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/Compiler.h"
78 #include "llvm/Support/ConstantRange.h"
79 #include "llvm/Support/ErrorHandling.h"
80 #include "llvm/Support/GetElementPtrTypeIterator.h"
81 #include "llvm/Support/InstIterator.h"
82 #include "llvm/Support/MathExtras.h"
83 #include "llvm/Support/raw_ostream.h"
84 #include "llvm/ADT/Statistic.h"
85 #include "llvm/ADT/STLExtras.h"
86 #include "llvm/ADT/SmallPtrSet.h"
90 STATISTIC(NumArrayLenItCounts,
91 "Number of trip counts computed with array length");
92 STATISTIC(NumTripCountsComputed,
93 "Number of loops with predictable loop counts");
94 STATISTIC(NumTripCountsNotComputed,
95 "Number of loops without predictable loop counts");
96 STATISTIC(NumBruteForceTripCountsComputed,
97 "Number of loops with trip counts computed by force");
99 static cl::opt<unsigned>
100 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
101 cl::desc("Maximum number of iterations SCEV will "
102 "symbolically execute a constant "
106 static RegisterPass<ScalarEvolution>
107 R("scalar-evolution", "Scalar Evolution Analysis", false, true);
108 char ScalarEvolution::ID = 0;
110 //===----------------------------------------------------------------------===//
111 // SCEV class definitions
112 //===----------------------------------------------------------------------===//
114 //===----------------------------------------------------------------------===//
115 // Implementation of the SCEV class.
120 void SCEV::dump() const {
125 bool SCEV::isZero() const {
126 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
127 return SC->getValue()->isZero();
131 bool SCEV::isOne() const {
132 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
133 return SC->getValue()->isOne();
137 bool SCEV::isAllOnesValue() const {
138 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
139 return SC->getValue()->isAllOnesValue();
143 SCEVCouldNotCompute::SCEVCouldNotCompute() :
144 SCEV(FoldingSetNodeID(), scCouldNotCompute) {}
146 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
147 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
151 const Type *SCEVCouldNotCompute::getType() const {
152 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
156 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
157 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
161 bool SCEVCouldNotCompute::hasOperand(const SCEV *) const {
162 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
166 void SCEVCouldNotCompute::print(raw_ostream &OS) const {
167 OS << "***COULDNOTCOMPUTE***";
170 bool SCEVCouldNotCompute::classof(const SCEV *S) {
171 return S->getSCEVType() == scCouldNotCompute;
174 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
176 ID.AddInteger(scConstant);
179 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
180 SCEV *S = SCEVAllocator.Allocate<SCEVConstant>();
181 new (S) SCEVConstant(ID, V);
182 UniqueSCEVs.InsertNode(S, IP);
186 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
187 return getConstant(ConstantInt::get(getContext(), Val));
191 ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
193 ConstantInt::get(cast<IntegerType>(Ty), V, isSigned));
196 const Type *SCEVConstant::getType() const { return V->getType(); }
198 void SCEVConstant::print(raw_ostream &OS) const {
199 WriteAsOperand(OS, V, false);
202 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeID &ID,
203 unsigned SCEVTy, const SCEV *op, const Type *ty)
204 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
206 bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
207 return Op->dominates(BB, DT);
210 bool SCEVCastExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
211 return Op->properlyDominates(BB, DT);
214 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeID &ID,
215 const SCEV *op, const Type *ty)
216 : SCEVCastExpr(ID, scTruncate, op, ty) {
217 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
218 (Ty->isInteger() || isa<PointerType>(Ty)) &&
219 "Cannot truncate non-integer value!");
222 void SCEVTruncateExpr::print(raw_ostream &OS) const {
223 OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
226 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeID &ID,
227 const SCEV *op, const Type *ty)
228 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
229 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
230 (Ty->isInteger() || isa<PointerType>(Ty)) &&
231 "Cannot zero extend non-integer value!");
234 void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
235 OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
238 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeID &ID,
239 const SCEV *op, const Type *ty)
240 : SCEVCastExpr(ID, scSignExtend, op, ty) {
241 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
242 (Ty->isInteger() || isa<PointerType>(Ty)) &&
243 "Cannot sign extend non-integer value!");
246 void SCEVSignExtendExpr::print(raw_ostream &OS) const {
247 OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
250 void SCEVCommutativeExpr::print(raw_ostream &OS) const {
251 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
252 const char *OpStr = getOperationStr();
253 OS << "(" << *Operands[0];
254 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
255 OS << OpStr << *Operands[i];
259 bool SCEVNAryExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
260 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
261 if (!getOperand(i)->dominates(BB, DT))
267 bool SCEVNAryExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
268 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
269 if (!getOperand(i)->properlyDominates(BB, DT))
275 bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
276 return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
279 bool SCEVUDivExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
280 return LHS->properlyDominates(BB, DT) && RHS->properlyDominates(BB, DT);
283 void SCEVUDivExpr::print(raw_ostream &OS) const {
284 OS << "(" << *LHS << " /u " << *RHS << ")";
287 const Type *SCEVUDivExpr::getType() const {
288 // In most cases the types of LHS and RHS will be the same, but in some
289 // crazy cases one or the other may be a pointer. ScalarEvolution doesn't
290 // depend on the type for correctness, but handling types carefully can
291 // avoid extra casts in the SCEVExpander. The LHS is more likely to be
292 // a pointer type than the RHS, so use the RHS' type here.
293 return RHS->getType();
296 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
297 // Add recurrences are never invariant in the function-body (null loop).
301 // This recurrence is variant w.r.t. QueryLoop if QueryLoop contains L.
302 if (QueryLoop->contains(L->getHeader()))
305 // This recurrence is variant w.r.t. QueryLoop if any of its operands
307 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
308 if (!getOperand(i)->isLoopInvariant(QueryLoop))
311 // Otherwise it's loop-invariant.
315 void SCEVAddRecExpr::print(raw_ostream &OS) const {
316 OS << "{" << *Operands[0];
317 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
318 OS << ",+," << *Operands[i];
319 OS << "}<" << L->getHeader()->getName() + ">";
322 void SCEVFieldOffsetExpr::print(raw_ostream &OS) const {
323 // LLVM struct fields don't have names, so just print the field number.
324 OS << "offsetof(" << *STy << ", " << FieldNo << ")";
327 void SCEVAllocSizeExpr::print(raw_ostream &OS) const {
328 OS << "sizeof(" << *AllocTy << ")";
331 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
332 // All non-instruction values are loop invariant. All instructions are loop
333 // invariant if they are not contained in the specified loop.
334 // Instructions are never considered invariant in the function body
335 // (null loop) because they are defined within the "loop".
336 if (Instruction *I = dyn_cast<Instruction>(V))
337 return L && !L->contains(I->getParent());
341 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
342 if (Instruction *I = dyn_cast<Instruction>(getValue()))
343 return DT->dominates(I->getParent(), BB);
347 bool SCEVUnknown::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
348 if (Instruction *I = dyn_cast<Instruction>(getValue()))
349 return DT->properlyDominates(I->getParent(), BB);
353 const Type *SCEVUnknown::getType() const {
357 void SCEVUnknown::print(raw_ostream &OS) const {
358 WriteAsOperand(OS, V, false);
361 //===----------------------------------------------------------------------===//
363 //===----------------------------------------------------------------------===//
365 static bool CompareTypes(const Type *A, const Type *B) {
366 if (A->getTypeID() != B->getTypeID())
367 return A->getTypeID() < B->getTypeID();
368 if (const IntegerType *AI = dyn_cast<IntegerType>(A)) {
369 const IntegerType *BI = cast<IntegerType>(B);
370 return AI->getBitWidth() < BI->getBitWidth();
372 if (const PointerType *AI = dyn_cast<PointerType>(A)) {
373 const PointerType *BI = cast<PointerType>(B);
374 return CompareTypes(AI->getElementType(), BI->getElementType());
376 if (const ArrayType *AI = dyn_cast<ArrayType>(A)) {
377 const ArrayType *BI = cast<ArrayType>(B);
378 if (AI->getNumElements() != BI->getNumElements())
379 return AI->getNumElements() < BI->getNumElements();
380 return CompareTypes(AI->getElementType(), BI->getElementType());
382 if (const VectorType *AI = dyn_cast<VectorType>(A)) {
383 const VectorType *BI = cast<VectorType>(B);
384 if (AI->getNumElements() != BI->getNumElements())
385 return AI->getNumElements() < BI->getNumElements();
386 return CompareTypes(AI->getElementType(), BI->getElementType());
388 if (const StructType *AI = dyn_cast<StructType>(A)) {
389 const StructType *BI = cast<StructType>(B);
390 if (AI->getNumElements() != BI->getNumElements())
391 return AI->getNumElements() < BI->getNumElements();
392 for (unsigned i = 0, e = AI->getNumElements(); i != e; ++i)
393 if (CompareTypes(AI->getElementType(i), BI->getElementType(i)) ||
394 CompareTypes(BI->getElementType(i), AI->getElementType(i)))
395 return CompareTypes(AI->getElementType(i), BI->getElementType(i));
401 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
402 /// than the complexity of the RHS. This comparator is used to canonicalize
404 class VISIBILITY_HIDDEN SCEVComplexityCompare {
407 explicit SCEVComplexityCompare(LoopInfo *li) : LI(li) {}
409 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
410 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
414 // Primarily, sort the SCEVs by their getSCEVType().
415 if (LHS->getSCEVType() != RHS->getSCEVType())
416 return LHS->getSCEVType() < RHS->getSCEVType();
418 // Aside from the getSCEVType() ordering, the particular ordering
419 // isn't very important except that it's beneficial to be consistent,
420 // so that (a + b) and (b + a) don't end up as different expressions.
422 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
423 // not as complete as it could be.
424 if (const SCEVUnknown *LU = dyn_cast<SCEVUnknown>(LHS)) {
425 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
427 // Order pointer values after integer values. This helps SCEVExpander
429 if (isa<PointerType>(LU->getType()) && !isa<PointerType>(RU->getType()))
431 if (isa<PointerType>(RU->getType()) && !isa<PointerType>(LU->getType()))
434 // Compare getValueID values.
435 if (LU->getValue()->getValueID() != RU->getValue()->getValueID())
436 return LU->getValue()->getValueID() < RU->getValue()->getValueID();
438 // Sort arguments by their position.
439 if (const Argument *LA = dyn_cast<Argument>(LU->getValue())) {
440 const Argument *RA = cast<Argument>(RU->getValue());
441 return LA->getArgNo() < RA->getArgNo();
444 // For instructions, compare their loop depth, and their opcode.
445 // This is pretty loose.
446 if (Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
447 Instruction *RV = cast<Instruction>(RU->getValue());
449 // Compare loop depths.
450 if (LI->getLoopDepth(LV->getParent()) !=
451 LI->getLoopDepth(RV->getParent()))
452 return LI->getLoopDepth(LV->getParent()) <
453 LI->getLoopDepth(RV->getParent());
456 if (LV->getOpcode() != RV->getOpcode())
457 return LV->getOpcode() < RV->getOpcode();
459 // Compare the number of operands.
460 if (LV->getNumOperands() != RV->getNumOperands())
461 return LV->getNumOperands() < RV->getNumOperands();
467 // Compare constant values.
468 if (const SCEVConstant *LC = dyn_cast<SCEVConstant>(LHS)) {
469 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
470 if (LC->getValue()->getBitWidth() != RC->getValue()->getBitWidth())
471 return LC->getValue()->getBitWidth() < RC->getValue()->getBitWidth();
472 return LC->getValue()->getValue().ult(RC->getValue()->getValue());
475 // Compare addrec loop depths.
476 if (const SCEVAddRecExpr *LA = dyn_cast<SCEVAddRecExpr>(LHS)) {
477 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
478 if (LA->getLoop()->getLoopDepth() != RA->getLoop()->getLoopDepth())
479 return LA->getLoop()->getLoopDepth() < RA->getLoop()->getLoopDepth();
482 // Lexicographically compare n-ary expressions.
483 if (const SCEVNAryExpr *LC = dyn_cast<SCEVNAryExpr>(LHS)) {
484 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
485 for (unsigned i = 0, e = LC->getNumOperands(); i != e; ++i) {
486 if (i >= RC->getNumOperands())
488 if (operator()(LC->getOperand(i), RC->getOperand(i)))
490 if (operator()(RC->getOperand(i), LC->getOperand(i)))
493 return LC->getNumOperands() < RC->getNumOperands();
496 // Lexicographically compare udiv expressions.
497 if (const SCEVUDivExpr *LC = dyn_cast<SCEVUDivExpr>(LHS)) {
498 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
499 if (operator()(LC->getLHS(), RC->getLHS()))
501 if (operator()(RC->getLHS(), LC->getLHS()))
503 if (operator()(LC->getRHS(), RC->getRHS()))
505 if (operator()(RC->getRHS(), LC->getRHS()))
510 // Compare cast expressions by operand.
511 if (const SCEVCastExpr *LC = dyn_cast<SCEVCastExpr>(LHS)) {
512 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
513 return operator()(LC->getOperand(), RC->getOperand());
516 // Compare offsetof expressions.
517 if (const SCEVFieldOffsetExpr *LA = dyn_cast<SCEVFieldOffsetExpr>(LHS)) {
518 const SCEVFieldOffsetExpr *RA = cast<SCEVFieldOffsetExpr>(RHS);
519 if (CompareTypes(LA->getStructType(), RA->getStructType()) ||
520 CompareTypes(RA->getStructType(), LA->getStructType()))
521 return CompareTypes(LA->getStructType(), RA->getStructType());
522 return LA->getFieldNo() < RA->getFieldNo();
525 // Compare sizeof expressions by the allocation type.
526 if (const SCEVAllocSizeExpr *LA = dyn_cast<SCEVAllocSizeExpr>(LHS)) {
527 const SCEVAllocSizeExpr *RA = cast<SCEVAllocSizeExpr>(RHS);
528 return CompareTypes(LA->getAllocType(), RA->getAllocType());
531 llvm_unreachable("Unknown SCEV kind!");
537 /// GroupByComplexity - Given a list of SCEV objects, order them by their
538 /// complexity, and group objects of the same complexity together by value.
539 /// When this routine is finished, we know that any duplicates in the vector are
540 /// consecutive and that complexity is monotonically increasing.
542 /// Note that we go take special precautions to ensure that we get determinstic
543 /// results from this routine. In other words, we don't want the results of
544 /// this to depend on where the addresses of various SCEV objects happened to
547 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
549 if (Ops.size() < 2) return; // Noop
550 if (Ops.size() == 2) {
551 // This is the common case, which also happens to be trivially simple.
553 if (SCEVComplexityCompare(LI)(Ops[1], Ops[0]))
554 std::swap(Ops[0], Ops[1]);
558 // Do the rough sort by complexity.
559 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
561 // Now that we are sorted by complexity, group elements of the same
562 // complexity. Note that this is, at worst, N^2, but the vector is likely to
563 // be extremely short in practice. Note that we take this approach because we
564 // do not want to depend on the addresses of the objects we are grouping.
565 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
566 const SCEV *S = Ops[i];
567 unsigned Complexity = S->getSCEVType();
569 // If there are any objects of the same complexity and same value as this
571 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
572 if (Ops[j] == S) { // Found a duplicate.
573 // Move it to immediately after i'th element.
574 std::swap(Ops[i+1], Ops[j]);
575 ++i; // no need to rescan it.
576 if (i == e-2) return; // Done!
584 //===----------------------------------------------------------------------===//
585 // Simple SCEV method implementations
586 //===----------------------------------------------------------------------===//
588 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
590 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
592 const Type* ResultTy) {
593 // Handle the simplest case efficiently.
595 return SE.getTruncateOrZeroExtend(It, ResultTy);
597 // We are using the following formula for BC(It, K):
599 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
601 // Suppose, W is the bitwidth of the return value. We must be prepared for
602 // overflow. Hence, we must assure that the result of our computation is
603 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
604 // safe in modular arithmetic.
606 // However, this code doesn't use exactly that formula; the formula it uses
607 // is something like the following, where T is the number of factors of 2 in
608 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
611 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
613 // This formula is trivially equivalent to the previous formula. However,
614 // this formula can be implemented much more efficiently. The trick is that
615 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
616 // arithmetic. To do exact division in modular arithmetic, all we have
617 // to do is multiply by the inverse. Therefore, this step can be done at
620 // The next issue is how to safely do the division by 2^T. The way this
621 // is done is by doing the multiplication step at a width of at least W + T
622 // bits. This way, the bottom W+T bits of the product are accurate. Then,
623 // when we perform the division by 2^T (which is equivalent to a right shift
624 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
625 // truncated out after the division by 2^T.
627 // In comparison to just directly using the first formula, this technique
628 // is much more efficient; using the first formula requires W * K bits,
629 // but this formula less than W + K bits. Also, the first formula requires
630 // a division step, whereas this formula only requires multiplies and shifts.
632 // It doesn't matter whether the subtraction step is done in the calculation
633 // width or the input iteration count's width; if the subtraction overflows,
634 // the result must be zero anyway. We prefer here to do it in the width of
635 // the induction variable because it helps a lot for certain cases; CodeGen
636 // isn't smart enough to ignore the overflow, which leads to much less
637 // efficient code if the width of the subtraction is wider than the native
640 // (It's possible to not widen at all by pulling out factors of 2 before
641 // the multiplication; for example, K=2 can be calculated as
642 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
643 // extra arithmetic, so it's not an obvious win, and it gets
644 // much more complicated for K > 3.)
646 // Protection from insane SCEVs; this bound is conservative,
647 // but it probably doesn't matter.
649 return SE.getCouldNotCompute();
651 unsigned W = SE.getTypeSizeInBits(ResultTy);
653 // Calculate K! / 2^T and T; we divide out the factors of two before
654 // multiplying for calculating K! / 2^T to avoid overflow.
