1 //===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the implementation of the scalar evolution analysis
11 // engine, which is used primarily to analyze expressions involving induction
12 // variables in loops.
14 // There are several aspects to this library. First is the representation of
15 // scalar expressions, which are represented as subclasses of the SCEV class.
16 // These classes are used to represent certain types of subexpressions that we
17 // can handle. We only create one SCEV of a particular shape, so
18 // pointer-comparisons for equality are legal.
20 // One important aspect of the SCEV objects is that they are never cyclic, even
21 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
22 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
23 // recurrence) then we represent it directly as a recurrence node, otherwise we
24 // represent it as a SCEVUnknown node.
26 // In addition to being able to represent expressions of various types, we also
27 // have folders that are used to build the *canonical* representation for a
28 // particular expression. These folders are capable of using a variety of
29 // rewrite rules to simplify the expressions.
31 // Once the folders are defined, we can implement the more interesting
32 // higher-level code, such as the code that recognizes PHI nodes of various
33 // types, computes the execution count of a loop, etc.
35 // TODO: We should use these routines and value representations to implement
36 // dependence analysis!
38 //===----------------------------------------------------------------------===//
40 // There are several good references for the techniques used in this analysis.
42 // Chains of recurrences -- a method to expedite the evaluation
43 // of closed-form functions
44 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
46 // On computational properties of chains of recurrences
49 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
50 // Robert A. van Engelen
52 // Efficient Symbolic Analysis for Optimizing Compilers
53 // Robert A. van Engelen
55 // Using the chains of recurrences algebra for data dependence testing and
56 // induction variable substitution
57 // MS Thesis, Johnie Birch
59 //===----------------------------------------------------------------------===//
61 #define DEBUG_TYPE "scalar-evolution"
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/Constants.h"
64 #include "llvm/DerivedTypes.h"
65 #include "llvm/GlobalVariable.h"
66 #include "llvm/GlobalAlias.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/LLVMContext.h"
69 #include "llvm/Operator.h"
70 #include "llvm/Analysis/ConstantFolding.h"
71 #include "llvm/Analysis/Dominators.h"
72 #include "llvm/Analysis/LoopInfo.h"
73 #include "llvm/Analysis/ValueTracking.h"
74 #include "llvm/Assembly/Writer.h"
75 #include "llvm/Target/TargetData.h"
76 #include "llvm/Support/CommandLine.h"
77 #include "llvm/Support/ConstantRange.h"
78 #include "llvm/Support/Debug.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))
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];
320 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
324 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
325 // All non-instruction values are loop invariant. All instructions are loop
326 // invariant if they are not contained in the specified loop.
327 // Instructions are never considered invariant in the function body
328 // (null loop) because they are defined within the "loop".
329 if (Instruction *I = dyn_cast<Instruction>(V))
330 return L && !L->contains(I);
334 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
335 if (Instruction *I = dyn_cast<Instruction>(getValue()))
336 return DT->dominates(I->getParent(), BB);
340 bool SCEVUnknown::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
341 if (Instruction *I = dyn_cast<Instruction>(getValue()))
342 return DT->properlyDominates(I->getParent(), BB);
346 const Type *SCEVUnknown::getType() const {
350 bool SCEVUnknown::isOffsetOf(const StructType *&STy, Constant *&FieldNo) const {
351 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
352 if (VCE->getOpcode() == Instruction::PtrToInt)
353 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
354 if (CE->getOpcode() == Instruction::GetElementPtr)
355 if (CE->getOperand(0)->isNullValue()) {
357 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
358 if (const StructType *StructTy = dyn_cast<StructType>(Ty))
359 if (CE->getNumOperands() == 3 &&
360 CE->getOperand(1)->isNullValue()) {
362 FieldNo = CE->getOperand(2);
370 bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
371 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
372 if (VCE->getOpcode() == Instruction::PtrToInt)
373 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
374 if (CE->getOpcode() == Instruction::GetElementPtr)
375 if (CE->getOperand(0)->isNullValue()) {
377 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
378 if (CE->getNumOperands() == 2)
379 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
389 bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
390 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
391 if (VCE->getOpcode() == Instruction::PtrToInt)
392 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
393 if (CE->getOpcode() == Instruction::GetElementPtr)
394 if (CE->getOperand(0)->isNullValue()) {
396 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
397 if (const StructType *STy = dyn_cast<StructType>(Ty))
398 if (CE->getNumOperands() == 3 &&
399 CE->getOperand(1)->isNullValue()) {
400 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
402 STy->getNumElements() == 2 &&
403 STy->getElementType(0)->isInteger(1)) {
404 AllocTy = STy->getElementType(1);
413 void SCEVUnknown::print(raw_ostream &OS) const {
415 if (isSizeOf(AllocTy)) {
416 OS << "sizeof(" << *AllocTy << ")";
419 if (isAlignOf(AllocTy)) {
420 OS << "alignof(" << *AllocTy << ")";
424 const StructType *STy;
426 if (isOffsetOf(STy, FieldNo)) {
427 OS << "offsetof(" << *STy << ", ";
428 WriteAsOperand(OS, FieldNo, false);
433 // Otherwise just print it normally.
434 WriteAsOperand(OS, V, false);
437 //===----------------------------------------------------------------------===//
439 //===----------------------------------------------------------------------===//
441 static bool CompareTypes(const Type *A, const Type *B) {
442 if (A->getTypeID() != B->getTypeID())
443 return A->getTypeID() < B->getTypeID();
444 if (const IntegerType *AI = dyn_cast<IntegerType>(A)) {
445 const IntegerType *BI = cast<IntegerType>(B);
446 return AI->getBitWidth() < BI->getBitWidth();
448 if (const PointerType *AI = dyn_cast<PointerType>(A)) {
449 const PointerType *BI = cast<PointerType>(B);
450 return CompareTypes(AI->getElementType(), BI->getElementType());
452 if (const ArrayType *AI = dyn_cast<ArrayType>(A)) {
453 const ArrayType *BI = cast<ArrayType>(B);
454 if (AI->getNumElements() != BI->getNumElements())
455 return AI->getNumElements() < BI->getNumElements();
456 return CompareTypes(AI->getElementType(), BI->getElementType());
458 if (const VectorType *AI = dyn_cast<VectorType>(A)) {
459 const VectorType *BI = cast<VectorType>(B);
460 if (AI->getNumElements() != BI->getNumElements())
461 return AI->getNumElements() < BI->getNumElements();
462 return CompareTypes(AI->getElementType(), BI->getElementType());
464 if (const StructType *AI = dyn_cast<StructType>(A)) {
465 const StructType *BI = cast<StructType>(B);
466 if (AI->getNumElements() != BI->getNumElements())
467 return AI->getNumElements() < BI->getNumElements();
468 for (unsigned i = 0, e = AI->getNumElements(); i != e; ++i)
469 if (CompareTypes(AI->getElementType(i), BI->getElementType(i)) ||
470 CompareTypes(BI->getElementType(i), AI->getElementType(i)))
471 return CompareTypes(AI->getElementType(i), BI->getElementType(i));
477 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
478 /// than the complexity of the RHS. This comparator is used to canonicalize
480 class SCEVComplexityCompare {
483 explicit SCEVComplexityCompare(LoopInfo *li) : LI(li) {}
485 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
486 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
490 // Primarily, sort the SCEVs by their getSCEVType().
491 if (LHS->getSCEVType() != RHS->getSCEVType())
492 return LHS->getSCEVType() < RHS->getSCEVType();
494 // Aside from the getSCEVType() ordering, the particular ordering
495 // isn't very important except that it's beneficial to be consistent,
496 // so that (a + b) and (b + a) don't end up as different expressions.
498 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
499 // not as complete as it could be.
500 if (const SCEVUnknown *LU = dyn_cast<SCEVUnknown>(LHS)) {
501 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
503 // Order pointer values after integer values. This helps SCEVExpander
505 if (isa<PointerType>(LU->getType()) && !isa<PointerType>(RU->getType()))
507 if (isa<PointerType>(RU->getType()) && !isa<PointerType>(LU->getType()))
510 // Compare getValueID values.
511 if (LU->getValue()->getValueID() != RU->getValue()->getValueID())
512 return LU->getValue()->getValueID() < RU->getValue()->getValueID();
514 // Sort arguments by their position.
515 if (const Argument *LA = dyn_cast<Argument>(LU->getValue())) {
516 const Argument *RA = cast<Argument>(RU->getValue());
517 return LA->getArgNo() < RA->getArgNo();
520 // For instructions, compare their loop depth, and their opcode.
521 // This is pretty loose.
522 if (Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
523 Instruction *RV = cast<Instruction>(RU->getValue());
525 // Compare loop depths.
526 if (LI->getLoopDepth(LV->getParent()) !=
527 LI->getLoopDepth(RV->getParent()))
528 return LI->getLoopDepth(LV->getParent()) <
529 LI->getLoopDepth(RV->getParent());
532 if (LV->getOpcode() != RV->getOpcode())
533 return LV->getOpcode() < RV->getOpcode();
535 // Compare the number of operands.
536 if (LV->getNumOperands() != RV->getNumOperands())
537 return LV->getNumOperands() < RV->getNumOperands();
543 // Compare constant values.
544 if (const SCEVConstant *LC = dyn_cast<SCEVConstant>(LHS)) {
545 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
546 if (LC->getValue()->getBitWidth() != RC->getValue()->getBitWidth())
547 return LC->getValue()->getBitWidth() < RC->getValue()->getBitWidth();
548 return LC->getValue()->getValue().ult(RC->getValue()->getValue());
551 // Compare addrec loop depths.
552 if (const SCEVAddRecExpr *LA = dyn_cast<SCEVAddRecExpr>(LHS)) {
553 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
554 if (LA->getLoop()->getLoopDepth() != RA->getLoop()->getLoopDepth())
555 return LA->getLoop()->getLoopDepth() < RA->getLoop()->getLoopDepth();
558 // Lexicographically compare n-ary expressions.
559 if (const SCEVNAryExpr *LC = dyn_cast<SCEVNAryExpr>(LHS)) {
560 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
561 for (unsigned i = 0, e = LC->getNumOperands(); i != e; ++i) {
562 if (i >= RC->getNumOperands())
564 if (operator()(LC->getOperand(i), RC->getOperand(i)))
566 if (operator()(RC->getOperand(i), LC->getOperand(i)))
569 return LC->getNumOperands() < RC->getNumOperands();
572 // Lexicographically compare udiv expressions.
573 if (const SCEVUDivExpr *LC = dyn_cast<SCEVUDivExpr>(LHS)) {
574 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
575 if (operator()(LC->getLHS(), RC->getLHS()))
577 if (operator()(RC->getLHS(), LC->getLHS()))
579 if (operator()(LC->getRHS(), RC->getRHS()))
581 if (operator()(RC->getRHS(), LC->getRHS()))
586 // Compare cast expressions by operand.
587 if (const SCEVCastExpr *LC = dyn_cast<SCEVCastExpr>(LHS)) {
588 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
589 return operator()(LC->getOperand(), RC->getOperand());
592 llvm_unreachable("Unknown SCEV kind!");
598 /// GroupByComplexity - Given a list of SCEV objects, order them by their
599 /// complexity, and group objects of the same complexity together by value.
600 /// When this routine is finished, we know that any duplicates in the vector are
601 /// consecutive and that complexity is monotonically increasing.
603 /// Note that we go take special precautions to ensure that we get determinstic
604 /// results from this routine. In other words, we don't want the results of
605 /// this to depend on where the addresses of various SCEV objects happened to
608 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
610 if (Ops.size() < 2) return; // Noop
611 if (Ops.size() == 2) {
612 // This is the common case, which also happens to be trivially simple.
614 if (SCEVComplexityCompare(LI)(Ops[1], Ops[0]))
615 std::swap(Ops[0], Ops[1]);
619 // Do the rough sort by complexity.
620 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
622 // Now that we are sorted by complexity, group elements of the same
623 // complexity. Note that this is, at worst, N^2, but the vector is likely to
624 // be extremely short in practice. Note that we take this approach because we
625 // do not want to depend on the addresses of the objects we are grouping.
626 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
627 const SCEV *S = Ops[i];
628 unsigned Complexity = S->getSCEVType();
630 // If there are any objects of the same complexity and same value as this
632 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
633 if (Ops[j] == S) { // Found a duplicate.
634 // Move it to immediately after i'th element.
635 std::swap(Ops[i+1], Ops[j]);
636 ++i; // no need to rescan it.
637 if (i == e-2) return; // Done!
645 //===----------------------------------------------------------------------===//
646 // Simple SCEV method implementations
647 //===----------------------------------------------------------------------===//
649 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
651 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
653 const Type* ResultTy) {
654 // Handle the simplest case efficiently.
656 return SE.getTruncateOrZeroExtend(It, ResultTy);
658 // We are using the following formula for BC(It, K):
660 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
662 // Suppose, W is the bitwidth of the return value. We must be prepared for
663 // overflow. Hence, we must assure that the result of our computation is
664 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
665 // safe in modular arithmetic.
667 // However, this code doesn't use exactly that formula; the formula it uses
668 // is something like the following, where T is the number of factors of 2 in
669 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
672 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
674 // This formula is trivially equivalent to the previous formula. However,
675 // this formula can be implemented much more efficiently. The trick is that
676 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
677 // arithmetic. To do exact division in modular arithmetic, all we have
678 // to do is multiply by the inverse. Therefore, this step can be done at
681 // The next issue is how to safely do the division by 2^T. The way this
682 // is done is by doing the multiplication step at a width of at least W + T
683 // bits. This way, the bottom W+T bits of the product are accurate. Then,
684 // when we perform the division by 2^T (which is equivalent to a right shift
685 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
686 // truncated out after the division by 2^T.
