstantInt::get(getContext(),
1560 LHSC->getValue()->getValue() *
1561 RHSC->getValue()->getValue());
1562 Ops[0] = getConstant(Fold);
1563 Ops.erase(Ops.begin()+1); // Erase the folded element
1564 if (Ops.size() == 1) return Ops[0];
1565 LHSC = cast<SCEVConstant>(Ops[0]);
1568 // If we are left with a constant one being multiplied, strip it off.
1569 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1570 Ops.erase(Ops.begin());
1572 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1573 // If we have a multiply of zero, it will always be zero.
1578 // Skip over the add expression until we get to a multiply.
1579 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1582 if (Ops.size() == 1)
1585 // If there are mul operands inline them all into this expression.
1586 if (Idx < Ops.size()) {
1587 bool DeletedMul = false;
1588 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1589 // If we have an mul, expand the mul operands onto the end of the operands
1591 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1592 Ops.erase(Ops.begin()+Idx);
1596 // If we deleted at least one mul, we added operands to the end of the list,
1597 // and they are not necessarily sorted. Recurse to resort and resimplify
1598 // any operands we just aquired.
1600 return getMulExpr(Ops);
1603 // If there are any add recurrences in the operands list, see if any other
1604 // added values are loop invariant. If so, we can fold them into the
1606 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1609 // Scan over all recurrences, trying to fold loop invariants into them.
1610 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1611 // Scan all of the other operands to this mul and add them to the vector if
1612 // they are loop invariant w.r.t. the recurrence.
1613 SmallVector<const SCEV *, 8> LIOps;
1614 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1615 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1616 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1617 LIOps.push_back(Ops[i]);
1618 Ops.erase(Ops.begin()+i);
1622 // If we found some loop invariants, fold them into the recurrence.
1623 if (!LIOps.empty()) {
1624 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1625 SmallVector<const SCEV *, 4> NewOps;
1626 NewOps.reserve(AddRec->getNumOperands());
1627 if (LIOps.size() == 1) {
1628 const SCEV *Scale = LIOps[0];
1629 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1630 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1632 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1633 SmallVector<const SCEV *, 4> MulOps(LIOps.begin(), LIOps.end());
1634 MulOps.push_back(AddRec->getOperand(i));
1635 NewOps.push_back(getMulExpr(MulOps));
1639 const SCEV *NewRec = getAddRecExpr(NewOps, AddRec->getLoop(),
1640 AddRec->hasNoUnsignedWrap() && HasNUW,
1641 AddRec->hasNoSignedWrap() && HasNSW);
1643 // If all of the other operands were loop invariant, we are done.
1644 if (Ops.size() == 1) return NewRec;
1646 // Otherwise, multiply the folded AddRec by the non-liv parts.
1647 for (unsigned i = 0;; ++i)
1648 if (Ops[i] == AddRec) {
1652 return getMulExpr(Ops);
1655 // Okay, if there weren't any loop invariants to be folded, check to see if
1656 // there are multiple AddRec's with the same loop induction variable being
1657 // multiplied together. If so, we can fold them.
1658 for (unsigned OtherIdx = Idx+1;
1659 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1660 if (OtherIdx != Idx) {
1661 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1662 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1663 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1664 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1665 const SCEV *NewStart = getMulExpr(F->getStart(),
1667 const SCEV *B = F->getStepRecurrence(*this);
1668 const SCEV *D = G->getStepRecurrence(*this);
1669 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1672 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1674 if (Ops.size() == 2) return NewAddRec;
1676 Ops.erase(Ops.begin()+Idx);
1677 Ops.erase(Ops.begin()+OtherIdx-1);
1678 Ops.push_back(NewAddRec);
1679 return getMulExpr(Ops);
1683 // Otherwise couldn't fold anything into this recurrence. Move onto the
1687 // Okay, it looks like we really DO need an mul expr. Check to see if we
1688 // already have one, otherwise create a new one.
1689 FoldingSetNodeID ID;
1690 ID.AddInteger(scMulExpr);
1691 ID.AddInteger(Ops.size());
1692 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1693 ID.AddPointer(Ops[i]);
1695 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1696 SCEVMulExpr *S = SCEVAllocator.Allocate<SCEVMulExpr>();
1697 new (S) SCEVMulExpr(ID, Ops);
1698 UniqueSCEVs.InsertNode(S, IP);
1699 if (HasNUW) S->setHasNoUnsignedWrap(true);
1700 if (HasNSW) S->setHasNoSignedWrap(true);
1704 /// getUDivExpr - Get a canonical unsigned division expression, or something
1705 /// simpler if possible.
1706 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1708 assert(getEffectiveSCEVType(LHS->getType()) ==
1709 getEffectiveSCEVType(RHS->getType()) &&
1710 "SCEVUDivExpr operand types don't match!");
1712 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1713 if (RHSC->getValue()->equalsInt(1))
1714 return LHS; // X udiv 1 --> x
1716 return getIntegerSCEV(0, LHS->getType()); // value is undefined
1718 // Determine if the division can be folded into the operands of
1720 // TODO: Generalize this to non-constants by using known-bits information.
1721 const Type *Ty = LHS->getType();
1722 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1723 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ;
1724 // For non-power-of-two values, effectively round the value up to the
1725 // nearest power of two.
1726 if (!RHSC->getValue()->getValue().isPowerOf2())
1728 const IntegerType *ExtTy =
1729 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1730 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1731 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1732 if (const SCEVConstant *Step =
1733 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1734 if (!Step->getValue()->getValue()
1735 .urem(RHSC->getValue()->getValue()) &&
1736 getZeroExtendExpr(AR, ExtTy) ==
1737 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1738 getZeroExtendExpr(Step, ExtTy),
1740 SmallVector<const SCEV *, 4> Operands;
1741 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1742 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1743 return getAddRecExpr(Operands, AR->getLoop());
1745 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1746 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1747 SmallVector<const SCEV *, 4> Operands;
1748 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1749 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1750 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1751 // Find an operand that's safely divisible.
1752 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1753 const SCEV *Op = M->getOperand(i);
1754 const SCEV *Div = getUDivExpr(Op, RHSC);
1755 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1756 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
1757 Operands = SmallVector<const SCEV *, 4>(MOperands.begin(),
1760 return getMulExpr(Operands);
1764 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1765 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1766 SmallVector<const SCEV *, 4> Operands;
1767 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1768 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1769 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1771 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1772 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
1773 if (isa<SCEVUDivExpr>(Op) || getMulExpr(Op, RHS) != A->getOperand(i))
1775 Operands.push_back(Op);
1777 if (Operands.size() == A->getNumOperands())
1778 return getAddExpr(Operands);
1782 // Fold if both operands are constant.
1783 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1784 Constant *LHSCV = LHSC->getValue();
1785 Constant *RHSCV = RHSC->getValue();
1786 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
1791 FoldingSetNodeID ID;
1792 ID.AddInteger(scUDivExpr);
1796 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1797 SCEV *S = SCEVAllocator.Allocate<SCEVUDivExpr>();
1798 new (S) SCEVUDivExpr(ID, LHS, RHS);
1799 UniqueSCEVs.InsertNode(S, IP);
1804 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1805 /// Simplify the expression as much as possible.
1806 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
1807 const SCEV *Step, const Loop *L,
1808 bool HasNUW, bool HasNSW) {
1809 SmallVector<const SCEV *, 4> Operands;
1810 Operands.push_back(Start);
1811 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1812 if (StepChrec->getLoop() == L) {
1813 Operands.insert(Operands.end(), StepChrec->op_begin(),
1814 StepChrec->op_end());
1815 return getAddRecExpr(Operands, L);
1818 Operands.push_back(Step);
1819 return getAddRecExpr(Operands, L, HasNUW, HasNSW);
1822 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1823 /// Simplify the expression as much as possible.
1825 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
1827 bool HasNUW, bool HasNSW) {
1828 if (Operands.size() == 1) return Operands[0];
1830 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
1831 assert(getEffectiveSCEVType(Operands[i]->getType()) ==
1832 getEffectiveSCEVType(Operands[0]->getType()) &&
1833 "SCEVAddRecExpr operand types don't match!");
1836 if (Operands.back()->isZero()) {
1837 Operands.pop_back();
1838 return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
1841 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
1842 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
1843 const Loop *NestedLoop = NestedAR->getLoop();
1844 if (L->getLoopDepth() < NestedLoop->getLoopDepth()) {
1845 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
1846 NestedAR->op_end());
1847 Operands[0] = NestedAR->getStart();
1848 // AddRecs require their operands be loop-invariant with respect to their
1849 // loops. Don't perform this transformation if it would break this
1851 bool AllInvariant = true;
1852 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1853 if (!Operands[i]->isLoopInvariant(L)) {
1854 AllInvariant = false;
1858 NestedOperands[0] = getAddRecExpr(Operands, L);
1859 AllInvariant = true;
1860 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
1861 if (!NestedOperands[i]->isLoopInvariant(NestedLoop)) {
1862 AllInvariant = false;
1866 // Ok, both add recurrences are valid after the transformation.
1867 return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
1869 // Reset Operands to its original state.
1870 Operands[0] = NestedAR;
1874 FoldingSetNodeID ID;
1875 ID.AddInteger(scAddRecExpr);
1876 ID.AddInteger(Operands.size());
1877 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1878 ID.AddPointer(Operands[i]);
1881 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1882 SCEVAddRecExpr *S = SCEVAllocator.Allocate<SCEVAddRecExpr>();
1883 new (S) SCEVAddRecExpr(ID, Operands, L);
1884 UniqueSCEVs.InsertNode(S, IP);
1885 if (HasNUW) S->setHasNoUnsignedWrap(true);
1886 if (HasNSW) S->setHasNoSignedWrap(true);
1890 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
1892 SmallVector<const SCEV *, 2> Ops;
1895 return getSMaxExpr(Ops);
1899 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
1900 assert(!Ops.empty() && "Cannot get empty smax!");
1901 if (Ops.size() == 1) return Ops[0];
1903 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1904 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1905 getEffectiveSCEVType(Ops[0]->getType()) &&
1906 "SCEVSMaxExpr operand types don't match!");
1909 // Sort by complexity, this groups all similar expression types together.
1910 GroupByComplexity(Ops, LI);
1912 // If there are any constants, fold them together.
1914 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1916 assert(Idx < Ops.size());
1917 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1918 // We found two constants, fold them together!
1919 ConstantInt *Fold = ConstantInt::get(getContext(),
1920 APIntOps::smax(LHSC->getValue()->getValue(),
1921 RHSC->getValue()->getValue()));
1922 Ops[0] = getConstant(Fold);
1923 Ops.erase(Ops.begin()+1); // Erase the folded element
1924 if (Ops.size() == 1) return Ops[0];
1925 LHSC = cast<SCEVConstant>(Ops[0]);
1928 // If we are left with a constant minimum-int, strip it off.
1929 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
1930 Ops.erase(Ops.begin());
1932 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
1933 // If we have an smax with a constant maximum-int, it will always be
1939 if (Ops.size() == 1) return Ops[0];
1941 // Find the first SMax
1942 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
1945 // Check to see if one of the operands is an SMax. If so, expand its operands
1946 // onto our operand list, and recurse to simplify.
1947 if (Idx < Ops.size()) {
1948 bool DeletedSMax = false;
1949 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
1950 Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
1951 Ops.erase(Ops.begin()+Idx);
1956 return getSMaxExpr(Ops);
1959 // Okay, check to see if the same value occurs in the operand list twice. If
1960 // so, delete one. Since we sorted the list, these values are required to
1962 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1963 if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
1964 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1968 if (Ops.size() == 1) return Ops[0];
1970 assert(!Ops.empty() && "Reduced smax down to nothing!");
1972 // Okay, it looks like we really DO need an smax expr. Check to see if we
1973 // already have one, otherwise create a new one.
1974 FoldingSetNodeID ID;
1975 ID.AddInteger(scSMaxExpr);
1976 ID.AddInteger(Ops.size());
1977 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1978 ID.AddPointer(Ops[i]);
1980 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1981 SCEV *S = SCEVAllocator.Allocate<SCEVSMaxExpr>();
1982 new (S) SCEVSMaxExpr(ID, Ops);
1983 UniqueSCEVs.InsertNode(S, IP);
1987 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
1989 SmallVector<const SCEV *, 2> Ops;
1992 return getUMaxExpr(Ops);
1996 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
1997 assert(!Ops.empty() && "Cannot get empty umax!");
1998 if (Ops.size() == 1) return Ops[0];
2000 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2001 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
2002 getEffectiveSCEVType(Ops[0]->getType()) &&
2003 "SCEVUMaxExpr operand types don't match!");
2006 // Sort by complexity, this groups all similar expression types together.
2007 GroupByComplexity(Ops, LI);
2009 // If there are any constants, fold them together.
2011 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2013 assert(Idx < Ops.size());
2014 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2015 // We found two constants, fold them together!
2016 ConstantInt *Fold = ConstantInt::get(getContext(),
2017 APIntOps::umax(LHSC->getValue()->getValue(),
2018 RHSC->getValue()->getValue()));
2019 Ops[0] = getConstant(Fold);
2020 Ops.erase(Ops.begin()+1); // Erase the folded element
2021 if (Ops.size() == 1) return Ops[0];
2022 LHSC = cast<SCEVConstant>(Ops[0]);
2025 // If we are left with a constant minimum-int, strip it off.
2026 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2027 Ops.erase(Ops.begin());
2029 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2030 // If we have an umax with a constant maximum-int, it will always be
2036 if (Ops.size() == 1) return Ops[0];
2038 // Find the first UMax
2039 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2042 // Check to see if one of the operands is a UMax. If so, expand its operands
2043 // onto our operand list, and recurse to simplify.