655 // Other overflow doesn't matter because we only care about the bottom
656 // W bits of the result.
657 APInt OddFactorial(W, 1);
659 for (unsigned i = 3; i <= K; ++i) {
661 unsigned TwoFactors = Mult.countTrailingZeros();
663 Mult = Mult.lshr(TwoFactors);
664 OddFactorial *= Mult;
667 // We need at least W + T bits for the multiplication step
668 unsigned CalculationBits = W + T;
670 // Calcuate 2^T, at width T+W.
671 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
673 // Calculate the multiplicative inverse of K! / 2^T;
674 // this multiplication factor will perform the exact division by
676 APInt Mod = APInt::getSignedMinValue(W+1);
677 APInt MultiplyFactor = OddFactorial.zext(W+1);
678 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
679 MultiplyFactor = MultiplyFactor.trunc(W);
681 // Calculate the product, at width T+W
682 const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
684 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
685 for (unsigned i = 1; i != K; ++i) {
686 const SCEV *S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
687 Dividend = SE.getMulExpr(Dividend,
688 SE.getTruncateOrZeroExtend(S, CalculationTy));
692 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
694 // Truncate the result, and divide by K! / 2^T.
696 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
697 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
700 /// evaluateAtIteration - Return the value of this chain of recurrences at
701 /// the specified iteration number. We can evaluate this recurrence by
702 /// multiplying each element in the chain by the binomial coefficient
703 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
705 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
707 /// where BC(It, k) stands for binomial coefficient.
709 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
710 ScalarEvolution &SE) const {
711 const SCEV *Result = getStart();
712 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
713 // The computation is correct in the face of overflow provided that the
714 // multiplication is performed _after_ the evaluation of the binomial
716 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
717 if (isa<SCEVCouldNotCompute>(Coeff))
720 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
725 //===----------------------------------------------------------------------===//
726 // SCEV Expression folder implementations
727 //===----------------------------------------------------------------------===//
729 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
731 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
732 "This is not a truncating conversion!");
733 assert(isSCEVable(Ty) &&
734 "This is not a conversion to a SCEVable type!");
735 Ty = getEffectiveSCEVType(Ty);
738 ID.AddInteger(scTruncate);
742 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
744 // Fold if the operand is constant.
745 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
747 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
749 // trunc(trunc(x)) --> trunc(x)
750 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
751 return getTruncateExpr(ST->getOperand(), Ty);
753 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
754 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
755 return getTruncateOrSignExtend(SS->getOperand(), Ty);
757 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
758 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
759 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
761 // If the input value is a chrec scev, truncate the chrec's operands.
762 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
763 SmallVector<const SCEV *, 4> Operands;
764 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
765 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
766 return getAddRecExpr(Operands, AddRec->getLoop());
769 // The cast wasn't folded; create an explicit cast node.
770 // Recompute the insert position, as it may have been invalidated.
771 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
772 SCEV *S = SCEVAllocator.Allocate<SCEVTruncateExpr>();
773 new (S) SCEVTruncateExpr(ID, Op, Ty);
774 UniqueSCEVs.InsertNode(S, IP);
778 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
780 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
781 "This is not an extending conversion!");
782 assert(isSCEVable(Ty) &&
783 "This is not a conversion to a SCEVable type!");
784 Ty = getEffectiveSCEVType(Ty);
786 // Fold if the operand is constant.
787 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
788 const Type *IntTy = getEffectiveSCEVType(Ty);
789 Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
790 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
791 return getConstant(cast<ConstantInt>(C));
794 // zext(zext(x)) --> zext(x)
795 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
796 return getZeroExtendExpr(SZ->getOperand(), Ty);
798 // Before doing any expensive analysis, check to see if we've already
799 // computed a SCEV for this Op and Ty.
801 ID.AddInteger(scZeroExtend);
805 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
807 // If the input value is a chrec scev, and we can prove that the value
808 // did not overflow the old, smaller, value, we can zero extend all of the
809 // operands (often constants). This allows analysis of something like
810 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
811 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
812 if (AR->isAffine()) {
813 const SCEV *Start = AR->getStart();
814 const SCEV *Step = AR->getStepRecurrence(*this);
815 unsigned BitWidth = getTypeSizeInBits(AR->getType());
816 const Loop *L = AR->getLoop();
818 // If we have special knowledge that this addrec won't overflow,
819 // we don't need to do any further analysis.
820 if (AR->hasNoUnsignedWrap())
821 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
822 getZeroExtendExpr(Step, Ty),
825 // Check whether the backedge-taken count is SCEVCouldNotCompute.
826 // Note that this serves two purposes: It filters out loops that are
827 // simply not analyzable, and it covers the case where this code is
828 // being called from within backedge-taken count analysis, such that
829 // attempting to ask for the backedge-taken count would likely result
830 // in infinite recursion. In the later case, the analysis code will
831 // cope with a conservative value, and it will take care to purge
832 // that value once it has finished.
833 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
834 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
835 // Manually compute the final value for AR, checking for
838 // Check whether the backedge-taken count can be losslessly casted to
839 // the addrec's type. The count is always unsigned.
840 const SCEV *CastedMaxBECount =
841 getTruncateOrZeroExtend(MaxBECount, Start->getType());
842 const SCEV *RecastedMaxBECount =
843 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
844 if (MaxBECount == RecastedMaxBECount) {
845 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
846 // Check whether Start+Step*MaxBECount has no unsigned overflow.
848 getMulExpr(CastedMaxBECount,
849 getTruncateOrZeroExtend(Step, Start->getType()));
850 const SCEV *Add = getAddExpr(Start, ZMul);
851 const SCEV *OperandExtendedAdd =
852 getAddExpr(getZeroExtendExpr(Start, WideTy),
853 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
854 getZeroExtendExpr(Step, WideTy)));
855 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
856 // Return the expression with the addrec on the outside.
857 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
858 getZeroExtendExpr(Step, Ty),
861 // Similar to above, only this time treat the step value as signed.
862 // This covers loops that count down.
864 getMulExpr(CastedMaxBECount,
865 getTruncateOrSignExtend(Step, Start->getType()));
866 Add = getAddExpr(Start, SMul);
868 getAddExpr(getZeroExtendExpr(Start, WideTy),
869 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
870 getSignExtendExpr(Step, WideTy)));
871 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
872 // Return the expression with the addrec on the outside.
873 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
874 getSignExtendExpr(Step, Ty),
878 // If the backedge is guarded by a comparison with the pre-inc value
879 // the addrec is safe. Also, if the entry is guarded by a comparison
880 // with the start value and the backedge is guarded by a comparison
881 // with the post-inc value, the addrec is safe.
882 if (isKnownPositive(Step)) {
883 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
884 getUnsignedRange(Step).getUnsignedMax());
885 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
886 (isLoopGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
887 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
888 AR->getPostIncExpr(*this), N)))
889 // Return the expression with the addrec on the outside.
890 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
891 getZeroExtendExpr(Step, Ty),
893 } else if (isKnownNegative(Step)) {
894 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
895 getSignedRange(Step).getSignedMin());
896 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) &&
897 (isLoopGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) ||
898 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
899 AR->getPostIncExpr(*this), N)))
900 // Return the expression with the addrec on the outside.
901 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
902 getSignExtendExpr(Step, Ty),
908 // The cast wasn't folded; create an explicit cast node.
909 // Recompute the insert position, as it may have been invalidated.
910 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
911 SCEV *S = SCEVAllocator.Allocate<SCEVZeroExtendExpr>();
912 new (S) SCEVZeroExtendExpr(ID, Op, Ty);
913 UniqueSCEVs.InsertNode(S, IP);
917 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
919 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
920 "This is not an extending conversion!");
921 assert(isSCEVable(Ty) &&
922 "This is not a conversion to a SCEVable type!");
923 Ty = getEffectiveSCEVType(Ty);
925 // Fold if the operand is constant.
926 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
927 const Type *IntTy = getEffectiveSCEVType(Ty);
928 Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
929 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
930 return getConstant(cast<ConstantInt>(C));
933 // sext(sext(x)) --> sext(x)
934 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
935 return getSignExtendExpr(SS->getOperand(), Ty);
937 // Before doing any expensive analysis, check to see if we've already
938 // computed a SCEV for this Op and Ty.
940 ID.AddInteger(scSignExtend);
944 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
946 // If the input value is a chrec scev, and we can prove that the value
947 // did not overflow the old, smaller, value, we can sign extend all of the
948 // operands (often constants). This allows analysis of something like
949 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
950 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
951 if (AR->isAffine()) {
952 const SCEV *Start = AR->getStart();
953 const SCEV *Step = AR->getStepRecurrence(*this);
954 unsigned BitWidth = getTypeSizeInBits(AR->getType());
955 const Loop *L = AR->getLoop();
957 // If we have special knowledge that this addrec won't overflow,
958 // we don't need to do any further analysis.
959 if (AR->hasNoSignedWrap())
960 return getAddRecExpr(getSignExtendExpr(Start, Ty),
961 getSignExtendExpr(Step, Ty),
964 // Check whether the backedge-taken count is SCEVCouldNotCompute.
965 // Note that this serves two purposes: It filters out loops that are
966 // simply not analyzable, and it covers the case where this code is
967 // being called from within backedge-taken count analysis, such that
968 // attempting to ask for the backedge-taken count would likely result
969 // in infinite recursion. In the later case, the analysis code will
970 // cope with a conservative value, and it will take care to purge
971 // that value once it has finished.
972 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
973 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
974 // Manually compute the final value for AR, checking for
977 // Check whether the backedge-taken count can be losslessly casted to
978 // the addrec's type. The count is always unsigned.
979 const SCEV *CastedMaxBECount =
980 getTruncateOrZeroExtend(MaxBECount, Start->getType());
981 const SCEV *RecastedMaxBECount =
982 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
983 if (MaxBECount == RecastedMaxBECount) {
984 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
985 // Check whether Start+Step*MaxBECount has no signed overflow.
987 getMulExpr(CastedMaxBECount,
988 getTruncateOrSignExtend(Step, Start->getType()));
989 const SCEV *Add = getAddExpr(Start, SMul);
990 const SCEV *OperandExtendedAdd =
991 getAddExpr(getSignExtendExpr(Start, WideTy),
992 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
993 getSignExtendExpr(Step, WideTy)));
994 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
995 // Return the expression with the addrec on the outside.
996 return getAddRecExpr(getSignExtendExpr(Start, Ty),
997 getSignExtendExpr(Step, Ty),
1000 // Similar to above, only this time treat the step value as unsigned.
1001 // This covers loops that count up with an unsigned step.
1003 getMulExpr(CastedMaxBECount,
1004 getTruncateOrZeroExtend(Step, Start->getType()));
1005 Add = getAddExpr(Start, UMul);
1006 OperandExtendedAdd =
1007 getAddExpr(getSignExtendExpr(Start, WideTy),
1008 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1009 getZeroExtendExpr(Step, WideTy)));
1010 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1011 // Return the expression with the addrec on the outside.
1012 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1013 getZeroExtendExpr(Step, Ty),
1017 // If the backedge is guarded by a comparison with the pre-inc value
1018 // the addrec is safe. Also, if the entry is guarded by a comparison
1019 // with the start value and the backedge is guarded by a comparison
1020 // with the post-inc value, the addrec is safe.
1021 if (isKnownPositive(Step)) {
1022 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1023 getSignedRange(Step).getSignedMax());
1024 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1025 (isLoopGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1026 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1027 AR->getPostIncExpr(*this), N)))
1028 // Return the expression with the addrec on the outside.
1029 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1030 getSignExtendExpr(Step, Ty),
1032 } else if (isKnownNegative(Step)) {
1033 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1034 getSignedRange(Step).getSignedMin());
1035 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1036 (isLoopGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1037 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1038 AR->getPostIncExpr(*this), N)))
1039 // Return the expression with the addrec on the outside.
1040 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1041 getSignExtendExpr(Step, Ty),
1047 // The cast wasn't folded; create an explicit cast node.
1048 // Recompute the insert position, as it may have been invalidated.
1049 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1050 SCEV *S = SCEVAllocator.Allocate<SCEVSignExtendExpr>();
1051 new (S) SCEVSignExtendExpr(ID, Op, Ty);
1052 UniqueSCEVs.InsertNode(S, IP);
1056 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1057 /// unspecified bits out to the given type.
1059 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1061 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1062 "This is not an extending conversion!");
1063 assert(isSCEVable(Ty) &&
1064 "This is not a conversion to a SCEVable type!");
1065 Ty = getEffectiveSCEVType(Ty);
1067 // Sign-extend negative constants.
1068 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1069 if (SC->getValue()->getValue().isNegative())
1070 return getSignExtendExpr(Op, Ty);
1072 // Peel off a truncate cast.
1073 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1074 const SCEV *NewOp = T->getOperand();
1075 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1076 return getAnyExtendExpr(NewOp, Ty);
1077 return getTruncateOrNoop(NewOp, Ty);
1080 // Next try a zext cast. If the cast is folded, use it.
1081 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1082 if (!isa<SCEVZeroExtendExpr>(ZExt))
1085 // Next try a sext cast. If the cast is folded, use it.
1086 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1087 if (!isa<SCEVSignExtendExpr>(SExt))
1090 // If the expression is obviously signed, use the sext cast value.
1091 if (isa<SCEVSMaxExpr>(Op))
1094 // Absent any other information, use the zext cast value.
1098 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1099 /// a list of operands to be added under the given scale, update the given
1100 /// map. This is a helper function for getAddRecExpr. As an example of
1101 /// what it does, given a sequence of operands that would form an add
1102 /// expression like this:
1104 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1106 /// where A and B are constants, update the map with these values:
1108 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1110 /// and add 13 + A*B*29 to AccumulatedConstant.
1111 /// This will allow getAddRecExpr to produce this:
1113 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1115 /// This form often exposes folding opportunities that are hidden in
1116 /// the original operand list.
1118 /// Return true iff it appears that any interesting folding opportunities
1119 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1120 /// the common case where no interesting opportunities are present, and
1121 /// is also used as a check to avoid infinite recursion.
1124 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1125 SmallVector<const SCEV *, 8> &NewOps,
1126 APInt &AccumulatedConstant,
1127 const SmallVectorImpl<const SCEV *> &Ops,
1129 ScalarEvolution &SE) {
1130 bool Interesting = false;
1132 // Iterate over the add operands.
1133 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1134 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1135 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1137 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1138 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1139 // A multiplication of a constant with another add; recurse.
1141 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1142 cast<SCEVAddExpr>(Mul->getOperand(1))
1146 // A multiplication of a constant with some other value. Update
1148 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1149 const SCEV *Key = SE.getMulExpr(MulOps);
1150 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1151 M.insert(std::make_pair(Key, NewScale));
1153 NewOps.push_back(Pair.first->first);
1155 Pair.first->second += NewScale;
1156 // The map already had an entry for this value, which may indicate
1157 // a folding opportunity.
1161 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1162 // Pull a buried constant out to the outside.
1163 if (Scale != 1 || AccumulatedConstant != 0 || C->isZero())
1165 AccumulatedConstant += Scale * C->getValue()->getValue();
1167 // An ordinary operand. Update the map.
1168 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1169 M.insert(std::make_pair(Ops[i], Scale));
1171 NewOps.push_back(Pair.first->first);
1173 Pair.first->second += Scale;
1174 // The map already had an entry for this value, which may indicate
1175 // a folding opportunity.
1185 struct APIntCompare {
1186 bool operator()(const APInt &LHS, const APInt &RHS) const {
1187 return LHS.ult(RHS);
1192 /// getAddExpr - Get a canonical add expression, or something simpler if
1194 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops) {
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);
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);
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 SCEV *S = SCEVAllocator.Allocate<SCEVAddExpr>();
1520 new (S) SCEVAddExpr(ID, Ops);
1521 UniqueSCEVs.InsertNode(S, IP);
1526 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1528 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops) {
1529 assert(!Ops.empty() && "Cannot get empty mul!");
1531 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1532 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1533 getEffectiveSCEVType(Ops[0]->getType()) &&
1534 "SCEVMulExpr operand types don't match!");
1537 // Sort by complexity, this groups all similar expression types together.
1538 GroupByComplexity(Ops, LI);
1540 // If there are any constants, fold them together.
1542 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1544 // C1*(C2+V) -> C1*C2 + C1*V
1545 if (Ops.size() == 2)
1546 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1547 if (Add->getNumOperands() == 2 &&
1548 isa<SCEVConstant>(Add->getOperand(0)))
1549 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1550 getMulExpr(LHSC, Add->getOperand(1)));
1554 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1555 // We found two constants, fold them together!
1556 ConstantInt *Fold = ConstantInt::get(getContext(),
1557 LHSC->getValue()->getValue() *
1558 RHSC->getValue()->getValue());
1559 Ops[0] = getConstant(Fold);
1560 Ops.erase(Ops.begin()+1); // Erase the folded element
1561 if (Ops.size() == 1) return Ops[0];
1562 LHSC = cast<SCEVConstant>(Ops[0]);
1565 // If we are left with a constant one being multiplied, strip it off.
1566 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1567 Ops.erase(Ops.begin());
1569 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1570 // If we have a multiply of zero, it will always be zero.
1575 // Skip over the add expression until we get to a multiply.
1576 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1579 if (Ops.size() == 1)
1582 // If there are mul operands inline them all into this expression.
1583 if (Idx < Ops.size()) {
1584 bool DeletedMul = false;
1585 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1586 // If we have an mul, expand the mul operands onto the end of the operands
1588 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1589 Ops.erase(Ops.begin()+Idx);
1593 // If we deleted at least one mul, we added operands to the end of the list,
1594 // and they are not necessarily sorted. Recurse to resort and resimplify
1595 // any operands we just aquired.
1597 return getMulExpr(Ops);
1600 // If there are any add recurrences in the operands list, see if any other
1601 // added values are loop invariant. If so, we can fold them into the
1603 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1606 // Scan over all recurrences, trying to fold loop invariants into them.