688 // In comparison to just directly using the first formula, this technique
689 // is much more efficient; using the first formula requires W * K bits,
690 // but this formula less than W + K bits. Also, the first formula requires
691 // a division step, whereas this formula only requires multiplies and shifts.
693 // It doesn't matter whether the subtraction step is done in the calculation
694 // width or the input iteration count's width; if the subtraction overflows,
695 // the result must be zero anyway. We prefer here to do it in the width of
696 // the induction variable because it helps a lot for certain cases; CodeGen
697 // isn't smart enough to ignore the overflow, which leads to much less
698 // efficient code if the width of the subtraction is wider than the native
701 // (It's possible to not widen at all by pulling out factors of 2 before
702 // the multiplication; for example, K=2 can be calculated as
703 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
704 // extra arithmetic, so it's not an obvious win, and it gets
705 // much more complicated for K > 3.)
707 // Protection from insane SCEVs; this bound is conservative,
708 // but it probably doesn't matter.
710 return SE.getCouldNotCompute();
712 unsigned W = SE.getTypeSizeInBits(ResultTy);
714 // Calculate K! / 2^T and T; we divide out the factors of two before
715 // multiplying for calculating K! / 2^T to avoid overflow.
716 // Other overflow doesn't matter because we only care about the bottom
717 // W bits of the result.
718 APInt OddFactorial(W, 1);
720 for (unsigned i = 3; i <= K; ++i) {
722 unsigned TwoFactors = Mult.countTrailingZeros();
724 Mult = Mult.lshr(TwoFactors);
725 OddFactorial *= Mult;
728 // We need at least W + T bits for the multiplication step
729 unsigned CalculationBits = W + T;
731 // Calcuate 2^T, at width T+W.
732 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
734 // Calculate the multiplicative inverse of K! / 2^T;
735 // this multiplication factor will perform the exact division by
737 APInt Mod = APInt::getSignedMinValue(W+1);
738 APInt MultiplyFactor = OddFactorial.zext(W+1);
739 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
740 MultiplyFactor = MultiplyFactor.trunc(W);
742 // Calculate the product, at width T+W
743 const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
745 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
746 for (unsigned i = 1; i != K; ++i) {
747 const SCEV *S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
748 Dividend = SE.getMulExpr(Dividend,
749 SE.getTruncateOrZeroExtend(S, CalculationTy));
753 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
755 // Truncate the result, and divide by K! / 2^T.
757 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
758 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
761 /// evaluateAtIteration - Return the value of this chain of recurrences at
762 /// the specified iteration number. We can evaluate this recurrence by
763 /// multiplying each element in the chain by the binomial coefficient
764 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
766 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
768 /// where BC(It, k) stands for binomial coefficient.
770 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
771 ScalarEvolution &SE) const {
772 const SCEV *Result = getStart();
773 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
774 // The computation is correct in the face of overflow provided that the
775 // multiplication is performed _after_ the evaluation of the binomial
777 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
778 if (isa<SCEVCouldNotCompute>(Coeff))
781 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
786 //===----------------------------------------------------------------------===//
787 // SCEV Expression folder implementations
788 //===----------------------------------------------------------------------===//
790 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
792 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
793 "This is not a truncating conversion!");
794 assert(isSCEVable(Ty) &&
795 "This is not a conversion to a SCEVable type!");
796 Ty = getEffectiveSCEVType(Ty);
799 ID.AddInteger(scTruncate);
803 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
805 // Fold if the operand is constant.
806 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
808 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
810 // trunc(trunc(x)) --> trunc(x)
811 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
812 return getTruncateExpr(ST->getOperand(), Ty);
814 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
815 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
816 return getTruncateOrSignExtend(SS->getOperand(), Ty);
818 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
819 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
820 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
822 // If the input value is a chrec scev, truncate the chrec's operands.
823 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
824 SmallVector<const SCEV *, 4> Operands;
825 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
826 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
827 return getAddRecExpr(Operands, AddRec->getLoop());
830 // The cast wasn't folded; create an explicit cast node.
831 // Recompute the insert position, as it may have been invalidated.
832 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
833 SCEV *S = SCEVAllocator.Allocate<SCEVTruncateExpr>();
834 new (S) SCEVTruncateExpr(ID, Op, Ty);
835 UniqueSCEVs.InsertNode(S, IP);
839 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
841 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
842 "This is not an extending conversion!");
843 assert(isSCEVable(Ty) &&
844 "This is not a conversion to a SCEVable type!");
845 Ty = getEffectiveSCEVType(Ty);
847 // Fold if the operand is constant.
848 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
849 const Type *IntTy = getEffectiveSCEVType(Ty);
850 Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
851 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
852 return getConstant(cast<ConstantInt>(C));
855 // zext(zext(x)) --> zext(x)
856 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
857 return getZeroExtendExpr(SZ->getOperand(), Ty);
859 // Before doing any expensive analysis, check to see if we've already
860 // computed a SCEV for this Op and Ty.
862 ID.AddInteger(scZeroExtend);
866 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
868 // If the input value is a chrec scev, and we can prove that the value
869 // did not overflow the old, smaller, value, we can zero extend all of the
870 // operands (often constants). This allows analysis of something like
871 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
872 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
873 if (AR->isAffine()) {
874 const SCEV *Start = AR->getStart();
875 const SCEV *Step = AR->getStepRecurrence(*this);
876 unsigned BitWidth = getTypeSizeInBits(AR->getType());
877 const Loop *L = AR->getLoop();
879 // If we have special knowledge that this addrec won't overflow,
880 // we don't need to do any further analysis.
881 if (AR->hasNoUnsignedWrap())
882 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
883 getZeroExtendExpr(Step, Ty),
886 // Check whether the backedge-taken count is SCEVCouldNotCompute.
887 // Note that this serves two purposes: It filters out loops that are
888 // simply not analyzable, and it covers the case where this code is
889 // being called from within backedge-taken count analysis, such that
890 // attempting to ask for the backedge-taken count would likely result
891 // in infinite recursion. In the later case, the analysis code will
892 // cope with a conservative value, and it will take care to purge
893 // that value once it has finished.
894 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
895 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
896 // Manually compute the final value for AR, checking for
899 // Check whether the backedge-taken count can be losslessly casted to
900 // the addrec's type. The count is always unsigned.
901 const SCEV *CastedMaxBECount =
902 getTruncateOrZeroExtend(MaxBECount, Start->getType());
903 const SCEV *RecastedMaxBECount =
904 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
905 if (MaxBECount == RecastedMaxBECount) {
906 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
907 // Check whether Start+Step*MaxBECount has no unsigned overflow.
909 getMulExpr(CastedMaxBECount,
910 getTruncateOrZeroExtend(Step, Start->getType()));
911 const SCEV *Add = getAddExpr(Start, ZMul);
912 const SCEV *OperandExtendedAdd =
913 getAddExpr(getZeroExtendExpr(Start, WideTy),
914 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
915 getZeroExtendExpr(Step, WideTy)));
916 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
917 // Return the expression with the addrec on the outside.
918 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
919 getZeroExtendExpr(Step, Ty),
922 // Similar to above, only this time treat the step value as signed.
923 // This covers loops that count down.
925 getMulExpr(CastedMaxBECount,
926 getTruncateOrSignExtend(Step, Start->getType()));
927 Add = getAddExpr(Start, SMul);
929 getAddExpr(getZeroExtendExpr(Start, WideTy),
930 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
931 getSignExtendExpr(Step, WideTy)));
932 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
933 // Return the expression with the addrec on the outside.
934 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
935 getSignExtendExpr(Step, Ty),
939 // If the backedge is guarded by a comparison with the pre-inc value
940 // the addrec is safe. Also, if the entry is guarded by a comparison
941 // with the start value and the backedge is guarded by a comparison
942 // with the post-inc value, the addrec is safe.
943 if (isKnownPositive(Step)) {
944 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
945 getUnsignedRange(Step).getUnsignedMax());
946 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
947 (isLoopGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
948 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
949 AR->getPostIncExpr(*this), N)))
950 // Return the expression with the addrec on the outside.
951 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
952 getZeroExtendExpr(Step, Ty),
954 } else if (isKnownNegative(Step)) {
955 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
956 getSignedRange(Step).getSignedMin());
957 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) &&
958 (isLoopGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) ||
959 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
960 AR->getPostIncExpr(*this), N)))
961 // Return the expression with the addrec on the outside.
962 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
963 getSignExtendExpr(Step, Ty),
969 // The cast wasn't folded; create an explicit cast node.
970 // Recompute the insert position, as it may have been invalidated.
971 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
972 SCEV *S = SCEVAllocator.Allocate<SCEVZeroExtendExpr>();
973 new (S) SCEVZeroExtendExpr(ID, Op, Ty);
974 UniqueSCEVs.InsertNode(S, IP);
978 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
980 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
981 "This is not an extending conversion!");
982 assert(isSCEVable(Ty) &&
983 "This is not a conversion to a SCEVable type!");
984 Ty = getEffectiveSCEVType(Ty);
986 // Fold if the operand is constant.
987 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
988 const Type *IntTy = getEffectiveSCEVType(Ty);
989 Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
990 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
991 return getConstant(cast<ConstantInt>(C));
994 // sext(sext(x)) --> sext(x)
995 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
996 return getSignExtendExpr(SS->getOperand(), Ty);
998 // Before doing any expensive analysis, check to see if we've already
999 // computed a SCEV for this Op and Ty.
1000 FoldingSetNodeID ID;
1001 ID.AddInteger(scSignExtend);
1005 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1007 // If the input value is a chrec scev, and we can prove that the value
1008 // did not overflow the old, smaller, value, we can sign extend all of the
1009 // operands (often constants). This allows analysis of something like
1010 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1011 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1012 if (AR->isAffine()) {
1013 const SCEV *Start = AR->getStart();
1014 const SCEV *Step = AR->getStepRecurrence(*this);
1015 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1016 const Loop *L = AR->getLoop();
1018 // If we have special knowledge that this addrec won't overflow,
1019 // we don't need to do any further analysis.
1020 if (AR->hasNoSignedWrap())
1021 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1022 getSignExtendExpr(Step, Ty),
1025 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1026 // Note that this serves two purposes: It filters out loops that are
1027 // simply not analyzable, and it covers the case where this code is
1028 // being called from within backedge-taken count analysis, such that
1029 // attempting to ask for the backedge-taken count would likely result
1030 // in infinite recursion. In the later case, the analysis code will
1031 // cope with a conservative value, and it will take care to purge
1032 // that value once it has finished.
1033 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1034 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1035 // Manually compute the final value for AR, checking for
1038 // Check whether the backedge-taken count can be losslessly casted to
1039 // the addrec's type. The count is always unsigned.
1040 const SCEV *CastedMaxBECount =
1041 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1042 const SCEV *RecastedMaxBECount =
1043 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1044 if (MaxBECount == RecastedMaxBECount) {
1045 const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1046 // Check whether Start+Step*MaxBECount has no signed overflow.
1048 getMulExpr(CastedMaxBECount,
1049 getTruncateOrSignExtend(Step, Start->getType()));
1050 const SCEV *Add = getAddExpr(Start, SMul);
1051 const SCEV *OperandExtendedAdd =
1052 getAddExpr(getSignExtendExpr(Start, WideTy),
1053 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1054 getSignExtendExpr(Step, WideTy)));
1055 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1056 // Return the expression with the addrec on the outside.
1057 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1058 getSignExtendExpr(Step, Ty),
1061 // Similar to above, only this time treat the step value as unsigned.
1062 // This covers loops that count up with an unsigned step.
1064 getMulExpr(CastedMaxBECount,
1065 getTruncateOrZeroExtend(Step, Start->getType()));
1066 Add = getAddExpr(Start, UMul);
1067 OperandExtendedAdd =
1068 getAddExpr(getSignExtendExpr(Start, WideTy),
1069 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1070 getZeroExtendExpr(Step, WideTy)));
1071 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
1072 // Return the expression with the addrec on the outside.
1073 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1074 getZeroExtendExpr(Step, Ty),
1078 // If the backedge is guarded by a comparison with the pre-inc value
1079 // the addrec is safe. Also, if the entry is guarded by a comparison
1080 // with the start value and the backedge is guarded by a comparison
1081 // with the post-inc value, the addrec is safe.
1082 if (isKnownPositive(Step)) {
1083 const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
1084 getSignedRange(Step).getSignedMax());
1085 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
1086 (isLoopGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
1087 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
1088 AR->getPostIncExpr(*this), N)))
1089 // Return the expression with the addrec on the outside.
1090 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1091 getSignExtendExpr(Step, Ty),
1093 } else if (isKnownNegative(Step)) {
1094 const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
1095 getSignedRange(Step).getSignedMin());
1096 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
1097 (isLoopGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
1098 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
1099 AR->getPostIncExpr(*this), N)))
1100 // Return the expression with the addrec on the outside.
1101 return getAddRecExpr(getSignExtendExpr(Start, Ty),
1102 getSignExtendExpr(Step, Ty),
1108 // The cast wasn't folded; create an explicit cast node.
1109 // Recompute the insert position, as it may have been invalidated.
1110 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1111 SCEV *S = SCEVAllocator.Allocate<SCEVSignExtendExpr>();
1112 new (S) SCEVSignExtendExpr(ID, Op, Ty);
1113 UniqueSCEVs.InsertNode(S, IP);
1117 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1118 /// unspecified bits out to the given type.
1120 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1122 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1123 "This is not an extending conversion!");
1124 assert(isSCEVable(Ty) &&
1125 "This is not a conversion to a SCEVable type!");
1126 Ty = getEffectiveSCEVType(Ty);
1128 // Sign-extend negative constants.