2044 if (Idx < Ops.size()) {
2045 bool DeletedUMax = false;
2046 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2047 Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
2048 Ops.erase(Ops.begin()+Idx);
2053 return getUMaxExpr(Ops);
2056 // Okay, check to see if the same value occurs in the operand list twice. If
2057 // so, delete one. Since we sorted the list, these values are required to
2059 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2060 if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
2061 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2065 if (Ops.size() == 1) return Ops[0];
2067 assert(!Ops.empty() && "Reduced umax down to nothing!");
2069 // Okay, it looks like we really DO need a umax expr. Check to see if we
2070 // already have one, otherwise create a new one.
2071 FoldingSetNodeID ID;
2072 ID.AddInteger(scUMaxExpr);
2073 ID.AddInteger(Ops.size());
2074 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2075 ID.AddPointer(Ops[i]);
2077 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2078 SCEV *S = SCEVAllocator.Allocate<SCEVUMaxExpr>();
2079 new (S) SCEVUMaxExpr(ID, Ops);
2080 UniqueSCEVs.InsertNode(S, IP);
2084 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2086 // ~smax(~x, ~y) == smin(x, y).
2087 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2090 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2092 // ~umax(~x, ~y) == umin(x, y)
2093 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2096 const SCEV *ScalarEvolution::getFieldOffsetExpr(const StructType *STy,
2098 // If we have TargetData we can determine the constant offset.
2100 const Type *IntPtrTy = TD->getIntPtrType(getContext());
2101 const StructLayout &SL = *TD->getStructLayout(STy);
2102 uint64_t Offset = SL.getElementOffset(FieldNo);
2103 return getIntegerSCEV(Offset, IntPtrTy);
2106 // Field 0 is always at offset 0.
2108 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2109 return getIntegerSCEV(0, Ty);
2112 // Okay, it looks like we really DO need an offsetof expr. Check to see if we
2113 // already have one, otherwise create a new one.
2114 FoldingSetNodeID ID;
2115 ID.AddInteger(scFieldOffset);
2117 ID.AddInteger(FieldNo);
2119 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2120 SCEV *S = SCEVAllocator.Allocate<SCEVFieldOffsetExpr>();
2121 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2122 new (S) SCEVFieldOffsetExpr(ID, Ty, STy, FieldNo);
2123 UniqueSCEVs.InsertNode(S, IP);
2127 const SCEV *ScalarEvolution::getAllocSizeExpr(const Type *AllocTy) {
2128 // If we have TargetData we can determine the constant size.
2129 if (TD && AllocTy->isSized()) {
2130 const Type *IntPtrTy = TD->getIntPtrType(getContext());
2131 return getIntegerSCEV(TD->getTypeAllocSize(AllocTy), IntPtrTy);
2134 // Expand an array size into the element size times the number
2136 if (const ArrayType *ATy = dyn_cast<ArrayType>(AllocTy)) {
2137 const SCEV *E = getAllocSizeExpr(ATy->getElementType());
2139 E, getConstant(ConstantInt::get(cast<IntegerType>(E->getType()),
2140 ATy->getNumElements())));
2143 // Expand a vector size into the element size times the number
2145 if (const VectorType *VTy = dyn_cast<VectorType>(AllocTy)) {
2146 const SCEV *E = getAllocSizeExpr(VTy->getElementType());
2148 E, getConstant(ConstantInt::get(cast<IntegerType>(E->getType()),
2149 VTy->getNumElements())));
2152 // Okay, it looks like we really DO need a sizeof expr. Check to see if we
2153 // already have one, otherwise create a new one.
2154 FoldingSetNodeID ID;
2155 ID.AddInteger(scAllocSize);
2156 ID.AddPointer(AllocTy);
2158 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2159 SCEV *S = SCEVAllocator.Allocate<SCEVAllocSizeExpr>();
2160 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2161 new (S) SCEVAllocSizeExpr(ID, Ty, AllocTy);
2162 UniqueSCEVs.InsertNode(S, IP);
2166 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2167 // Don't attempt to do anything other than create a SCEVUnknown object
2168 // here. createSCEV only calls getUnknown after checking for all other
2169 // interesting possibilities, and any other code that calls getUnknown
2170 // is doing so in order to hide a value from SCEV canonicalization.
2172 FoldingSetNodeID ID;
2173 ID.AddInteger(scUnknown);
2176 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2177 SCEV *S = SCEVAllocator.Allocate<SCEVUnknown>();
2178 new (S) SCEVUnknown(ID, V);
2179 UniqueSCEVs.InsertNode(S, IP);
2183 //===----------------------------------------------------------------------===//
2184 // Basic SCEV Analysis and PHI Idiom Recognition Code
2187 /// isSCEVable - Test if values of the given type are analyzable within
2188 /// the SCEV framework. This primarily includes integer types, and it
2189 /// can optionally include pointer types if the ScalarEvolution class
2190 /// has access to target-specific information.
2191 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2192 // Integers and pointers are always SCEVable.
2193 return Ty->isInteger() || isa<PointerType>(Ty);
2196 /// getTypeSizeInBits - Return the size in bits of the specified type,
2197 /// for which isSCEVable must return true.
2198 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2199 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2201 // If we have a TargetData, use it!
2203 return TD->getTypeSizeInBits(Ty);
2205 // Integer types have fixed sizes.
2206 if (Ty->isInteger())
2207 return Ty->getPrimitiveSizeInBits();
2209 // The only other support type is pointer. Without TargetData, conservatively
2210 // assume pointers are 64-bit.
2211 assert(isa<PointerType>(Ty) && "isSCEVable permitted a non-SCEVable type!");
2215 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2216 /// the given type and which represents how SCEV will treat the given
2217 /// type, for which isSCEVable must return true. For pointer types,
2218 /// this is the pointer-sized integer type.
2219 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2220 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2222 if (Ty->isInteger())
2225 // The only other support type is pointer.
2226 assert(isa<PointerType>(Ty) && "Unexpected non-pointer non-integer type!");
2227 if (TD) return TD->getIntPtrType(getContext());
2229 // Without TargetData, conservatively assume pointers are 64-bit.
2230 return Type::getInt64Ty(getContext());
2233 const SCEV *ScalarEvolution::getCouldNotCompute() {
2234 return &CouldNotCompute;
2237 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2238 /// expression and create a new one.
2239 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2240 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2242 std::map<SCEVCallbackVH, const SCEV *>::iterator I = Scalars.find(V);
2243 if (I != Scalars.end()) return I->second;
2244 const SCEV *S = createSCEV(V);
2245 Scalars.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2249 /// getIntegerSCEV - Given a SCEVable type, create a constant for the
2250 /// specified signed integer value and return a SCEV for the constant.
2251 const SCEV *ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
2252 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
2253 return getConstant(ConstantInt::get(ITy, Val));
2256 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2258 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2259 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2261 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2263 const Type *Ty = V->getType();
2264 Ty = getEffectiveSCEVType(Ty);
2265 return getMulExpr(V,
2266 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2269 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2270 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2271 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2273 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2275 const Type *Ty = V->getType();
2276 Ty = getEffectiveSCEVType(Ty);
2277 const SCEV *AllOnes =
2278 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2279 return getMinusSCEV(AllOnes, V);
2282 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
2284 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS,
2287 return getAddExpr(LHS, getNegativeSCEV(RHS));
2290 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2291 /// input value to the specified type. If the type must be extended, it is zero
2294 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V,
2296 const Type *SrcTy = V->getType();
2297 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2298 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2299 "Cannot truncate or zero extend with non-integer arguments!");
2300 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2301 return V; // No conversion
2302 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2303 return getTruncateExpr(V, Ty);
2304 return getZeroExtendExpr(V, Ty);
2307 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2308 /// input value to the specified type. If the type must be extended, it is sign
2311 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2313 const Type *SrcTy = V->getType();
2314 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2315 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2316 "Cannot truncate or zero extend with non-integer arguments!");
2317 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2318 return V; // No conversion
2319 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2320 return getTruncateExpr(V, Ty);
2321 return getSignExtendExpr(V, Ty);
2324 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2325 /// input value to the specified type. If the type must be extended, it is zero
2326 /// extended. The conversion must not be narrowing.
2328 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2329 const Type *SrcTy = V->getType();
2330 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2331 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2332 "Cannot noop or zero extend with non-integer arguments!");
2333 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2334 "getNoopOrZeroExtend cannot truncate!");
2335 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2336 return V; // No conversion
2337 return getZeroExtendExpr(V, Ty);
2340 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2341 /// input value to the specified type. If the type must be extended, it is sign
2342 /// extended. The conversion must not be narrowing.
2344 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
2345 const Type *SrcTy = V->getType();
2346 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2347 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2348 "Cannot noop or sign extend with non-integer arguments!");
2349 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2350 "getNoopOrSignExtend cannot truncate!");
2351 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2352 return V; // No conversion
2353 return getSignExtendExpr(V, Ty);
2356 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2357 /// the input value to the specified type. If the type must be extended,
2358 /// it is extended with unspecified bits. The conversion must not be
2361 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2362 const Type *SrcTy = V->getType();
2363 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2364 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2365 "Cannot noop or any extend with non-integer arguments!");
2366 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2367 "getNoopOrAnyExtend cannot truncate!");
2368 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2369 return V; // No conversion
2370 return getAnyExtendExpr(V, Ty);
2373 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2374 /// input value to the specified type. The conversion must not be widening.
2376 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2377 const Type *SrcTy = V->getType();
2378 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2379 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2380 "Cannot truncate or noop with non-integer arguments!");
2381 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2382 "getTruncateOrNoop cannot extend!");
2383 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2384 return V; // No conversion
2385 return getTruncateExpr(V, Ty);
2388 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2389 /// the types using zero-extension, and then perform a umax operation
2391 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2393 const SCEV *PromotedLHS = LHS;
2394 const SCEV *PromotedRHS = RHS;
2396 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2397 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2399 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2401 return getUMaxExpr(PromotedLHS, PromotedRHS);
2404 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2405 /// the types using zero-extension, and then perform a umin operation
2407 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2409 const SCEV *PromotedLHS = LHS;
2410 const SCEV *PromotedRHS = RHS;
2412 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2413 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2415 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2417 return getUMinExpr(PromotedLHS, PromotedRHS);
2420 /// PushDefUseChildren - Push users of the given Instruction
2421 /// onto the given Worklist.
2423 PushDefUseChildren(Instruction *I,
2424 SmallVectorImpl<Instruction *> &Worklist) {
2425 // Push the def-use children onto the Worklist stack.
2426 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2428 Worklist.push_back(cast<Instruction>(UI));
2431 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2432 /// instructions that depend on the given instruction and removes them from
2433 /// the Scalars map if they reference SymName. This is used during PHI
2436 ScalarEvolution::ForgetSymbolicName(Instruction *I, const SCEV *SymName) {
2437 SmallVector<Instruction *, 16> Worklist;
2438 PushDefUseChildren(I, Worklist);
2440 SmallPtrSet<Instruction *, 8> Visited;
2442 while (!Worklist.empty()) {
2443 Instruction *I = Worklist.pop_back_val();
2444 if (!Visited.insert(I)) continue;
2446 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
2447 Scalars.find(static_cast<Value *>(I));
2448 if (It != Scalars.end()) {
2449 // Short-circuit the def-use traversal if the symbolic name
2450 // ceases to appear in expressions.
2451 if (!It->second->hasOperand(SymName))
2454 // SCEVUnknown for a PHI either means that it has an unrecognized
2455 // structure, or it's a PHI that's in the progress of being computed
2456 // by createNodeForPHI. In the former case, additional loop trip
2457 // count information isn't going to change anything. In the later
2458 // case, createNodeForPHI will perform the necessary updates on its
2459 // own when it gets to that point.
2460 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
2461 ValuesAtScopes.erase(It->second);
2466 PushDefUseChildren(I, Worklist);
2470 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2471 /// a loop header, making it a potential recurrence, or it doesn't.
2473 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2474 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
2475 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2476 if (L->getHeader() == PN->getParent()) {
2477 // If it lives in the loop header, it has two incoming values, one
2478 // from outside the loop, and one from inside.
2479 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
2480 unsigned BackEdge = IncomingEdge^1;
2482 // While we are analyzing this PHI node, handle its value symbolically.
2483 const SCEV *SymbolicName = getUnknown(PN);
2484 assert(Scalars.find(PN) == Scalars.end() &&
2485 "PHI node already processed?");
2486 Scalars.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2488 // Using this symbolic name for the PHI, analyze the value coming around
2490 Value *BEValueV = PN->getIncomingValue(BackEdge);
2491 const SCEV *BEValue = getSCEV(BEValueV);
2493 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2494 // has a special value for the first iteration of the loop.
2496 // If the value coming around the backedge is an add with the symbolic
2497 // value we just inserted, then we found a simple induction variable!
2498 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2499 // If there is a single occurrence of the symbolic value, replace it
2500 // with a recurrence.
2501 unsigned FoundIndex = Add->getNumOperands();
2502 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2503 if (Add->getOperand(i) == SymbolicName)
2504 if (FoundIndex == e) {
2509 if (FoundIndex != Add->getNumOperands()) {
2510 // Create an add with everything but the specified operand.
2511 SmallVector<const SCEV *, 8> Ops;
2512 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2513 if (i != FoundIndex)
2514 Ops.push_back(Add->getOperand(i));
2515 const SCEV *Accum = getAddExpr(Ops);
2517 // This is not a valid addrec if the step amount is varying each
2518 // loop iteration, but is not itself an addrec in this loop.