1607 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1608 // Scan all of the other operands to this mul and add them to the vector if
1609 // they are loop invariant w.r.t. the recurrence.
1610 SmallVector<const SCEV *, 8> LIOps;
1611 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1612 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1613 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1614 LIOps.push_back(Ops[i]);
1615 Ops.erase(Ops.begin()+i);
1619 // If we found some loop invariants, fold them into the recurrence.
1620 if (!LIOps.empty()) {
1621 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1622 SmallVector<const SCEV *, 4> NewOps;
1623 NewOps.reserve(AddRec->getNumOperands());
1624 if (LIOps.size() == 1) {
1625 const SCEV *Scale = LIOps[0];
1626 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1627 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1629 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1630 SmallVector<const SCEV *, 4> MulOps(LIOps.begin(), LIOps.end());
1631 MulOps.push_back(AddRec->getOperand(i));
1632 NewOps.push_back(getMulExpr(MulOps));
1636 const SCEV *NewRec = getAddRecExpr(NewOps, AddRec->getLoop());
1638 // If all of the other operands were loop invariant, we are done.
1639 if (Ops.size() == 1) return NewRec;
1641 // Otherwise, multiply the folded AddRec by the non-liv parts.
1642 for (unsigned i = 0;; ++i)
1643 if (Ops[i] == AddRec) {
1647 return getMulExpr(Ops);
1650 // Okay, if there weren't any loop invariants to be folded, check to see if
1651 // there are multiple AddRec's with the same loop induction variable being
1652 // multiplied together. If so, we can fold them.
1653 for (unsigned OtherIdx = Idx+1;
1654 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1655 if (OtherIdx != Idx) {
1656 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1657 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1658 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1659 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1660 const SCEV *NewStart = getMulExpr(F->getStart(),
1662 const SCEV *B = F->getStepRecurrence(*this);
1663 const SCEV *D = G->getStepRecurrence(*this);
1664 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1667 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1669 if (Ops.size() == 2) return NewAddRec;
1671 Ops.erase(Ops.begin()+Idx);
1672 Ops.erase(Ops.begin()+OtherIdx-1);
1673 Ops.push_back(NewAddRec);
1674 return getMulExpr(Ops);
1678 // Otherwise couldn't fold anything into this recurrence. Move onto the
1682 // Okay, it looks like we really DO need an mul expr. Check to see if we
1683 // already have one, otherwise create a new one.
1684 FoldingSetNodeID ID;
1685 ID.AddInteger(scMulExpr);
1686 ID.AddInteger(Ops.size());
1687 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1688 ID.AddPointer(Ops[i]);
1690 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1691 SCEV *S = SCEVAllocator.Allocate<SCEVMulExpr>();
1692 new (S) SCEVMulExpr(ID, Ops);
1693 UniqueSCEVs.InsertNode(S, IP);
1697 /// getUDivExpr - Get a canonical unsigned division expression, or something
1698 /// simpler if possible.
1699 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1701 assert(getEffectiveSCEVType(LHS->getType()) ==
1702 getEffectiveSCEVType(RHS->getType()) &&
1703 "SCEVUDivExpr operand types don't match!");
1705 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1706 if (RHSC->getValue()->equalsInt(1))
1707 return LHS; // X udiv 1 --> x
1709 return getIntegerSCEV(0, LHS->getType()); // value is undefined
1711 // Determine if the division can be folded into the operands of
1713 // TODO: Generalize this to non-constants by using known-bits information.
1714 const Type *Ty = LHS->getType();
1715 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1716 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ;
1717 // For non-power-of-two values, effectively round the value up to the
1718 // nearest power of two.
1719 if (!RHSC->getValue()->getValue().isPowerOf2())
1721 const IntegerType *ExtTy =
1722 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1723 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1724 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1725 if (const SCEVConstant *Step =
1726 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1727 if (!Step->getValue()->getValue()
1728 .urem(RHSC->getValue()->getValue()) &&
1729 getZeroExtendExpr(AR, ExtTy) ==
1730 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1731 getZeroExtendExpr(Step, ExtTy),
1733 SmallVector<const SCEV *, 4> Operands;
1734 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1735 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1736 return getAddRecExpr(Operands, AR->getLoop());
1738 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1739 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1740 SmallVector<const SCEV *, 4> Operands;
1741 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1742 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1743 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1744 // Find an operand that's safely divisible.
1745 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1746 const SCEV *Op = M->getOperand(i);
1747 const SCEV *Div = getUDivExpr(Op, RHSC);
1748 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1749 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
1750 Operands = SmallVector<const SCEV *, 4>(MOperands.begin(),
1753 return getMulExpr(Operands);
1757 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1758 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1759 SmallVector<const SCEV *, 4> Operands;
1760 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1761 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1762 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1764 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1765 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
1766 if (isa<SCEVUDivExpr>(Op) || getMulExpr(Op, RHS) != A->getOperand(i))
1768 Operands.push_back(Op);
1770 if (Operands.size() == A->getNumOperands())
1771 return getAddExpr(Operands);
1775 // Fold if both operands are constant.
1776 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1777 Constant *LHSCV = LHSC->getValue();
1778 Constant *RHSCV = RHSC->getValue();
1779 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
1784 FoldingSetNodeID ID;
1785 ID.AddInteger(scUDivExpr);
1789 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1790 SCEV *S = SCEVAllocator.Allocate<SCEVUDivExpr>();
1791 new (S) SCEVUDivExpr(ID, LHS, RHS);
1792 UniqueSCEVs.InsertNode(S, IP);
1797 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1798 /// Simplify the expression as much as possible.
1799 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
1800 const SCEV *Step, const Loop *L) {
1801 SmallVector<const SCEV *, 4> Operands;
1802 Operands.push_back(Start);
1803 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1804 if (StepChrec->getLoop() == L) {
1805 Operands.insert(Operands.end(), StepChrec->op_begin(),
1806 StepChrec->op_end());
1807 return getAddRecExpr(Operands, L);
1810 Operands.push_back(Step);
1811 return getAddRecExpr(Operands, L);
1814 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1815 /// Simplify the expression as much as possible.
1817 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
1819 if (Operands.size() == 1) return Operands[0];
1821 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
1822 assert(getEffectiveSCEVType(Operands[i]->getType()) ==
1823 getEffectiveSCEVType(Operands[0]->getType()) &&
1824 "SCEVAddRecExpr operand types don't match!");
1827 if (Operands.back()->isZero()) {
1828 Operands.pop_back();
1829 return getAddRecExpr(Operands, L); // {X,+,0} --> X
1832 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
1833 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
1834 const Loop* NestedLoop = NestedAR->getLoop();
1835 if (L->getLoopDepth() < NestedLoop->getLoopDepth()) {
1836 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
1837 NestedAR->op_end());
1838 Operands[0] = NestedAR->getStart();
1839 // AddRecs require their operands be loop-invariant with respect to their
1840 // loops. Don't perform this transformation if it would break this
1842 bool AllInvariant = true;
1843 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1844 if (!Operands[i]->isLoopInvariant(L)) {
1845 AllInvariant = false;
1849 NestedOperands[0] = getAddRecExpr(Operands, L);
1850 AllInvariant = true;
1851 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
1852 if (!NestedOperands[i]->isLoopInvariant(NestedLoop)) {
1853 AllInvariant = false;
1857 // Ok, both add recurrences are valid after the transformation.
1858 return getAddRecExpr(NestedOperands, NestedLoop);
1860 // Reset Operands to its original state.
1861 Operands[0] = NestedAR;
1865 FoldingSetNodeID ID;
1866 ID.AddInteger(scAddRecExpr);
1867 ID.AddInteger(Operands.size());
1868 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1869 ID.AddPointer(Operands[i]);
1872 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1873 SCEV *S = SCEVAllocator.Allocate<SCEVAddRecExpr>();
1874 new (S) SCEVAddRecExpr(ID, Operands, L);
1875 UniqueSCEVs.InsertNode(S, IP);
1879 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
1881 SmallVector<const SCEV *, 2> Ops;
1884 return getSMaxExpr(Ops);
1888 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
1889 assert(!Ops.empty() && "Cannot get empty smax!");
1890 if (Ops.size() == 1) return Ops[0];
1892 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1893 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1894 getEffectiveSCEVType(Ops[0]->getType()) &&
1895 "SCEVSMaxExpr operand types don't match!");
1898 // Sort by complexity, this groups all similar expression types together.
1899 GroupByComplexity(Ops, LI);
1901 // If there are any constants, fold them together.
1903 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1905 assert(Idx < Ops.size());
1906 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1907 // We found two constants, fold them together!
1908 ConstantInt *Fold = ConstantInt::get(getContext(),
1909 APIntOps::smax(LHSC->getValue()->getValue(),
1910 RHSC->getValue()->getValue()));
1911 Ops[0] = getConstant(Fold);
1912 Ops.erase(Ops.begin()+1); // Erase the folded element
1913 if (Ops.size() == 1) return Ops[0];
1914 LHSC = cast<SCEVConstant>(Ops[0]);
1917 // If we are left with a constant minimum-int, strip it off.
1918 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
1919 Ops.erase(Ops.begin());
1921 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
1922 // If we have an smax with a constant maximum-int, it will always be
1928 if (Ops.size() == 1) return Ops[0];
1930 // Find the first SMax
1931 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
1934 // Check to see if one of the operands is an SMax. If so, expand its operands
1935 // onto our operand list, and recurse to simplify.
1936 if (Idx < Ops.size()) {
1937 bool DeletedSMax = false;
1938 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
1939 Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
1940 Ops.erase(Ops.begin()+Idx);
1945 return getSMaxExpr(Ops);
1948 // Okay, check to see if the same value occurs in the operand list twice. If
1949 // so, delete one. Since we sorted the list, these values are required to
1951 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1952 if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
1953 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1957 if (Ops.size() == 1) return Ops[0];
1959 assert(!Ops.empty() && "Reduced smax down to nothing!");
1961 // Okay, it looks like we really DO need an smax expr. Check to see if we
1962 // already have one, otherwise create a new one.
1963 FoldingSetNodeID ID;
1964 ID.AddInteger(scSMaxExpr);
1965 ID.AddInteger(Ops.size());
1966 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1967 ID.AddPointer(Ops[i]);
1969 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1970 SCEV *S = SCEVAllocator.Allocate<SCEVSMaxExpr>();
1971 new (S) SCEVSMaxExpr(ID, Ops);
1972 UniqueSCEVs.InsertNode(S, IP);
1976 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
1978 SmallVector<const SCEV *, 2> Ops;
1981 return getUMaxExpr(Ops);
1985 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
1986 assert(!Ops.empty() && "Cannot get empty umax!");
1987 if (Ops.size() == 1) return Ops[0];
1989 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1990 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1991 getEffectiveSCEVType(Ops[0]->getType()) &&
1992 "SCEVUMaxExpr operand types don't match!");
1995 // Sort by complexity, this groups all similar expression types together.
1996 GroupByComplexity(Ops, LI);
1998 // If there are any constants, fold them together.
2000 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2002 assert(Idx < Ops.size());
2003 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2004 // We found two constants, fold them together!
2005 ConstantInt *Fold = ConstantInt::get(getContext(),
2006 APIntOps::umax(LHSC->getValue()->getValue(),
2007 RHSC->getValue()->getValue()));
2008 Ops[0] = getConstant(Fold);
2009 Ops.erase(Ops.begin()+1); // Erase the folded element
2010 if (Ops.size() == 1) return Ops[0];
2011 LHSC = cast<SCEVConstant>(Ops[0]);
2014 // If we are left with a constant minimum-int, strip it off.
2015 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2016 Ops.erase(Ops.begin());
2018 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2019 // If we have an umax with a constant maximum-int, it will always be
2025 if (Ops.size() == 1) return Ops[0];
2027 // Find the first UMax
2028 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2031 // Check to see if one of the operands is a UMax. If so, expand its operands
2032 // onto our operand list, and recurse to simplify.
2033 if (Idx < Ops.size()) {
2034 bool DeletedUMax = false;
2035 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2036 Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
2037 Ops.erase(Ops.begin()+Idx);
2042 return getUMaxExpr(Ops);
2045 // Okay, check to see if the same value occurs in the operand list twice. If
2046 // so, delete one. Since we sorted the list, these values are required to
2048 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2049 if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
2050 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2054 if (Ops.size() == 1) return Ops[0];
2056 assert(!Ops.empty() && "Reduced umax down to nothing!");
2058 // Okay, it looks like we really DO need a umax expr. Check to see if we
2059 // already have one, otherwise create a new one.
2060 FoldingSetNodeID ID;
2061 ID.AddInteger(scUMaxExpr);
2062 ID.AddInteger(Ops.size());
2063 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2064 ID.AddPointer(Ops[i]);
2066 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2067 SCEV *S = SCEVAllocator.Allocate<SCEVUMaxExpr>();
2068 new (S) SCEVUMaxExpr(ID, Ops);
2069 UniqueSCEVs.InsertNode(S, IP);
2073 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2075 // ~smax(~x, ~y) == smin(x, y).
2076 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2079 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2081 // ~umax(~x, ~y) == umin(x, y)
2082 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2085 const SCEV *ScalarEvolution::getFieldOffsetExpr(const StructType *STy,
2087 // If we have TargetData we can determine the constant offset.
2089 const Type *IntPtrTy = TD->getIntPtrType(getContext());
2090 const StructLayout &SL = *TD->getStructLayout(STy);
2091 uint64_t Offset = SL.getElementOffset(FieldNo);
2092 return getIntegerSCEV(Offset, IntPtrTy);
2095 // Field 0 is always at offset 0.
2097 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2098 return getIntegerSCEV(0, Ty);
2101 // Okay, it looks like we really DO need an offsetof expr. Check to see if we
2102 // already have one, otherwise create a new one.
2103 FoldingSetNodeID ID;
2104 ID.AddInteger(scFieldOffset);
2106 ID.AddInteger(FieldNo);
2108 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2109 SCEV *S = SCEVAllocator.Allocate<SCEVFieldOffsetExpr>();
2110 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2111 new (S) SCEVFieldOffsetExpr(ID, Ty, STy, FieldNo);
2112 UniqueSCEVs.InsertNode(S, IP);
2116 const SCEV *ScalarEvolution::getAllocSizeExpr(const Type *AllocTy) {
2117 // If we have TargetData we can determine the constant size.
2118 if (TD && AllocTy->isSized()) {
2119 const Type *IntPtrTy = TD->getIntPtrType(getContext());
2120 return getIntegerSCEV(TD->getTypeAllocSize(AllocTy), IntPtrTy);
2123 // Expand an array size into the element size times the number
2125 if (const ArrayType *ATy = dyn_cast<ArrayType>(AllocTy)) {
2126 const SCEV *E = getAllocSizeExpr(ATy->getElementType());
2128 E, getConstant(ConstantInt::get(cast<IntegerType>(E->getType()),
2129 ATy->getNumElements())));
2132 // Expand a vector size into the element size times the number
2134 if (const VectorType *VTy = dyn_cast<VectorType>(AllocTy)) {
2135 const SCEV *E = getAllocSizeExpr(VTy->getElementType());
2137 E, getConstant(ConstantInt::get(cast<IntegerType>(E->getType()),
2138 VTy->getNumElements())));
2141 // Okay, it looks like we really DO need a sizeof expr. Check to see if we
2142 // already have one, otherwise create a new one.
2143 FoldingSetNodeID ID;
2144 ID.AddInteger(scAllocSize);
2145 ID.AddPointer(AllocTy);
2147 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2148 SCEV *S = SCEVAllocator.Allocate<SCEVAllocSizeExpr>();
2149 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2150 new (S) SCEVAllocSizeExpr(ID, Ty, AllocTy);
2151 UniqueSCEVs.InsertNode(S, IP);
2155 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2156 // Don't attempt to do anything other than create a SCEVUnknown object
2157 // here. createSCEV only calls getUnknown after checking for all other
2158 // interesting possibilities, and any other code that calls getUnknown
2159 // is doing so in order to hide a value from SCEV canonicalization.
2161 FoldingSetNodeID ID;
2162 ID.AddInteger(scUnknown);
2165 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2166 SCEV *S = SCEVAllocator.Allocate<SCEVUnknown>();
2167 new (S) SCEVUnknown(ID, V);
2168 UniqueSCEVs.InsertNode(S, IP);
2172 //===----------------------------------------------------------------------===//
2173 // Basic SCEV Analysis and PHI Idiom Recognition Code
2176 /// isSCEVable - Test if values of the given type are analyzable within
2177 /// the SCEV framework. This primarily includes integer types, and it
2178 /// can optionally include pointer types if the ScalarEvolution class
2179 /// has access to target-specific information.
2180 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2181 // Integers and pointers are always SCEVable.
2182 return Ty->isInteger() || isa<PointerType>(Ty);
2185 /// getTypeSizeInBits - Return the size in bits of the specified type,
2186 /// for which isSCEVable must return true.
2187 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2188 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2190 // If we have a TargetData, use it!
2192 return TD->getTypeSizeInBits(Ty);
2194 // Integer types have fixed sizes.
2195 if (Ty->isInteger())
2196 return Ty->getPrimitiveSizeInBits();
2198 // The only other support type is pointer. Without TargetData, conservatively
2199 // assume pointers are 64-bit.
2200 assert(isa<PointerType>(Ty) && "isSCEVable permitted a non-SCEVable type!");
2204 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2205 /// the given type and which represents how SCEV will treat the given
2206 /// type, for which isSCEVable must return true. For pointer types,
2207 /// this is the pointer-sized integer type.
2208 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2209 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2211 if (Ty->isInteger())
2214 // The only other support type is pointer.
2215 assert(isa<PointerType>(Ty) && "Unexpected non-pointer non-integer type!");
2216 if (TD) return TD->getIntPtrType(getContext());
2218 // Without TargetData, conservatively assume pointers are 64-bit.