1129 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1130 if (SC->getValue()->getValue().isNegative())
1131 return getSignExtendExpr(Op, Ty);
1133 // Peel off a truncate cast.
1134 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1135 const SCEV *NewOp = T->getOperand();
1136 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1137 return getAnyExtendExpr(NewOp, Ty);
1138 return getTruncateOrNoop(NewOp, Ty);
1141 // Next try a zext cast. If the cast is folded, use it.
1142 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1143 if (!isa<SCEVZeroExtendExpr>(ZExt))
1146 // Next try a sext cast. If the cast is folded, use it.
1147 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1148 if (!isa<SCEVSignExtendExpr>(SExt))
1151 // Force the cast to be folded into the operands of an addrec.
1152 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1153 SmallVector<const SCEV *, 4> Ops;
1154 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1156 Ops.push_back(getAnyExtendExpr(*I, Ty));
1157 return getAddRecExpr(Ops, AR->getLoop());
1160 // If the expression is obviously signed, use the sext cast value.
1161 if (isa<SCEVSMaxExpr>(Op))
1164 // Absent any other information, use the zext cast value.
1168 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1169 /// a list of operands to be added under the given scale, update the given
1170 /// map. This is a helper function for getAddRecExpr. As an example of
1171 /// what it does, given a sequence of operands that would form an add
1172 /// expression like this:
1174 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1176 /// where A and B are constants, update the map with these values:
1178 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1180 /// and add 13 + A*B*29 to AccumulatedConstant.
1181 /// This will allow getAddRecExpr to produce this:
1183 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1185 /// This form often exposes folding opportunities that are hidden in
1186 /// the original operand list.
1188 /// Return true iff it appears that any interesting folding opportunities
1189 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1190 /// the common case where no interesting opportunities are present, and
1191 /// is also used as a check to avoid infinite recursion.
1194 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1195 SmallVector<const SCEV *, 8> &NewOps,
1196 APInt &AccumulatedConstant,
1197 const SmallVectorImpl<const SCEV *> &Ops,
1199 ScalarEvolution &SE) {
1200 bool Interesting = false;
1202 // Iterate over the add operands.
1203 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1204 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1205 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1207 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1208 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1209 // A multiplication of a constant with another add; recurse.
1211 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1212 cast<SCEVAddExpr>(Mul->getOperand(1))
1216 // A multiplication of a constant with some other value. Update
1218 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1219 const SCEV *Key = SE.getMulExpr(MulOps);
1220 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1221 M.insert(std::make_pair(Key, NewScale));
1223 NewOps.push_back(Pair.first->first);
1225 Pair.first->second += NewScale;
1226 // The map already had an entry for this value, which may indicate
1227 // a folding opportunity.
1231 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1232 // Pull a buried constant out to the outside.
1233 if (Scale != 1 || AccumulatedConstant != 0 || C->isZero())
1235 AccumulatedConstant += Scale * C->getValue()->getValue();
1237 // An ordinary operand. Update the map.
1238 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1239 M.insert(std::make_pair(Ops[i], Scale));
1241 NewOps.push_back(Pair.first->first);
1243 Pair.first->second += Scale;
1244 // The map already had an entry for this value, which may indicate
1245 // a folding opportunity.
1255 struct APIntCompare {
1256 bool operator()(const APInt &LHS, const APInt &RHS) const {
1257 return LHS.ult(RHS);
1262 /// getAddExpr - Get a canonical add expression, or something simpler if
1264 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1265 bool HasNUW, bool HasNSW) {
1266 assert(!Ops.empty() && "Cannot get empty add!");
1267 if (Ops.size() == 1) return Ops[0];
1269 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1270 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1271 getEffectiveSCEVType(Ops[0]->getType()) &&
1272 "SCEVAddExpr operand types don't match!");
1275 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1276 if (!HasNUW && HasNSW) {
1278 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1279 if (!isKnownNonNegative(Ops[i])) {
1283 if (All) HasNUW = true;
1286 // Sort by complexity, this groups all similar expression types together.
1287 GroupByComplexity(Ops, LI);
1289 // If there are any constants, fold them together.
1291 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1293 assert(Idx < Ops.size());
1294 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1295 // We found two constants, fold them together!
1296 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1297 RHSC->getValue()->getValue());
1298 if (Ops.size() == 2) return Ops[0];
1299 Ops.erase(Ops.begin()+1); // Erase the folded element
1300 LHSC = cast<SCEVConstant>(Ops[0]);
1303 // If we are left with a constant zero being added, strip it off.
1304 if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1305 Ops.erase(Ops.begin());
1310 if (Ops.size() == 1) return Ops[0];
1312 // Okay, check to see if the same value occurs in the operand list twice. If
1313 // so, merge them together into an multiply expression. Since we sorted the
1314 // list, these values are required to be adjacent.
1315 const Type *Ty = Ops[0]->getType();
1316 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1317 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1318 // Found a match, merge the two values into a multiply, and add any
1319 // remaining values to the result.
1320 const SCEV *Two = getIntegerSCEV(2, Ty);
1321 const SCEV *Mul = getMulExpr(Ops[i], Two);
1322 if (Ops.size() == 2)
1324 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
1326 return getAddExpr(Ops, HasNUW, HasNSW);
1329 // Check for truncates. If all the operands are truncated from the same
1330 // type, see if factoring out the truncate would permit the result to be
1331 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1332 // if the contents of the resulting outer trunc fold to something simple.
1333 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1334 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1335 const Type *DstType = Trunc->getType();
1336 const Type *SrcType = Trunc->getOperand()->getType();
1337 SmallVector<const SCEV *, 8> LargeOps;
1339 // Check all the operands to see if they can be represented in the
1340 // source type of the truncate.
1341 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1342 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1343 if (T->getOperand()->getType() != SrcType) {
1347 LargeOps.push_back(T->getOperand());
1348 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1349 // This could be either sign or zero extension, but sign extension
1350 // is much more likely to be foldable here.
1351 LargeOps.push_back(getSignExtendExpr(C, SrcType));
1352 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1353 SmallVector<const SCEV *, 8> LargeMulOps;
1354 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1355 if (const SCEVTruncateExpr *T =
1356 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1357 if (T->getOperand()->getType() != SrcType) {
1361 LargeMulOps.push_back(T->getOperand());
1362 } else if (const SCEVConstant *C =
1363 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1364 // This could be either sign or zero extension, but sign extension
1365 // is much more likely to be foldable here.
1366 LargeMulOps.push_back(getSignExtendExpr(C, SrcType));
1373 LargeOps.push_back(getMulExpr(LargeMulOps));
1380 // Evaluate the expression in the larger type.
1381 const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
1382 // If it folds to something simple, use it. Otherwise, don't.
1383 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1384 return getTruncateExpr(Fold, DstType);
1388 // Skip past any other cast SCEVs.
1389 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1392 // If there are add operands they would be next.
1393 if (Idx < Ops.size()) {
1394 bool DeletedAdd = false;
1395 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1396 // If we have an add, expand the add operands onto the end of the operands
1398 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
1399 Ops.erase(Ops.begin()+Idx);
1403 // If we deleted at least one add, we added operands to the end of the list,
1404 // and they are not necessarily sorted. Recurse to resort and resimplify
1405 // any operands we just aquired.
1407 return getAddExpr(Ops);
1410 // Skip over the add expression until we get to a multiply.
1411 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1414 // Check to see if there are any folding opportunities present with
1415 // operands multiplied by constant values.
1416 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1417 uint64_t BitWidth = getTypeSizeInBits(Ty);
1418 DenseMap<const SCEV *, APInt> M;
1419 SmallVector<const SCEV *, 8> NewOps;
1420 APInt AccumulatedConstant(BitWidth, 0);
1421 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1422 Ops, APInt(BitWidth, 1), *this)) {
1423 // Some interesting folding opportunity is present, so its worthwhile to
1424 // re-generate the operands list. Group the operands by constant scale,
1425 // to avoid multiplying by the same constant scale multiple times.
1426 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1427 for (SmallVector<const SCEV *, 8>::iterator I = NewOps.begin(),
1428 E = NewOps.end(); I != E; ++I)
1429 MulOpLists[M.find(*I)->second].push_back(*I);
1430 // Re-generate the operands list.
1432 if (AccumulatedConstant != 0)
1433 Ops.push_back(getConstant(AccumulatedConstant));
1434 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1435 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1437 Ops.push_back(getMulExpr(getConstant(I->first),
1438 getAddExpr(I->second)));
1440 return getIntegerSCEV(0, Ty);
1441 if (Ops.size() == 1)
1443 return getAddExpr(Ops);
1447 // If we are adding something to a multiply expression, make sure the
1448 // something is not already an operand of the multiply. If so, merge it into
1450 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1451 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1452 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1453 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1454 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1455 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(Ops[AddOp])) {
1456 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1457 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1458 if (Mul->getNumOperands() != 2) {
1459 // If the multiply has more than two operands, we must get the
1461 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(), Mul->op_end());
1462 MulOps.erase(MulOps.begin()+MulOp);
1463 InnerMul = getMulExpr(MulOps);
1465 const SCEV *One = getIntegerSCEV(1, Ty);
1466 const SCEV *AddOne = getAddExpr(InnerMul, One);
1467 const SCEV *OuterMul = getMulExpr(AddOne, Ops[AddOp]);
1468 if (Ops.size() == 2) return OuterMul;
1470 Ops.erase(Ops.begin()+AddOp);
1471 Ops.erase(Ops.begin()+Idx-1);
1473 Ops.erase(Ops.begin()+Idx);
1474 Ops.erase(Ops.begin()+AddOp-1);
1476 Ops.push_back(OuterMul);
1477 return getAddExpr(Ops);
1480 // Check this multiply against other multiplies being added together.
1481 for (unsigned OtherMulIdx = Idx+1;
1482 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1484 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1485 // If MulOp occurs in OtherMul, we can fold the two multiplies
1487 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1488 OMulOp != e; ++OMulOp)
1489 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1490 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1491 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1492 if (Mul->getNumOperands() != 2) {
1493 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1495 MulOps.erase(MulOps.begin()+MulOp);
1496 InnerMul1 = getMulExpr(MulOps);
1498 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1499 if (OtherMul->getNumOperands() != 2) {
1500 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1501 OtherMul->op_end());
1502 MulOps.erase(MulOps.begin()+OMulOp);
1503 InnerMul2 = getMulExpr(MulOps);
1505 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1506 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1507 if (Ops.size() == 2) return OuterMul;
1508 Ops.erase(Ops.begin()+Idx);
1509 Ops.erase(Ops.begin()+OtherMulIdx-1);
1510 Ops.push_back(OuterMul);
1511 return getAddExpr(Ops);
1517 // If there are any add recurrences in the operands list, see if any other
1518 // added values are loop invariant. If so, we can fold them into the
1520 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1523 // Scan over all recurrences, trying to fold loop invariants into them.
1524 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1525 // Scan all of the other operands to this add and add them to the vector if
1526 // they are loop invariant w.r.t. the recurrence.
1527 SmallVector<const SCEV *, 8> LIOps;
1528 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1529 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1530 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1531 LIOps.push_back(Ops[i]);
1532 Ops.erase(Ops.begin()+i);
1536 // If we found some loop invariants, fold them into the recurrence.
1537 if (!LIOps.empty()) {
1538 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1539 LIOps.push_back(AddRec->getStart());
1541 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1543 AddRecOps[0] = getAddExpr(LIOps);
1545 // It's tempting to propagate NUW/NSW flags here, but nuw/nsw addition
1546 // is not associative so this isn't necessarily safe.
1547 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
1549 // If all of the other operands were loop invariant, we are done.
1550 if (Ops.size() == 1) return NewRec;
1552 // Otherwise, add the folded AddRec by the non-liv parts.
1553 for (unsigned i = 0;; ++i)
1554 if (Ops[i] == AddRec) {
1558 return getAddExpr(Ops);
1561 // Okay, if there weren't any loop invariants to be folded, check to see if
1562 // there are multiple AddRec's with the same loop induction variable being
1563 // added together. If so, we can fold them.
1564 for (unsigned OtherIdx = Idx+1;
1565 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1566 if (OtherIdx != Idx) {
1567 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1568 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1569 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
1570 SmallVector<const SCEV *, 4> NewOps(AddRec->op_begin(),
1572 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
1573 if (i >= NewOps.size()) {
1574 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
1575 OtherAddRec->op_end());
1578 NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
1580 const SCEV *NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
1582 if (Ops.size() == 2) return NewAddRec;
1584 Ops.erase(Ops.begin()+Idx);
1585 Ops.erase(Ops.begin()+OtherIdx-1);
1586 Ops.push_back(NewAddRec);
1587 return getAddExpr(Ops);
1591 // Otherwise couldn't fold anything into this recurrence. Move onto the
1595 // Okay, it looks like we really DO need an add expr. Check to see if we
1596 // already have one, otherwise create a new one.
1597 FoldingSetNodeID ID;
1598 ID.AddInteger(scAddExpr);
1599 ID.AddInteger(Ops.size());
1600 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1601 ID.AddPointer(Ops[i]);
1604 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1606 S = SCEVAllocator.Allocate<SCEVAddExpr>();
1607 new (S) SCEVAddExpr(ID, Ops);
1608 UniqueSCEVs.InsertNode(S, IP);
1610 if (HasNUW) S->setHasNoUnsignedWrap(true);
1611 if (HasNSW) S->setHasNoSignedWrap(true);
1615 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1617 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1618 bool HasNUW, bool HasNSW) {
1619 assert(!Ops.empty() && "Cannot get empty mul!");
1620 if (Ops.size() == 1) return Ops[0];
1622 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1623 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1624 getEffectiveSCEVType(Ops[0]->getType()) &&
1625 "SCEVMulExpr operand types don't match!");
1628 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1629 if (!HasNUW && HasNSW) {
1631 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1632 if (!isKnownNonNegative(Ops[i])) {
1636 if (All) HasNUW = true;
1639 // Sort by complexity, this groups all similar expression types together.