2519 if (Accum->isLoopInvariant(L) ||
2520 (isa<SCEVAddRecExpr>(Accum) &&
2521 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2522 const SCEV *StartVal =
2523 getSCEV(PN->getIncomingValue(IncomingEdge));
2524 const SCEVAddRecExpr *PHISCEV =
2525 cast<SCEVAddRecExpr>(getAddRecExpr(StartVal, Accum, L));
2527 // If the increment doesn't overflow, then neither the addrec nor the
2528 // post-increment will overflow.
2529 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV))
2530 if (OBO->getOperand(0) == PN &&
2531 getSCEV(OBO->getOperand(1)) ==
2532 PHISCEV->getStepRecurrence(*this)) {
2533 const SCEVAddRecExpr *PostInc = PHISCEV->getPostIncExpr(*this);
2534 if (OBO->hasNoUnsignedWrap()) {
2535 const_cast<SCEVAddRecExpr *>(PHISCEV)
2536 ->setHasNoUnsignedWrap(true);
2537 const_cast<SCEVAddRecExpr *>(PostInc)
2538 ->setHasNoUnsignedWrap(true);
2540 if (OBO->hasNoSignedWrap()) {
2541 const_cast<SCEVAddRecExpr *>(PHISCEV)
2542 ->setHasNoSignedWrap(true);
2543 const_cast<SCEVAddRecExpr *>(PostInc)
2544 ->setHasNoSignedWrap(true);
2548 // Okay, for the entire analysis of this edge we assumed the PHI
2549 // to be symbolic. We now need to go back and purge all of the
2550 // entries for the scalars that use the symbolic expression.
2551 ForgetSymbolicName(PN, SymbolicName);
2552 Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
2556 } else if (const SCEVAddRecExpr *AddRec =
2557 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2558 // Otherwise, this could be a loop like this:
2559 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2560 // In this case, j = {1,+,1} and BEValue is j.
2561 // Because the other in-value of i (0) fits the evolution of BEValue
2562 // i really is an addrec evolution.
2563 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2564 const SCEV *StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
2566 // If StartVal = j.start - j.stride, we can use StartVal as the
2567 // initial step of the addrec evolution.
2568 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2569 AddRec->getOperand(1))) {
2570 const SCEV *PHISCEV =
2571 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2573 // Okay, for the entire analysis of this edge we assumed the PHI
2574 // to be symbolic. We now need to go back and purge all of the
2575 // entries for the scalars that use the symbolic expression.
2576 ForgetSymbolicName(PN, SymbolicName);
2577 Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
2583 return SymbolicName;
2586 // It's tempting to recognize PHIs with a unique incoming value, however
2587 // this leads passes like indvars to break LCSSA form. Fortunately, such
2588 // PHIs are rare, as instcombine zaps them.
2590 // If it's not a loop phi, we can't handle it yet.
2591 return getUnknown(PN);
2594 /// createNodeForGEP - Expand GEP instructions into add and multiply
2595 /// operations. This allows them to be analyzed by regular SCEV code.
2597 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
2599 bool InBounds = GEP->isInBounds();
2600 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2601 Value *Base = GEP->getOperand(0);
2602 // Don't attempt to analyze GEPs over unsized objects.
2603 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2604 return getUnknown(GEP);
2605 const SCEV *TotalOffset = getIntegerSCEV(0, IntPtrTy);
2606 gep_type_iterator GTI = gep_type_begin(GEP);
2607 for (GetElementPtrInst::op_iterator I = next(GEP->op_begin()),
2611 // Compute the (potentially symbolic) offset in bytes for this index.
2612 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2613 // For a struct, add the member offset.
2614 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2615 TotalOffset = getAddExpr(TotalOffset,
2616 getFieldOffsetExpr(STy, FieldNo),
2617 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2619 // For an array, add the element offset, explicitly scaled.
2620 const SCEV *LocalOffset = getSCEV(Index);
2621 if (!isa<PointerType>(LocalOffset->getType()))
2622 // Getelementptr indicies are signed.
2623 LocalOffset = getTruncateOrSignExtend(LocalOffset, IntPtrTy);
2624 // Lower "inbounds" GEPs to NSW arithmetic.
2625 LocalOffset = getMulExpr(LocalOffset, getAllocSizeExpr(*GTI),
2626 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2627 TotalOffset = getAddExpr(TotalOffset, LocalOffset,
2628 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2631 return getAddExpr(getSCEV(Base), TotalOffset,
2632 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2635 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2636 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2637 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2638 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2640 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2641 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2642 return C->getValue()->getValue().countTrailingZeros();
2644 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2645 return std::min(GetMinTrailingZeros(T->getOperand()),
2646 (uint32_t)getTypeSizeInBits(T->getType()));
2648 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2649 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2650 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2651 getTypeSizeInBits(E->getType()) : OpRes;
2654 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2655 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2656 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2657 getTypeSizeInBits(E->getType()) : OpRes;
2660 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2661 // The result is the min of all operands results.
2662 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2663 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2664 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2668 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2669 // The result is the sum of all operands results.
2670 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2671 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2672 for (unsigned i = 1, e = M->getNumOperands();
2673 SumOpRes != BitWidth && i != e; ++i)
2674 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2679 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2680 // The result is the min of all operands results.
2681 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2682 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2683 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2687 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2688 // The result is the min of all operands results.
2689 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2690 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2691 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2695 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2696 // The result is the min of all operands results.
2697 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2698 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2699 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2703 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2704 // For a SCEVUnknown, ask ValueTracking.
2705 unsigned BitWidth = getTypeSizeInBits(U->getType());
2706 APInt Mask = APInt::getAllOnesValue(BitWidth);
2707 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2708 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
2709 return Zeros.countTrailingOnes();
2716 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
2719 ScalarEvolution::getUnsignedRange(const SCEV *S) {
2721 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2722 return ConstantRange(C->getValue()->getValue());
2724 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2725 ConstantRange X = getUnsignedRange(Add->getOperand(0));
2726 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2727 X = X.add(getUnsignedRange(Add->getOperand(i)));
2731 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2732 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
2733 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2734 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
2738 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2739 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
2740 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2741 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
2745 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2746 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
2747 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2748 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
2752 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2753 ConstantRange X = getUnsignedRange(UDiv->getLHS());
2754 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
2758 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2759 ConstantRange X = getUnsignedRange(ZExt->getOperand());
2760 return X.zeroExtend(cast<IntegerType>(ZExt->getType())->getBitWidth());
2763 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2764 ConstantRange X = getUnsignedRange(SExt->getOperand());
2765 return X.signExtend(cast<IntegerType>(SExt->getType())->getBitWidth());
2768 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
2769 ConstantRange X = getUnsignedRange(Trunc->getOperand());
2770 return X.truncate(cast<IntegerType>(Trunc->getType())->getBitWidth());
2773 ConstantRange FullSet(getTypeSizeInBits(S->getType()), true);
2775 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
2776 const SCEV *T = getBackedgeTakenCount(AddRec->getLoop());
2777 const SCEVConstant *Trip = dyn_cast<SCEVConstant>(T);
2778 if (!Trip) return FullSet;
2780 // TODO: non-affine addrec
2781 if (AddRec->isAffine()) {
2782 const Type *Ty = AddRec->getType();
2783 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
2784 if (getTypeSizeInBits(MaxBECount->getType()) <= getTypeSizeInBits(Ty)) {
2785 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
2787 const SCEV *Start = AddRec->getStart();
2788 const SCEV *Step = AddRec->getStepRecurrence(*this);
2789 const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
2791 // Check for overflow.
2792 // TODO: This is very conservative.
2793 if (!(Step->isOne() &&
2794 isKnownPredicate(ICmpInst::ICMP_ULT, Start, End)) &&
2795 !(Step->isAllOnesValue() &&
2796 isKnownPredicate(ICmpInst::ICMP_UGT, Start, End)))
2799 ConstantRange StartRange = getUnsignedRange(Start);
2800 ConstantRange EndRange = getUnsignedRange(End);
2801 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
2802 EndRange.getUnsignedMin());
2803 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
2804 EndRange.getUnsignedMax());
2805 if (Min.isMinValue() && Max.isMaxValue())
2807 return ConstantRange(Min, Max+1);
2812 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2813 // For a SCEVUnknown, ask ValueTracking.
2814 unsigned BitWidth = getTypeSizeInBits(U->getType());
2815 APInt Mask = APInt::getAllOnesValue(BitWidth);
2816 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2817 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
2818 if (Ones == ~Zeros + 1)
2820 return ConstantRange(Ones, ~Zeros + 1);
2826 /// getSignedRange - Determine the signed range for a particular SCEV.
2829 ScalarEvolution::getSignedRange(const SCEV *S) {
2831 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2832 return ConstantRange(C->getValue()->getValue());
2834 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2835 ConstantRange X = getSignedRange(Add->getOperand(0));
2836 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2837 X = X.add(getSignedRange(Add->getOperand(i)));
2841 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2842 ConstantRange X = getSignedRange(Mul->getOperand(0));
2843 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2844 X = X.multiply(getSignedRange(Mul->getOperand(i)));
2848 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2849 ConstantRange X = getSignedRange(SMax->getOperand(0));
2850 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2851 X = X.smax(getSignedRange(SMax->getOperand(i)));
2855 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2856 ConstantRange X = getSignedRange(UMax->getOperand(0));
2857 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2858 X = X.umax(getSignedRange(UMax->getOperand(i)));
2862 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2863 ConstantRange X = getSignedRange(UDiv->getLHS());
2864 ConstantRange Y = getSignedRange(UDiv->getRHS());
2868 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2869 ConstantRange X = getSignedRange(ZExt->getOperand());
2870 return X.zeroExtend(cast<IntegerType>(ZExt->getType())->getBitWidth());
2873 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2874 ConstantRange X = getSignedRange(SExt->getOperand());
2875 return X.signExtend(cast<IntegerType>(SExt->getType())->getBitWidth());
2878 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
2879 ConstantRange X = getSignedRange(Trunc->getOperand());
2880 return X.truncate(cast<IntegerType>(Trunc->getType())->getBitWidth());
2883 ConstantRange FullSet(getTypeSizeInBits(S->getType()), true);
2885 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
2886 const SCEV *T = getBackedgeTakenCount(AddRec->getLoop());
2887 const SCEVConstant *Trip = dyn_cast<SCEVConstant>(T);
2888 if (!Trip) return FullSet;
2890 // TODO: non-affine addrec
2891 if (AddRec->isAffine()) {
2892 const Type *Ty = AddRec->getType();
2893 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
2894 if (getTypeSizeInBits(MaxBECount->getType()) <= getTypeSizeInBits(Ty)) {
2895 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
2897 const SCEV *Start = AddRec->getStart();
2898 const SCEV *Step = AddRec->getStepRecurrence(*this);
2899 const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
2901 // Check for overflow.
2902 // TODO: This is very conservative.
2903 if (!(Step->isOne() &&
2904 isKnownPredicate(ICmpInst::ICMP_SLT, Start, End)) &&
2905 !(Step->isAllOnesValue() &&
2906 isKnownPredicate(ICmpInst::ICMP_SGT, Start, End)))
2909 ConstantRange StartRange = getSignedRange(Start);
2910 ConstantRange EndRange = getSignedRange(End);
2911 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
2912 EndRange.getSignedMin());
2913 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
2914 EndRange.getSignedMax());
2915 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
2917 return ConstantRange(Min, Max+1);
2922 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2923 // For a SCEVUnknown, ask ValueTracking.
2924 unsigned BitWidth = getTypeSizeInBits(U->getType());
2925 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
2929 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
2930 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1);
2936 /// createSCEV - We know that there is no SCEV for the specified value.
2937 /// Analyze the expression.
2939 const SCEV *ScalarEvolution::createSCEV(Value *V) {
2940 if (!isSCEVable(V->getType()))
2941 return getUnknown(V);
2943 unsigned Opcode = Instruction::UserOp1;
2944 if (Instruction *I = dyn_cast<Instruction>(V))
2945 Opcode = I->getOpcode();
2946 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2947 Opcode = CE->getOpcode();
2948 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2949 return getConstant(CI);
2950 else if (isa<ConstantPointerNull>(V))
2951 return getIntegerSCEV(0, V->getType());
2952 else if (isa<UndefValue>(V))
2953 return getIntegerSCEV(0, V->getType());
2954 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
2955 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
2957 return getUnknown(V);
2959 Operator *U = cast<Operator>(V);
2961 case Instruction::Add:
2962 // Don't transfer the NSW and NUW bits from the Add instruction to the
2963 // Add expression, because the Instruction may be guarded by control
2964 // flow and the no-overflow bits may not be valid for the expression in
2966 return getAddExpr(getSCEV(U->getOperand(0)),
2967 getSCEV(U->getOperand(1)));
2968 case Instruction::Mul:
2969 // Don't transfer the NSW and NUW bits from the Mul instruction to the
2970 // Mul expression, as with Add.
2971 return getMulExpr(getSCEV(U->getOperand(0)),
2972 getSCEV(U->getOperand(1)));
2973 case Instruction::UDiv:
2974 return getUDivExpr(getSCEV(U->getOperand(0)),
2975 getSCEV(U->getOperand(1)));
2976 case Instruction::Sub:
2977 return getMinusSCEV(getSCEV(U->getOperand(0)),
2978 getSCEV(U->getOperand(1)));
2979 case Instruction::And:
2980 // For an expression like x&255 that merely masks off the high bits,
2981 // use zext(trunc(x)) as the SCEV expression.
2982 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2983 if (CI->isNullValue())
2984 return getSCEV(U->getOperand(1));
2985 if (CI->isAllOnesValue())
2986 return getSCEV(U->getOperand(0));
2987 const APInt &A = CI->getValue();
2989 // Instcombine's ShrinkDemandedConstant may strip bits out of
2990 // constants, obscuring what would otherwise be a low-bits mask.