2219 return Type::getInt64Ty(getContext());
2222 const SCEV *ScalarEvolution::getCouldNotCompute() {
2223 return &CouldNotCompute;
2226 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2227 /// expression and create a new one.
2228 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2229 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2231 std::map<SCEVCallbackVH, const SCEV *>::iterator I = Scalars.find(V);
2232 if (I != Scalars.end()) return I->second;
2233 const SCEV *S = createSCEV(V);
2234 Scalars.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2238 /// getIntegerSCEV - Given a SCEVable type, create a constant for the
2239 /// specified signed integer value and return a SCEV for the constant.
2240 const SCEV *ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
2241 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
2242 return getConstant(ConstantInt::get(ITy, Val));
2245 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2247 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2248 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2250 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2252 const Type *Ty = V->getType();
2253 Ty = getEffectiveSCEVType(Ty);
2254 return getMulExpr(V,
2255 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2258 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2259 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2260 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2262 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2264 const Type *Ty = V->getType();
2265 Ty = getEffectiveSCEVType(Ty);
2266 const SCEV *AllOnes =
2267 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2268 return getMinusSCEV(AllOnes, V);
2271 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
2273 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS,
2276 return getAddExpr(LHS, getNegativeSCEV(RHS));
2279 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2280 /// input value to the specified type. If the type must be extended, it is zero
2283 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V,
2285 const Type *SrcTy = V->getType();
2286 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2287 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2288 "Cannot truncate or zero extend with non-integer arguments!");
2289 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2290 return V; // No conversion
2291 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2292 return getTruncateExpr(V, Ty);
2293 return getZeroExtendExpr(V, Ty);
2296 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2297 /// input value to the specified type. If the type must be extended, it is sign
2300 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2302 const Type *SrcTy = V->getType();
2303 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2304 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2305 "Cannot truncate or zero extend with non-integer arguments!");
2306 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2307 return V; // No conversion
2308 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2309 return getTruncateExpr(V, Ty);
2310 return getSignExtendExpr(V, Ty);
2313 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2314 /// input value to the specified type. If the type must be extended, it is zero
2315 /// extended. The conversion must not be narrowing.
2317 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2318 const Type *SrcTy = V->getType();
2319 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2320 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2321 "Cannot noop or zero extend with non-integer arguments!");
2322 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2323 "getNoopOrZeroExtend cannot truncate!");
2324 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2325 return V; // No conversion
2326 return getZeroExtendExpr(V, Ty);
2329 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2330 /// input value to the specified type. If the type must be extended, it is sign
2331 /// extended. The conversion must not be narrowing.
2333 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
2334 const Type *SrcTy = V->getType();
2335 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2336 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2337 "Cannot noop or sign extend with non-integer arguments!");
2338 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2339 "getNoopOrSignExtend cannot truncate!");
2340 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2341 return V; // No conversion
2342 return getSignExtendExpr(V, Ty);
2345 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2346 /// the input value to the specified type. If the type must be extended,
2347 /// it is extended with unspecified bits. The conversion must not be
2350 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2351 const Type *SrcTy = V->getType();
2352 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2353 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2354 "Cannot noop or any extend with non-integer arguments!");
2355 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2356 "getNoopOrAnyExtend cannot truncate!");
2357 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2358 return V; // No conversion
2359 return getAnyExtendExpr(V, Ty);
2362 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2363 /// input value to the specified type. The conversion must not be widening.
2365 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2366 const Type *SrcTy = V->getType();
2367 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2368 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2369 "Cannot truncate or noop with non-integer arguments!");
2370 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2371 "getTruncateOrNoop cannot extend!");
2372 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2373 return V; // No conversion
2374 return getTruncateExpr(V, Ty);
2377 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2378 /// the types using zero-extension, and then perform a umax operation
2380 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2382 const SCEV *PromotedLHS = LHS;
2383 const SCEV *PromotedRHS = RHS;
2385 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2386 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2388 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2390 return getUMaxExpr(PromotedLHS, PromotedRHS);
2393 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2394 /// the types using zero-extension, and then perform a umin operation
2396 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2398 const SCEV *PromotedLHS = LHS;
2399 const SCEV *PromotedRHS = RHS;
2401 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2402 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2404 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2406 return getUMinExpr(PromotedLHS, PromotedRHS);
2409 /// PushDefUseChildren - Push users of the given Instruction
2410 /// onto the given Worklist.
2412 PushDefUseChildren(Instruction *I,
2413 SmallVectorImpl<Instruction *> &Worklist) {
2414 // Push the def-use children onto the Worklist stack.
2415 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2417 Worklist.push_back(cast<Instruction>(UI));
2420 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2421 /// instructions that depend on the given instruction and removes them from
2422 /// the Scalars map if they reference SymName. This is used during PHI
2425 ScalarEvolution::ForgetSymbolicName(Instruction *I, const SCEV *SymName) {
2426 SmallVector<Instruction *, 16> Worklist;
2427 PushDefUseChildren(I, Worklist);
2429 SmallPtrSet<Instruction *, 8> Visited;
2431 while (!Worklist.empty()) {
2432 Instruction *I = Worklist.pop_back_val();
2433 if (!Visited.insert(I)) continue;
2435 std::map<SCEVCallbackVH, const SCEV*>::iterator It =
2436 Scalars.find(static_cast<Value *>(I));
2437 if (It != Scalars.end()) {
2438 // Short-circuit the def-use traversal if the symbolic name
2439 // ceases to appear in expressions.
2440 if (!It->second->hasOperand(SymName))
2443 // SCEVUnknown for a PHI either means that it has an unrecognized
2444 // structure, or it's a PHI that's in the progress of being computed
2445 // by createNodeForPHI. In the former case, additional loop trip
2446 // count information isn't going to change anything. In the later
2447 // case, createNodeForPHI will perform the necessary updates on its
2448 // own when it gets to that point.
2449 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
2450 ValuesAtScopes.erase(It->second);
2455 PushDefUseChildren(I, Worklist);
2459 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2460 /// a loop header, making it a potential recurrence, or it doesn't.
2462 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2463 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
2464 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2465 if (L->getHeader() == PN->getParent()) {
2466 // If it lives in the loop header, it has two incoming values, one
2467 // from outside the loop, and one from inside.
2468 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
2469 unsigned BackEdge = IncomingEdge^1;
2471 // While we are analyzing this PHI node, handle its value symbolically.
2472 const SCEV *SymbolicName = getUnknown(PN);
2473 assert(Scalars.find(PN) == Scalars.end() &&
2474 "PHI node already processed?");
2475 Scalars.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2477 // Using this symbolic name for the PHI, analyze the value coming around
2479 Value *BEValueV = PN->getIncomingValue(BackEdge);
2480 const SCEV *BEValue = getSCEV(BEValueV);
2482 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2483 // has a special value for the first iteration of the loop.
2485 // If the value coming around the backedge is an add with the symbolic
2486 // value we just inserted, then we found a simple induction variable!
2487 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2488 // If there is a single occurrence of the symbolic value, replace it
2489 // with a recurrence.
2490 unsigned FoundIndex = Add->getNumOperands();
2491 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2492 if (Add->getOperand(i) == SymbolicName)
2493 if (FoundIndex == e) {
2498 if (FoundIndex != Add->getNumOperands()) {
2499 // Create an add with everything but the specified operand.
2500 SmallVector<const SCEV *, 8> Ops;
2501 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2502 if (i != FoundIndex)
2503 Ops.push_back(Add->getOperand(i));
2504 const SCEV *Accum = getAddExpr(Ops);
2506 // This is not a valid addrec if the step amount is varying each
2507 // loop iteration, but is not itself an addrec in this loop.
2508 if (Accum->isLoopInvariant(L) ||
2509 (isa<SCEVAddRecExpr>(Accum) &&
2510 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2511 const SCEV *StartVal =
2512 getSCEV(PN->getIncomingValue(IncomingEdge));
2513 const SCEVAddRecExpr *PHISCEV =
2514 cast<SCEVAddRecExpr>(getAddRecExpr(StartVal, Accum, L));
2516 // If the increment doesn't overflow, then neither the addrec nor the
2517 // post-increment will overflow.
2518 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV))
2519 if (OBO->getOperand(0) == PN &&
2520 getSCEV(OBO->getOperand(1)) ==
2521 PHISCEV->getStepRecurrence(*this)) {
2522 const SCEVAddRecExpr *PostInc = PHISCEV->getPostIncExpr(*this);
2523 if (OBO->hasNoUnsignedWrap()) {
2524 const_cast<SCEVAddRecExpr *>(PHISCEV)
2525 ->setHasNoUnsignedWrap(true);
2526 const_cast<SCEVAddRecExpr *>(PostInc)
2527 ->setHasNoUnsignedWrap(true);
2529 if (OBO->hasNoSignedWrap()) {
2530 const_cast<SCEVAddRecExpr *>(PHISCEV)
2531 ->setHasNoSignedWrap(true);
2532 const_cast<SCEVAddRecExpr *>(PostInc)
2533 ->setHasNoSignedWrap(true);
2537 // Okay, for the entire analysis of this edge we assumed the PHI
2538 // to be symbolic. We now need to go back and purge all of the
2539 // entries for the scalars that use the symbolic expression.
2540 ForgetSymbolicName(PN, SymbolicName);
2541 Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
2545 } else if (const SCEVAddRecExpr *AddRec =
2546 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2547 // Otherwise, this could be a loop like this:
2548 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2549 // In this case, j = {1,+,1} and BEValue is j.
2550 // Because the other in-value of i (0) fits the evolution of BEValue
2551 // i really is an addrec evolution.
2552 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2553 const SCEV *StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
2555 // If StartVal = j.start - j.stride, we can use StartVal as the
2556 // initial step of the addrec evolution.
2557 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2558 AddRec->getOperand(1))) {
2559 const SCEV *PHISCEV =
2560 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2562 // Okay, for the entire analysis of this edge we assumed the PHI
2563 // to be symbolic. We now need to go back and purge all of the
2564 // entries for the scalars that use the symbolic expression.
2565 ForgetSymbolicName(PN, SymbolicName);
2566 Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
2572 return SymbolicName;
2575 // It's tempting to recognize PHIs with a unique incoming value, however
2576 // this leads passes like indvars to break LCSSA form. Fortunately, such
2577 // PHIs are rare, as instcombine zaps them.
2579 // If it's not a loop phi, we can't handle it yet.
2580 return getUnknown(PN);
2583 /// createNodeForGEP - Expand GEP instructions into add and multiply
2584 /// operations. This allows them to be analyzed by regular SCEV code.
2586 const SCEV *ScalarEvolution::createNodeForGEP(Operator *GEP) {
2588 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2589 Value *Base = GEP->getOperand(0);
2590 // Don't attempt to analyze GEPs over unsized objects.
2591 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2592 return getUnknown(GEP);
2593 const SCEV *TotalOffset = getIntegerSCEV(0, IntPtrTy);
2594 gep_type_iterator GTI = gep_type_begin(GEP);
2595 for (GetElementPtrInst::op_iterator I = next(GEP->op_begin()),
2599 // Compute the (potentially symbolic) offset in bytes for this index.
2600 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2601 // For a struct, add the member offset.
2602 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2603 TotalOffset = getAddExpr(TotalOffset,
2604 getFieldOffsetExpr(STy, FieldNo));
2606 // For an array, add the element offset, explicitly scaled.
2607 const SCEV *LocalOffset = getSCEV(Index);
2608 if (!isa<PointerType>(LocalOffset->getType()))
2609 // Getelementptr indicies are signed.
2610 LocalOffset = getTruncateOrSignExtend(LocalOffset, IntPtrTy);
2611 LocalOffset = getMulExpr(LocalOffset, getAllocSizeExpr(*GTI));
2612 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2615 return getAddExpr(getSCEV(Base), TotalOffset);
2618 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2619 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2620 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2621 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2623 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2624 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2625 return C->getValue()->getValue().countTrailingZeros();
2627 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2628 return std::min(GetMinTrailingZeros(T->getOperand()),
2629 (uint32_t)getTypeSizeInBits(T->getType()));
2631 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2632 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2633 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2634 getTypeSizeInBits(E->getType()) : OpRes;
2637 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2638 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2639 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2640 getTypeSizeInBits(E->getType()) : OpRes;
2643 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2644 // The result is the min of all operands results.
2645 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2646 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2647 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2651 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2652 // The result is the sum of all operands results.
2653 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2654 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2655 for (unsigned i = 1, e = M->getNumOperands();
2656 SumOpRes != BitWidth && i != e; ++i)
2657 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2662 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2663 // The result is the min of all operands results.
2664 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2665 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2666 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2670 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2671 // The result is the min of all operands results.
2672 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2673 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2674 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2678 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2679 // The result is the min of all operands results.
2680 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2681 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2682 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2686 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2687 // For a SCEVUnknown, ask ValueTracking.
2688 unsigned BitWidth = getTypeSizeInBits(U->getType());
2689 APInt Mask = APInt::getAllOnesValue(BitWidth);
2690 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2691 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
2692 return Zeros.countTrailingOnes();
2699 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
2702 ScalarEvolution::getUnsignedRange(const SCEV *S) {
2704 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2705 return ConstantRange(C->getValue()->getValue());
2707 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2708 ConstantRange X = getUnsignedRange(Add->getOperand(0));
2709 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2710 X = X.add(getUnsignedRange(Add->getOperand(i)));
2714 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2715 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
2716 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2717 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
2721 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2722 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
2723 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2724 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
2728 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2729 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
2730 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2731 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
2735 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2736 ConstantRange X = getUnsignedRange(UDiv->getLHS());
2737 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
2741 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2742 ConstantRange X = getUnsignedRange(ZExt->getOperand());
2743 return X.zeroExtend(cast<IntegerType>(ZExt->getType())->getBitWidth());
2746 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2747 ConstantRange X = getUnsignedRange(SExt->getOperand());
2748 return X.signExtend(cast<IntegerType>(SExt->getType())->getBitWidth());
2751 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
2752 ConstantRange X = getUnsignedRange(Trunc->getOperand());
2753 return X.truncate(cast<IntegerType>(Trunc->getType())->getBitWidth());
2756 ConstantRange FullSet(getTypeSizeInBits(S->getType()), true);
2758 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
2759 const SCEV *T = getBackedgeTakenCount(AddRec->getLoop());
2760 const SCEVConstant *Trip = dyn_cast<SCEVConstant>(T);
2761 if (!Trip) return FullSet;
2763 // TODO: non-affine addrec
2764 if (AddRec->isAffine()) {
2765 const Type *Ty = AddRec->getType();
2766 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
2767 if (getTypeSizeInBits(MaxBECount->getType()) <= getTypeSizeInBits(Ty)) {
2768 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
2770 const SCEV *Start = AddRec->getStart();
2771 const SCEV *Step = AddRec->getStepRecurrence(*this);
2772 const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
2774 // Check for overflow.
2775 // TODO: This is very conservative.
2776 if (!(Step->isOne() &&
2777 isKnownPredicate(ICmpInst::ICMP_ULT, Start, End)) &&
2778 !(Step->isAllOnesValue() &&
2779 isKnownPredicate(ICmpInst::ICMP_UGT, Start, End)))
2782 ConstantRange StartRange = getUnsignedRange(Start);
2783 ConstantRange EndRange = getUnsignedRange(End);
2784 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
2785 EndRange.getUnsignedMin());
2786 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
2787 EndRange.getUnsignedMax());
2788 if (Min.isMinValue() && Max.isMaxValue())
2790 return ConstantRange(Min, Max+1);
2795 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2796 // For a SCEVUnknown, ask ValueTracking.
2797 unsigned BitWidth = getTypeSizeInBits(U->getType());
2798 APInt Mask = APInt::getAllOnesValue(BitWidth);
2799 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2800 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
2801 if (Ones == ~Zeros + 1)
2803 return ConstantRange(Ones, ~Zeros + 1);
2809 /// getSignedRange - Determine the signed range for a particular SCEV.
2812 ScalarEvolution::getSignedRange(const SCEV *S) {
2814 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2815 return ConstantRange(C->getValue()->getValue());
2817 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2818 ConstantRange X = getSignedRange(Add->getOperand(0));
2819 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2820 X = X.add(getSignedRange(Add->getOperand(i)));
2824 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2825 ConstantRange X = getSignedRange(Mul->getOperand(0));
2826 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2827 X = X.multiply(getSignedRange(Mul->getOperand(i)));
2831 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2832 ConstantRange X = getSignedRange(SMax->getOperand(0));
2833 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2834 X = X.smax(getSignedRange(SMax->getOperand(i)));
2838 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2839 ConstantRange X = getSignedRange(UMax->getOperand(0));
2840 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2841 X = X.umax(getSignedRange(UMax->getOperand(i)));
2845 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2846 ConstantRange X = getSignedRange(UDiv->getLHS());
2847 ConstantRange Y = getSignedRange(UDiv->getRHS());
2851 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2852 ConstantRange X = getSignedRange(ZExt->getOperand());
2853 return X.zeroExtend(cast<IntegerType>(ZExt->getType())->getBitWidth());
2856 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2857 ConstantRange X = getSignedRange(SExt->getOperand());
2858 return X.signExtend(cast<IntegerType>(SExt->getType())->getBitWidth());
2861 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
2862 ConstantRange X = getSignedRange(Trunc->getOperand());
2863 return X.truncate(cast<IntegerType>(Trunc->getType())->getBitWidth());
2866 ConstantRange FullSet(getTypeSizeInBits(S->getType()), true);
2868 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
2869 const SCEV *T = getBackedgeTakenCount(AddRec->getLoop());
2870 const SCEVConstant *Trip = dyn_cast<SCEVConstant>(T);
2871 if (!Trip) return FullSet;
2873 // TODO: non-affine addrec
2874 if (AddRec->isAffine()) {
2875 const Type *Ty = AddRec->getType();
2876 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
2877 if (getTypeSizeInBits(MaxBECount->getType()) <= getTypeSizeInBits(Ty)) {
2878 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
2880 const SCEV *Start = AddRec->getStart();
2881 const SCEV *Step = AddRec->getStepRecurrence(*this);
2882 const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
2884 // Check for overflow.