1640 GroupByComplexity(Ops, LI);
1642 // If there are any constants, fold them together.
1644 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1646 // C1*(C2+V) -> C1*C2 + C1*V
1647 if (Ops.size() == 2)
1648 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1649 if (Add->getNumOperands() == 2 &&
1650 isa<SCEVConstant>(Add->getOperand(0)))
1651 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1652 getMulExpr(LHSC, Add->getOperand(1)));
1655 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1656 // We found two constants, fold them together!
1657 ConstantInt *Fold = ConstantInt::get(getContext(),
1658 LHSC->getValue()->getValue() *
1659 RHSC->getValue()->getValue());
1660 Ops[0] = getConstant(Fold);
1661 Ops.erase(Ops.begin()+1); // Erase the folded element
1662 if (Ops.size() == 1) return Ops[0];
1663 LHSC = cast<SCEVConstant>(Ops[0]);
1666 // If we are left with a constant one being multiplied, strip it off.
1667 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1668 Ops.erase(Ops.begin());
1670 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1671 // If we have a multiply of zero, it will always be zero.
1673 } else if (Ops[0]->isAllOnesValue()) {
1674 // If we have a mul by -1 of an add, try distributing the -1 among the
1676 if (Ops.size() == 2)
1677 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1678 SmallVector<const SCEV *, 4> NewOps;
1679 bool AnyFolded = false;
1680 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1682 const SCEV *Mul = getMulExpr(Ops[0], *I);
1683 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1684 NewOps.push_back(Mul);
1687 return getAddExpr(NewOps);
1692 // Skip over the add expression until we get to a multiply.
1693 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1696 if (Ops.size() == 1)
1699 // If there are mul operands inline them all into this expression.
1700 if (Idx < Ops.size()) {
1701 bool DeletedMul = false;
1702 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1703 // If we have an mul, expand the mul operands onto the end of the operands
1705 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1706 Ops.erase(Ops.begin()+Idx);
1710 // If we deleted at least one mul, we added operands to the end of the list,
1711 // and they are not necessarily sorted. Recurse to resort and resimplify
1712 // any operands we just aquired.
1714 return getMulExpr(Ops);
1717 // If there are any add recurrences in the operands list, see if any other
1718 // added values are loop invariant. If so, we can fold them into the
1720 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1723 // Scan over all recurrences, trying to fold loop invariants into them.
1724 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1725 // Scan all of the other operands to this mul and add them to the vector if
1726 // they are loop invariant w.r.t. the recurrence.
1727 SmallVector<const SCEV *, 8> LIOps;
1728 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1729 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1730 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1731 LIOps.push_back(Ops[i]);
1732 Ops.erase(Ops.begin()+i);
1736 // If we found some loop invariants, fold them into the recurrence.
1737 if (!LIOps.empty()) {
1738 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1739 SmallVector<const SCEV *, 4> NewOps;
1740 NewOps.reserve(AddRec->getNumOperands());
1741 if (LIOps.size() == 1) {
1742 const SCEV *Scale = LIOps[0];
1743 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1744 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1746 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1747 SmallVector<const SCEV *, 4> MulOps(LIOps.begin(), LIOps.end());
1748 MulOps.push_back(AddRec->getOperand(i));
1749 NewOps.push_back(getMulExpr(MulOps));
1753 // It's tempting to propagate the NSW flag here, but nsw multiplication
1754 // is not associative so this isn't necessarily safe.
1755 const SCEV *NewRec = getAddRecExpr(NewOps, AddRec->getLoop(),
1756 HasNUW && AddRec->hasNoUnsignedWrap(),
1759 // If all of the other operands were loop invariant, we are done.
1760 if (Ops.size() == 1) return NewRec;
1762 // Otherwise, multiply the folded AddRec by the non-liv parts.
1763 for (unsigned i = 0;; ++i)
1764 if (Ops[i] == AddRec) {
1768 return getMulExpr(Ops);
1771 // Okay, if there weren't any loop invariants to be folded, check to see if
1772 // there are multiple AddRec's with the same loop induction variable being
1773 // multiplied together. If so, we can fold them.
1774 for (unsigned OtherIdx = Idx+1;
1775 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1776 if (OtherIdx != Idx) {
1777 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1778 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1779 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1780 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1781 const SCEV *NewStart = getMulExpr(F->getStart(),
1783 const SCEV *B = F->getStepRecurrence(*this);
1784 const SCEV *D = G->getStepRecurrence(*this);
1785 const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
1788 const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
1790 if (Ops.size() == 2) return NewAddRec;
1792 Ops.erase(Ops.begin()+Idx);
1793 Ops.erase(Ops.begin()+OtherIdx-1);
1794 Ops.push_back(NewAddRec);
1795 return getMulExpr(Ops);
1799 // Otherwise couldn't fold anything into this recurrence. Move onto the
1803 // Okay, it looks like we really DO need an mul expr. Check to see if we
1804 // already have one, otherwise create a new one.
1805 FoldingSetNodeID ID;
1806 ID.AddInteger(scMulExpr);
1807 ID.AddInteger(Ops.size());
1808 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1809 ID.AddPointer(Ops[i]);
1812 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1814 S = SCEVAllocator.Allocate<SCEVMulExpr>();
1815 new (S) SCEVMulExpr(ID, Ops);
1816 UniqueSCEVs.InsertNode(S, IP);
1818 if (HasNUW) S->setHasNoUnsignedWrap(true);
1819 if (HasNSW) S->setHasNoSignedWrap(true);
1823 /// getUDivExpr - Get a canonical unsigned division expression, or something
1824 /// simpler if possible.
1825 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
1827 assert(getEffectiveSCEVType(LHS->getType()) ==
1828 getEffectiveSCEVType(RHS->getType()) &&
1829 "SCEVUDivExpr operand types don't match!");
1831 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1832 if (RHSC->getValue()->equalsInt(1))
1833 return LHS; // X udiv 1 --> x
1835 return getIntegerSCEV(0, LHS->getType()); // value is undefined
1837 // Determine if the division can be folded into the operands of
1839 // TODO: Generalize this to non-constants by using known-bits information.
1840 const Type *Ty = LHS->getType();
1841 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1842 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ;
1843 // For non-power-of-two values, effectively round the value up to the
1844 // nearest power of two.
1845 if (!RHSC->getValue()->getValue().isPowerOf2())
1847 const IntegerType *ExtTy =
1848 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
1849 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1850 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1851 if (const SCEVConstant *Step =
1852 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1853 if (!Step->getValue()->getValue()
1854 .urem(RHSC->getValue()->getValue()) &&
1855 getZeroExtendExpr(AR, ExtTy) ==
1856 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1857 getZeroExtendExpr(Step, ExtTy),
1859 SmallVector<const SCEV *, 4> Operands;
1860 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1861 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1862 return getAddRecExpr(Operands, AR->getLoop());
1864 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1865 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1866 SmallVector<const SCEV *, 4> Operands;
1867 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1868 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1869 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1870 // Find an operand that's safely divisible.
1871 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1872 const SCEV *Op = M->getOperand(i);
1873 const SCEV *Div = getUDivExpr(Op, RHSC);
1874 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1875 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
1876 Operands = SmallVector<const SCEV *, 4>(MOperands.begin(),
1879 return getMulExpr(Operands);
1883 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1884 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1885 SmallVector<const SCEV *, 4> Operands;
1886 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1887 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1888 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1890 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1891 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
1892 if (isa<SCEVUDivExpr>(Op) || getMulExpr(Op, RHS) != A->getOperand(i))
1894 Operands.push_back(Op);
1896 if (Operands.size() == A->getNumOperands())
1897 return getAddExpr(Operands);
1901 // Fold if both operands are constant.
1902 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1903 Constant *LHSCV = LHSC->getValue();
1904 Constant *RHSCV = RHSC->getValue();
1905 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
1910 FoldingSetNodeID ID;
1911 ID.AddInteger(scUDivExpr);
1915 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1916 SCEV *S = SCEVAllocator.Allocate<SCEVUDivExpr>();
1917 new (S) SCEVUDivExpr(ID, LHS, RHS);
1918 UniqueSCEVs.InsertNode(S, IP);
1923 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1924 /// Simplify the expression as much as possible.
1925 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
1926 const SCEV *Step, const Loop *L,
1927 bool HasNUW, bool HasNSW) {
1928 SmallVector<const SCEV *, 4> Operands;
1929 Operands.push_back(Start);
1930 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1931 if (StepChrec->getLoop() == L) {
1932 Operands.insert(Operands.end(), StepChrec->op_begin(),
1933 StepChrec->op_end());
1934 return getAddRecExpr(Operands, L);
1937 Operands.push_back(Step);
1938 return getAddRecExpr(Operands, L, HasNUW, HasNSW);
1941 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1942 /// Simplify the expression as much as possible.
1944 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
1946 bool HasNUW, bool HasNSW) {
1947 if (Operands.size() == 1) return Operands[0];
1949 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
1950 assert(getEffectiveSCEVType(Operands[i]->getType()) ==
1951 getEffectiveSCEVType(Operands[0]->getType()) &&
1952 "SCEVAddRecExpr operand types don't match!");
1955 if (Operands.back()->isZero()) {
1956 Operands.pop_back();
1957 return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
1960 // If HasNSW is true and all the operands are non-negative, infer HasNUW.
1961 if (!HasNUW && HasNSW) {
1963 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1964 if (!isKnownNonNegative(Operands[i])) {
1968 if (All) HasNUW = true;
1971 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
1972 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
1973 const Loop *NestedLoop = NestedAR->getLoop();
1974 if (L->contains(NestedLoop->getHeader()) ?
1975 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
1976 (!NestedLoop->contains(L->getHeader()) &&
1977 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
1978 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
1979 NestedAR->op_end());
1980 Operands[0] = NestedAR->getStart();
1981 // AddRecs require their operands be loop-invariant with respect to their
1982 // loops. Don't perform this transformation if it would break this
1984 bool AllInvariant = true;
1985 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
1986 if (!Operands[i]->isLoopInvariant(L)) {
1987 AllInvariant = false;
1991 NestedOperands[0] = getAddRecExpr(Operands, L);
1992 AllInvariant = true;
1993 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
1994 if (!NestedOperands[i]->isLoopInvariant(NestedLoop)) {
1995 AllInvariant = false;
1999 // Ok, both add recurrences are valid after the transformation.
2000 return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
2002 // Reset Operands to its original state.
2003 Operands[0] = NestedAR;
2007 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2008 // already have one, otherwise create a new one.
2009 FoldingSetNodeID ID;
2010 ID.AddInteger(scAddRecExpr);
2011 ID.AddInteger(Operands.size());
2012 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2013 ID.AddPointer(Operands[i]);
2017 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2019 S = SCEVAllocator.Allocate<SCEVAddRecExpr>();
2020 new (S) SCEVAddRecExpr(ID, Operands, L);
2021 UniqueSCEVs.InsertNode(S, IP);
2023 if (HasNUW) S->setHasNoUnsignedWrap(true);
2024 if (HasNSW) S->setHasNoSignedWrap(true);
2028 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2030 SmallVector<const SCEV *, 2> Ops;
2033 return getSMaxExpr(Ops);
2037 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2038 assert(!Ops.empty() && "Cannot get empty smax!");
2039 if (Ops.size() == 1) return Ops[0];
2041 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2042 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
2043 getEffectiveSCEVType(Ops[0]->getType()) &&
2044 "SCEVSMaxExpr operand types don't match!");
2047 // Sort by complexity, this groups all similar expression types together.
2048 GroupByComplexity(Ops, LI);
2050 // If there are any constants, fold them together.
2052 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2054 assert(Idx < Ops.size());
2055 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2056 // We found two constants, fold them together!
2057 ConstantInt *Fold = ConstantInt::get(getContext(),
2058 APIntOps::smax(LHSC->getValue()->getValue(),
2059 RHSC->getValue()->getValue()));
2060 Ops[0] = getConstant(Fold);
2061 Ops.erase(Ops.begin()+1); // Erase the folded element
2062 if (Ops.size() == 1) return Ops[0];
2063 LHSC = cast<SCEVConstant>(Ops[0]);
2066 // If we are left with a constant minimum-int, strip it off.
2067 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2068 Ops.erase(Ops.begin());
2070 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2071 // If we have an smax with a constant maximum-int, it will always be
2077 if (Ops.size() == 1) return Ops[0];
2079 // Find the first SMax
2080 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2083 // Check to see if one of the operands is an SMax. If so, expand its operands
2084 // onto our operand list, and recurse to simplify.
2085 if (Idx < Ops.size()) {
2086 bool DeletedSMax = false;
2087 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2088 Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
2089 Ops.erase(Ops.begin()+Idx);
2094 return getSMaxExpr(Ops);
2097 // Okay, check to see if the same value occurs in the operand list twice. If
2098 // so, delete one. Since we sorted the list, these values are required to
2100 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2101 if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
2102 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2106 if (Ops.size() == 1) return Ops[0];
2108 assert(!Ops.empty() && "Reduced smax down to nothing!");
2110 // Okay, it looks like we really DO need an smax expr. Check to see if we
2111 // already have one, otherwise create a new one.