2991 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
2992 // knew about to reconstruct a low-bits mask value.
2993 unsigned LZ = A.countLeadingZeros();
2994 unsigned BitWidth = A.getBitWidth();
2995 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
2996 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2997 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
2999 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3001 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3003 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3004 IntegerType::get(getContext(), BitWidth - LZ)),
3009 case Instruction::Or:
3010 // If the RHS of the Or is a constant, we may have something like:
3011 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3012 // optimizations will transparently handle this case.
3014 // In order for this transformation to be safe, the LHS must be of the
3015 // form X*(2^n) and the Or constant must be less than 2^n.
3016 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3017 const SCEV *LHS = getSCEV(U->getOperand(0));
3018 const APInt &CIVal = CI->getValue();
3019 if (GetMinTrailingZeros(LHS) >=
3020 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3021 // Build a plain add SCEV.
3022 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3023 // If the LHS of the add was an addrec and it has no-wrap flags,
3024 // transfer the no-wrap flags, since an or won't introduce a wrap.
3025 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3026 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3027 if (OldAR->hasNoUnsignedWrap())
3028 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3029 if (OldAR->hasNoSignedWrap())
3030 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3036 case Instruction::Xor:
3037 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3038 // If the RHS of the xor is a signbit, then this is just an add.
3039 // Instcombine turns add of signbit into xor as a strength reduction step.
3040 if (CI->getValue().isSignBit())
3041 return getAddExpr(getSCEV(U->getOperand(0)),
3042 getSCEV(U->getOperand(1)));
3044 // If the RHS of xor is -1, then this is a not operation.
3045 if (CI->isAllOnesValue())
3046 return getNotSCEV(getSCEV(U->getOperand(0)));
3048 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3049 // This is a variant of the check for xor with -1, and it handles
3050 // the case where instcombine has trimmed non-demanded bits out
3051 // of an xor with -1.
3052 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3053 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3054 if (BO->getOpcode() == Instruction::And &&
3055 LCI->getValue() == CI->getValue())
3056 if (const SCEVZeroExtendExpr *Z =
3057 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3058 const Type *UTy = U->getType();
3059 const SCEV *Z0 = Z->getOperand();
3060 const Type *Z0Ty = Z0->getType();
3061 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3063 // If C is a low-bits mask, the zero extend is zerving to
3064 // mask off the high bits. Complement the operand and
3065 // re-apply the zext.
3066 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3067 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3069 // If C is a single bit, it may be in the sign-bit position
3070 // before the zero-extend. In this case, represent the xor
3071 // using an add, which is equivalent, and re-apply the zext.
3072 APInt Trunc = APInt(CI->getValue()).trunc(Z0TySize);
3073 if (APInt(Trunc).zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3075 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3081 case Instruction::Shl:
3082 // Turn shift left of a constant amount into a multiply.
3083 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3084 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
3085 Constant *X = ConstantInt::get(getContext(),
3086 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
3087 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3091 case Instruction::LShr:
3092 // Turn logical shift right of a constant into a unsigned divide.
3093 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3094 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
3095 Constant *X = ConstantInt::get(getContext(),
3096 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
3097 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3101 case Instruction::AShr:
3102 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3103 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3104 if (Instruction *L = dyn_cast<Instruction>(U->getOperand(0)))
3105 if (L->getOpcode() == Instruction::Shl &&
3106 L->getOperand(1) == U->getOperand(1)) {
3107 unsigned BitWidth = getTypeSizeInBits(U->getType());
3108 uint64_t Amt = BitWidth - CI->getZExtValue();
3109 if (Amt == BitWidth)
3110 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3112 return getIntegerSCEV(0, U->getType()); // value is undefined
3114 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3115 IntegerType::get(getContext(), Amt)),
3120 case Instruction::Trunc:
3121 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3123 case Instruction::ZExt:
3124 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3126 case Instruction::SExt:
3127 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3129 case Instruction::BitCast:
3130 // BitCasts are no-op casts so we just eliminate the cast.
3131 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3132 return getSCEV(U->getOperand(0));
3135 // It's tempting to handle inttoptr and ptrtoint, however this can
3136 // lead to pointer expressions which cannot be expanded to GEPs
3137 // (because they may overflow). For now, the only pointer-typed
3138 // expressions we handle are GEPs and address literals.
3140 case Instruction::GetElementPtr:
3141 return createNodeForGEP(cast<GEPOperator>(U));
3143 case Instruction::PHI:
3144 return createNodeForPHI(cast<PHINode>(U));
3146 case Instruction::Select:
3147 // This could be a smax or umax that was lowered earlier.
3148 // Try to recover it.
3149 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3150 Value *LHS = ICI->getOperand(0);
3151 Value *RHS = ICI->getOperand(1);
3152 switch (ICI->getPredicate()) {
3153 case ICmpInst::ICMP_SLT:
3154 case ICmpInst::ICMP_SLE:
3155 std::swap(LHS, RHS);
3157 case ICmpInst::ICMP_SGT:
3158 case ICmpInst::ICMP_SGE:
3159 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
3160 return getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
3161 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
3162 return getSMinExpr(getSCEV(LHS), getSCEV(RHS));
3164 case ICmpInst::ICMP_ULT:
3165 case ICmpInst::ICMP_ULE:
3166 std::swap(LHS, RHS);
3168 case ICmpInst::ICMP_UGT:
3169 case ICmpInst::ICMP_UGE:
3170 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
3171 return getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
3172 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
3173 return getUMinExpr(getSCEV(LHS), getSCEV(RHS));
3175 case ICmpInst::ICMP_NE:
3176 // n != 0 ? n : 1 -> umax(n, 1)
3177 if (LHS == U->getOperand(1) &&
3178 isa<ConstantInt>(U->getOperand(2)) &&
3179 cast<ConstantInt>(U->getOperand(2))->isOne() &&
3180 isa<ConstantInt>(RHS) &&
3181 cast<ConstantInt>(RHS)->isZero())
3182 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(2)));
3184 case ICmpInst::ICMP_EQ:
3185 // n == 0 ? 1 : n -> umax(n, 1)
3186 if (LHS == U->getOperand(2) &&
3187 isa<ConstantInt>(U->getOperand(1)) &&
3188 cast<ConstantInt>(U->getOperand(1))->isOne() &&
3189 isa<ConstantInt>(RHS) &&
3190 cast<ConstantInt>(RHS)->isZero())
3191 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(1)));
3198 default: // We cannot analyze this expression.
3202 return getUnknown(V);
3207 //===----------------------------------------------------------------------===//
3208 // Iteration Count Computation Code
3211 /// getBackedgeTakenCount - If the specified loop has a predictable
3212 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3213 /// object. The backedge-taken count is the number of times the loop header
3214 /// will be branched to from within the loop. This is one less than the
3215 /// trip count of the loop, since it doesn't count the first iteration,
3216 /// when the header is branched to from outside the loop.
3218 /// Note that it is not valid to call this method on a loop without a
3219 /// loop-invariant backedge-taken count (see
3220 /// hasLoopInvariantBackedgeTakenCount).
3222 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3223 return getBackedgeTakenInfo(L).Exact;
3226 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3227 /// return the least SCEV value that is known never to be less than the
3228 /// actual backedge taken count.
3229 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3230 return getBackedgeTakenInfo(L).Max;
3233 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3234 /// onto the given Worklist.
3236 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3237 BasicBlock *Header = L->getHeader();
3239 // Push all Loop-header PHIs onto the Worklist stack.
3240 for (BasicBlock::iterator I = Header->begin();
3241 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3242 Worklist.push_back(PN);
3245 const ScalarEvolution::BackedgeTakenInfo &
3246 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3247 // Initially insert a CouldNotCompute for this loop. If the insertion
3248 // succeeds, procede to actually compute a backedge-taken count and
3249 // update the value. The temporary CouldNotCompute value tells SCEV
3250 // code elsewhere that it shouldn't attempt to request a new
3251 // backedge-taken count, which could result in infinite recursion.
3252 std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3253 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3255 BackedgeTakenInfo ItCount = ComputeBackedgeTakenCount(L);
3256 if (ItCount.Exact != getCouldNotCompute()) {
3257 assert(ItCount.Exact->isLoopInvariant(L) &&
3258 ItCount.Max->isLoopInvariant(L) &&
3259 "Computed trip count isn't loop invariant for loop!");
3260 ++NumTripCountsComputed;
3262 // Update the value in the map.
3263 Pair.first->second = ItCount;
3265 if (ItCount.Max != getCouldNotCompute())
3266 // Update the value in the map.
3267 Pair.first->second = ItCount;
3268 if (isa<PHINode>(L->getHeader()->begin()))
3269 // Only count loops that have phi nodes as not being computable.
3270 ++NumTripCountsNotComputed;
3273 // Now that we know more about the trip count for this loop, forget any
3274 // existing SCEV values for PHI nodes in this loop since they are only
3275 // conservative estimates made without the benefit of trip count
3276 // information. This is similar to the code in forgetLoop, except that
3277 // it handles SCEVUnknown PHI nodes specially.
3278 if (ItCount.hasAnyInfo()) {
3279 SmallVector<Instruction *, 16> Worklist;
3280 PushLoopPHIs(L, Worklist);
3282 SmallPtrSet<Instruction *, 8> Visited;
3283 while (!Worklist.empty()) {
3284 Instruction *I = Worklist.pop_back_val();
3285 if (!Visited.insert(I)) continue;
3287 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
3288 Scalars.find(static_cast<Value *>(I));
3289 if (It != Scalars.end()) {
3290 // SCEVUnknown for a PHI either means that it has an unrecognized
3291 // structure, or it's a PHI that's in the progress of being computed
3292 // by createNodeForPHI. In the former case, additional loop trip
3293 // count information isn't going to change anything. In the later
3294 // case, createNodeForPHI will perform the necessary updates on its
3295 // own when it gets to that point.
3296 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
3297 ValuesAtScopes.erase(It->second);
3300 if (PHINode *PN = dyn_cast<PHINode>(I))
3301 ConstantEvolutionLoopExitValue.erase(PN);
3304 PushDefUseChildren(I, Worklist);
3308 return Pair.first->second;
3311 /// forgetLoop - This method should be called by the client when it has
3312 /// changed a loop in a way that may effect ScalarEvolution's ability to
3313 /// compute a trip count, or if the loop is deleted.
3314 void ScalarEvolution::forgetLoop(const Loop *L) {
3315 // Drop any stored trip count value.
3316 BackedgeTakenCounts.erase(L);
3318 // Drop information about expressions based on loop-header PHIs.
3319 SmallVector<Instruction *, 16> Worklist;
3320 PushLoopPHIs(L, Worklist);
3322 SmallPtrSet<Instruction *, 8> Visited;
3323 while (!Worklist.empty()) {
3324 Instruction *I = Worklist.pop_back_val();
3325 if (!Visited.insert(I)) continue;
3327 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
3328 Scalars.find(static_cast<Value *>(I));
3329 if (It != Scalars.end()) {
3330 ValuesAtScopes.erase(It->second);
3332 if (PHINode *PN = dyn_cast<PHINode>(I))
3333 ConstantEvolutionLoopExitValue.erase(PN);
3336 PushDefUseChildren(I, Worklist);
3340 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3341 /// of the specified loop will execute.
3342 ScalarEvolution::BackedgeTakenInfo
3343 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3344 SmallVector<BasicBlock *, 8> ExitingBlocks;
3345 L->getExitingBlocks(ExitingBlocks);
3347 // Examine all exits and pick the most conservative values.
3348 const SCEV *BECount = getCouldNotCompute();
3349 const SCEV *MaxBECount = getCouldNotCompute();
3350 bool CouldNotComputeBECount = false;
3351 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3352 BackedgeTakenInfo NewBTI =
3353 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3355 if (NewBTI.Exact == getCouldNotCompute()) {
3356 // We couldn't compute an exact value for this exit, so
3357 // we won't be able to compute an exact value for the loop.
3358 CouldNotComputeBECount = true;
3359 BECount = getCouldNotCompute();
3360 } else if (!CouldNotComputeBECount) {
3361 if (BECount == getCouldNotCompute())
3362 BECount = NewBTI.Exact;
3364 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3366 if (MaxBECount == getCouldNotCompute())
3367 MaxBECount = NewBTI.Max;
3368 else if (NewBTI.Max != getCouldNotCompute())
3369 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3372 return BackedgeTakenInfo(BECount, MaxBECount);
3375 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3376 /// of the specified loop will execute if it exits via the specified block.
3377 ScalarEvolution::BackedgeTakenInfo
3378 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3379 BasicBlock *ExitingBlock) {
3381 // Okay, we've chosen an exiting block. See what condition causes us to
3382 // exit at this block.
3384 // FIXME: we should be able to handle switch instructions (with a single exit)
3385 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3386 if (ExitBr == 0) return getCouldNotCompute();
3387 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3389 // At this point, we know we have a conditional branch that determines whether
3390 // the loop is exited. However, we don't know if the branch is executed each
3391 // time through the loop. If not, then the execution count of the branch will
3392 // not be equal to the trip count of the loop.
3394 // Currently we check for this by checking to see if the Exit branch goes to
3395 // the loop header. If so, we know it will always execute the same number of
3396 // times as the loop. We also handle the case where the exit block *is* the
3397 // loop header. This is common for un-rotated loops.
3399 // If both of those tests fail, walk up the unique predecessor chain to the
3400 // header, stopping if there is an edge that doesn't exit the loop. If the
3401 // header is reached, the execution count of the branch will be equal to the
3402 // trip count of the loop.