2885 // TODO: This is very conservative.
2886 if (!(Step->isOne() &&
2887 isKnownPredicate(ICmpInst::ICMP_SLT, Start, End)) &&
2888 !(Step->isAllOnesValue() &&
2889 isKnownPredicate(ICmpInst::ICMP_SGT, Start, End)))
2892 ConstantRange StartRange = getSignedRange(Start);
2893 ConstantRange EndRange = getSignedRange(End);
2894 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
2895 EndRange.getSignedMin());
2896 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
2897 EndRange.getSignedMax());
2898 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
2900 return ConstantRange(Min, Max+1);
2905 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2906 // For a SCEVUnknown, ask ValueTracking.
2907 unsigned BitWidth = getTypeSizeInBits(U->getType());
2908 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
2912 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
2913 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1);
2919 /// createSCEV - We know that there is no SCEV for the specified value.
2920 /// Analyze the expression.
2922 const SCEV *ScalarEvolution::createSCEV(Value *V) {
2923 if (!isSCEVable(V->getType()))
2924 return getUnknown(V);
2926 unsigned Opcode = Instruction::UserOp1;
2927 if (Instruction *I = dyn_cast<Instruction>(V))
2928 Opcode = I->getOpcode();
2929 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2930 Opcode = CE->getOpcode();
2931 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2932 return getConstant(CI);
2933 else if (isa<ConstantPointerNull>(V))
2934 return getIntegerSCEV(0, V->getType());
2935 else if (isa<UndefValue>(V))
2936 return getIntegerSCEV(0, V->getType());
2937 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
2938 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
2940 return getUnknown(V);
2942 Operator *U = cast<Operator>(V);
2944 case Instruction::Add:
2945 return getAddExpr(getSCEV(U->getOperand(0)),
2946 getSCEV(U->getOperand(1)));
2947 case Instruction::Mul:
2948 return getMulExpr(getSCEV(U->getOperand(0)),
2949 getSCEV(U->getOperand(1)));
2950 case Instruction::UDiv:
2951 return getUDivExpr(getSCEV(U->getOperand(0)),
2952 getSCEV(U->getOperand(1)));
2953 case Instruction::Sub:
2954 return getMinusSCEV(getSCEV(U->getOperand(0)),
2955 getSCEV(U->getOperand(1)));
2956 case Instruction::And:
2957 // For an expression like x&255 that merely masks off the high bits,
2958 // use zext(trunc(x)) as the SCEV expression.
2959 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2960 if (CI->isNullValue())
2961 return getSCEV(U->getOperand(1));
2962 if (CI->isAllOnesValue())
2963 return getSCEV(U->getOperand(0));
2964 const APInt &A = CI->getValue();
2966 // Instcombine's ShrinkDemandedConstant may strip bits out of
2967 // constants, obscuring what would otherwise be a low-bits mask.
2968 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
2969 // knew about to reconstruct a low-bits mask value.
2970 unsigned LZ = A.countLeadingZeros();
2971 unsigned BitWidth = A.getBitWidth();
2972 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
2973 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2974 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
2976 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
2978 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
2980 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
2981 IntegerType::get(getContext(), BitWidth - LZ)),
2986 case Instruction::Or:
2987 // If the RHS of the Or is a constant, we may have something like:
2988 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
2989 // optimizations will transparently handle this case.
2991 // In order for this transformation to be safe, the LHS must be of the
2992 // form X*(2^n) and the Or constant must be less than 2^n.
2993 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2994 const SCEV *LHS = getSCEV(U->getOperand(0));
2995 const APInt &CIVal = CI->getValue();
2996 if (GetMinTrailingZeros(LHS) >=
2997 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
2998 // Build a plain add SCEV.
2999 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3000 // If the LHS of the add was an addrec and it has no-wrap flags,
3001 // transfer the no-wrap flags, since an or won't introduce a wrap.
3002 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3003 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3004 if (OldAR->hasNoUnsignedWrap())
3005 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3006 if (OldAR->hasNoSignedWrap())
3007 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3013 case Instruction::Xor:
3014 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3015 // If the RHS of the xor is a signbit, then this is just an add.
3016 // Instcombine turns add of signbit into xor as a strength reduction step.
3017 if (CI->getValue().isSignBit())
3018 return getAddExpr(getSCEV(U->getOperand(0)),
3019 getSCEV(U->getOperand(1)));
3021 // If the RHS of xor is -1, then this is a not operation.
3022 if (CI->isAllOnesValue())
3023 return getNotSCEV(getSCEV(U->getOperand(0)));
3025 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3026 // This is a variant of the check for xor with -1, and it handles
3027 // the case where instcombine has trimmed non-demanded bits out
3028 // of an xor with -1.
3029 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3030 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3031 if (BO->getOpcode() == Instruction::And &&
3032 LCI->getValue() == CI->getValue())
3033 if (const SCEVZeroExtendExpr *Z =
3034 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3035 const Type *UTy = U->getType();
3036 const SCEV *Z0 = Z->getOperand();
3037 const Type *Z0Ty = Z0->getType();
3038 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3040 // If C is a low-bits mask, the zero extend is zerving to
3041 // mask off the high bits. Complement the operand and
3042 // re-apply the zext.
3043 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3044 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3046 // If C is a single bit, it may be in the sign-bit position
3047 // before the zero-extend. In this case, represent the xor
3048 // using an add, which is equivalent, and re-apply the zext.
3049 APInt Trunc = APInt(CI->getValue()).trunc(Z0TySize);
3050 if (APInt(Trunc).zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3052 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3058 case Instruction::Shl:
3059 // Turn shift left of a constant amount into a multiply.
3060 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3061 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
3062 Constant *X = ConstantInt::get(getContext(),
3063 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
3064 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3068 case Instruction::LShr:
3069 // Turn logical shift right of a constant into a unsigned divide.
3070 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3071 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
3072 Constant *X = ConstantInt::get(getContext(),
3073 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
3074 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3078 case Instruction::AShr:
3079 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3080 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3081 if (Instruction *L = dyn_cast<Instruction>(U->getOperand(0)))
3082 if (L->getOpcode() == Instruction::Shl &&
3083 L->getOperand(1) == U->getOperand(1)) {
3084 unsigned BitWidth = getTypeSizeInBits(U->getType());
3085 uint64_t Amt = BitWidth - CI->getZExtValue();
3086 if (Amt == BitWidth)
3087 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3089 return getIntegerSCEV(0, U->getType()); // value is undefined
3091 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3092 IntegerType::get(getContext(), Amt)),
3097 case Instruction::Trunc:
3098 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3100 case Instruction::ZExt:
3101 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3103 case Instruction::SExt:
3104 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3106 case Instruction::BitCast:
3107 // BitCasts are no-op casts so we just eliminate the cast.
3108 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3109 return getSCEV(U->getOperand(0));
3112 // It's tempting to handle inttoptr and ptrtoint, however this can
3113 // lead to pointer expressions which cannot be expanded to GEPs
3114 // (because they may overflow). For now, the only pointer-typed
3115 // expressions we handle are GEPs and address literals.
3117 case Instruction::GetElementPtr:
3118 return createNodeForGEP(U);
3120 case Instruction::PHI:
3121 return createNodeForPHI(cast<PHINode>(U));
3123 case Instruction::Select:
3124 // This could be a smax or umax that was lowered earlier.
3125 // Try to recover it.
3126 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3127 Value *LHS = ICI->getOperand(0);
3128 Value *RHS = ICI->getOperand(1);
3129 switch (ICI->getPredicate()) {
3130 case ICmpInst::ICMP_SLT:
3131 case ICmpInst::ICMP_SLE:
3132 std::swap(LHS, RHS);
3134 case ICmpInst::ICMP_SGT:
3135 case ICmpInst::ICMP_SGE:
3136 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
3137 return getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
3138 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
3139 return getSMinExpr(getSCEV(LHS), getSCEV(RHS));
3141 case ICmpInst::ICMP_ULT:
3142 case ICmpInst::ICMP_ULE:
3143 std::swap(LHS, RHS);
3145 case ICmpInst::ICMP_UGT:
3146 case ICmpInst::ICMP_UGE:
3147 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
3148 return getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
3149 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
3150 return getUMinExpr(getSCEV(LHS), getSCEV(RHS));
3152 case ICmpInst::ICMP_NE:
3153 // n != 0 ? n : 1 -> umax(n, 1)
3154 if (LHS == U->getOperand(1) &&
3155 isa<ConstantInt>(U->getOperand(2)) &&
3156 cast<ConstantInt>(U->getOperand(2))->isOne() &&
3157 isa<ConstantInt>(RHS) &&
3158 cast<ConstantInt>(RHS)->isZero())
3159 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(2)));
3161 case ICmpInst::ICMP_EQ:
3162 // n == 0 ? 1 : n -> umax(n, 1)
3163 if (LHS == U->getOperand(2) &&
3164 isa<ConstantInt>(U->getOperand(1)) &&
3165 cast<ConstantInt>(U->getOperand(1))->isOne() &&
3166 isa<ConstantInt>(RHS) &&
3167 cast<ConstantInt>(RHS)->isZero())
3168 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(1)));
3175 default: // We cannot analyze this expression.
3179 return getUnknown(V);
3184 //===----------------------------------------------------------------------===//
3185 // Iteration Count Computation Code
3188 /// getBackedgeTakenCount - If the specified loop has a predictable
3189 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3190 /// object. The backedge-taken count is the number of times the loop header
3191 /// will be branched to from within the loop. This is one less than the
3192 /// trip count of the loop, since it doesn't count the first iteration,
3193 /// when the header is branched to from outside the loop.
3195 /// Note that it is not valid to call this method on a loop without a
3196 /// loop-invariant backedge-taken count (see
3197 /// hasLoopInvariantBackedgeTakenCount).
3199 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3200 return getBackedgeTakenInfo(L).Exact;
3203 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3204 /// return the least SCEV value that is known never to be less than the
3205 /// actual backedge taken count.
3206 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3207 return getBackedgeTakenInfo(L).Max;
3210 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3211 /// onto the given Worklist.
3213 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3214 BasicBlock *Header = L->getHeader();
3216 // Push all Loop-header PHIs onto the Worklist stack.
3217 for (BasicBlock::iterator I = Header->begin();
3218 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3219 Worklist.push_back(PN);
3222 const ScalarEvolution::BackedgeTakenInfo &
3223 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3224 // Initially insert a CouldNotCompute for this loop. If the insertion
3225 // succeeds, procede to actually compute a backedge-taken count and
3226 // update the value. The temporary CouldNotCompute value tells SCEV
3227 // code elsewhere that it shouldn't attempt to request a new
3228 // backedge-taken count, which could result in infinite recursion.
3229 std::pair<std::map<const Loop*, BackedgeTakenInfo>::iterator, bool> Pair =
3230 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3232 BackedgeTakenInfo ItCount = ComputeBackedgeTakenCount(L);
3233 if (ItCount.Exact != getCouldNotCompute()) {
3234 assert(ItCount.Exact->isLoopInvariant(L) &&
3235 ItCount.Max->isLoopInvariant(L) &&
3236 "Computed trip count isn't loop invariant for loop!");
3237 ++NumTripCountsComputed;
3239 // Update the value in the map.
3240 Pair.first->second = ItCount;
3242 if (ItCount.Max != getCouldNotCompute())
3243 // Update the value in the map.
3244 Pair.first->second = ItCount;
3245 if (isa<PHINode>(L->getHeader()->begin()))
3246 // Only count loops that have phi nodes as not being computable.
3247 ++NumTripCountsNotComputed;
3250 // Now that we know more about the trip count for this loop, forget any
3251 // existing SCEV values for PHI nodes in this loop since they are only
3252 // conservative estimates made without the benefit of trip count
3253 // information. This is similar to the code in
3254 // forgetLoopBackedgeTakenCount, except that it handles SCEVUnknown PHI
3256 if (ItCount.hasAnyInfo()) {
3257 SmallVector<Instruction *, 16> Worklist;
3258 PushLoopPHIs(L, Worklist);
3260 SmallPtrSet<Instruction *, 8> Visited;
3261 while (!Worklist.empty()) {
3262 Instruction *I = Worklist.pop_back_val();
3263 if (!Visited.insert(I)) continue;
3265 std::map<SCEVCallbackVH, const SCEV*>::iterator It =
3266 Scalars.find(static_cast<Value *>(I));
3267 if (It != Scalars.end()) {
3268 // SCEVUnknown for a PHI either means that it has an unrecognized
3269 // structure, or it's a PHI that's in the progress of being computed
3270 // by createNodeForPHI. In the former case, additional loop trip
3271 // count information isn't going to change anything. In the later
3272 // case, createNodeForPHI will perform the necessary updates on its
3273 // own when it gets to that point.
3274 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
3275 ValuesAtScopes.erase(It->second);
3278 if (PHINode *PN = dyn_cast<PHINode>(I))
3279 ConstantEvolutionLoopExitValue.erase(PN);
3282 PushDefUseChildren(I, Worklist);
3286 return Pair.first->second;
3289 /// forgetLoopBackedgeTakenCount - This method should be called by the
3290 /// client when it has changed a loop in a way that may effect
3291 /// ScalarEvolution's ability to compute a trip count, or if the loop
3293 void ScalarEvolution::forgetLoopBackedgeTakenCount(const Loop *L) {
3294 BackedgeTakenCounts.erase(L);
3296 SmallVector<Instruction *, 16> Worklist;
3297 PushLoopPHIs(L, Worklist);
3299 SmallPtrSet<Instruction *, 8> Visited;
3300 while (!Worklist.empty()) {
3301 Instruction *I = Worklist.pop_back_val();
3302 if (!Visited.insert(I)) continue;
3304 std::map<SCEVCallbackVH, const SCEV*>::iterator It =
3305 Scalars.find(static_cast<Value *>(I));
3306 if (It != Scalars.end()) {
3307 ValuesAtScopes.erase(It->second);
3309 if (PHINode *PN = dyn_cast<PHINode>(I))
3310 ConstantEvolutionLoopExitValue.erase(PN);
3313 PushDefUseChildren(I, Worklist);
3317 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3318 /// of the specified loop will execute.
3319 ScalarEvolution::BackedgeTakenInfo
3320 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3321 SmallVector<BasicBlock*, 8> ExitingBlocks;
3322 L->getExitingBlocks(ExitingBlocks);
3324 // Examine all exits and pick the most conservative values.
3325 const SCEV *BECount = getCouldNotCompute();
3326 const SCEV *MaxBECount = getCouldNotCompute();
3327 bool CouldNotComputeBECount = false;
3328 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3329 BackedgeTakenInfo NewBTI =
3330 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3332 if (NewBTI.Exact == getCouldNotCompute()) {
3333 // We couldn't compute an exact value for this exit, so
3334 // we won't be able to compute an exact value for the loop.
3335 CouldNotComputeBECount = true;
3336 BECount = getCouldNotCompute();
3337 } else if (!CouldNotComputeBECount) {
3338 if (BECount == getCouldNotCompute())
3339 BECount = NewBTI.Exact;
3341 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3343 if (MaxBECount == getCouldNotCompute())
3344 MaxBECount = NewBTI.Max;
3345 else if (NewBTI.Max != getCouldNotCompute())
3346 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3349 return BackedgeTakenInfo(BECount, MaxBECount);
3352 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3353 /// of the specified loop will execute if it exits via the specified block.
3354 ScalarEvolution::BackedgeTakenInfo
3355 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3356 BasicBlock *ExitingBlock) {
3358 // Okay, we've chosen an exiting block. See what condition causes us to
3359 // exit at this block.
3361 // FIXME: we should be able to handle switch instructions (with a single exit)
3362 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3363 if (ExitBr == 0) return getCouldNotCompute();
3364 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3366 // At this point, we know we have a conditional branch that determines whether
3367 // the loop is exited. However, we don't know if the branch is executed each
3368 // time through the loop. If not, then the execution count of the branch will
3369 // not be equal to the trip count of the loop.
3371 // Currently we check for this by checking to see if the Exit branch goes to
3372 // the loop header. If so, we know it will always execute the same number of
3373 // times as the loop. We also handle the case where the exit block *is* the
3374 // loop header. This is common for un-rotated loops.
3376 // If both of those tests fail, walk up the unique predecessor chain to the
3377 // header, stopping if there is an edge that doesn't exit the loop. If the
3378 // header is reached, the execution count of the branch will be equal to the
3379 // trip count of the loop.
3381 // More extensive analysis could be done to handle more cases here.
3383 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3384 ExitBr->getSuccessor(1) != L->getHeader() &&
3385 ExitBr->getParent() != L->getHeader()) {
3386 // The simple checks failed, try climbing the unique predecessor chain
3387 // up to the header.
3389 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3390 BasicBlock *Pred = BB->getUniquePredecessor();
3392 return getCouldNotCompute();
3393 TerminatorInst *PredTerm = Pred->getTerminator();
3394 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3395 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3398 // If the predecessor has a successor that isn't BB and isn't
3399 // outside the loop, assume the worst.
3400 if (L->contains(PredSucc))
3401 return getCouldNotCompute();
3403 if (Pred == L->getHeader()) {
3410 return getCouldNotCompute();
3413 // Procede to the next level to examine the exit condition expression.
3414 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3415 ExitBr->getSuccessor(0),
3416 ExitBr->getSuccessor(1));
3419 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3420 /// backedge of the specified loop will execute if its exit condition
3421 /// were a conditional branch of ExitCond, TBB, and FBB.
3422 ScalarEvolution::BackedgeTakenInfo
3423 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3427 // Check if the controlling expression for this loop is an And or Or.