2112 FoldingSetNodeID ID;
2113 ID.AddInteger(scSMaxExpr);
2114 ID.AddInteger(Ops.size());
2115 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2116 ID.AddPointer(Ops[i]);
2118 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2119 SCEV *S = SCEVAllocator.Allocate<SCEVSMaxExpr>();
2120 new (S) SCEVSMaxExpr(ID, Ops);
2121 UniqueSCEVs.InsertNode(S, IP);
2125 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2127 SmallVector<const SCEV *, 2> Ops;
2130 return getUMaxExpr(Ops);
2134 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2135 assert(!Ops.empty() && "Cannot get empty umax!");
2136 if (Ops.size() == 1) return Ops[0];
2138 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2139 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
2140 getEffectiveSCEVType(Ops[0]->getType()) &&
2141 "SCEVUMaxExpr operand types don't match!");
2144 // Sort by complexity, this groups all similar expression types together.
2145 GroupByComplexity(Ops, LI);
2147 // If there are any constants, fold them together.
2149 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2151 assert(Idx < Ops.size());
2152 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2153 // We found two constants, fold them together!
2154 ConstantInt *Fold = ConstantInt::get(getContext(),
2155 APIntOps::umax(LHSC->getValue()->getValue(),
2156 RHSC->getValue()->getValue()));
2157 Ops[0] = getConstant(Fold);
2158 Ops.erase(Ops.begin()+1); // Erase the folded element
2159 if (Ops.size() == 1) return Ops[0];
2160 LHSC = cast<SCEVConstant>(Ops[0]);
2163 // If we are left with a constant minimum-int, strip it off.
2164 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2165 Ops.erase(Ops.begin());
2167 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2168 // If we have an umax with a constant maximum-int, it will always be
2174 if (Ops.size() == 1) return Ops[0];
2176 // Find the first UMax
2177 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2180 // Check to see if one of the operands is a UMax. If so, expand its operands
2181 // onto our operand list, and recurse to simplify.
2182 if (Idx < Ops.size()) {
2183 bool DeletedUMax = false;
2184 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2185 Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
2186 Ops.erase(Ops.begin()+Idx);
2191 return getUMaxExpr(Ops);
2194 // Okay, check to see if the same value occurs in the operand list twice. If
2195 // so, delete one. Since we sorted the list, these values are required to
2197 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2198 if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
2199 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2203 if (Ops.size() == 1) return Ops[0];
2205 assert(!Ops.empty() && "Reduced umax down to nothing!");
2207 // Okay, it looks like we really DO need a umax expr. Check to see if we
2208 // already have one, otherwise create a new one.
2209 FoldingSetNodeID ID;
2210 ID.AddInteger(scUMaxExpr);
2211 ID.AddInteger(Ops.size());
2212 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2213 ID.AddPointer(Ops[i]);
2215 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2216 SCEV *S = SCEVAllocator.Allocate<SCEVUMaxExpr>();
2217 new (S) SCEVUMaxExpr(ID, Ops);
2218 UniqueSCEVs.InsertNode(S, IP);
2222 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2224 // ~smax(~x, ~y) == smin(x, y).
2225 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2228 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2230 // ~umax(~x, ~y) == umin(x, y)
2231 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2234 const SCEV *ScalarEvolution::getFieldOffsetExpr(const StructType *STy,
2236 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2237 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2238 C = ConstantFoldConstantExpression(CE, TD);
2239 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2240 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2243 const SCEV *ScalarEvolution::getAllocSizeExpr(const Type *AllocTy) {
2244 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2245 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2246 C = ConstantFoldConstantExpression(CE, TD);
2247 const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2248 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2251 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2252 // Don't attempt to do anything other than create a SCEVUnknown object
2253 // here. createSCEV only calls getUnknown after checking for all other
2254 // interesting possibilities, and any other code that calls getUnknown
2255 // is doing so in order to hide a value from SCEV canonicalization.
2257 FoldingSetNodeID ID;
2258 ID.AddInteger(scUnknown);
2261 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2262 SCEV *S = SCEVAllocator.Allocate<SCEVUnknown>();
2263 new (S) SCEVUnknown(ID, V);
2264 UniqueSCEVs.InsertNode(S, IP);
2268 //===----------------------------------------------------------------------===//
2269 // Basic SCEV Analysis and PHI Idiom Recognition Code
2272 /// isSCEVable - Test if values of the given type are analyzable within
2273 /// the SCEV framework. This primarily includes integer types, and it
2274 /// can optionally include pointer types if the ScalarEvolution class
2275 /// has access to target-specific information.
2276 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
2277 // Integers and pointers are always SCEVable.
2278 return Ty->isInteger() || isa<PointerType>(Ty);
2281 /// getTypeSizeInBits - Return the size in bits of the specified type,
2282 /// for which isSCEVable must return true.
2283 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
2284 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2286 // If we have a TargetData, use it!
2288 return TD->getTypeSizeInBits(Ty);
2290 // Integer types have fixed sizes.
2291 if (Ty->isInteger())
2292 return Ty->getPrimitiveSizeInBits();
2294 // The only other support type is pointer. Without TargetData, conservatively
2295 // assume pointers are 64-bit.
2296 assert(isa<PointerType>(Ty) && "isSCEVable permitted a non-SCEVable type!");
2300 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2301 /// the given type and which represents how SCEV will treat the given
2302 /// type, for which isSCEVable must return true. For pointer types,
2303 /// this is the pointer-sized integer type.
2304 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
2305 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2307 if (Ty->isInteger())
2310 // The only other support type is pointer.
2311 assert(isa<PointerType>(Ty) && "Unexpected non-pointer non-integer type!");
2312 if (TD) return TD->getIntPtrType(getContext());
2314 // Without TargetData, conservatively assume pointers are 64-bit.
2315 return Type::getInt64Ty(getContext());
2318 const SCEV *ScalarEvolution::getCouldNotCompute() {
2319 return &CouldNotCompute;
2322 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2323 /// expression and create a new one.
2324 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2325 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2327 std::map<SCEVCallbackVH, const SCEV *>::iterator I = Scalars.find(V);
2328 if (I != Scalars.end()) return I->second;
2329 const SCEV *S = createSCEV(V);
2330 Scalars.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2334 /// getIntegerSCEV - Given a SCEVable type, create a constant for the
2335 /// specified signed integer value and return a SCEV for the constant.
2336 const SCEV *ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
2337 const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
2338 return getConstant(ConstantInt::get(ITy, Val));
2341 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2343 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2344 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2346 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2348 const Type *Ty = V->getType();
2349 Ty = getEffectiveSCEVType(Ty);
2350 return getMulExpr(V,
2351 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2354 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2355 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2356 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2358 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2360 const Type *Ty = V->getType();
2361 Ty = getEffectiveSCEVType(Ty);
2362 const SCEV *AllOnes =
2363 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2364 return getMinusSCEV(AllOnes, V);
2367 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
2369 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS,
2372 return getAddExpr(LHS, getNegativeSCEV(RHS));
2375 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2376 /// input value to the specified type. If the type must be extended, it is zero
2379 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V,
2381 const Type *SrcTy = V->getType();
2382 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2383 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2384 "Cannot truncate or zero extend with non-integer arguments!");
2385 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2386 return V; // No conversion
2387 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2388 return getTruncateExpr(V, Ty);
2389 return getZeroExtendExpr(V, Ty);
2392 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2393 /// input value to the specified type. If the type must be extended, it is sign
2396 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2398 const Type *SrcTy = V->getType();
2399 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2400 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2401 "Cannot truncate or zero extend with non-integer arguments!");
2402 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2403 return V; // No conversion
2404 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2405 return getTruncateExpr(V, Ty);
2406 return getSignExtendExpr(V, Ty);
2409 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2410 /// input value to the specified type. If the type must be extended, it is zero
2411 /// extended. The conversion must not be narrowing.
2413 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
2414 const Type *SrcTy = V->getType();
2415 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2416 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2417 "Cannot noop or zero extend with non-integer arguments!");
2418 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2419 "getNoopOrZeroExtend cannot truncate!");
2420 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2421 return V; // No conversion
2422 return getZeroExtendExpr(V, Ty);
2425 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2426 /// input value to the specified type. If the type must be extended, it is sign
2427 /// extended. The conversion must not be narrowing.
2429 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
2430 const Type *SrcTy = V->getType();
2431 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2432 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2433 "Cannot noop or sign extend with non-integer arguments!");
2434 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2435 "getNoopOrSignExtend cannot truncate!");
2436 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2437 return V; // No conversion
2438 return getSignExtendExpr(V, Ty);
2441 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2442 /// the input value to the specified type. If the type must be extended,
2443 /// it is extended with unspecified bits. The conversion must not be
2446 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
2447 const Type *SrcTy = V->getType();
2448 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2449 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2450 "Cannot noop or any extend with non-integer arguments!");
2451 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2452 "getNoopOrAnyExtend cannot truncate!");
2453 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2454 return V; // No conversion
2455 return getAnyExtendExpr(V, Ty);
2458 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2459 /// input value to the specified type. The conversion must not be widening.
2461 ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
2462 const Type *SrcTy = V->getType();
2463 assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
2464 (Ty->isInteger() || isa<PointerType>(Ty)) &&
2465 "Cannot truncate or noop with non-integer arguments!");
2466 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2467 "getTruncateOrNoop cannot extend!");
2468 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2469 return V; // No conversion
2470 return getTruncateExpr(V, Ty);
2473 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2474 /// the types using zero-extension, and then perform a umax operation
2476 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2478 const SCEV *PromotedLHS = LHS;
2479 const SCEV *PromotedRHS = RHS;
2481 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2482 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2484 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2486 return getUMaxExpr(PromotedLHS, PromotedRHS);
2489 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2490 /// the types using zero-extension, and then perform a umin operation
2492 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2494 const SCEV *PromotedLHS = LHS;
2495 const SCEV *PromotedRHS = RHS;
2497 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2498 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2500 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2502 return getUMinExpr(PromotedLHS, PromotedRHS);
2505 /// PushDefUseChildren - Push users of the given Instruction
2506 /// onto the given Worklist.
2508 PushDefUseChildren(Instruction *I,
2509 SmallVectorImpl<Instruction *> &Worklist) {
2510 // Push the def-use children onto the Worklist stack.
2511 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2513 Worklist.push_back(cast<Instruction>(UI));
2516 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2517 /// instructions that depend on the given instruction and removes them from
2518 /// the Scalars map if they reference SymName. This is used during PHI
2521 ScalarEvolution::ForgetSymbolicName(Instruction *I, const SCEV *SymName) {
2522 SmallVector<Instruction *, 16> Worklist;
2523 PushDefUseChildren(I, Worklist);
2525 SmallPtrSet<Instruction *, 8> Visited;
2527 while (!Worklist.empty()) {
2528 Instruction *I = Worklist.pop_back_val();
2529 if (!Visited.insert(I)) continue;
2531 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
2532 Scalars.find(static_cast<Value *>(I));
2533 if (It != Scalars.end()) {
2534 // Short-circuit the def-use traversal if the symbolic name
2535 // ceases to appear in expressions.
2536 if (!It->second->hasOperand(SymName))
2539 // SCEVUnknown for a PHI either means that it has an unrecognized
2540 // structure, or it's a PHI that's in the progress of being computed
2541 // by createNodeForPHI. In the former case, additional loop trip
2542 // count information isn't going to change anything. In the later
2543 // case, createNodeForPHI will perform the necessary updates on its
2544 // own when it gets to that point.
2545 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
2546 ValuesAtScopes.erase(It->second);
2551 PushDefUseChildren(I, Worklist);
2555 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2556 /// a loop header, making it a potential recurrence, or it doesn't.
2558 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2559 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
2560 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2561 if (L->getHeader() == PN->getParent()) {
2562 // If it lives in the loop header, it has two incoming values, one
2563 // from outside the loop, and one from inside.
2564 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
2565 unsigned BackEdge = IncomingEdge^1;
2567 // While we are analyzing this PHI node, handle its value symbolically.
2568 const SCEV *SymbolicName = getUnknown(PN);
2569 assert(Scalars.find(PN) == Scalars.end() &&
2570 "PHI node already processed?");
2571 Scalars.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2573 // Using this symbolic name for the PHI, analyze the value coming around
2575 Value *BEValueV = PN->getIncomingValue(BackEdge);
2576 const SCEV *BEValue = getSCEV(BEValueV);
2578 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2579 // has a special value for the first iteration of the loop.
2581 // If the value coming around the backedge is an add with the symbolic
2582 // value we just inserted, then we found a simple induction variable!
2583 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2584 // If there is a single occurrence of the symbolic value, replace it
2585 // with a recurrence.
2586 unsigned FoundIndex = Add->getNumOperands();
2587 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2588 if (Add->getOperand(i) == SymbolicName)
2589 if (FoundIndex == e) {
2594 if (FoundIndex != Add->getNumOperands()) {
2595 // Create an add with everything but the specified operand.
2596 SmallVector<const SCEV *, 8> Ops;
2597 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2598 if (i != FoundIndex)
2599 Ops.push_back(Add->getOperand(i));
2600 const SCEV *Accum = getAddExpr(Ops);
2602 // This is not a valid addrec if the step amount is varying each
2603 // loop iteration, but is not itself an addrec in this loop.
2604 if (Accum->isLoopInvariant(L) ||
2605 (isa<SCEVAddRecExpr>(Accum) &&
2606 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2607 bool HasNUW = false;
2608 bool HasNSW = false;
2610 // If the increment doesn't overflow, then neither the addrec nor
2611 // the post-increment will overflow.
2612 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
2613 if (OBO->hasNoUnsignedWrap())
2615 if (OBO->hasNoSignedWrap())
2619 const SCEV *StartVal =
2620 getSCEV(PN->getIncomingValue(IncomingEdge));
2621 const SCEV *PHISCEV =
2622 getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
2624 // Since the no-wrap flags are on the increment, they apply to the
2625 // post-incremented value as well.
2626 if (Accum->isLoopInvariant(L))
2627 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
2628 Accum, L, HasNUW, HasNSW);
2630 // Okay, for the entire analysis of this edge we assumed the PHI
2631 // to be symbolic. We now need to go back and purge all of the
2632 // entries for the scalars that use the symbolic expression.