3404 // More extensive analysis could be done to handle more cases here.
3406 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3407 ExitBr->getSuccessor(1) != L->getHeader() &&
3408 ExitBr->getParent() != L->getHeader()) {
3409 // The simple checks failed, try climbing the unique predecessor chain
3410 // up to the header.
3412 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3413 BasicBlock *Pred = BB->getUniquePredecessor();
3415 return getCouldNotCompute();
3416 TerminatorInst *PredTerm = Pred->getTerminator();
3417 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3418 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3421 // If the predecessor has a successor that isn't BB and isn't
3422 // outside the loop, assume the worst.
3423 if (L->contains(PredSucc))
3424 return getCouldNotCompute();
3426 if (Pred == L->getHeader()) {
3433 return getCouldNotCompute();
3436 // Procede to the next level to examine the exit condition expression.
3437 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3438 ExitBr->getSuccessor(0),
3439 ExitBr->getSuccessor(1));
3442 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3443 /// backedge of the specified loop will execute if its exit condition
3444 /// were a conditional branch of ExitCond, TBB, and FBB.
3445 ScalarEvolution::BackedgeTakenInfo
3446 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3450 // Check if the controlling expression for this loop is an And or Or.
3451 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3452 if (BO->getOpcode() == Instruction::And) {
3453 // Recurse on the operands of the and.
3454 BackedgeTakenInfo BTI0 =
3455 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3456 BackedgeTakenInfo BTI1 =
3457 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3458 const SCEV *BECount = getCouldNotCompute();
3459 const SCEV *MaxBECount = getCouldNotCompute();
3460 if (L->contains(TBB)) {
3461 // Both conditions must be true for the loop to continue executing.
3462 // Choose the less conservative count.
3463 if (BTI0.Exact == getCouldNotCompute() ||
3464 BTI1.Exact == getCouldNotCompute())
3465 BECount = getCouldNotCompute();
3467 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3468 if (BTI0.Max == getCouldNotCompute())
3469 MaxBECount = BTI1.Max;
3470 else if (BTI1.Max == getCouldNotCompute())
3471 MaxBECount = BTI0.Max;
3473 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3475 // Both conditions must be true for the loop to exit.
3476 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3477 if (BTI0.Exact != getCouldNotCompute() &&
3478 BTI1.Exact != getCouldNotCompute())
3479 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3480 if (BTI0.Max != getCouldNotCompute() &&
3481 BTI1.Max != getCouldNotCompute())
3482 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
3485 return BackedgeTakenInfo(BECount, MaxBECount);
3487 if (BO->getOpcode() == Instruction::Or) {
3488 // Recurse on the operands of the or.
3489 BackedgeTakenInfo BTI0 =
3490 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3491 BackedgeTakenInfo BTI1 =
3492 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3493 const SCEV *BECount = getCouldNotCompute();
3494 const SCEV *MaxBECount = getCouldNotCompute();
3495 if (L->contains(FBB)) {
3496 // Both conditions must be false for the loop to continue executing.
3497 // Choose the less conservative count.
3498 if (BTI0.Exact == getCouldNotCompute() ||
3499 BTI1.Exact == getCouldNotCompute())
3500 BECount = getCouldNotCompute();
3502 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3503 if (BTI0.Max == getCouldNotCompute())
3504 MaxBECount = BTI1.Max;
3505 else if (BTI1.Max == getCouldNotCompute())
3506 MaxBECount = BTI0.Max;
3508 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3510 // Both conditions must be false for the loop to exit.
3511 assert(L->contains(TBB) && "Loop block has no successor in loop!");
3512 if (BTI0.Exact != getCouldNotCompute() &&
3513 BTI1.Exact != getCouldNotCompute())
3514 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3515 if (BTI0.Max != getCouldNotCompute() &&
3516 BTI1.Max != getCouldNotCompute())
3517 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
3520 return BackedgeTakenInfo(BECount, MaxBECount);
3524 // With an icmp, it may be feasible to compute an exact backedge-taken count.
3525 // Procede to the next level to examine the icmp.
3526 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
3527 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
3529 // If it's not an integer or pointer comparison then compute it the hard way.
3530 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3533 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
3534 /// backedge of the specified loop will execute if its exit condition
3535 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
3536 ScalarEvolution::BackedgeTakenInfo
3537 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
3542 // If the condition was exit on true, convert the condition to exit on false
3543 ICmpInst::Predicate Cond;
3544 if (!L->contains(FBB))
3545 Cond = ExitCond->getPredicate();
3547 Cond = ExitCond->getInversePredicate();
3549 // Handle common loops like: for (X = "string"; *X; ++X)
3550 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
3551 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
3553 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
3554 if (!isa<SCEVCouldNotCompute>(ItCnt)) {
3555 unsigned BitWidth = getTypeSizeInBits(ItCnt->getType());
3556 return BackedgeTakenInfo(ItCnt,
3557 isa<SCEVConstant>(ItCnt) ? ItCnt :
3558 getConstant(APInt::getMaxValue(BitWidth)-1));
3562 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
3563 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
3565 // Try to evaluate any dependencies out of the loop.
3566 LHS = getSCEVAtScope(LHS, L);
3567 RHS = getSCEVAtScope(RHS, L);
3569 // At this point, we would like to compute how many iterations of the
3570 // loop the predicate will return true for these inputs.
3571 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
3572 // If there is a loop-invariant, force it into the RHS.
3573 std::swap(LHS, RHS);
3574 Cond = ICmpInst::getSwappedPredicate(Cond);
3577 // If we have a comparison of a chrec against a constant, try to use value
3578 // ranges to answer this query.
3579 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
3580 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
3581 if (AddRec->getLoop() == L) {
3582 // Form the constant range.
3583 ConstantRange CompRange(
3584 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
3586 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
3587 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
3591 case ICmpInst::ICMP_NE: { // while (X != Y)
3592 // Convert to: while (X-Y != 0)
3593 const SCEV *TC = HowFarToZero(getMinusSCEV(LHS, RHS), L);
3594 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3597 case ICmpInst::ICMP_EQ: { // while (X == Y)
3598 // Convert to: while (X-Y == 0)
3599 const SCEV *TC = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
3600 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3603 case ICmpInst::ICMP_SLT: {
3604 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
3605 if (BTI.hasAnyInfo()) return BTI;
3608 case ICmpInst::ICMP_SGT: {
3609 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3610 getNotSCEV(RHS), L, true);
3611 if (BTI.hasAnyInfo()) return BTI;
3614 case ICmpInst::ICMP_ULT: {
3615 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
3616 if (BTI.hasAnyInfo()) return BTI;
3619 case ICmpInst::ICMP_UGT: {
3620 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3621 getNotSCEV(RHS), L, false);
3622 if (BTI.hasAnyInfo()) return BTI;
3627 errs() << "ComputeBackedgeTakenCount ";
3628 if (ExitCond->getOperand(0)->getType()->isUnsigned())
3629 errs() << "[unsigned] ";
3630 errs() << *LHS << " "
3631 << Instruction::getOpcodeName(Instruction::ICmp)
3632 << " " << *RHS << "\n";
3637 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3640 static ConstantInt *
3641 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
3642 ScalarEvolution &SE) {
3643 const SCEV *InVal = SE.getConstant(C);
3644 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
3645 assert(isa<SCEVConstant>(Val) &&
3646 "Evaluation of SCEV at constant didn't fold correctly?");
3647 return cast<SCEVConstant>(Val)->getValue();
3650 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
3651 /// and a GEP expression (missing the pointer index) indexing into it, return
3652 /// the addressed element of the initializer or null if the index expression is
3655 GetAddressedElementFromGlobal(GlobalVariable *GV,
3656 const std::vector<ConstantInt*> &Indices) {
3657 Constant *Init = GV->getInitializer();
3658 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
3659 uint64_t Idx = Indices[i]->getZExtValue();
3660 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
3661 assert(Idx < CS->getNumOperands() && "Bad struct index!");
3662 Init = cast<Constant>(CS->getOperand(Idx));
3663 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
3664 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
3665 Init = cast<Constant>(CA->getOperand(Idx));
3666 } else if (isa<ConstantAggregateZero>(Init)) {
3667 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
3668 assert(Idx < STy->getNumElements() && "Bad struct index!");
3669 Init = Constant::getNullValue(STy->getElementType(Idx));
3670 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
3671 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
3672 Init = Constant::getNullValue(ATy->getElementType());
3674 llvm_unreachable("Unknown constant aggregate type!");
3678 return 0; // Unknown initializer type
3684 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
3685 /// 'icmp op load X, cst', try to see if we can compute the backedge
3686 /// execution count.
3688 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
3692 ICmpInst::Predicate predicate) {
3693 if (LI->isVolatile()) return getCouldNotCompute();
3695 // Check to see if the loaded pointer is a getelementptr of a global.
3696 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
3697 if (!GEP) return getCouldNotCompute();
3699 // Make sure that it is really a constant global we are gepping, with an
3700 // initializer, and make sure the first IDX is really 0.
3701 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
3702 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
3703 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
3704 !cast<Constant>(GEP->getOperand(1))->isNullValue())
3705 return getCouldNotCompute();
3707 // Okay, we allow one non-constant index into the GEP instruction.
3709 std::vector<ConstantInt*> Indexes;
3710 unsigned VarIdxNum = 0;
3711 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
3712 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
3713 Indexes.push_back(CI);
3714 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
3715 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
3716 VarIdx = GEP->getOperand(i);
3718 Indexes.push_back(0);
3721 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
3722 // Check to see if X is a loop variant variable value now.
3723 const SCEV *Idx = getSCEV(VarIdx);
3724 Idx = getSCEVAtScope(Idx, L);
3726 // We can only recognize very limited forms of loop index expressions, in
3727 // particular, only affine AddRec's like {C1,+,C2}.
3728 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
3729 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
3730 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
3731 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
3732 return getCouldNotCompute();
3734 unsigned MaxSteps = MaxBruteForceIterations;
3735 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
3736 ConstantInt *ItCst = ConstantInt::get(
3737 cast<IntegerType>(IdxExpr->getType()), IterationNum);
3738 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
3740 // Form the GEP offset.
3741 Indexes[VarIdxNum] = Val;
3743 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
3744 if (Result == 0) break; // Cannot compute!
3746 // Evaluate the condition for this iteration.
3747 Result = ConstantExpr::getICmp(predicate, Result, RHS);
3748 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
3749 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
3751 errs() << "\n***\n*** Computed loop count " << *ItCst
3752 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
3755 ++NumArrayLenItCounts;
3756 return getConstant(ItCst); // Found terminating iteration!
3759 return getCouldNotCompute();
3763 /// CanConstantFold - Return true if we can constant fold an instruction of the
3764 /// specified type, assuming that all operands were constants.
3765 static bool CanConstantFold(const Instruction *I) {
3766 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
3767 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
3770 if (const CallInst *CI = dyn_cast<CallInst>(I))
3771 if (const Function *F = CI->getCalledFunction())
3772 return canConstantFoldCallTo(F);
3776 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
3777 /// in the loop that V is derived from. We allow arbitrary operations along the
3778 /// way, but the operands of an operation must either be constants or a value
3779 /// derived from a constant PHI. If this expression does not fit with these
3780 /// constraints, return null.
3781 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
3782 // If this is not an instruction, or if this is an instruction outside of the
3783 // loop, it can't be derived from a loop PHI.
3784 Instruction *I = dyn_cast<Instruction>(V);
3785 if (I == 0 || !L->contains(I)) return 0;
3787 if (PHINode *PN = dyn_cast<PHINode>(I)) {
3788 if (L->getHeader() == I->getParent())
3791 // We don't currently keep track of the control flow needed to evaluate
3792 // PHIs, so we cannot handle PHIs inside of loops.
3796 // If we won't be able to constant fold this expression even if the operands
3797 // are constants, return early.
3798 if (!CanConstantFold(I)) return 0;
3800 // Otherwise, we can evaluate this instruction if all of its operands are
3801 // constant or derived from a PHI node themselves.
3803 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
3804 if (!(isa<Constant>(I->getOperand(Op)) ||
3805 isa<GlobalValue>(I->getOperand(Op)))) {
3806 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
3807 if (P == 0) return 0; // Not evolving from PHI
3811 return 0; // Evolving from multiple different PHIs.
3814 // This is a expression evolving from a constant PHI!
3818 /// EvaluateExpression - Given an expression that passes the
3819 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
3820 /// in the loop has the value PHIVal. If we can't fold this expression for some
3821 /// reason, return null.
3822 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
3823 const TargetData *TD) {
3824 if (isa<PHINode>(V)) return PHIVal;
3825 if (Constant *C = dyn_cast<Constant>(V)) return C;
3826 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
3827 Instruction *I = cast<Instruction>(V);
3829 std::vector<Constant*> Operands;
3830 Operands.resize(I->getNumOperands());
3832 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3833 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
3834 if (Operands[i] == 0) return 0;
3837 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
3838 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
3840 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
3841 &Operands[0], Operands.size(), TD);
3844 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
3845 /// in the header of its containing loop, we know the loop executes a
3846 /// constant number of times, and the PHI node is just a recurrence
3847 /// involving constants, fold it.
3849 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
3852 std::map<PHINode*, Constant*>::iterator I =
3853 ConstantEvolutionLoopExitValue.find(PN);
3854 if (I != ConstantEvolutionLoopExitValue.end())
3857 if (BEs.ugt(APInt(BEs.getBitWidth(),MaxBruteForceIterations)))
3858 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
3860 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
3862 // Since the loop is canonicalized, the PHI node must have two entries. One
3863 // entry must be a constant (coming in from outside of the loop), and the
3864 // second must be derived from the same PHI.