3428 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3429 if (BO->getOpcode() == Instruction::And) {
3430 // Recurse on the operands of the and.
3431 BackedgeTakenInfo BTI0 =
3432 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3433 BackedgeTakenInfo BTI1 =
3434 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3435 const SCEV *BECount = getCouldNotCompute();
3436 const SCEV *MaxBECount = getCouldNotCompute();
3437 if (L->contains(TBB)) {
3438 // Both conditions must be true for the loop to continue executing.
3439 // Choose the less conservative count.
3440 if (BTI0.Exact == getCouldNotCompute() ||
3441 BTI1.Exact == getCouldNotCompute())
3442 BECount = getCouldNotCompute();
3444 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3445 if (BTI0.Max == getCouldNotCompute())
3446 MaxBECount = BTI1.Max;
3447 else if (BTI1.Max == getCouldNotCompute())
3448 MaxBECount = BTI0.Max;
3450 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3452 // Both conditions must be true for the loop to exit.
3453 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3454 if (BTI0.Exact != getCouldNotCompute() &&
3455 BTI1.Exact != getCouldNotCompute())
3456 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3457 if (BTI0.Max != getCouldNotCompute() &&
3458 BTI1.Max != getCouldNotCompute())
3459 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
3462 return BackedgeTakenInfo(BECount, MaxBECount);
3464 if (BO->getOpcode() == Instruction::Or) {
3465 // Recurse on the operands of the or.
3466 BackedgeTakenInfo BTI0 =
3467 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3468 BackedgeTakenInfo BTI1 =
3469 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3470 const SCEV *BECount = getCouldNotCompute();
3471 const SCEV *MaxBECount = getCouldNotCompute();
3472 if (L->contains(FBB)) {
3473 // Both conditions must be false for the loop to continue executing.
3474 // Choose the less conservative count.
3475 if (BTI0.Exact == getCouldNotCompute() ||
3476 BTI1.Exact == getCouldNotCompute())
3477 BECount = getCouldNotCompute();
3479 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3480 if (BTI0.Max == getCouldNotCompute())
3481 MaxBECount = BTI1.Max;
3482 else if (BTI1.Max == getCouldNotCompute())
3483 MaxBECount = BTI0.Max;
3485 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3487 // Both conditions must be false for the loop to exit.
3488 assert(L->contains(TBB) && "Loop block has no successor in loop!");
3489 if (BTI0.Exact != getCouldNotCompute() &&
3490 BTI1.Exact != getCouldNotCompute())
3491 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3492 if (BTI0.Max != getCouldNotCompute() &&
3493 BTI1.Max != getCouldNotCompute())
3494 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
3497 return BackedgeTakenInfo(BECount, MaxBECount);
3501 // With an icmp, it may be feasible to compute an exact backedge-taken count.
3502 // Procede to the next level to examine the icmp.
3503 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
3504 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
3506 // If it's not an integer or pointer comparison then compute it the hard way.
3507 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3510 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
3511 /// backedge of the specified loop will execute if its exit condition
3512 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
3513 ScalarEvolution::BackedgeTakenInfo
3514 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
3519 // If the condition was exit on true, convert the condition to exit on false
3520 ICmpInst::Predicate Cond;
3521 if (!L->contains(FBB))
3522 Cond = ExitCond->getPredicate();
3524 Cond = ExitCond->getInversePredicate();
3526 // Handle common loops like: for (X = "string"; *X; ++X)
3527 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
3528 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
3530 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
3531 if (!isa<SCEVCouldNotCompute>(ItCnt)) {
3532 unsigned BitWidth = getTypeSizeInBits(ItCnt->getType());
3533 return BackedgeTakenInfo(ItCnt,
3534 isa<SCEVConstant>(ItCnt) ? ItCnt :
3535 getConstant(APInt::getMaxValue(BitWidth)-1));
3539 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
3540 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
3542 // Try to evaluate any dependencies out of the loop.
3543 LHS = getSCEVAtScope(LHS, L);
3544 RHS = getSCEVAtScope(RHS, L);
3546 // At this point, we would like to compute how many iterations of the
3547 // loop the predicate will return true for these inputs.
3548 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
3549 // If there is a loop-invariant, force it into the RHS.
3550 std::swap(LHS, RHS);
3551 Cond = ICmpInst::getSwappedPredicate(Cond);
3554 // If we have a comparison of a chrec against a constant, try to use value
3555 // ranges to answer this query.
3556 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
3557 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
3558 if (AddRec->getLoop() == L) {
3559 // Form the constant range.
3560 ConstantRange CompRange(
3561 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
3563 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
3564 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
3568 case ICmpInst::ICMP_NE: { // while (X != Y)
3569 // Convert to: while (X-Y != 0)
3570 const SCEV *TC = HowFarToZero(getMinusSCEV(LHS, RHS), L);
3571 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3574 case ICmpInst::ICMP_EQ: { // while (X == Y)
3575 // Convert to: while (X-Y == 0)
3576 const SCEV *TC = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
3577 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3580 case ICmpInst::ICMP_SLT: {
3581 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
3582 if (BTI.hasAnyInfo()) return BTI;
3585 case ICmpInst::ICMP_SGT: {
3586 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3587 getNotSCEV(RHS), L, true);
3588 if (BTI.hasAnyInfo()) return BTI;
3591 case ICmpInst::ICMP_ULT: {
3592 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
3593 if (BTI.hasAnyInfo()) return BTI;
3596 case ICmpInst::ICMP_UGT: {
3597 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3598 getNotSCEV(RHS), L, false);
3599 if (BTI.hasAnyInfo()) return BTI;
3604 errs() << "ComputeBackedgeTakenCount ";
3605 if (ExitCond->getOperand(0)->getType()->isUnsigned())
3606 errs() << "[unsigned] ";
3607 errs() << *LHS << " "
3608 << Instruction::getOpcodeName(Instruction::ICmp)
3609 << " " << *RHS << "\n";
3614 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3617 static ConstantInt *
3618 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
3619 ScalarEvolution &SE) {
3620 const SCEV *InVal = SE.getConstant(C);
3621 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
3622 assert(isa<SCEVConstant>(Val) &&
3623 "Evaluation of SCEV at constant didn't fold correctly?");
3624 return cast<SCEVConstant>(Val)->getValue();
3627 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
3628 /// and a GEP expression (missing the pointer index) indexing into it, return
3629 /// the addressed element of the initializer or null if the index expression is
3632 GetAddressedElementFromGlobal(LLVMContext &Context, GlobalVariable *GV,
3633 const std::vector<ConstantInt*> &Indices) {
3634 Constant *Init = GV->getInitializer();
3635 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
3636 uint64_t Idx = Indices[i]->getZExtValue();
3637 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
3638 assert(Idx < CS->getNumOperands() && "Bad struct index!");
3639 Init = cast<Constant>(CS->getOperand(Idx));
3640 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
3641 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
3642 Init = cast<Constant>(CA->getOperand(Idx));
3643 } else if (isa<ConstantAggregateZero>(Init)) {
3644 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
3645 assert(Idx < STy->getNumElements() && "Bad struct index!");
3646 Init = Constant::getNullValue(STy->getElementType(Idx));
3647 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
3648 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
3649 Init = Constant::getNullValue(ATy->getElementType());
3651 llvm_unreachable("Unknown constant aggregate type!");
3655 return 0; // Unknown initializer type
3661 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
3662 /// 'icmp op load X, cst', try to see if we can compute the backedge
3663 /// execution count.
3665 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
3669 ICmpInst::Predicate predicate) {
3670 if (LI->isVolatile()) return getCouldNotCompute();
3672 // Check to see if the loaded pointer is a getelementptr of a global.
3673 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
3674 if (!GEP) return getCouldNotCompute();
3676 // Make sure that it is really a constant global we are gepping, with an
3677 // initializer, and make sure the first IDX is really 0.
3678 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
3679 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
3680 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
3681 !cast<Constant>(GEP->getOperand(1))->isNullValue())
3682 return getCouldNotCompute();
3684 // Okay, we allow one non-constant index into the GEP instruction.
3686 std::vector<ConstantInt*> Indexes;
3687 unsigned VarIdxNum = 0;
3688 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
3689 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
3690 Indexes.push_back(CI);
3691 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
3692 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
3693 VarIdx = GEP->getOperand(i);
3695 Indexes.push_back(0);
3698 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
3699 // Check to see if X is a loop variant variable value now.
3700 const SCEV *Idx = getSCEV(VarIdx);
3701 Idx = getSCEVAtScope(Idx, L);
3703 // We can only recognize very limited forms of loop index expressions, in
3704 // particular, only affine AddRec's like {C1,+,C2}.
3705 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
3706 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
3707 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
3708 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
3709 return getCouldNotCompute();
3711 unsigned MaxSteps = MaxBruteForceIterations;
3712 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
3713 ConstantInt *ItCst = ConstantInt::get(
3714 cast<IntegerType>(IdxExpr->getType()), IterationNum);
3715 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
3717 // Form the GEP offset.
3718 Indexes[VarIdxNum] = Val;
3720 Constant *Result = GetAddressedElementFromGlobal(getContext(), GV, Indexes);
3721 if (Result == 0) break; // Cannot compute!
3723 // Evaluate the condition for this iteration.
3724 Result = ConstantExpr::getICmp(predicate, Result, RHS);
3725 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
3726 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
3728 errs() << "\n***\n*** Computed loop count " << *ItCst
3729 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
3732 ++NumArrayLenItCounts;
3733 return getConstant(ItCst); // Found terminating iteration!
3736 return getCouldNotCompute();
3740 /// CanConstantFold - Return true if we can constant fold an instruction of the
3741 /// specified type, assuming that all operands were constants.
3742 static bool CanConstantFold(const Instruction *I) {
3743 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
3744 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
3747 if (const CallInst *CI = dyn_cast<CallInst>(I))
3748 if (const Function *F = CI->getCalledFunction())
3749 return canConstantFoldCallTo(F);
3753 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
3754 /// in the loop that V is derived from. We allow arbitrary operations along the
3755 /// way, but the operands of an operation must either be constants or a value
3756 /// derived from a constant PHI. If this expression does not fit with these
3757 /// constraints, return null.
3758 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
3759 // If this is not an instruction, or if this is an instruction outside of the
3760 // loop, it can't be derived from a loop PHI.
3761 Instruction *I = dyn_cast<Instruction>(V);
3762 if (I == 0 || !L->contains(I->getParent())) return 0;
3764 if (PHINode *PN = dyn_cast<PHINode>(I)) {
3765 if (L->getHeader() == I->getParent())
3768 // We don't currently keep track of the control flow needed to evaluate
3769 // PHIs, so we cannot handle PHIs inside of loops.
3773 // If we won't be able to constant fold this expression even if the operands
3774 // are constants, return early.
3775 if (!CanConstantFold(I)) return 0;
3777 // Otherwise, we can evaluate this instruction if all of its operands are
3778 // constant or derived from a PHI node themselves.
3780 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
3781 if (!(isa<Constant>(I->getOperand(Op)) ||
3782 isa<GlobalValue>(I->getOperand(Op)))) {
3783 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
3784 if (P == 0) return 0; // Not evolving from PHI
3788 return 0; // Evolving from multiple different PHIs.
3791 // This is a expression evolving from a constant PHI!
3795 /// EvaluateExpression - Given an expression that passes the
3796 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
3797 /// in the loop has the value PHIVal. If we can't fold this expression for some
3798 /// reason, return null.
3799 static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
3800 if (isa<PHINode>(V)) return PHIVal;
3801 if (Constant *C = dyn_cast<Constant>(V)) return C;
3802 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
3803 Instruction *I = cast<Instruction>(V);
3804 LLVMContext &Context = I->getParent()->getContext();
3806 std::vector<Constant*> Operands;
3807 Operands.resize(I->getNumOperands());
3809 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3810 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal);
3811 if (Operands[i] == 0) return 0;
3814 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
3815 return ConstantFoldCompareInstOperands(CI->getPredicate(),
3816 &Operands[0], Operands.size(),
3819 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
3820 &Operands[0], Operands.size(),
3824 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
3825 /// in the header of its containing loop, we know the loop executes a
3826 /// constant number of times, and the PHI node is just a recurrence
3827 /// involving constants, fold it.
3829 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
3832 std::map<PHINode*, Constant*>::iterator I =
3833 ConstantEvolutionLoopExitValue.find(PN);
3834 if (I != ConstantEvolutionLoopExitValue.end())
3837 if (BEs.ugt(APInt(BEs.getBitWidth(),MaxBruteForceIterations)))
3838 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
3840 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
3842 // Since the loop is canonicalized, the PHI node must have two entries. One
3843 // entry must be a constant (coming in from outside of the loop), and the
3844 // second must be derived from the same PHI.
3845 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
3846 Constant *StartCST =
3847 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
3849 return RetVal = 0; // Must be a constant.
3851 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
3852 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
3854 return RetVal = 0; // Not derived from same PHI.
3856 // Execute the loop symbolically to determine the exit value.
3857 if (BEs.getActiveBits() >= 32)
3858 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
3860 unsigned NumIterations = BEs.getZExtValue(); // must be in range
3861 unsigned IterationNum = 0;
3862 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
3863 if (IterationNum == NumIterations)
3864 return RetVal = PHIVal; // Got exit value!
3866 // Compute the value of the PHI node for the next iteration.
3867 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
3868 if (NextPHI == PHIVal)
3869 return RetVal = NextPHI; // Stopped evolving!
3871 return 0; // Couldn't evaluate!
3876 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
3877 /// constant number of times (the condition evolves only from constants),
3878 /// try to evaluate a few iterations of the loop until we get the exit
3879 /// condition gets a value of ExitWhen (true or false). If we cannot
3880 /// evaluate the trip count of the loop, return getCouldNotCompute().
3882 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
3885 PHINode *PN = getConstantEvolvingPHI(Cond, L);
3886 if (PN == 0) return getCouldNotCompute();
3888 // Since the loop is canonicalized, the PHI node must have two entries. One
3889 // entry must be a constant (coming in from outside of the loop), and the
3890 // second must be derived from the same PHI.
3891 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
3892 Constant *StartCST =
3893 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
3894 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
3896 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
3897 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
3898 if (PN2 != PN) return getCouldNotCompute(); // Not derived from same PHI.
3900 // Okay, we find a PHI node that defines the trip count of this loop. Execute
3901 // the loop symbolically to determine when the condition gets a value of
3903 unsigned IterationNum = 0;
3904 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
3905 for (Constant *PHIVal = StartCST;
3906 IterationNum != MaxIterations; ++IterationNum) {
3907 ConstantInt *CondVal =
3908 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal));
3910 // Couldn't symbolically evaluate.
3911 if (!CondVal) return getCouldNotCompute();
3913 if (CondVal->getValue() == uint64_t(ExitWhen)) {
3914 ++NumBruteForceTripCountsComputed;
3915 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
3918 // Compute the value of the PHI node for the next iteration.
3919 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
3920 if (NextPHI == 0 || NextPHI == PHIVal)
3921 return getCouldNotCompute();// Couldn't evaluate or not making progress...
3925 // Too many iterations were needed to evaluate.
3926 return getCouldNotCompute();
3929 /// getSCEVAtScope - Return a SCEV expression for the specified value
3930 /// at the specified scope in the program. The L value specifies a loop
3931 /// nest to evaluate the expression at, where null is the top-level or a
3932 /// specified loop is immediately inside of the loop.
3934 /// This method can be used to compute the exit value for a variable defined
3935 /// in a loop by querying what the value will hold in the parent loop.
3937 /// In the case that a relevant loop exit value cannot be computed, the
3938 /// original value V is returned.
3939 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
3940 // Check to see if we've folded this expression at this loop before.
3941 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
3942 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
3943 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
3945 return Pair.first->second ? Pair.first->second : V;
3947 // Otherwise compute it.
3948 const SCEV *C = computeSCEVAtScope(V, L);
3949 ValuesAtScopes[V][L] = C;
3953 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
3954 if (isa<SCEVConstant>(V)) return V;
3956 // If this instruction is evolved from a constant-evolving PHI, compute the
3957 // exit value from the loop without using SCEVs.
3958 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
3959 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
3960 const Loop *LI = (*this->LI)[I->getParent()];
3961 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
3962 if (PHINode *PN = dyn_cast<PHINode>(I))
3963 if (PN->getParent() == LI->getHeader()) {
3964 // Okay, there is no closed form solution for the PHI node. Check
3965 // to see if the loop that contains it has a known backedge-taken
3966 // count. If so, we may be able to force computation of the exit
3968 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
3969 if (const SCEVConstant *BTCC =
3970 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
3971 // Okay, we know how many times the containing loop executes. If
3972 // this is a constant evolving PHI node, get the final value at
3973 // the specified iteration number.
3974 Constant *RV = getConstantEvolutionLoopExitValue(PN,
3975 BTCC->getValue()->getValue(),
3977 if (RV) return getSCEV(RV);
3981 // Okay, this is an expression that we cannot symbolically evaluate
3982 // into a SCEV. Check to see if it's possible to symbolically evaluate
3983 // the arguments into constants, and if so, try to constant propagate the
3984 // result. This is particularly useful for computing loop exit values.
3985 if (CanConstantFold(I)) {
3986 std::vector<Constant*> Operands;
3987 Operands.reserve(I->getNumOperands());
3988 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3989 Value *Op = I->getOperand(i);
3990 if (Constant *C = dyn_cast<Constant>(Op)) {
3991 Operands.push_back(C);
3993 // If any of the operands is non-constant and if they are
3994 // non-integer and non-pointer, don't even try to analyze them
3995 // with scev techniques.