2633 ForgetSymbolicName(PN, SymbolicName);
2634 Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
2638 } else if (const SCEVAddRecExpr *AddRec =
2639 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2640 // Otherwise, this could be a loop like this:
2641 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2642 // In this case, j = {1,+,1} and BEValue is j.
2643 // Because the other in-value of i (0) fits the evolution of BEValue
2644 // i really is an addrec evolution.
2645 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2646 const SCEV *StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
2648 // If StartVal = j.start - j.stride, we can use StartVal as the
2649 // initial step of the addrec evolution.
2650 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2651 AddRec->getOperand(1))) {
2652 const SCEV *PHISCEV =
2653 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2655 // Okay, for the entire analysis of this edge we assumed the PHI
2656 // to be symbolic. We now need to go back and purge all of the
2657 // entries for the scalars that use the symbolic expression.
2658 ForgetSymbolicName(PN, SymbolicName);
2659 Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
2665 return SymbolicName;
2668 // It's tempting to recognize PHIs with a unique incoming value, however
2669 // this leads passes like indvars to break LCSSA form. Fortunately, such
2670 // PHIs are rare, as instcombine zaps them.
2672 // If it's not a loop phi, we can't handle it yet.
2673 return getUnknown(PN);
2676 /// createNodeForGEP - Expand GEP instructions into add and multiply
2677 /// operations. This allows them to be analyzed by regular SCEV code.
2679 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
2681 bool InBounds = GEP->isInBounds();
2682 const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
2683 Value *Base = GEP->getOperand(0);
2684 // Don't attempt to analyze GEPs over unsized objects.
2685 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2686 return getUnknown(GEP);
2687 const SCEV *TotalOffset = getIntegerSCEV(0, IntPtrTy);
2688 gep_type_iterator GTI = gep_type_begin(GEP);
2689 for (GetElementPtrInst::op_iterator I = next(GEP->op_begin()),
2693 // Compute the (potentially symbolic) offset in bytes for this index.
2694 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2695 // For a struct, add the member offset.
2696 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2697 TotalOffset = getAddExpr(TotalOffset,
2698 getFieldOffsetExpr(STy, FieldNo),
2699 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2701 // For an array, add the element offset, explicitly scaled.
2702 const SCEV *LocalOffset = getSCEV(Index);
2703 if (!isa<PointerType>(LocalOffset->getType()))
2704 // Getelementptr indicies are signed.
2705 LocalOffset = getTruncateOrSignExtend(LocalOffset, IntPtrTy);
2706 // Lower "inbounds" GEPs to NSW arithmetic.
2707 LocalOffset = getMulExpr(LocalOffset, getAllocSizeExpr(*GTI),
2708 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2709 TotalOffset = getAddExpr(TotalOffset, LocalOffset,
2710 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2713 return getAddExpr(getSCEV(Base), TotalOffset,
2714 /*HasNUW=*/false, /*HasNSW=*/InBounds);
2717 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2718 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2719 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2720 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2722 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
2723 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2724 return C->getValue()->getValue().countTrailingZeros();
2726 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2727 return std::min(GetMinTrailingZeros(T->getOperand()),
2728 (uint32_t)getTypeSizeInBits(T->getType()));
2730 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2731 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2732 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2733 getTypeSizeInBits(E->getType()) : OpRes;
2736 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2737 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
2738 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
2739 getTypeSizeInBits(E->getType()) : OpRes;
2742 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2743 // The result is the min of all operands results.
2744 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2745 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2746 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2750 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2751 // The result is the sum of all operands results.
2752 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
2753 uint32_t BitWidth = getTypeSizeInBits(M->getType());
2754 for (unsigned i = 1, e = M->getNumOperands();
2755 SumOpRes != BitWidth && i != e; ++i)
2756 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
2761 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2762 // The result is the min of all operands results.
2763 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
2764 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2765 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
2769 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2770 // The result is the min of all operands results.
2771 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2772 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2773 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2777 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2778 // The result is the min of all operands results.
2779 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
2780 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2781 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
2785 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2786 // For a SCEVUnknown, ask ValueTracking.
2787 unsigned BitWidth = getTypeSizeInBits(U->getType());
2788 APInt Mask = APInt::getAllOnesValue(BitWidth);
2789 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2790 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
2791 return Zeros.countTrailingOnes();
2798 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
2801 ScalarEvolution::getUnsignedRange(const SCEV *S) {
2803 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2804 return ConstantRange(C->getValue()->getValue());
2806 unsigned BitWidth = getTypeSizeInBits(S->getType());
2807 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
2809 // If the value has known zeros, the maximum unsigned value will have those
2810 // known zeros as well.
2811 uint32_t TZ = GetMinTrailingZeros(S);
2813 ConservativeResult =
2814 ConstantRange(APInt::getMinValue(BitWidth),
2815 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
2817 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2818 ConstantRange X = getUnsignedRange(Add->getOperand(0));
2819 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2820 X = X.add(getUnsignedRange(Add->getOperand(i)));
2821 return ConservativeResult.intersectWith(X);
2824 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2825 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
2826 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2827 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
2828 return ConservativeResult.intersectWith(X);
2831 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2832 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
2833 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2834 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
2835 return ConservativeResult.intersectWith(X);
2838 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2839 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
2840 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2841 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
2842 return ConservativeResult.intersectWith(X);
2845 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2846 ConstantRange X = getUnsignedRange(UDiv->getLHS());
2847 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
2848 return ConservativeResult.intersectWith(X.udiv(Y));
2851 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2852 ConstantRange X = getUnsignedRange(ZExt->getOperand());
2853 return ConservativeResult.intersectWith(X.zeroExtend(BitWidth));
2856 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2857 ConstantRange X = getUnsignedRange(SExt->getOperand());
2858 return ConservativeResult.intersectWith(X.signExtend(BitWidth));
2861 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
2862 ConstantRange X = getUnsignedRange(Trunc->getOperand());
2863 return ConservativeResult.intersectWith(X.truncate(BitWidth));
2866 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
2867 // If there's no unsigned wrap, the value will never be less than its
2869 if (AddRec->hasNoUnsignedWrap())
2870 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
2871 ConservativeResult =
2872 ConstantRange(C->getValue()->getValue(),
2873 APInt(getTypeSizeInBits(C->getType()), 0));
2875 // TODO: non-affine addrec
2876 if (AddRec->isAffine()) {
2877 const Type *Ty = AddRec->getType();
2878 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
2879 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
2880 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
2881 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
2883 const SCEV *Start = AddRec->getStart();
2884 const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
2886 // Check for overflow.
2887 if (!AddRec->hasNoUnsignedWrap())
2888 return ConservativeResult;
2890 ConstantRange StartRange = getUnsignedRange(Start);
2891 ConstantRange EndRange = getUnsignedRange(End);
2892 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
2893 EndRange.getUnsignedMin());
2894 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
2895 EndRange.getUnsignedMax());
2896 if (Min.isMinValue() && Max.isMaxValue())
2897 return ConservativeResult;
2898 return ConservativeResult.intersectWith(ConstantRange(Min, Max+1));
2902 return ConservativeResult;
2905 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2906 // For a SCEVUnknown, ask ValueTracking.
2907 unsigned BitWidth = getTypeSizeInBits(U->getType());
2908 APInt Mask = APInt::getAllOnesValue(BitWidth);
2909 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
2910 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
2911 if (Ones == ~Zeros + 1)
2912 return ConservativeResult;
2913 return ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1));
2916 return ConservativeResult;
2919 /// getSignedRange - Determine the signed range for a particular SCEV.
2922 ScalarEvolution::getSignedRange(const SCEV *S) {
2924 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2925 return ConstantRange(C->getValue()->getValue());
2927 unsigned BitWidth = getTypeSizeInBits(S->getType());
2928 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
2930 // If the value has known zeros, the maximum signed value will have those
2931 // known zeros as well.
2932 uint32_t TZ = GetMinTrailingZeros(S);
2934 ConservativeResult =
2935 ConstantRange(APInt::getSignedMinValue(BitWidth),
2936 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
2938 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2939 ConstantRange X = getSignedRange(Add->getOperand(0));
2940 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
2941 X = X.add(getSignedRange(Add->getOperand(i)));
2942 return ConservativeResult.intersectWith(X);
2945 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2946 ConstantRange X = getSignedRange(Mul->getOperand(0));
2947 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
2948 X = X.multiply(getSignedRange(Mul->getOperand(i)));
2949 return ConservativeResult.intersectWith(X);
2952 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
2953 ConstantRange X = getSignedRange(SMax->getOperand(0));
2954 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
2955 X = X.smax(getSignedRange(SMax->getOperand(i)));
2956 return ConservativeResult.intersectWith(X);
2959 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
2960 ConstantRange X = getSignedRange(UMax->getOperand(0));
2961 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
2962 X = X.umax(getSignedRange(UMax->getOperand(i)));
2963 return ConservativeResult.intersectWith(X);
2966 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
2967 ConstantRange X = getSignedRange(UDiv->getLHS());
2968 ConstantRange Y = getSignedRange(UDiv->getRHS());
2969 return ConservativeResult.intersectWith(X.udiv(Y));
2972 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
2973 ConstantRange X = getSignedRange(ZExt->getOperand());
2974 return ConservativeResult.intersectWith(X.zeroExtend(BitWidth));
2977 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
2978 ConstantRange X = getSignedRange(SExt->getOperand());
2979 return ConservativeResult.intersectWith(X.signExtend(BitWidth));
2982 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
2983 ConstantRange X = getSignedRange(Trunc->getOperand());
2984 return ConservativeResult.intersectWith(X.truncate(BitWidth));
2987 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
2988 // If there's no signed wrap, and all the operands have the same sign or
2989 // zero, the value won't ever change sign.
2990 if (AddRec->hasNoSignedWrap()) {
2991 bool AllNonNeg = true;
2992 bool AllNonPos = true;
2993 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
2994 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
2995 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
2998 ConservativeResult = ConservativeResult.intersectWith(
2999 ConstantRange(APInt(BitWidth, 0),
3000 APInt::getSignedMinValue(BitWidth)));
3002 ConservativeResult = ConservativeResult.intersectWith(
3003 ConstantRange(APInt::getSignedMinValue(BitWidth),
3004 APInt(BitWidth, 1)));
3007 // TODO: non-affine addrec
3008 if (AddRec->isAffine()) {
3009 const Type *Ty = AddRec->getType();
3010 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3011 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3012 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3013 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3015 const SCEV *Start = AddRec->getStart();
3016 const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
3018 // Check for overflow.
3019 if (!AddRec->hasNoSignedWrap())
3020 return ConservativeResult;
3022 ConstantRange StartRange = getSignedRange(Start);
3023 ConstantRange EndRange = getSignedRange(End);
3024 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3025 EndRange.getSignedMin());
3026 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3027 EndRange.getSignedMax());
3028 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3029 return ConservativeResult;
3030 return ConservativeResult.intersectWith(ConstantRange(Min, Max+1));
3034 return ConservativeResult;
3037 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3038 // For a SCEVUnknown, ask ValueTracking.
3039 if (!U->getValue()->getType()->isInteger() && !TD)
3040 return ConservativeResult;
3041 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3043 return ConservativeResult;
3044 return ConservativeResult.intersectWith(
3045 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3046 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1));
3049 return ConservativeResult;
3052 /// createSCEV - We know that there is no SCEV for the specified value.
3053 /// Analyze the expression.
3055 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3056 if (!isSCEVable(V->getType()))
3057 return getUnknown(V);
3059 unsigned Opcode = Instruction::UserOp1;
3060 if (Instruction *I = dyn_cast<Instruction>(V))
3061 Opcode = I->getOpcode();
3062 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3063 Opcode = CE->getOpcode();
3064 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3065 return getConstant(CI);
3066 else if (isa<ConstantPointerNull>(V))
3067 return getIntegerSCEV(0, V->getType());
3068 else if (isa<UndefValue>(V))
3069 return getIntegerSCEV(0, V->getType());
3070 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3071 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3073 return getUnknown(V);
3075 Operator *U = cast<Operator>(V);
3077 case Instruction::Add:
3078 // Don't transfer the NSW and NUW bits from the Add instruction to the
3079 // Add expression, because the Instruction may be guarded by control
3080 // flow and the no-overflow bits may not be valid for the expression in
3082 return getAddExpr(getSCEV(U->getOperand(0)),
3083 getSCEV(U->getOperand(1)));
3084 case Instruction::Mul:
3085 // Don't transfer the NSW and NUW bits from the Mul instruction to the
3086 // Mul expression, as with Add.
3087 return getMulExpr(getSCEV(U->getOperand(0)),
3088 getSCEV(U->getOperand(1)));
3089 case Instruction::UDiv:
3090 return getUDivExpr(getSCEV(U->getOperand(0)),
3091 getSCEV(U->getOperand(1)));
3092 case Instruction::Sub:
3093 return getMinusSCEV(getSCEV(U->getOperand(0)),
3094 getSCEV(U->getOperand(1)));
3095 case Instruction::And:
3096 // For an expression like x&255 that merely masks off the high bits,
3097 // use zext(trunc(x)) as the SCEV expression.
3098 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3099 if (CI->isNullValue())
3100 return getSCEV(U->getOperand(1));
3101 if (CI->isAllOnesValue())
3102 return getSCEV(U->getOperand(0));
3103 const APInt &A = CI->getValue();
3105 // Instcombine's ShrinkDemandedConstant may strip bits out of
3106 // constants, obscuring what would otherwise be a low-bits mask.
3107 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3108 // knew about to reconstruct a low-bits mask value.