3865 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
3866 Constant *StartCST =
3867 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
3869 return RetVal = 0; // Must be a constant.
3871 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
3872 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
3874 return RetVal = 0; // Not derived from same PHI.
3876 // Execute the loop symbolically to determine the exit value.
3877 if (BEs.getActiveBits() >= 32)
3878 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
3880 unsigned NumIterations = BEs.getZExtValue(); // must be in range
3881 unsigned IterationNum = 0;
3882 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
3883 if (IterationNum == NumIterations)
3884 return RetVal = PHIVal; // Got exit value!
3886 // Compute the value of the PHI node for the next iteration.
3887 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
3888 if (NextPHI == PHIVal)
3889 return RetVal = NextPHI; // Stopped evolving!
3891 return 0; // Couldn't evaluate!
3896 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
3897 /// constant number of times (the condition evolves only from constants),
3898 /// try to evaluate a few iterations of the loop until we get the exit
3899 /// condition gets a value of ExitWhen (true or false). If we cannot
3900 /// evaluate the trip count of the loop, return getCouldNotCompute().
3902 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
3905 PHINode *PN = getConstantEvolvingPHI(Cond, L);
3906 if (PN == 0) return getCouldNotCompute();
3908 // Since the loop is canonicalized, the PHI node must have two entries. One
3909 // entry must be a constant (coming in from outside of the loop), and the
3910 // second must be derived from the same PHI.
3911 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
3912 Constant *StartCST =
3913 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
3914 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
3916 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
3917 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
3918 if (PN2 != PN) return getCouldNotCompute(); // Not derived from same PHI.
3920 // Okay, we find a PHI node that defines the trip count of this loop. Execute
3921 // the loop symbolically to determine when the condition gets a value of
3923 unsigned IterationNum = 0;
3924 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
3925 for (Constant *PHIVal = StartCST;
3926 IterationNum != MaxIterations; ++IterationNum) {
3927 ConstantInt *CondVal =
3928 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
3930 // Couldn't symbolically evaluate.
3931 if (!CondVal) return getCouldNotCompute();
3933 if (CondVal->getValue() == uint64_t(ExitWhen)) {
3934 ++NumBruteForceTripCountsComputed;
3935 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
3938 // Compute the value of the PHI node for the next iteration.
3939 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
3940 if (NextPHI == 0 || NextPHI == PHIVal)
3941 return getCouldNotCompute();// Couldn't evaluate or not making progress...
3945 // Too many iterations were needed to evaluate.
3946 return getCouldNotCompute();
3949 /// getSCEVAtScope - Return a SCEV expression for the specified value
3950 /// at the specified scope in the program. The L value specifies a loop
3951 /// nest to evaluate the expression at, where null is the top-level or a
3952 /// specified loop is immediately inside of the loop.
3954 /// This method can be used to compute the exit value for a variable defined
3955 /// in a loop by querying what the value will hold in the parent loop.
3957 /// In the case that a relevant loop exit value cannot be computed, the
3958 /// original value V is returned.
3959 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
3960 // Check to see if we've folded this expression at this loop before.
3961 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
3962 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
3963 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
3965 return Pair.first->second ? Pair.first->second : V;
3967 // Otherwise compute it.
3968 const SCEV *C = computeSCEVAtScope(V, L);
3969 ValuesAtScopes[V][L] = C;
3973 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
3974 if (isa<SCEVConstant>(V)) return V;
3976 // If this instruction is evolved from a constant-evolving PHI, compute the
3977 // exit value from the loop without using SCEVs.
3978 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
3979 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
3980 const Loop *LI = (*this->LI)[I->getParent()];
3981 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
3982 if (PHINode *PN = dyn_cast<PHINode>(I))
3983 if (PN->getParent() == LI->getHeader()) {
3984 // Okay, there is no closed form solution for the PHI node. Check
3985 // to see if the loop that contains it has a known backedge-taken
3986 // count. If so, we may be able to force computation of the exit
3988 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
3989 if (const SCEVConstant *BTCC =
3990 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
3991 // Okay, we know how many times the containing loop executes. If
3992 // this is a constant evolving PHI node, get the final value at
3993 // the specified iteration number.
3994 Constant *RV = getConstantEvolutionLoopExitValue(PN,
3995 BTCC->getValue()->getValue(),
3997 if (RV) return getSCEV(RV);
4001 // Okay, this is an expression that we cannot symbolically evaluate
4002 // into a SCEV. Check to see if it's possible to symbolically evaluate
4003 // the arguments into constants, and if so, try to constant propagate the
4004 // result. This is particularly useful for computing loop exit values.
4005 if (CanConstantFold(I)) {
4006 std::vector<Constant*> Operands;
4007 Operands.reserve(I->getNumOperands());
4008 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4009 Value *Op = I->getOperand(i);
4010 if (Constant *C = dyn_cast<Constant>(Op)) {
4011 Operands.push_back(C);
4013 // If any of the operands is non-constant and if they are
4014 // non-integer and non-pointer, don't even try to analyze them
4015 // with scev techniques.
4016 if (!isSCEVable(Op->getType()))
4019 const SCEV *OpV = getSCEVAtScope(Op, L);
4020 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
4021 Constant *C = SC->getValue();
4022 if (C->getType() != Op->getType())
4023 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4027 Operands.push_back(C);
4028 } else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
4029 if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
4030 if (C->getType() != Op->getType())
4032 ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4036 Operands.push_back(C);
4046 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4047 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4048 Operands[0], Operands[1], TD);
4050 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4051 &Operands[0], Operands.size(), TD);
4056 // This is some other type of SCEVUnknown, just return it.
4060 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4061 // Avoid performing the look-up in the common case where the specified
4062 // expression has no loop-variant portions.
4063 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4064 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4065 if (OpAtScope != Comm->getOperand(i)) {
4066 // Okay, at least one of these operands is loop variant but might be
4067 // foldable. Build a new instance of the folded commutative expression.
4068 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4069 Comm->op_begin()+i);
4070 NewOps.push_back(OpAtScope);
4072 for (++i; i != e; ++i) {
4073 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4074 NewOps.push_back(OpAtScope);
4076 if (isa<SCEVAddExpr>(Comm))
4077 return getAddExpr(NewOps);
4078 if (isa<SCEVMulExpr>(Comm))
4079 return getMulExpr(NewOps);
4080 if (isa<SCEVSMaxExpr>(Comm))
4081 return getSMaxExpr(NewOps);
4082 if (isa<SCEVUMaxExpr>(Comm))
4083 return getUMaxExpr(NewOps);
4084 llvm_unreachable("Unknown commutative SCEV type!");
4087 // If we got here, all operands are loop invariant.
4091 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4092 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4093 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4094 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4095 return Div; // must be loop invariant
4096 return getUDivExpr(LHS, RHS);
4099 // If this is a loop recurrence for a loop that does not contain L, then we
4100 // are dealing with the final value computed by the loop.
4101 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4102 if (!L || !AddRec->getLoop()->contains(L)) {
4103 // To evaluate this recurrence, we need to know how many times the AddRec
4104 // loop iterates. Compute this now.
4105 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4106 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4108 // Then, evaluate the AddRec.
4109 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4114 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4115 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4116 if (Op == Cast->getOperand())
4117 return Cast; // must be loop invariant
4118 return getZeroExtendExpr(Op, Cast->getType());
4121 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4122 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4123 if (Op == Cast->getOperand())
4124 return Cast; // must be loop invariant
4125 return getSignExtendExpr(Op, Cast->getType());
4128 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4129 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4130 if (Op == Cast->getOperand())
4131 return Cast; // must be loop invariant
4132 return getTruncateExpr(Op, Cast->getType());
4135 if (isa<SCEVTargetDataConstant>(V))
4138 llvm_unreachable("Unknown SCEV type!");
4142 /// getSCEVAtScope - This is a convenience function which does
4143 /// getSCEVAtScope(getSCEV(V), L).
4144 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4145 return getSCEVAtScope(getSCEV(V), L);
4148 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4149 /// following equation:
4151 /// A * X = B (mod N)
4153 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4154 /// A and B isn't important.
4156 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4157 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4158 ScalarEvolution &SE) {
4159 uint32_t BW = A.getBitWidth();
4160 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4161 assert(A != 0 && "A must be non-zero.");
4165 // The gcd of A and N may have only one prime factor: 2. The number of
4166 // trailing zeros in A is its multiplicity
4167 uint32_t Mult2 = A.countTrailingZeros();
4170 // 2. Check if B is divisible by D.
4172 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4173 // is not less than multiplicity of this prime factor for D.
4174 if (B.countTrailingZeros() < Mult2)
4175 return SE.getCouldNotCompute();
4177 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4180 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4181 // bit width during computations.
4182 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4183 APInt Mod(BW + 1, 0);
4184 Mod.set(BW - Mult2); // Mod = N / D
4185 APInt I = AD.multiplicativeInverse(Mod);
4187 // 4. Compute the minimum unsigned root of the equation:
4188 // I * (B / D) mod (N / D)
4189 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4191 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4193 return SE.getConstant(Result.trunc(BW));
4196 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4197 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4198 /// might be the same) or two SCEVCouldNotCompute objects.
4200 static std::pair<const SCEV *,const SCEV *>
4201 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4202 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4203 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4204 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4205 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4207 // We currently can only solve this if the coefficients are constants.
4208 if (!LC || !MC || !NC) {
4209 const SCEV *CNC = SE.getCouldNotCompute();
4210 return std::make_pair(CNC, CNC);
4213 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4214 const APInt &L = LC->getValue()->getValue();
4215 const APInt &M = MC->getValue()->getValue();
4216 const APInt &N = NC->getValue()->getValue();
4217 APInt Two(BitWidth, 2);
4218 APInt Four(BitWidth, 4);
4221 using namespace APIntOps;
4223 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4224 // The B coefficient is M-N/2
4228 // The A coefficient is N/2
4229 APInt A(N.sdiv(Two));
4231 // Compute the B^2-4ac term.
4234 SqrtTerm -= Four * (A * C);
4236 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4237 // integer value or else APInt::sqrt() will assert.
4238 APInt SqrtVal(SqrtTerm.sqrt());
4240 // Compute the two solutions for the quadratic formula.
4241 // The divisions must be performed as signed divisions.
4243 APInt TwoA( A << 1 );
4244 if (TwoA.isMinValue()) {
4245 const SCEV *CNC = SE.getCouldNotCompute();
4246 return std::make_pair(CNC, CNC);
4249 LLVMContext &Context = SE.getContext();
4251 ConstantInt *Solution1 =
4252 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4253 ConstantInt *Solution2 =
4254 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4256 return std::make_pair(SE.getConstant(Solution1),
4257 SE.getConstant(Solution2));
4258 } // end APIntOps namespace
4261 /// HowFarToZero - Return the number of times a backedge comparing the specified
4262 /// value to zero will execute. If not computable, return CouldNotCompute.
4263 const SCEV *ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4264 // If the value is a constant
4265 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4266 // If the value is already zero, the branch will execute zero times.
4267 if (C->getValue()->isZero()) return C;
4268 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4271 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4272 if (!AddRec || AddRec->getLoop() != L)
4273 return getCouldNotCompute();
4275 if (AddRec->isAffine()) {
4276 // If this is an affine expression, the execution count of this branch is
4277 // the minimum unsigned root of the following equation:
4279 // Start + Step*N = 0 (mod 2^BW)
4283 // Step*N = -Start (mod 2^BW)
4285 // where BW is the common bit width of Start and Step.
4287 // Get the initial value for the loop.
4288 const SCEV *Start = getSCEVAtScope(AddRec->getStart(),
4289 L->getParentLoop());
4290 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1),
4291 L->getParentLoop());
4293 if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
4294 // For now we handle only constant steps.
4296 // First, handle unitary steps.
4297 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4298 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4299 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4300 return Start; // N = Start (as unsigned)
4302 // Then, try to solve the above equation provided that Start is constant.
4303 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4304 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4305 -StartC->getValue()->getValue(),
4308 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
4309 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4310 // the quadratic equation to solve it.
4311 std::pair<const SCEV *,const SCEV *> Roots = SolveQuadraticEquation(AddRec,
4313 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4314 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4317 errs() << "HFTZ: " << *V << " - sol#1: " << *R1
4318 << " sol#2: " << *R2 << "\n";
4320 // Pick the smallest positive root value.
4321 if (ConstantInt *CB =
4322 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
4323 R1->getValue(), R2->getValue()))) {
4324 if (CB->getZExtValue() == false)
4325 std::swap(R1, R2); // R1 is the minimum root now.
4327 // We can only use this value if the chrec ends up with an exact zero
4328 // value at this index. When solving for "X*X != 5", for example, we
4329 // should not accept a root of 2.
4330 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4332 return R1; // We found a quadratic root!
4337 return getCouldNotCompute();
4340 /// HowFarToNonZero - Return the number of times a backedge checking the
4341 /// specified value for nonzero will execute. If not computable, return
4343 const SCEV *ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4344 // Loops that look like: while (X == 0) are very strange indeed. We don't
4345 // handle them yet except for the trivial case. This could be expanded in the
4346 // future as needed.
4348 // If the value is a constant, check to see if it is known to be non-zero
4349 // already. If so, the backedge will execute zero times.
4350 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4351 if (!C->getValue()->isNullValue())
4352 return getIntegerSCEV(0, C->getType());
4353 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4356 // We could implement others, but I really doubt anyone writes loops like
4357 // this, and if they did, they would already be constant folded.
4358 return getCouldNotCompute();
4361 /// getLoopPredecessor - If the given loop's header has exactly one unique
4362 /// predecessor outside the loop, return it. Otherwise return null.