3996 if (!isSCEVable(Op->getType()))
3999 const SCEV* OpV = getSCEVAtScope(Op, L);
4000 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
4001 Constant *C = SC->getValue();
4002 if (C->getType() != Op->getType())
4003 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4007 Operands.push_back(C);
4008 } else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
4009 if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
4010 if (C->getType() != Op->getType())
4012 ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4016 Operands.push_back(C);
4026 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4027 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4028 &Operands[0], Operands.size(),
4031 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4032 &Operands[0], Operands.size(),
4038 // This is some other type of SCEVUnknown, just return it.
4042 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4043 // Avoid performing the look-up in the common case where the specified
4044 // expression has no loop-variant portions.
4045 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4046 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4047 if (OpAtScope != Comm->getOperand(i)) {
4048 // Okay, at least one of these operands is loop variant but might be
4049 // foldable. Build a new instance of the folded commutative expression.
4050 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4051 Comm->op_begin()+i);
4052 NewOps.push_back(OpAtScope);
4054 for (++i; i != e; ++i) {
4055 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4056 NewOps.push_back(OpAtScope);
4058 if (isa<SCEVAddExpr>(Comm))
4059 return getAddExpr(NewOps);
4060 if (isa<SCEVMulExpr>(Comm))
4061 return getMulExpr(NewOps);
4062 if (isa<SCEVSMaxExpr>(Comm))
4063 return getSMaxExpr(NewOps);
4064 if (isa<SCEVUMaxExpr>(Comm))
4065 return getUMaxExpr(NewOps);
4066 llvm_unreachable("Unknown commutative SCEV type!");
4069 // If we got here, all operands are loop invariant.
4073 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4074 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4075 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4076 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4077 return Div; // must be loop invariant
4078 return getUDivExpr(LHS, RHS);
4081 // If this is a loop recurrence for a loop that does not contain L, then we
4082 // are dealing with the final value computed by the loop.
4083 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4084 if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
4085 // To evaluate this recurrence, we need to know how many times the AddRec
4086 // loop iterates. Compute this now.
4087 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4088 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4090 // Then, evaluate the AddRec.
4091 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4096 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4097 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4098 if (Op == Cast->getOperand())
4099 return Cast; // must be loop invariant
4100 return getZeroExtendExpr(Op, Cast->getType());
4103 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4104 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4105 if (Op == Cast->getOperand())
4106 return Cast; // must be loop invariant
4107 return getSignExtendExpr(Op, Cast->getType());
4110 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4111 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4112 if (Op == Cast->getOperand())
4113 return Cast; // must be loop invariant
4114 return getTruncateExpr(Op, Cast->getType());
4117 if (isa<SCEVTargetDataConstant>(V))
4120 llvm_unreachable("Unknown SCEV type!");
4124 /// getSCEVAtScope - This is a convenience function which does
4125 /// getSCEVAtScope(getSCEV(V), L).
4126 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4127 return getSCEVAtScope(getSCEV(V), L);
4130 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4131 /// following equation:
4133 /// A * X = B (mod N)
4135 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4136 /// A and B isn't important.
4138 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4139 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4140 ScalarEvolution &SE) {
4141 uint32_t BW = A.getBitWidth();
4142 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4143 assert(A != 0 && "A must be non-zero.");
4147 // The gcd of A and N may have only one prime factor: 2. The number of
4148 // trailing zeros in A is its multiplicity
4149 uint32_t Mult2 = A.countTrailingZeros();
4152 // 2. Check if B is divisible by D.
4154 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4155 // is not less than multiplicity of this prime factor for D.
4156 if (B.countTrailingZeros() < Mult2)
4157 return SE.getCouldNotCompute();
4159 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4162 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4163 // bit width during computations.
4164 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4165 APInt Mod(BW + 1, 0);
4166 Mod.set(BW - Mult2); // Mod = N / D
4167 APInt I = AD.multiplicativeInverse(Mod);
4169 // 4. Compute the minimum unsigned root of the equation:
4170 // I * (B / D) mod (N / D)
4171 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4173 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4175 return SE.getConstant(Result.trunc(BW));
4178 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4179 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4180 /// might be the same) or two SCEVCouldNotCompute objects.
4182 static std::pair<const SCEV *,const SCEV *>
4183 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4184 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4185 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4186 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4187 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4189 // We currently can only solve this if the coefficients are constants.
4190 if (!LC || !MC || !NC) {
4191 const SCEV *CNC = SE.getCouldNotCompute();
4192 return std::make_pair(CNC, CNC);
4195 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4196 const APInt &L = LC->getValue()->getValue();
4197 const APInt &M = MC->getValue()->getValue();
4198 const APInt &N = NC->getValue()->getValue();
4199 APInt Two(BitWidth, 2);
4200 APInt Four(BitWidth, 4);
4203 using namespace APIntOps;
4205 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4206 // The B coefficient is M-N/2
4210 // The A coefficient is N/2
4211 APInt A(N.sdiv(Two));
4213 // Compute the B^2-4ac term.
4216 SqrtTerm -= Four * (A * C);
4218 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4219 // integer value or else APInt::sqrt() will assert.
4220 APInt SqrtVal(SqrtTerm.sqrt());
4222 // Compute the two solutions for the quadratic formula.
4223 // The divisions must be performed as signed divisions.
4225 APInt TwoA( A << 1 );
4226 if (TwoA.isMinValue()) {
4227 const SCEV *CNC = SE.getCouldNotCompute();
4228 return std::make_pair(CNC, CNC);
4231 LLVMContext &Context = SE.getContext();
4233 ConstantInt *Solution1 =
4234 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4235 ConstantInt *Solution2 =
4236 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4238 return std::make_pair(SE.getConstant(Solution1),
4239 SE.getConstant(Solution2));
4240 } // end APIntOps namespace
4243 /// HowFarToZero - Return the number of times a backedge comparing the specified
4244 /// value to zero will execute. If not computable, return CouldNotCompute.
4245 const SCEV *ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4246 // If the value is a constant
4247 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4248 // If the value is already zero, the branch will execute zero times.
4249 if (C->getValue()->isZero()) return C;
4250 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4253 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4254 if (!AddRec || AddRec->getLoop() != L)
4255 return getCouldNotCompute();
4257 if (AddRec->isAffine()) {
4258 // If this is an affine expression, the execution count of this branch is
4259 // the minimum unsigned root of the following equation:
4261 // Start + Step*N = 0 (mod 2^BW)
4265 // Step*N = -Start (mod 2^BW)
4267 // where BW is the common bit width of Start and Step.
4269 // Get the initial value for the loop.
4270 const SCEV *Start = getSCEVAtScope(AddRec->getStart(),
4271 L->getParentLoop());
4272 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1),
4273 L->getParentLoop());
4275 if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
4276 // For now we handle only constant steps.
4278 // First, handle unitary steps.
4279 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4280 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4281 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4282 return Start; // N = Start (as unsigned)
4284 // Then, try to solve the above equation provided that Start is constant.
4285 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4286 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4287 -StartC->getValue()->getValue(),
4290 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
4291 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4292 // the quadratic equation to solve it.
4293 std::pair<const SCEV *,const SCEV *> Roots = SolveQuadraticEquation(AddRec,
4295 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4296 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4299 errs() << "HFTZ: " << *V << " - sol#1: " << *R1
4300 << " sol#2: " << *R2 << "\n";
4302 // Pick the smallest positive root value.
4303 if (ConstantInt *CB =
4304 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
4305 R1->getValue(), R2->getValue()))) {
4306 if (CB->getZExtValue() == false)
4307 std::swap(R1, R2); // R1 is the minimum root now.
4309 // We can only use this value if the chrec ends up with an exact zero
4310 // value at this index. When solving for "X*X != 5", for example, we
4311 // should not accept a root of 2.
4312 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4314 return R1; // We found a quadratic root!
4319 return getCouldNotCompute();
4322 /// HowFarToNonZero - Return the number of times a backedge checking the
4323 /// specified value for nonzero will execute. If not computable, return
4325 const SCEV *ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4326 // Loops that look like: while (X == 0) are very strange indeed. We don't
4327 // handle them yet except for the trivial case. This could be expanded in the
4328 // future as needed.
4330 // If the value is a constant, check to see if it is known to be non-zero
4331 // already. If so, the backedge will execute zero times.
4332 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4333 if (!C->getValue()->isNullValue())
4334 return getIntegerSCEV(0, C->getType());
4335 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4338 // We could implement others, but I really doubt anyone writes loops like
4339 // this, and if they did, they would already be constant folded.
4340 return getCouldNotCompute();
4343 /// getLoopPredecessor - If the given loop's header has exactly one unique
4344 /// predecessor outside the loop, return it. Otherwise return null.
4346 BasicBlock *ScalarEvolution::getLoopPredecessor(const Loop *L) {
4347 BasicBlock *Header = L->getHeader();
4348 BasicBlock *Pred = 0;
4349 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header);
4351 if (!L->contains(*PI)) {
4352 if (Pred && Pred != *PI) return 0; // Multiple predecessors.
4358 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4359 /// (which may not be an immediate predecessor) which has exactly one
4360 /// successor from which BB is reachable, or null if no such block is
4364 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4365 // If the block has a unique predecessor, then there is no path from the
4366 // predecessor to the block that does not go through the direct edge
4367 // from the predecessor to the block.
4368 if (BasicBlock *Pred = BB->getSinglePredecessor())
4371 // A loop's header is defined to be a block that dominates the loop.
4372 // If the header has a unique predecessor outside the loop, it must be
4373 // a block that has exactly one successor that can reach the loop.
4374 if (Loop *L = LI->getLoopFor(BB))
4375 return getLoopPredecessor(L);
4380 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4381 /// testing whether two expressions are equal, however for the purposes of
4382 /// looking for a condition guarding a loop, it can be useful to be a little
4383 /// more general, since a front-end may have replicated the controlling
4386 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4387 // Quick check to see if they are the same SCEV.
4388 if (A == B) return true;
4390 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4391 // two different instructions with the same value. Check for this case.
4392 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4393 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4394 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4395 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4396 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4399 // Otherwise assume they may have a different value.
4403 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
4404 return getSignedRange(S).getSignedMax().isNegative();
4407 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
4408 return getSignedRange(S).getSignedMin().isStrictlyPositive();
4411 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
4412 return !getSignedRange(S).getSignedMin().isNegative();
4415 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
4416 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
4419 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
4420 return isKnownNegative(S) || isKnownPositive(S);
4423 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
4424 const SCEV *LHS, const SCEV *RHS) {
4426 if (HasSameValue(LHS, RHS))
4427 return ICmpInst::isTrueWhenEqual(Pred);
4431 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4433 case ICmpInst::ICMP_SGT:
4434 Pred = ICmpInst::ICMP_SLT;
4435 std::swap(LHS, RHS);
4436 case ICmpInst::ICMP_SLT: {
4437 ConstantRange LHSRange = getSignedRange(LHS);
4438 ConstantRange RHSRange = getSignedRange(RHS);
4439 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
4441 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
4445 case ICmpInst::ICMP_SGE:
4446 Pred = ICmpInst::ICMP_SLE;
4447 std::swap(LHS, RHS);
4448 case ICmpInst::ICMP_SLE: {
4449 ConstantRange LHSRange = getSignedRange(LHS);
4450 ConstantRange RHSRange = getSignedRange(RHS);
4451 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
4453 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
4457 case ICmpInst::ICMP_UGT:
4458 Pred = ICmpInst::ICMP_ULT;
4459 std::swap(LHS, RHS);
4460 case ICmpInst::ICMP_ULT: {
4461 ConstantRange LHSRange = getUnsignedRange(LHS);
4462 ConstantRange RHSRange = getUnsignedRange(RHS);
4463 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
4465 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
4469 case ICmpInst::ICMP_UGE:
4470 Pred = ICmpInst::ICMP_ULE;
4471 std::swap(LHS, RHS);
4472 case ICmpInst::ICMP_ULE: {
4473 ConstantRange LHSRange = getUnsignedRange(LHS);
4474 ConstantRange RHSRange = getUnsignedRange(RHS);
4475 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
4477 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
4481 case ICmpInst::ICMP_NE: {
4482 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
4484 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
4487 const SCEV *Diff = getMinusSCEV(LHS, RHS);
4488 if (isKnownNonZero(Diff))
4492 case ICmpInst::ICMP_EQ:
4493 // The check at the top of the function catches the case where
4494 // the values are known to be equal.
4500 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
4501 /// protected by a conditional between LHS and RHS. This is used to
4502 /// to eliminate casts.
4504 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
4505 ICmpInst::Predicate Pred,
4506 const SCEV *LHS, const SCEV *RHS) {
4507 // Interpret a null as meaning no loop, where there is obviously no guard
4508 // (interprocedural conditions notwithstanding).
4509 if (!L) return true;
4511 BasicBlock *Latch = L->getLoopLatch();
4515 BranchInst *LoopContinuePredicate =
4516 dyn_cast<BranchInst>(Latch->getTerminator());
4517 if (!LoopContinuePredicate ||
4518 LoopContinuePredicate->isUnconditional())
4521 return isImpliedCond(LoopContinuePredicate->getCondition(), Pred, LHS, RHS,
4522 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
4525 /// isLoopGuardedByCond - Test whether entry to the loop is protected
4526 /// by a conditional between LHS and RHS. This is used to help avoid max
4527 /// expressions in loop trip counts, and to eliminate casts.
4529 ScalarEvolution::isLoopGuardedByCond(const Loop *L,
4530 ICmpInst::Predicate Pred,
4531 const SCEV *LHS, const SCEV *RHS) {
4532 // Interpret a null as meaning no loop, where there is obviously no guard
4533 // (interprocedural conditions notwithstanding).
4534 if (!L) return false;
4536 BasicBlock *Predecessor = getLoopPredecessor(L);
4537 BasicBlock *PredecessorDest = L->getHeader();
4539 // Starting at the loop predecessor, climb up the predecessor chain, as long
4540 // as there are predecessors that can be found that have unique successors
4541 // leading to the original header.
4543 PredecessorDest = Predecessor,
4544 Predecessor = getPredecessorWithUniqueSuccessorForBB(Predecessor)) {
4546 BranchInst *LoopEntryPredicate =
4547 dyn_cast<BranchInst>(Predecessor->getTerminator());
4548 if (!LoopEntryPredicate ||
4549 LoopEntryPredicate->isUnconditional())
4552 if (isImpliedCond(LoopEntryPredicate->getCondition(), Pred, LHS, RHS,
4553 LoopEntryPredicate->getSuccessor(0) != PredecessorDest))
4560 /// isImpliedCond - Test whether the condition described by Pred, LHS,
4561 /// and RHS is true whenever the given Cond value evaluates to true.
4562 bool ScalarEvolution::isImpliedCond(Value *CondValue,
4563 ICmpInst::Predicate Pred,
4564 const SCEV *LHS, const SCEV *RHS,
4566 // Recursivly handle And and Or conditions.
4567 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CondValue)) {
4568 if (BO->getOpcode() == Instruction::And) {
4570 return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
4571 isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
4572 } else if (BO->getOpcode() == Instruction::Or) {
4574 return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
4575 isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
4579 ICmpInst *ICI = dyn_cast<ICmpInst>(CondValue);
4580 if (!ICI) return false;
4582 // Bail if the ICmp's operands' types are wider than the needed type
4583 // before attempting to call getSCEV on them. This avoids infinite
4584 // recursion, since the analysis of widening casts can require loop
4585 // exit condition information for overflow checking, which would
4587 if (getTypeSizeInBits(LHS->getType()) <
4588 getTypeSizeInBits(ICI->getOperand(0)->getType()))
4591 // Now that we found a conditional branch that dominates the loop, check to
4592 // see if it is the comparison we are looking for.
4593 ICmpInst::Predicate FoundPred;
4595 FoundPred = ICI->getInversePredicate();
4597 FoundPred = ICI->getPredicate();
4599 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
4600 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
4602 // Balance the types. The case where FoundLHS' type is wider than
4603 // LHS' type is checked for above.
4604 if (getTypeSizeInBits(LHS->getType()) >
4605 getTypeSizeInBits(FoundLHS->getType())) {
4606 if (CmpInst::isSigned(Pred)) {
4607 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
4608 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
4610 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
4611 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
4615 // Canonicalize the query to match the way instcombine will have
4616 // canonicalized the comparison.
4617 // First, put a constant operand on the right.
4618 if (isa<SCEVConstant>(LHS)) {
4619 std::swap(LHS, RHS);
4620 Pred = ICmpInst::getSwappedPredicate(Pred);
4622 // Then, canonicalize comparisons with boundary cases.