3109 unsigned LZ = A.countLeadingZeros();
3110 unsigned BitWidth = A.getBitWidth();
3111 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
3112 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3113 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
3115 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3117 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3119 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3120 IntegerType::get(getContext(), BitWidth - LZ)),
3125 case Instruction::Or:
3126 // If the RHS of the Or is a constant, we may have something like:
3127 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3128 // optimizations will transparently handle this case.
3130 // In order for this transformation to be safe, the LHS must be of the
3131 // form X*(2^n) and the Or constant must be less than 2^n.
3132 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3133 const SCEV *LHS = getSCEV(U->getOperand(0));
3134 const APInt &CIVal = CI->getValue();
3135 if (GetMinTrailingZeros(LHS) >=
3136 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3137 // Build a plain add SCEV.
3138 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3139 // If the LHS of the add was an addrec and it has no-wrap flags,
3140 // transfer the no-wrap flags, since an or won't introduce a wrap.
3141 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3142 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3143 if (OldAR->hasNoUnsignedWrap())
3144 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
3145 if (OldAR->hasNoSignedWrap())
3146 const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
3152 case Instruction::Xor:
3153 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3154 // If the RHS of the xor is a signbit, then this is just an add.
3155 // Instcombine turns add of signbit into xor as a strength reduction step.
3156 if (CI->getValue().isSignBit())
3157 return getAddExpr(getSCEV(U->getOperand(0)),
3158 getSCEV(U->getOperand(1)));
3160 // If the RHS of xor is -1, then this is a not operation.
3161 if (CI->isAllOnesValue())
3162 return getNotSCEV(getSCEV(U->getOperand(0)));
3164 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3165 // This is a variant of the check for xor with -1, and it handles
3166 // the case where instcombine has trimmed non-demanded bits out
3167 // of an xor with -1.
3168 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3169 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3170 if (BO->getOpcode() == Instruction::And &&
3171 LCI->getValue() == CI->getValue())
3172 if (const SCEVZeroExtendExpr *Z =
3173 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3174 const Type *UTy = U->getType();
3175 const SCEV *Z0 = Z->getOperand();
3176 const Type *Z0Ty = Z0->getType();
3177 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3179 // If C is a low-bits mask, the zero extend is zerving to
3180 // mask off the high bits. Complement the operand and
3181 // re-apply the zext.
3182 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3183 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3185 // If C is a single bit, it may be in the sign-bit position
3186 // before the zero-extend. In this case, represent the xor
3187 // using an add, which is equivalent, and re-apply the zext.
3188 APInt Trunc = APInt(CI->getValue()).trunc(Z0TySize);
3189 if (APInt(Trunc).zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3191 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3197 case Instruction::Shl:
3198 // Turn shift left of a constant amount into a multiply.
3199 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3200 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
3201 Constant *X = ConstantInt::get(getContext(),
3202 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
3203 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3207 case Instruction::LShr:
3208 // Turn logical shift right of a constant into a unsigned divide.
3209 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3210 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
3211 Constant *X = ConstantInt::get(getContext(),
3212 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
3213 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3217 case Instruction::AShr:
3218 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3219 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3220 if (Instruction *L = dyn_cast<Instruction>(U->getOperand(0)))
3221 if (L->getOpcode() == Instruction::Shl &&
3222 L->getOperand(1) == U->getOperand(1)) {
3223 unsigned BitWidth = getTypeSizeInBits(U->getType());
3224 uint64_t Amt = BitWidth - CI->getZExtValue();
3225 if (Amt == BitWidth)
3226 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3228 return getIntegerSCEV(0, U->getType()); // value is undefined
3230 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3231 IntegerType::get(getContext(), Amt)),
3236 case Instruction::Trunc:
3237 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3239 case Instruction::ZExt:
3240 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3242 case Instruction::SExt:
3243 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3245 case Instruction::BitCast:
3246 // BitCasts are no-op casts so we just eliminate the cast.
3247 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3248 return getSCEV(U->getOperand(0));
3251 // It's tempting to handle inttoptr and ptrtoint, however this can
3252 // lead to pointer expressions which cannot be expanded to GEPs
3253 // (because they may overflow). For now, the only pointer-typed
3254 // expressions we handle are GEPs and address literals.
3256 case Instruction::GetElementPtr:
3257 return createNodeForGEP(cast<GEPOperator>(U));
3259 case Instruction::PHI:
3260 return createNodeForPHI(cast<PHINode>(U));
3262 case Instruction::Select:
3263 // This could be a smax or umax that was lowered earlier.
3264 // Try to recover it.
3265 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3266 Value *LHS = ICI->getOperand(0);
3267 Value *RHS = ICI->getOperand(1);
3268 switch (ICI->getPredicate()) {
3269 case ICmpInst::ICMP_SLT:
3270 case ICmpInst::ICMP_SLE:
3271 std::swap(LHS, RHS);
3273 case ICmpInst::ICMP_SGT:
3274 case ICmpInst::ICMP_SGE:
3275 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
3276 return getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
3277 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
3278 return getSMinExpr(getSCEV(LHS), getSCEV(RHS));
3280 case ICmpInst::ICMP_ULT:
3281 case ICmpInst::ICMP_ULE:
3282 std::swap(LHS, RHS);
3284 case ICmpInst::ICMP_UGT:
3285 case ICmpInst::ICMP_UGE:
3286 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
3287 return getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
3288 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
3289 return getUMinExpr(getSCEV(LHS), getSCEV(RHS));
3291 case ICmpInst::ICMP_NE:
3292 // n != 0 ? n : 1 -> umax(n, 1)
3293 if (LHS == U->getOperand(1) &&
3294 isa<ConstantInt>(U->getOperand(2)) &&
3295 cast<ConstantInt>(U->getOperand(2))->isOne() &&
3296 isa<ConstantInt>(RHS) &&
3297 cast<ConstantInt>(RHS)->isZero())
3298 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(2)));
3300 case ICmpInst::ICMP_EQ:
3301 // n == 0 ? 1 : n -> umax(n, 1)
3302 if (LHS == U->getOperand(2) &&
3303 isa<ConstantInt>(U->getOperand(1)) &&
3304 cast<ConstantInt>(U->getOperand(1))->isOne() &&
3305 isa<ConstantInt>(RHS) &&
3306 cast<ConstantInt>(RHS)->isZero())
3307 return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(1)));
3314 default: // We cannot analyze this expression.
3318 return getUnknown(V);
3323 //===----------------------------------------------------------------------===//
3324 // Iteration Count Computation Code
3327 /// getBackedgeTakenCount - If the specified loop has a predictable
3328 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3329 /// object. The backedge-taken count is the number of times the loop header
3330 /// will be branched to from within the loop. This is one less than the
3331 /// trip count of the loop, since it doesn't count the first iteration,
3332 /// when the header is branched to from outside the loop.
3334 /// Note that it is not valid to call this method on a loop without a
3335 /// loop-invariant backedge-taken count (see
3336 /// hasLoopInvariantBackedgeTakenCount).
3338 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3339 return getBackedgeTakenInfo(L).Exact;
3342 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3343 /// return the least SCEV value that is known never to be less than the
3344 /// actual backedge taken count.
3345 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3346 return getBackedgeTakenInfo(L).Max;
3349 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3350 /// onto the given Worklist.
3352 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3353 BasicBlock *Header = L->getHeader();
3355 // Push all Loop-header PHIs onto the Worklist stack.
3356 for (BasicBlock::iterator I = Header->begin();
3357 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3358 Worklist.push_back(PN);
3361 const ScalarEvolution::BackedgeTakenInfo &
3362 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3363 // Initially insert a CouldNotCompute for this loop. If the insertion
3364 // succeeds, procede to actually compute a backedge-taken count and
3365 // update the value. The temporary CouldNotCompute value tells SCEV
3366 // code elsewhere that it shouldn't attempt to request a new
3367 // backedge-taken count, which could result in infinite recursion.
3368 std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3369 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
3371 BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
3372 if (BECount.Exact != getCouldNotCompute()) {
3373 assert(BECount.Exact->isLoopInvariant(L) &&
3374 BECount.Max->isLoopInvariant(L) &&
3375 "Computed backedge-taken count isn't loop invariant for loop!");
3376 ++NumTripCountsComputed;
3378 // Update the value in the map.
3379 Pair.first->second = BECount;
3381 if (BECount.Max != getCouldNotCompute())
3382 // Update the value in the map.
3383 Pair.first->second = BECount;
3384 if (isa<PHINode>(L->getHeader()->begin()))
3385 // Only count loops that have phi nodes as not being computable.
3386 ++NumTripCountsNotComputed;
3389 // Now that we know more about the trip count for this loop, forget any
3390 // existing SCEV values for PHI nodes in this loop since they are only
3391 // conservative estimates made without the benefit of trip count
3392 // information. This is similar to the code in forgetLoop, except that
3393 // it handles SCEVUnknown PHI nodes specially.
3394 if (BECount.hasAnyInfo()) {
3395 SmallVector<Instruction *, 16> Worklist;
3396 PushLoopPHIs(L, Worklist);
3398 SmallPtrSet<Instruction *, 8> Visited;
3399 while (!Worklist.empty()) {
3400 Instruction *I = Worklist.pop_back_val();
3401 if (!Visited.insert(I)) continue;
3403 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
3404 Scalars.find(static_cast<Value *>(I));
3405 if (It != Scalars.end()) {
3406 // SCEVUnknown for a PHI either means that it has an unrecognized
3407 // structure, or it's a PHI that's in the progress of being computed
3408 // by createNodeForPHI. In the former case, additional loop trip
3409 // count information isn't going to change anything. In the later
3410 // case, createNodeForPHI will perform the necessary updates on its
3411 // own when it gets to that point.
3412 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
3413 ValuesAtScopes.erase(It->second);
3416 if (PHINode *PN = dyn_cast<PHINode>(I))
3417 ConstantEvolutionLoopExitValue.erase(PN);
3420 PushDefUseChildren(I, Worklist);
3424 return Pair.first->second;
3427 /// forgetLoop - This method should be called by the client when it has
3428 /// changed a loop in a way that may effect ScalarEvolution's ability to
3429 /// compute a trip count, or if the loop is deleted.
3430 void ScalarEvolution::forgetLoop(const Loop *L) {
3431 // Drop any stored trip count value.
3432 BackedgeTakenCounts.erase(L);
3434 // Drop information about expressions based on loop-header PHIs.
3435 SmallVector<Instruction *, 16> Worklist;
3436 PushLoopPHIs(L, Worklist);
3438 SmallPtrSet<Instruction *, 8> Visited;
3439 while (!Worklist.empty()) {
3440 Instruction *I = Worklist.pop_back_val();
3441 if (!Visited.insert(I)) continue;
3443 std::map<SCEVCallbackVH, const SCEV *>::iterator It =
3444 Scalars.find(static_cast<Value *>(I));
3445 if (It != Scalars.end()) {
3446 ValuesAtScopes.erase(It->second);
3448 if (PHINode *PN = dyn_cast<PHINode>(I))
3449 ConstantEvolutionLoopExitValue.erase(PN);
3452 PushDefUseChildren(I, Worklist);
3456 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
3457 /// of the specified loop will execute.
3458 ScalarEvolution::BackedgeTakenInfo
3459 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
3460 SmallVector<BasicBlock *, 8> ExitingBlocks;
3461 L->getExitingBlocks(ExitingBlocks);
3463 // Examine all exits and pick the most conservative values.
3464 const SCEV *BECount = getCouldNotCompute();
3465 const SCEV *MaxBECount = getCouldNotCompute();
3466 bool CouldNotComputeBECount = false;
3467 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
3468 BackedgeTakenInfo NewBTI =
3469 ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
3471 if (NewBTI.Exact == getCouldNotCompute()) {
3472 // We couldn't compute an exact value for this exit, so
3473 // we won't be able to compute an exact value for the loop.
3474 CouldNotComputeBECount = true;
3475 BECount = getCouldNotCompute();
3476 } else if (!CouldNotComputeBECount) {
3477 if (BECount == getCouldNotCompute())
3478 BECount = NewBTI.Exact;
3480 BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
3482 if (MaxBECount == getCouldNotCompute())
3483 MaxBECount = NewBTI.Max;
3484 else if (NewBTI.Max != getCouldNotCompute())
3485 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
3488 return BackedgeTakenInfo(BECount, MaxBECount);
3491 /// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
3492 /// of the specified loop will execute if it exits via the specified block.
3493 ScalarEvolution::BackedgeTakenInfo
3494 ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
3495 BasicBlock *ExitingBlock) {
3497 // Okay, we've chosen an exiting block. See what condition causes us to
3498 // exit at this block.
3500 // FIXME: we should be able to handle switch instructions (with a single exit)
3501 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
3502 if (ExitBr == 0) return getCouldNotCompute();
3503 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
3505 // At this point, we know we have a conditional branch that determines whether
3506 // the loop is exited. However, we don't know if the branch is executed each
3507 // time through the loop. If not, then the execution count of the branch will
3508 // not be equal to the trip count of the loop.
3510 // Currently we check for this by checking to see if the Exit branch goes to
3511 // the loop header. If so, we know it will always execute the same number of
3512 // times as the loop. We also handle the case where the exit block *is* the
3513 // loop header. This is common for un-rotated loops.
3515 // If both of those tests fail, walk up the unique predecessor chain to the
3516 // header, stopping if there is an edge that doesn't exit the loop. If the
3517 // header is reached, the execution count of the branch will be equal to the
3518 // trip count of the loop.
3520 // More extensive analysis could be done to handle more cases here.
3522 if (ExitBr->getSuccessor(0) != L->getHeader() &&
3523 ExitBr->getSuccessor(1) != L->getHeader() &&
3524 ExitBr->getParent() != L->getHeader()) {
3525 // The simple checks failed, try climbing the unique predecessor chain
3526 // up to the header.