4364 BasicBlock *ScalarEvolution::getLoopPredecessor(const Loop *L) {
4365 BasicBlock *Header = L->getHeader();
4366 BasicBlock *Pred = 0;
4367 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header);
4369 if (!L->contains(*PI)) {
4370 if (Pred && Pred != *PI) return 0; // Multiple predecessors.
4376 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4377 /// (which may not be an immediate predecessor) which has exactly one
4378 /// successor from which BB is reachable, or null if no such block is
4382 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4383 // If the block has a unique predecessor, then there is no path from the
4384 // predecessor to the block that does not go through the direct edge
4385 // from the predecessor to the block.
4386 if (BasicBlock *Pred = BB->getSinglePredecessor())
4389 // A loop's header is defined to be a block that dominates the loop.
4390 // If the header has a unique predecessor outside the loop, it must be
4391 // a block that has exactly one successor that can reach the loop.
4392 if (Loop *L = LI->getLoopFor(BB))
4393 return getLoopPredecessor(L);
4398 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4399 /// testing whether two expressions are equal, however for the purposes of
4400 /// looking for a condition guarding a loop, it can be useful to be a little
4401 /// more general, since a front-end may have replicated the controlling
4404 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4405 // Quick check to see if they are the same SCEV.
4406 if (A == B) return true;
4408 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4409 // two different instructions with the same value. Check for this case.
4410 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4411 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4412 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4413 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4414 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4417 // Otherwise assume they may have a different value.
4421 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
4422 return getSignedRange(S).getSignedMax().isNegative();
4425 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
4426 return getSignedRange(S).getSignedMin().isStrictlyPositive();
4429 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
4430 return !getSignedRange(S).getSignedMin().isNegative();
4433 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
4434 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
4437 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
4438 return isKnownNegative(S) || isKnownPositive(S);
4441 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
4442 const SCEV *LHS, const SCEV *RHS) {
4444 if (HasSameValue(LHS, RHS))
4445 return ICmpInst::isTrueWhenEqual(Pred);
4449 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4451 case ICmpInst::ICMP_SGT:
4452 Pred = ICmpInst::ICMP_SLT;
4453 std::swap(LHS, RHS);
4454 case ICmpInst::ICMP_SLT: {
4455 ConstantRange LHSRange = getSignedRange(LHS);
4456 ConstantRange RHSRange = getSignedRange(RHS);
4457 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
4459 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
4463 case ICmpInst::ICMP_SGE:
4464 Pred = ICmpInst::ICMP_SLE;
4465 std::swap(LHS, RHS);
4466 case ICmpInst::ICMP_SLE: {
4467 ConstantRange LHSRange = getSignedRange(LHS);
4468 ConstantRange RHSRange = getSignedRange(RHS);
4469 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
4471 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
4475 case ICmpInst::ICMP_UGT:
4476 Pred = ICmpInst::ICMP_ULT;
4477 std::swap(LHS, RHS);
4478 case ICmpInst::ICMP_ULT: {
4479 ConstantRange LHSRange = getUnsignedRange(LHS);
4480 ConstantRange RHSRange = getUnsignedRange(RHS);
4481 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
4483 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
4487 case ICmpInst::ICMP_UGE:
4488 Pred = ICmpInst::ICMP_ULE;
4489 std::swap(LHS, RHS);
4490 case ICmpInst::ICMP_ULE: {
4491 ConstantRange LHSRange = getUnsignedRange(LHS);
4492 ConstantRange RHSRange = getUnsignedRange(RHS);
4493 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
4495 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
4499 case ICmpInst::ICMP_NE: {
4500 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
4502 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
4505 const SCEV *Diff = getMinusSCEV(LHS, RHS);
4506 if (isKnownNonZero(Diff))
4510 case ICmpInst::ICMP_EQ:
4511 // The check at the top of the function catches the case where
4512 // the values are known to be equal.
4518 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
4519 /// protected by a conditional between LHS and RHS. This is used to
4520 /// to eliminate casts.
4522 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
4523 ICmpInst::Predicate Pred,
4524 const SCEV *LHS, const SCEV *RHS) {
4525 // Interpret a null as meaning no loop, where there is obviously no guard
4526 // (interprocedural conditions notwithstanding).
4527 if (!L) return true;
4529 BasicBlock *Latch = L->getLoopLatch();
4533 BranchInst *LoopContinuePredicate =
4534 dyn_cast<BranchInst>(Latch->getTerminator());
4535 if (!LoopContinuePredicate ||
4536 LoopContinuePredicate->isUnconditional())
4539 return isImpliedCond(LoopContinuePredicate->getCondition(), Pred, LHS, RHS,
4540 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
4543 /// isLoopGuardedByCond - Test whether entry to the loop is protected
4544 /// by a conditional between LHS and RHS. This is used to help avoid max
4545 /// expressions in loop trip counts, and to eliminate casts.
4547 ScalarEvolution::isLoopGuardedByCond(const Loop *L,
4548 ICmpInst::Predicate Pred,
4549 const SCEV *LHS, const SCEV *RHS) {
4550 // Interpret a null as meaning no loop, where there is obviously no guard
4551 // (interprocedural conditions notwithstanding).
4552 if (!L) return false;
4554 BasicBlock *Predecessor = getLoopPredecessor(L);
4555 BasicBlock *PredecessorDest = L->getHeader();
4557 // Starting at the loop predecessor, climb up the predecessor chain, as long
4558 // as there are predecessors that can be found that have unique successors
4559 // leading to the original header.
4561 PredecessorDest = Predecessor,
4562 Predecessor = getPredecessorWithUniqueSuccessorForBB(Predecessor)) {
4564 BranchInst *LoopEntryPredicate =
4565 dyn_cast<BranchInst>(Predecessor->getTerminator());
4566 if (!LoopEntryPredicate ||
4567 LoopEntryPredicate->isUnconditional())
4570 if (isImpliedCond(LoopEntryPredicate->getCondition(), Pred, LHS, RHS,
4571 LoopEntryPredicate->getSuccessor(0) != PredecessorDest))
4578 /// isImpliedCond - Test whether the condition described by Pred, LHS,
4579 /// and RHS is true whenever the given Cond value evaluates to true.
4580 bool ScalarEvolution::isImpliedCond(Value *CondValue,
4581 ICmpInst::Predicate Pred,
4582 const SCEV *LHS, const SCEV *RHS,
4584 // Recursivly handle And and Or conditions.
4585 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CondValue)) {
4586 if (BO->getOpcode() == Instruction::And) {
4588 return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
4589 isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
4590 } else if (BO->getOpcode() == Instruction::Or) {
4592 return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
4593 isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
4597 ICmpInst *ICI = dyn_cast<ICmpInst>(CondValue);
4598 if (!ICI) return false;
4600 // Bail if the ICmp's operands' types are wider than the needed type
4601 // before attempting to call getSCEV on them. This avoids infinite
4602 // recursion, since the analysis of widening casts can require loop
4603 // exit condition information for overflow checking, which would
4605 if (getTypeSizeInBits(LHS->getType()) <
4606 getTypeSizeInBits(ICI->getOperand(0)->getType()))
4609 // Now that we found a conditional branch that dominates the loop, check to
4610 // see if it is the comparison we are looking for.
4611 ICmpInst::Predicate FoundPred;
4613 FoundPred = ICI->getInversePredicate();
4615 FoundPred = ICI->getPredicate();
4617 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
4618 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
4620 // Balance the types. The case where FoundLHS' type is wider than
4621 // LHS' type is checked for above.
4622 if (getTypeSizeInBits(LHS->getType()) >
4623 getTypeSizeInBits(FoundLHS->getType())) {
4624 if (CmpInst::isSigned(Pred)) {
4625 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
4626 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
4628 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
4629 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
4633 // Canonicalize the query to match the way instcombine will have
4634 // canonicalized the comparison.
4635 // First, put a constant operand on the right.
4636 if (isa<SCEVConstant>(LHS)) {
4637 std::swap(LHS, RHS);
4638 Pred = ICmpInst::getSwappedPredicate(Pred);
4640 // Then, canonicalize comparisons with boundary cases.
4641 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
4642 const APInt &RA = RC->getValue()->getValue();
4644 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4645 case ICmpInst::ICMP_EQ:
4646 case ICmpInst::ICMP_NE:
4648 case ICmpInst::ICMP_UGE:
4649 if ((RA - 1).isMinValue()) {
4650 Pred = ICmpInst::ICMP_NE;
4651 RHS = getConstant(RA - 1);
4654 if (RA.isMaxValue()) {
4655 Pred = ICmpInst::ICMP_EQ;
4658 if (RA.isMinValue()) return true;
4660 case ICmpInst::ICMP_ULE:
4661 if ((RA + 1).isMaxValue()) {
4662 Pred = ICmpInst::ICMP_NE;
4663 RHS = getConstant(RA + 1);
4666 if (RA.isMinValue()) {
4667 Pred = ICmpInst::ICMP_EQ;
4670 if (RA.isMaxValue()) return true;
4672 case ICmpInst::ICMP_SGE:
4673 if ((RA - 1).isMinSignedValue()) {
4674 Pred = ICmpInst::ICMP_NE;
4675 RHS = getConstant(RA - 1);
4678 if (RA.isMaxSignedValue()) {
4679 Pred = ICmpInst::ICMP_EQ;
4682 if (RA.isMinSignedValue()) return true;
4684 case ICmpInst::ICMP_SLE:
4685 if ((RA + 1).isMaxSignedValue()) {
4686 Pred = ICmpInst::ICMP_NE;
4687 RHS = getConstant(RA + 1);
4690 if (RA.isMinSignedValue()) {
4691 Pred = ICmpInst::ICMP_EQ;
4694 if (RA.isMaxSignedValue()) return true;
4696 case ICmpInst::ICMP_UGT:
4697 if (RA.isMinValue()) {
4698 Pred = ICmpInst::ICMP_NE;
4701 if ((RA + 1).isMaxValue()) {
4702 Pred = ICmpInst::ICMP_EQ;
4703 RHS = getConstant(RA + 1);
4706 if (RA.isMaxValue()) return false;
4708 case ICmpInst::ICMP_ULT:
4709 if (RA.isMaxValue()) {
4710 Pred = ICmpInst::ICMP_NE;
4713 if ((RA - 1).isMinValue()) {
4714 Pred = ICmpInst::ICMP_EQ;
4715 RHS = getConstant(RA - 1);
4718 if (RA.isMinValue()) return false;
4720 case ICmpInst::ICMP_SGT:
4721 if (RA.isMinSignedValue()) {
4722 Pred = ICmpInst::ICMP_NE;
4725 if ((RA + 1).isMaxSignedValue()) {
4726 Pred = ICmpInst::ICMP_EQ;
4727 RHS = getConstant(RA + 1);
4730 if (RA.isMaxSignedValue()) return false;
4732 case ICmpInst::ICMP_SLT:
4733 if (RA.isMaxSignedValue()) {
4734 Pred = ICmpInst::ICMP_NE;
4737 if ((RA - 1).isMinSignedValue()) {
4738 Pred = ICmpInst::ICMP_EQ;
4739 RHS = getConstant(RA - 1);
4742 if (RA.isMinSignedValue()) return false;
4747 // Check to see if we can make the LHS or RHS match.
4748 if (LHS == FoundRHS || RHS == FoundLHS) {
4749 if (isa<SCEVConstant>(RHS)) {
4750 std::swap(FoundLHS, FoundRHS);
4751 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
4753 std::swap(LHS, RHS);
4754 Pred = ICmpInst::getSwappedPredicate(Pred);
4758 // Check whether the found predicate is the same as the desired predicate.
4759 if (FoundPred == Pred)
4760 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
4762 // Check whether swapping the found predicate makes it the same as the
4763 // desired predicate.
4764 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
4765 if (isa<SCEVConstant>(RHS))
4766 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
4768 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
4769 RHS, LHS, FoundLHS, FoundRHS);
4772 // Check whether the actual condition is beyond sufficient.
4773 if (FoundPred == ICmpInst::ICMP_EQ)
4774 if (ICmpInst::isTrueWhenEqual(Pred))
4775 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
4777 if (Pred == ICmpInst::ICMP_NE)
4778 if (!ICmpInst::isTrueWhenEqual(FoundPred))
4779 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
4782 // Otherwise assume the worst.
4786 /// isImpliedCondOperands - Test whether the condition described by Pred,
4787 /// LHS, and RHS is true whenever the condition desribed by Pred, FoundLHS,
4788 /// and FoundRHS is true.
4789 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
4790 const SCEV *LHS, const SCEV *RHS,
4791 const SCEV *FoundLHS,
4792 const SCEV *FoundRHS) {
4793 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
4794 FoundLHS, FoundRHS) ||
4795 // ~x < ~y --> x > y
4796 isImpliedCondOperandsHelper(Pred, LHS, RHS,
4797 getNotSCEV(FoundRHS),
4798 getNotSCEV(FoundLHS));
4801 /// isImpliedCondOperandsHelper - Test whether the condition described by
4802 /// Pred, LHS, and RHS is true whenever the condition desribed by Pred,
4803 /// FoundLHS, and FoundRHS is true.