4623 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
4624 const APInt &RA = RC->getValue()->getValue();
4626 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4627 case ICmpInst::ICMP_EQ:
4628 case ICmpInst::ICMP_NE:
4630 case ICmpInst::ICMP_UGE:
4631 if ((RA - 1).isMinValue()) {
4632 Pred = ICmpInst::ICMP_NE;
4633 RHS = getConstant(RA - 1);
4636 if (RA.isMaxValue()) {
4637 Pred = ICmpInst::ICMP_EQ;
4640 if (RA.isMinValue()) return true;
4642 case ICmpInst::ICMP_ULE:
4643 if ((RA + 1).isMaxValue()) {
4644 Pred = ICmpInst::ICMP_NE;
4645 RHS = getConstant(RA + 1);
4648 if (RA.isMinValue()) {
4649 Pred = ICmpInst::ICMP_EQ;
4652 if (RA.isMaxValue()) return true;
4654 case ICmpInst::ICMP_SGE:
4655 if ((RA - 1).isMinSignedValue()) {
4656 Pred = ICmpInst::ICMP_NE;
4657 RHS = getConstant(RA - 1);
4660 if (RA.isMaxSignedValue()) {
4661 Pred = ICmpInst::ICMP_EQ;
4664 if (RA.isMinSignedValue()) return true;
4666 case ICmpInst::ICMP_SLE:
4667 if ((RA + 1).isMaxSignedValue()) {
4668 Pred = ICmpInst::ICMP_NE;
4669 RHS = getConstant(RA + 1);
4672 if (RA.isMinSignedValue()) {
4673 Pred = ICmpInst::ICMP_EQ;
4676 if (RA.isMaxSignedValue()) return true;
4678 case ICmpInst::ICMP_UGT:
4679 if (RA.isMinValue()) {
4680 Pred = ICmpInst::ICMP_NE;
4683 if ((RA + 1).isMaxValue()) {
4684 Pred = ICmpInst::ICMP_EQ;
4685 RHS = getConstant(RA + 1);
4688 if (RA.isMaxValue()) return false;
4690 case ICmpInst::ICMP_ULT:
4691 if (RA.isMaxValue()) {
4692 Pred = ICmpInst::ICMP_NE;
4695 if ((RA - 1).isMinValue()) {
4696 Pred = ICmpInst::ICMP_EQ;
4697 RHS = getConstant(RA - 1);
4700 if (RA.isMinValue()) return false;
4702 case ICmpInst::ICMP_SGT:
4703 if (RA.isMinSignedValue()) {
4704 Pred = ICmpInst::ICMP_NE;
4707 if ((RA + 1).isMaxSignedValue()) {
4708 Pred = ICmpInst::ICMP_EQ;
4709 RHS = getConstant(RA + 1);
4712 if (RA.isMaxSignedValue()) return false;
4714 case ICmpInst::ICMP_SLT:
4715 if (RA.isMaxSignedValue()) {
4716 Pred = ICmpInst::ICMP_NE;
4719 if ((RA - 1).isMinSignedValue()) {
4720 Pred = ICmpInst::ICMP_EQ;
4721 RHS = getConstant(RA - 1);
4724 if (RA.isMinSignedValue()) return false;
4729 // Check to see if we can make the LHS or RHS match.
4730 if (LHS == FoundRHS || RHS == FoundLHS) {
4731 if (isa<SCEVConstant>(RHS)) {
4732 std::swap(FoundLHS, FoundRHS);
4733 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
4735 std::swap(LHS, RHS);
4736 Pred = ICmpInst::getSwappedPredicate(Pred);
4740 // Check whether the found predicate is the same as the desired predicate.
4741 if (FoundPred == Pred)
4742 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
4744 // Check whether swapping the found predicate makes it the same as the
4745 // desired predicate.
4746 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
4747 if (isa<SCEVConstant>(RHS))
4748 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
4750 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
4751 RHS, LHS, FoundLHS, FoundRHS);
4754 // Check whether the actual condition is beyond sufficient.
4755 if (FoundPred == ICmpInst::ICMP_EQ)
4756 if (ICmpInst::isTrueWhenEqual(Pred))
4757 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
4759 if (Pred == ICmpInst::ICMP_NE)
4760 if (!ICmpInst::isTrueWhenEqual(FoundPred))
4761 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
4764 // Otherwise assume the worst.
4768 /// isImpliedCondOperands - Test whether the condition described by Pred,
4769 /// LHS, and RHS is true whenever the condition desribed by Pred, FoundLHS,
4770 /// and FoundRHS is true.
4771 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
4772 const SCEV *LHS, const SCEV *RHS,
4773 const SCEV *FoundLHS,
4774 const SCEV *FoundRHS) {
4775 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
4776 FoundLHS, FoundRHS) ||
4777 // ~x < ~y --> x > y
4778 isImpliedCondOperandsHelper(Pred, LHS, RHS,
4779 getNotSCEV(FoundRHS),
4780 getNotSCEV(FoundLHS));
4783 /// isImpliedCondOperandsHelper - Test whether the condition described by
4784 /// Pred, LHS, and RHS is true whenever the condition desribed by Pred,
4785 /// FoundLHS, and FoundRHS is true.
4787 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
4788 const SCEV *LHS, const SCEV *RHS,
4789 const SCEV *FoundLHS,
4790 const SCEV *FoundRHS) {
4792 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4793 case ICmpInst::ICMP_EQ:
4794 case ICmpInst::ICMP_NE:
4795 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
4798 case ICmpInst::ICMP_SLT:
4799 case ICmpInst::ICMP_SLE:
4800 if (isKnownPredicate(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
4801 isKnownPredicate(ICmpInst::ICMP_SGE, RHS, FoundRHS))
4804 case ICmpInst::ICMP_SGT:
4805 case ICmpInst::ICMP_SGE:
4806 if (isKnownPredicate(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
4807 isKnownPredicate(ICmpInst::ICMP_SLE, RHS, FoundRHS))
4810 case ICmpInst::ICMP_ULT:
4811 case ICmpInst::ICMP_ULE:
4812 if (isKnownPredicate(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
4813 isKnownPredicate(ICmpInst::ICMP_UGE, RHS, FoundRHS))
4816 case ICmpInst::ICMP_UGT:
4817 case ICmpInst::ICMP_UGE:
4818 if (isKnownPredicate(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
4819 isKnownPredicate(ICmpInst::ICMP_ULE, RHS, FoundRHS))
4827 /// getBECount - Subtract the end and start values and divide by the step,
4828 /// rounding up, to get the number of times the backedge is executed. Return
4829 /// CouldNotCompute if an intermediate computation overflows.
4830 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
4834 const Type *Ty = Start->getType();
4835 const SCEV *NegOne = getIntegerSCEV(-1, Ty);
4836 const SCEV *Diff = getMinusSCEV(End, Start);
4837 const SCEV *RoundUp = getAddExpr(Step, NegOne);
4839 // Add an adjustment to the difference between End and Start so that
4840 // the division will effectively round up.
4841 const SCEV *Add = getAddExpr(Diff, RoundUp);
4844 // Check Add for unsigned overflow.
4845 // TODO: More sophisticated things could be done here.
4846 const Type *WideTy = IntegerType::get(getContext(),
4847 getTypeSizeInBits(Ty) + 1);
4848 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
4849 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
4850 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
4851 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
4852 return getCouldNotCompute();
4855 return getUDivExpr(Add, Step);
4858 /// HowManyLessThans - Return the number of times a backedge containing the
4859 /// specified less-than comparison will execute. If not computable, return
4860 /// CouldNotCompute.
4861 ScalarEvolution::BackedgeTakenInfo
4862 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
4863 const Loop *L, bool isSigned) {
4864 // Only handle: "ADDREC < LoopInvariant".
4865 if (!RHS->isLoopInvariant(L)) return getCouldNotCompute();
4867 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
4868 if (!AddRec || AddRec->getLoop() != L)
4869 return getCouldNotCompute();
4871 // Check to see if we have a flag which makes analysis easy.
4872 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
4873 AddRec->hasNoUnsignedWrap();
4875 if (AddRec->isAffine()) {
4876 // FORNOW: We only support unit strides.
4877 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
4878 const SCEV *Step = AddRec->getStepRecurrence(*this);
4880 // TODO: handle non-constant strides.
4881 const SCEVConstant *CStep = dyn_cast<SCEVConstant>(Step);
4882 if (!CStep || CStep->isZero())
4883 return getCouldNotCompute();
4884 if (CStep->isOne()) {
4885 // With unit stride, the iteration never steps past the limit value.
4886 } else if (CStep->getValue()->getValue().isStrictlyPositive()) {
4888 // We know the iteration won't step past the maximum value for its type.
4890 } else if (const SCEVConstant *CLimit = dyn_cast<SCEVConstant>(RHS)) {
4891 // Test whether a positive iteration iteration can step past the limit
4892 // value and past the maximum value for its type in a single step.
4894 APInt Max = APInt::getSignedMaxValue(BitWidth);
4895 if ((Max - CStep->getValue()->getValue())
4896 .slt(CLimit->getValue()->getValue()))
4897 return getCouldNotCompute();
4899 APInt Max = APInt::getMaxValue(BitWidth);
4900 if ((Max - CStep->getValue()->getValue())
4901 .ult(CLimit->getValue()->getValue()))
4902 return getCouldNotCompute();
4905 // TODO: handle non-constant limit values below.
4906 return getCouldNotCompute();
4908 // TODO: handle negative strides below.
4909 return getCouldNotCompute();
4911 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
4912 // m. So, we count the number of iterations in which {n,+,s} < m is true.
4913 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
4914 // treat m-n as signed nor unsigned due to overflow possibility.
4916 // First, we get the value of the LHS in the first iteration: n
4917 const SCEV *Start = AddRec->getOperand(0);
4919 // Determine the minimum constant start value.
4920 const SCEV *MinStart = getConstant(isSigned ?
4921 getSignedRange(Start).getSignedMin() :
4922 getUnsignedRange(Start).getUnsignedMin());
4924 // If we know that the condition is true in order to enter the loop,
4925 // then we know that it will run exactly (m-n)/s times. Otherwise, we
4926 // only know that it will execute (max(m,n)-n)/s times. In both cases,
4927 // the division must round up.
4928 const SCEV *End = RHS;
4929 if (!isLoopGuardedByCond(L,
4930 isSigned ? ICmpInst::ICMP_SLT :
4932 getMinusSCEV(Start, Step), RHS))
4933 End = isSigned ? getSMaxExpr(RHS, Start)
4934 : getUMaxExpr(RHS, Start);
4936 // Determine the maximum constant end value.
4937 const SCEV *MaxEnd = getConstant(isSigned ?
4938 getSignedRange(End).getSignedMax() :
4939 getUnsignedRange(End).getUnsignedMax());
4941 // Finally, we subtract these two values and divide, rounding up, to get
4942 // the number of times the backedge is executed.
4943 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
4945 // The maximum backedge count is similar, except using the minimum start
4946 // value and the maximum end value.
4947 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
4949 return BackedgeTakenInfo(BECount, MaxBECount);
4952 return getCouldNotCompute();
4955 /// getNumIterationsInRange - Return the number of iterations of this loop that
4956 /// produce values in the specified constant range. Another way of looking at
4957 /// this is that it returns the first iteration number where the value is not in
4958 /// the condition, thus computing the exit count. If the iteration count can't
4959 /// be computed, an instance of SCEVCouldNotCompute is returned.
4960 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
4961 ScalarEvolution &SE) const {
4962 if (Range.isFullSet()) // Infinite loop.
4963 return SE.getCouldNotCompute();
4965 // If the start is a non-zero constant, shift the range to simplify things.
4966 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
4967 if (!SC->getValue()->isZero()) {
4968 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
4969 Operands[0] = SE.getIntegerSCEV(0, SC->getType());
4970 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
4971 if (const SCEVAddRecExpr *ShiftedAddRec =
4972 dyn_cast<SCEVAddRecExpr>(Shifted))
4973 return ShiftedAddRec->getNumIterationsInRange(
4974 Range.subtract(SC->getValue()->getValue()), SE);
4975 // This is strange and shouldn't happen.
4976 return SE.getCouldNotCompute();
4979 // The only time we can solve this is when we have all constant indices.
4980 // Otherwise, we cannot determine the overflow conditions.
4981 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
4982 if (!isa<SCEVConstant>(getOperand(i)))
4983 return SE.getCouldNotCompute();
4986 // Okay at this point we know that all elements of the chrec are constants and
4987 // that the start element is zero.
4989 // First check to see if the range contains zero. If not, the first
4991 unsigned BitWidth = SE.getTypeSizeInBits(getType());
4992 if (!Range.contains(APInt(BitWidth, 0)))
4993 return SE.getIntegerSCEV(0, getType());
4996 // If this is an affine expression then we have this situation:
4997 // Solve {0,+,A} in Range === Ax in Range
4999 // We know that zero is in the range. If A is positive then we know that
5000 // the upper value of the range must be the first possible exit value.
5001 // If A is negative then the lower of the range is the last possible loop
5002 // value. Also note that we already checked for a full range.
5003 APInt One(BitWidth,1);
5004 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5005 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5007 // The exit value should be (End+A)/A.
5008 APInt ExitVal = (End + A).udiv(A);
5009 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5011 // Evaluate at the exit value. If we really did fall out of the valid
5012 // range, then we computed our trip count, otherwise wrap around or other
5013 // things must have happened.
5014 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5015 if (Range.contains(Val->getValue()))
5016 return SE.getCouldNotCompute(); // Something strange happened
5018 // Ensure that the previous value is in the range. This is a sanity check.
5019 assert(Range.contains(
5020 EvaluateConstantChrecAtConstant(this,
5021 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5022 "Linear scev computation is off in a bad way!");
5023 return SE.getConstant(ExitValue);
5024 } else if (isQuadratic()) {
5025 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5026 // quadratic equation to solve it. To do this, we must frame our problem in
5027 // terms of figuring out when zero is crossed, instead of when
5028 // Range.getUpper() is crossed.
5029 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5030 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5031 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5033 // Next, solve the constructed addrec
5034 std::pair<const SCEV *,const SCEV *> Roots =
5035 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5036 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5037 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5039 // Pick the smallest positive root value.
5040 if (ConstantInt *CB =
5041 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5042 R1->getValue(), R2->getValue()))) {
5043 if (CB->getZExtValue() == false)
5044 std::swap(R1, R2); // R1 is the minimum root now.
5046 // Make sure the root is not off by one. The returned iteration should
5047 // not be in the range, but the previous one should be. When solving
5048 // for "X*X < 5", for example, we should not return a root of 2.
5049 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5052 if (Range.contains(R1Val->getValue())) {
5053 // The next iteration must be out of the range...
5054 ConstantInt *NextVal =
5055 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5057 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5058 if (!Range.contains(R1Val->getValue()))
5059 return SE.getConstant(NextVal);
5060 return SE.getCouldNotCompute(); // Something strange happened
5063 // If R1 was not in the range, then it is a good return value. Make
5064 // sure that R1-1 WAS in the range though, just in case.
5065 ConstantInt *NextVal =
5066 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5067 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5068 if (Range.contains(R1Val->getValue()))
5070 return SE.getCouldNotCompute(); // Something strange happened
5075 return SE.getCouldNotCompute();
5080 //===----------------------------------------------------------------------===//
5081 // SCEVCallbackVH Class Implementation
5082 //===----------------------------------------------------------------------===//
5084 void ScalarEvolution::SCEVCallbackVH::deleted() {
5085 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5086 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5087 SE->ConstantEvolutionLoopExitValue.erase(PN);
5088 SE->Scalars.erase(getValPtr());
5089 // this now dangles!
5092 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *) {
5093 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5095 // Forget all the expressions associated with users of the old value,
5096 // so that future queries will recompute the expressions using the new
5098 SmallVector<User *, 16> Worklist;
5099 SmallPtrSet<User *, 8> Visited;
5100 Value *Old = getValPtr();
5101 bool DeleteOld = false;
5102 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5104 Worklist.push_back(*UI);
5105 while (!Worklist.empty()) {
5106 User *U = Worklist.pop_back_val();
5107 // Deleting the Old value will cause this to dangle. Postpone
5108 // that until everything else is done.
5113 if (!Visited.insert(U))
5115 if (PHINode *PN = dyn_cast<PHINode>(U))
5116 SE->ConstantEvolutionLoopExitValue.erase(PN);
5117 SE->Scalars.erase(U);
5118 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5120 Worklist.push_back(*UI);
5122 // Delete the Old value if it (indirectly) references itself.
5124 if (PHINode *PN = dyn_cast<PHINode>(Old))
5125 SE->ConstantEvolutionLoopExitValue.erase(PN);
5126 SE->Scalars.erase(Old);
5127 // this now dangles!
5132 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5133 : CallbackVH(V), SE(se) {}
5135 //===----------------------------------------------------------------------===//
5136 // ScalarEvolution Class Implementation
5137 //===----------------------------------------------------------------------===//
5139 ScalarEvolution::ScalarEvolution()
5140 : FunctionPass(&ID) {
5143 bool ScalarEvolution::runOnFunction(Function &F) {
5145 LI = &getAnalysis<LoopInfo>();
5146 TD = getAnalysisIfAvailable<TargetData>();
5150 void ScalarEvolution::releaseMemory() {
5152 BackedgeTakenCounts.clear();
5153 ConstantEvolutionLoopExitValue.clear();
5154 ValuesAtScopes.clear();
5155 UniqueSCEVs.clear();
5156 SCEVAllocator.Reset();
5159 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5160 AU.setPreservesAll();
5161 AU.addRequiredTransitive<LoopInfo>();
5164 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5165 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5168 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5170 // Print all inner loops first
5171 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5172 PrintLoopInfo(OS, SE, *I);
5174 OS << "Loop " << L->getHeader()->getName() << ": ";
5176 SmallVector<BasicBlock*, 8> ExitBlocks;
5177 L->getExitBlocks(ExitBlocks);
5178 if (ExitBlocks.size() != 1)
5179 OS << "<multiple exits> ";
5181 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5182 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5184 OS << "Unpredictable backedge-taken count. ";
5188 OS << "Loop " << L->getHeader()->getName() << ": ";
5190 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5191 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5193 OS << "Unpredictable max backedge-taken count. ";
5199 void ScalarEvolution::print(raw_ostream &OS, const Module* ) const {
5200 // ScalarEvolution's implementaiton of the print method is to print
5201 // out SCEV values of all instructions that are interesting. Doing
5202 // this potentially causes it to create new SCEV objects though,
5203 // which technically conflicts with the const qualifier. This isn't
5204 // observable from outside the class though, so casting away the
5205 // const isn't dangerous.
5206 ScalarEvolution &SE = *const_cast<ScalarEvolution*>(this);
5208 OS << "Classifying expressions for: " << F->getName() << "\n";
5209 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
5210 if (isSCEVable(I->getType())) {
5213 const SCEV *SV = SE.getSCEV(&*I);
5216 const Loop *L = LI->getLoopFor((*I).getParent());
5218 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
5225 OS << "\t\t" "Exits: ";
5226 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
5227 if (!ExitValue->isLoopInvariant(L)) {
5228 OS << "<<Unknown>>";
5237 OS << "Determining loop execution counts for: " << F->getName() << "\n";
5238 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
5239 PrintLoopInfo(OS, &SE, *I);