4805 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
4806 const SCEV *LHS, const SCEV *RHS,
4807 const SCEV *FoundLHS,
4808 const SCEV *FoundRHS) {
4810 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4811 case ICmpInst::ICMP_EQ:
4812 case ICmpInst::ICMP_NE:
4813 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
4816 case ICmpInst::ICMP_SLT:
4817 case ICmpInst::ICMP_SLE:
4818 if (isKnownPredicate(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
4819 isKnownPredicate(ICmpInst::ICMP_SGE, RHS, FoundRHS))
4822 case ICmpInst::ICMP_SGT:
4823 case ICmpInst::ICMP_SGE:
4824 if (isKnownPredicate(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
4825 isKnownPredicate(ICmpInst::ICMP_SLE, RHS, FoundRHS))
4828 case ICmpInst::ICMP_ULT:
4829 case ICmpInst::ICMP_ULE:
4830 if (isKnownPredicate(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
4831 isKnownPredicate(ICmpInst::ICMP_UGE, RHS, FoundRHS))
4834 case ICmpInst::ICMP_UGT:
4835 case ICmpInst::ICMP_UGE:
4836 if (isKnownPredicate(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
4837 isKnownPredicate(ICmpInst::ICMP_ULE, RHS, FoundRHS))
4845 /// getBECount - Subtract the end and start values and divide by the step,
4846 /// rounding up, to get the number of times the backedge is executed. Return
4847 /// CouldNotCompute if an intermediate computation overflows.
4848 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
4852 const Type *Ty = Start->getType();
4853 const SCEV *NegOne = getIntegerSCEV(-1, Ty);
4854 const SCEV *Diff = getMinusSCEV(End, Start);
4855 const SCEV *RoundUp = getAddExpr(Step, NegOne);
4857 // Add an adjustment to the difference between End and Start so that
4858 // the division will effectively round up.
4859 const SCEV *Add = getAddExpr(Diff, RoundUp);
4862 // Check Add for unsigned overflow.
4863 // TODO: More sophisticated things could be done here.
4864 const Type *WideTy = IntegerType::get(getContext(),
4865 getTypeSizeInBits(Ty) + 1);
4866 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
4867 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
4868 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
4869 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
4870 return getCouldNotCompute();
4873 return getUDivExpr(Add, Step);
4876 /// HowManyLessThans - Return the number of times a backedge containing the
4877 /// specified less-than comparison will execute. If not computable, return
4878 /// CouldNotCompute.
4879 ScalarEvolution::BackedgeTakenInfo
4880 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
4881 const Loop *L, bool isSigned) {
4882 // Only handle: "ADDREC < LoopInvariant".
4883 if (!RHS->isLoopInvariant(L)) return getCouldNotCompute();
4885 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
4886 if (!AddRec || AddRec->getLoop() != L)
4887 return getCouldNotCompute();
4889 // Check to see if we have a flag which makes analysis easy.
4890 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
4891 AddRec->hasNoUnsignedWrap();
4893 if (AddRec->isAffine()) {
4894 // FORNOW: We only support unit strides.
4895 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
4896 const SCEV *Step = AddRec->getStepRecurrence(*this);
4898 // TODO: handle non-constant strides.
4899 const SCEVConstant *CStep = dyn_cast<SCEVConstant>(Step);
4900 if (!CStep || CStep->isZero())
4901 return getCouldNotCompute();
4902 if (CStep->isOne()) {
4903 // With unit stride, the iteration never steps past the limit value.
4904 } else if (CStep->getValue()->getValue().isStrictlyPositive()) {
4906 // We know the iteration won't step past the maximum value for its type.
4908 } else if (const SCEVConstant *CLimit = dyn_cast<SCEVConstant>(RHS)) {
4909 // Test whether a positive iteration iteration can step past the limit
4910 // value and past the maximum value for its type in a single step.
4912 APInt Max = APInt::getSignedMaxValue(BitWidth);
4913 if ((Max - CStep->getValue()->getValue())
4914 .slt(CLimit->getValue()->getValue()))
4915 return getCouldNotCompute();
4917 APInt Max = APInt::getMaxValue(BitWidth);
4918 if ((Max - CStep->getValue()->getValue())
4919 .ult(CLimit->getValue()->getValue()))
4920 return getCouldNotCompute();
4923 // TODO: handle non-constant limit values below.
4924 return getCouldNotCompute();
4926 // TODO: handle negative strides below.
4927 return getCouldNotCompute();
4929 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
4930 // m. So, we count the number of iterations in which {n,+,s} < m is true.
4931 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
4932 // treat m-n as signed nor unsigned due to overflow possibility.
4934 // First, we get the value of the LHS in the first iteration: n
4935 const SCEV *Start = AddRec->getOperand(0);
4937 // Determine the minimum constant start value.
4938 const SCEV *MinStart = getConstant(isSigned ?
4939 getSignedRange(Start).getSignedMin() :
4940 getUnsignedRange(Start).getUnsignedMin());
4942 // If we know that the condition is true in order to enter the loop,
4943 // then we know that it will run exactly (m-n)/s times. Otherwise, we
4944 // only know that it will execute (max(m,n)-n)/s times. In both cases,
4945 // the division must round up.
4946 const SCEV *End = RHS;
4947 if (!isLoopGuardedByCond(L,
4948 isSigned ? ICmpInst::ICMP_SLT :
4950 getMinusSCEV(Start, Step), RHS))
4951 End = isSigned ? getSMaxExpr(RHS, Start)
4952 : getUMaxExpr(RHS, Start);
4954 // Determine the maximum constant end value.
4955 const SCEV *MaxEnd = getConstant(isSigned ?
4956 getSignedRange(End).getSignedMax() :
4957 getUnsignedRange(End).getUnsignedMax());
4959 // Finally, we subtract these two values and divide, rounding up, to get
4960 // the number of times the backedge is executed.
4961 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
4963 // The maximum backedge count is similar, except using the minimum start
4964 // value and the maximum end value.
4965 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
4967 return BackedgeTakenInfo(BECount, MaxBECount);
4970 return getCouldNotCompute();
4973 /// getNumIterationsInRange - Return the number of iterations of this loop that
4974 /// produce values in the specified constant range. Another way of looking at
4975 /// this is that it returns the first iteration number where the value is not in
4976 /// the condition, thus computing the exit count. If the iteration count can't
4977 /// be computed, an instance of SCEVCouldNotCompute is returned.
4978 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
4979 ScalarEvolution &SE) const {
4980 if (Range.isFullSet()) // Infinite loop.
4981 return SE.getCouldNotCompute();
4983 // If the start is a non-zero constant, shift the range to simplify things.
4984 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
4985 if (!SC->getValue()->isZero()) {
4986 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
4987 Operands[0] = SE.getIntegerSCEV(0, SC->getType());
4988 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
4989 if (const SCEVAddRecExpr *ShiftedAddRec =
4990 dyn_cast<SCEVAddRecExpr>(Shifted))
4991 return ShiftedAddRec->getNumIterationsInRange(
4992 Range.subtract(SC->getValue()->getValue()), SE);
4993 // This is strange and shouldn't happen.
4994 return SE.getCouldNotCompute();
4997 // The only time we can solve this is when we have all constant indices.
4998 // Otherwise, we cannot determine the overflow conditions.
4999 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5000 if (!isa<SCEVConstant>(getOperand(i)))
5001 return SE.getCouldNotCompute();
5004 // Okay at this point we know that all elements of the chrec are constants and
5005 // that the start element is zero.
5007 // First check to see if the range contains zero. If not, the first
5009 unsigned BitWidth = SE.getTypeSizeInBits(getType());
5010 if (!Range.contains(APInt(BitWidth, 0)))
5011 return SE.getIntegerSCEV(0, getType());
5014 // If this is an affine expression then we have this situation:
5015 // Solve {0,+,A} in Range === Ax in Range
5017 // We know that zero is in the range. If A is positive then we know that
5018 // the upper value of the range must be the first possible exit value.
5019 // If A is negative then the lower of the range is the last possible loop
5020 // value. Also note that we already checked for a full range.
5021 APInt One(BitWidth,1);
5022 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5023 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5025 // The exit value should be (End+A)/A.
5026 APInt ExitVal = (End + A).udiv(A);
5027 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5029 // Evaluate at the exit value. If we really did fall out of the valid
5030 // range, then we computed our trip count, otherwise wrap around or other
5031 // things must have happened.
5032 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5033 if (Range.contains(Val->getValue()))
5034 return SE.getCouldNotCompute(); // Something strange happened
5036 // Ensure that the previous value is in the range. This is a sanity check.
5037 assert(Range.contains(
5038 EvaluateConstantChrecAtConstant(this,
5039 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5040 "Linear scev computation is off in a bad way!");
5041 return SE.getConstant(ExitValue);
5042 } else if (isQuadratic()) {
5043 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5044 // quadratic equation to solve it. To do this, we must frame our problem in
5045 // terms of figuring out when zero is crossed, instead of when
5046 // Range.getUpper() is crossed.
5047 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5048 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5049 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5051 // Next, solve the constructed addrec
5052 std::pair<const SCEV *,const SCEV *> Roots =
5053 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5054 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5055 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5057 // Pick the smallest positive root value.
5058 if (ConstantInt *CB =
5059 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5060 R1->getValue(), R2->getValue()))) {
5061 if (CB->getZExtValue() == false)
5062 std::swap(R1, R2); // R1 is the minimum root now.
5064 // Make sure the root is not off by one. The returned iteration should
5065 // not be in the range, but the previous one should be. When solving
5066 // for "X*X < 5", for example, we should not return a root of 2.
5067 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5070 if (Range.contains(R1Val->getValue())) {
5071 // The next iteration must be out of the range...
5072 ConstantInt *NextVal =
5073 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5075 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5076 if (!Range.contains(R1Val->getValue()))
5077 return SE.getConstant(NextVal);
5078 return SE.getCouldNotCompute(); // Something strange happened
5081 // If R1 was not in the range, then it is a good return value. Make
5082 // sure that R1-1 WAS in the range though, just in case.
5083 ConstantInt *NextVal =
5084 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5085 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5086 if (Range.contains(R1Val->getValue()))
5088 return SE.getCouldNotCompute(); // Something strange happened
5093 return SE.getCouldNotCompute();
5098 //===----------------------------------------------------------------------===//
5099 // SCEVCallbackVH Class Implementation
5100 //===----------------------------------------------------------------------===//
5102 void ScalarEvolution::SCEVCallbackVH::deleted() {
5103 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5104 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5105 SE->ConstantEvolutionLoopExitValue.erase(PN);
5106 SE->Scalars.erase(getValPtr());
5107 // this now dangles!
5110 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *) {
5111 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5113 // Forget all the expressions associated with users of the old value,
5114 // so that future queries will recompute the expressions using the new
5116 SmallVector<User *, 16> Worklist;
5117 SmallPtrSet<User *, 8> Visited;
5118 Value *Old = getValPtr();
5119 bool DeleteOld = false;
5120 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5122 Worklist.push_back(*UI);
5123 while (!Worklist.empty()) {
5124 User *U = Worklist.pop_back_val();
5125 // Deleting the Old value will cause this to dangle. Postpone
5126 // that until everything else is done.
5131 if (!Visited.insert(U))
5133 if (PHINode *PN = dyn_cast<PHINode>(U))
5134 SE->ConstantEvolutionLoopExitValue.erase(PN);
5135 SE->Scalars.erase(U);
5136 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5138 Worklist.push_back(*UI);
5140 // Delete the Old value if it (indirectly) references itself.
5142 if (PHINode *PN = dyn_cast<PHINode>(Old))
5143 SE->ConstantEvolutionLoopExitValue.erase(PN);
5144 SE->Scalars.erase(Old);
5145 // this now dangles!
5150 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5151 : CallbackVH(V), SE(se) {}
5153 //===----------------------------------------------------------------------===//
5154 // ScalarEvolution Class Implementation
5155 //===----------------------------------------------------------------------===//
5157 ScalarEvolution::ScalarEvolution()
5158 : FunctionPass(&ID) {
5161 bool ScalarEvolution::runOnFunction(Function &F) {
5163 LI = &getAnalysis<LoopInfo>();
5164 TD = getAnalysisIfAvailable<TargetData>();
5168 void ScalarEvolution::releaseMemory() {
5170 BackedgeTakenCounts.clear();
5171 ConstantEvolutionLoopExitValue.clear();
5172 ValuesAtScopes.clear();
5173 UniqueSCEVs.clear();
5174 SCEVAllocator.Reset();
5177 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5178 AU.setPreservesAll();
5179 AU.addRequiredTransitive<LoopInfo>();
5182 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5183 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5186 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5188 // Print all inner loops first
5189 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5190 PrintLoopInfo(OS, SE, *I);
5192 OS << "Loop " << L->getHeader()->getName() << ": ";
5194 SmallVector<BasicBlock *, 8> ExitBlocks;
5195 L->getExitBlocks(ExitBlocks);
5196 if (ExitBlocks.size() != 1)
5197 OS << "<multiple exits> ";
5199 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5200 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5202 OS << "Unpredictable backedge-taken count. ";
5206 OS << "Loop " << L->getHeader()->getName() << ": ";
5208 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5209 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5211 OS << "Unpredictable max backedge-taken count. ";
5217 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
5218 // ScalarEvolution's implementaiton of the print method is to print
5219 // out SCEV values of all instructions that are interesting. Doing
5220 // this potentially causes it to create new SCEV objects though,
5221 // which technically conflicts with the const qualifier. This isn't
5222 // observable from outside the class though, so casting away the
5223 // const isn't dangerous.
5224 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
5226 OS << "Classifying expressions for: " << F->getName() << "\n";
5227 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
5228 if (isSCEVable(I->getType())) {
5231 const SCEV *SV = SE.getSCEV(&*I);
5234 const Loop *L = LI->getLoopFor((*I).getParent());
5236 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
5243 OS << "\t\t" "Exits: ";
5244 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
5245 if (!ExitValue->isLoopInvariant(L)) {
5246 OS << "<<Unknown>>";
5255 OS << "Determining loop execution counts for: " << F->getName() << "\n";
5256 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
5257 PrintLoopInfo(OS, &SE, *I);