3528 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
3529 BasicBlock *Pred = BB->getUniquePredecessor();
3531 return getCouldNotCompute();
3532 TerminatorInst *PredTerm = Pred->getTerminator();
3533 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
3534 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
3537 // If the predecessor has a successor that isn't BB and isn't
3538 // outside the loop, assume the worst.
3539 if (L->contains(PredSucc))
3540 return getCouldNotCompute();
3542 if (Pred == L->getHeader()) {
3549 return getCouldNotCompute();
3552 // Procede to the next level to examine the exit condition expression.
3553 return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
3554 ExitBr->getSuccessor(0),
3555 ExitBr->getSuccessor(1));
3558 /// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
3559 /// backedge of the specified loop will execute if its exit condition
3560 /// were a conditional branch of ExitCond, TBB, and FBB.
3561 ScalarEvolution::BackedgeTakenInfo
3562 ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
3566 // Check if the controlling expression for this loop is an And or Or.
3567 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
3568 if (BO->getOpcode() == Instruction::And) {
3569 // Recurse on the operands of the and.
3570 BackedgeTakenInfo BTI0 =
3571 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3572 BackedgeTakenInfo BTI1 =
3573 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3574 const SCEV *BECount = getCouldNotCompute();
3575 const SCEV *MaxBECount = getCouldNotCompute();
3576 if (L->contains(TBB)) {
3577 // Both conditions must be true for the loop to continue executing.
3578 // Choose the less conservative count.
3579 if (BTI0.Exact == getCouldNotCompute() ||
3580 BTI1.Exact == getCouldNotCompute())
3581 BECount = getCouldNotCompute();
3583 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3584 if (BTI0.Max == getCouldNotCompute())
3585 MaxBECount = BTI1.Max;
3586 else if (BTI1.Max == getCouldNotCompute())
3587 MaxBECount = BTI0.Max;
3589 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3591 // Both conditions must be true for the loop to exit.
3592 assert(L->contains(FBB) && "Loop block has no successor in loop!");
3593 if (BTI0.Exact != getCouldNotCompute() &&
3594 BTI1.Exact != getCouldNotCompute())
3595 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3596 if (BTI0.Max != getCouldNotCompute() &&
3597 BTI1.Max != getCouldNotCompute())
3598 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
3601 return BackedgeTakenInfo(BECount, MaxBECount);
3603 if (BO->getOpcode() == Instruction::Or) {
3604 // Recurse on the operands of the or.
3605 BackedgeTakenInfo BTI0 =
3606 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
3607 BackedgeTakenInfo BTI1 =
3608 ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
3609 const SCEV *BECount = getCouldNotCompute();
3610 const SCEV *MaxBECount = getCouldNotCompute();
3611 if (L->contains(FBB)) {
3612 // Both conditions must be false for the loop to continue executing.
3613 // Choose the less conservative count.
3614 if (BTI0.Exact == getCouldNotCompute() ||
3615 BTI1.Exact == getCouldNotCompute())
3616 BECount = getCouldNotCompute();
3618 BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3619 if (BTI0.Max == getCouldNotCompute())
3620 MaxBECount = BTI1.Max;
3621 else if (BTI1.Max == getCouldNotCompute())
3622 MaxBECount = BTI0.Max;
3624 MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
3626 // Both conditions must be false for the loop to exit.
3627 assert(L->contains(TBB) && "Loop block has no successor in loop!");
3628 if (BTI0.Exact != getCouldNotCompute() &&
3629 BTI1.Exact != getCouldNotCompute())
3630 BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
3631 if (BTI0.Max != getCouldNotCompute() &&
3632 BTI1.Max != getCouldNotCompute())
3633 MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
3636 return BackedgeTakenInfo(BECount, MaxBECount);
3640 // With an icmp, it may be feasible to compute an exact backedge-taken count.
3641 // Procede to the next level to examine the icmp.
3642 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
3643 return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
3645 // If it's not an integer or pointer comparison then compute it the hard way.
3646 return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3649 /// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
3650 /// backedge of the specified loop will execute if its exit condition
3651 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
3652 ScalarEvolution::BackedgeTakenInfo
3653 ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
3658 // If the condition was exit on true, convert the condition to exit on false
3659 ICmpInst::Predicate Cond;
3660 if (!L->contains(FBB))
3661 Cond = ExitCond->getPredicate();
3663 Cond = ExitCond->getInversePredicate();
3665 // Handle common loops like: for (X = "string"; *X; ++X)
3666 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
3667 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
3669 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
3670 if (!isa<SCEVCouldNotCompute>(ItCnt)) {
3671 unsigned BitWidth = getTypeSizeInBits(ItCnt->getType());
3672 return BackedgeTakenInfo(ItCnt,
3673 isa<SCEVConstant>(ItCnt) ? ItCnt :
3674 getConstant(APInt::getMaxValue(BitWidth)-1));
3678 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
3679 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
3681 // Try to evaluate any dependencies out of the loop.
3682 LHS = getSCEVAtScope(LHS, L);
3683 RHS = getSCEVAtScope(RHS, L);
3685 // At this point, we would like to compute how many iterations of the
3686 // loop the predicate will return true for these inputs.
3687 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
3688 // If there is a loop-invariant, force it into the RHS.
3689 std::swap(LHS, RHS);
3690 Cond = ICmpInst::getSwappedPredicate(Cond);
3693 // If we have a comparison of a chrec against a constant, try to use value
3694 // ranges to answer this query.
3695 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
3696 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
3697 if (AddRec->getLoop() == L) {
3698 // Form the constant range.
3699 ConstantRange CompRange(
3700 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
3702 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
3703 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
3707 case ICmpInst::ICMP_NE: { // while (X != Y)
3708 // Convert to: while (X-Y != 0)
3709 const SCEV *TC = HowFarToZero(getMinusSCEV(LHS, RHS), L);
3710 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3713 case ICmpInst::ICMP_EQ: { // while (X == Y)
3714 // Convert to: while (X-Y == 0)
3715 const SCEV *TC = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
3716 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
3719 case ICmpInst::ICMP_SLT: {
3720 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
3721 if (BTI.hasAnyInfo()) return BTI;
3724 case ICmpInst::ICMP_SGT: {
3725 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3726 getNotSCEV(RHS), L, true);
3727 if (BTI.hasAnyInfo()) return BTI;
3730 case ICmpInst::ICMP_ULT: {
3731 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
3732 if (BTI.hasAnyInfo()) return BTI;
3735 case ICmpInst::ICMP_UGT: {
3736 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
3737 getNotSCEV(RHS), L, false);
3738 if (BTI.hasAnyInfo()) return BTI;
3743 dbgs() << "ComputeBackedgeTakenCount ";
3744 if (ExitCond->getOperand(0)->getType()->isUnsigned())
3745 dbgs() << "[unsigned] ";
3746 dbgs() << *LHS << " "
3747 << Instruction::getOpcodeName(Instruction::ICmp)
3748 << " " << *RHS << "\n";
3753 ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
3756 static ConstantInt *
3757 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
3758 ScalarEvolution &SE) {
3759 const SCEV *InVal = SE.getConstant(C);
3760 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
3761 assert(isa<SCEVConstant>(Val) &&
3762 "Evaluation of SCEV at constant didn't fold correctly?");
3763 return cast<SCEVConstant>(Val)->getValue();
3766 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
3767 /// and a GEP expression (missing the pointer index) indexing into it, return
3768 /// the addressed element of the initializer or null if the index expression is
3771 GetAddressedElementFromGlobal(GlobalVariable *GV,
3772 const std::vector<ConstantInt*> &Indices) {
3773 Constant *Init = GV->getInitializer();
3774 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
3775 uint64_t Idx = Indices[i]->getZExtValue();
3776 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
3777 assert(Idx < CS->getNumOperands() && "Bad struct index!");
3778 Init = cast<Constant>(CS->getOperand(Idx));
3779 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
3780 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
3781 Init = cast<Constant>(CA->getOperand(Idx));
3782 } else if (isa<ConstantAggregateZero>(Init)) {
3783 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
3784 assert(Idx < STy->getNumElements() && "Bad struct index!");
3785 Init = Constant::getNullValue(STy->getElementType(Idx));
3786 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
3787 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
3788 Init = Constant::getNullValue(ATy->getElementType());
3790 llvm_unreachable("Unknown constant aggregate type!");
3794 return 0; // Unknown initializer type
3800 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
3801 /// 'icmp op load X, cst', try to see if we can compute the backedge
3802 /// execution count.
3804 ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
3808 ICmpInst::Predicate predicate) {
3809 if (LI->isVolatile()) return getCouldNotCompute();
3811 // Check to see if the loaded pointer is a getelementptr of a global.
3812 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
3813 if (!GEP) return getCouldNotCompute();
3815 // Make sure that it is really a constant global we are gepping, with an
3816 // initializer, and make sure the first IDX is really 0.
3817 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
3818 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
3819 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
3820 !cast<Constant>(GEP->getOperand(1))->isNullValue())
3821 return getCouldNotCompute();
3823 // Okay, we allow one non-constant index into the GEP instruction.
3825 std::vector<ConstantInt*> Indexes;
3826 unsigned VarIdxNum = 0;
3827 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
3828 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
3829 Indexes.push_back(CI);
3830 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
3831 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
3832 VarIdx = GEP->getOperand(i);
3834 Indexes.push_back(0);
3837 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
3838 // Check to see if X is a loop variant variable value now.
3839 const SCEV *Idx = getSCEV(VarIdx);
3840 Idx = getSCEVAtScope(Idx, L);
3842 // We can only recognize very limited forms of loop index expressions, in
3843 // particular, only affine AddRec's like {C1,+,C2}.
3844 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
3845 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
3846 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
3847 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
3848 return getCouldNotCompute();
3850 unsigned MaxSteps = MaxBruteForceIterations;
3851 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
3852 ConstantInt *ItCst = ConstantInt::get(
3853 cast<IntegerType>(IdxExpr->getType()), IterationNum);
3854 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
3856 // Form the GEP offset.
3857 Indexes[VarIdxNum] = Val;
3859 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
3860 if (Result == 0) break; // Cannot compute!
3862 // Evaluate the condition for this iteration.
3863 Result = ConstantExpr::getICmp(predicate, Result, RHS);
3864 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
3865 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
3867 dbgs() << "\n***\n*** Computed loop count " << *ItCst
3868 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
3871 ++NumArrayLenItCounts;
3872 return getConstant(ItCst); // Found terminating iteration!
3875 return getCouldNotCompute();
3879 /// CanConstantFold - Return true if we can constant fold an instruction of the
3880 /// specified type, assuming that all operands were constants.
3881 static bool CanConstantFold(const Instruction *I) {
3882 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
3883 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
3886 if (const CallInst *CI = dyn_cast<CallInst>(I))
3887 if (const Function *F = CI->getCalledFunction())
3888 return canConstantFoldCallTo(F);
3892 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
3893 /// in the loop that V is derived from. We allow arbitrary operations along the
3894 /// way, but the operands of an operation must either be constants or a value
3895 /// derived from a constant PHI. If this expression does not fit with these
3896 /// constraints, return null.
3897 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
3898 // If this is not an instruction, or if this is an instruction outside of the
3899 // loop, it can't be derived from a loop PHI.
3900 Instruction *I = dyn_cast<Instruction>(V);
3901 if (I == 0 || !L->contains(I)) return 0;
3903 if (PHINode *PN = dyn_cast<PHINode>(I)) {
3904 if (L->getHeader() == I->getParent())
3907 // We don't currently keep track of the control flow needed to evaluate
3908 // PHIs, so we cannot handle PHIs inside of loops.
3912 // If we won't be able to constant fold this expression even if the operands
3913 // are constants, return early.
3914 if (!CanConstantFold(I)) return 0;
3916 // Otherwise, we can evaluate this instruction if all of its operands are
3917 // constant or derived from a PHI node themselves.
3919 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
3920 if (!(isa<Constant>(I->getOperand(Op)) ||
3921 isa<GlobalValue>(I->getOperand(Op)))) {
3922 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
3923 if (P == 0) return 0; // Not evolving from PHI
3927 return 0; // Evolving from multiple different PHIs.
3930 // This is a expression evolving from a constant PHI!
3934 /// EvaluateExpression - Given an expression that passes the
3935 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
3936 /// in the loop has the value PHIVal. If we can't fold this expression for some
3937 /// reason, return null.
3938 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
3939 const TargetData *TD) {
3940 if (isa<PHINode>(V)) return PHIVal;
3941 if (Constant *C = dyn_cast<Constant>(V)) return C;
3942 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
3943 Instruction *I = cast<Instruction>(V);
3945 std::vector<Constant*> Operands;
3946 Operands.resize(I->getNumOperands());
3948 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3949 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
3950 if (Operands[i] == 0) return 0;
3953 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
3954 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
3956 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
3957 &Operands[0], Operands.size(), TD);
3960 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
3961 /// in the header of its containing loop, we know the loop executes a
3962 /// constant number of times, and the PHI node is just a recurrence
3963 /// involving constants, fold it.
3965 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
3968 std::map<PHINode*, Constant*>::iterator I =
3969 ConstantEvolutionLoopExitValue.find(PN);
3970 if (I != ConstantEvolutionLoopExitValue.end())
3973 if (BEs.ugt(APInt(BEs.getBitWidth(),MaxBruteForceIterations)))
3974 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
3976 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
3978 // Since the loop is canonicalized, the PHI node must have two entries. One
3979 // entry must be a constant (coming in from outside of the loop), and the
3980 // second must be derived from the same PHI.
3981 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
3982 Constant *StartCST =
3983 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
3985 return RetVal = 0; // Must be a constant.
3987 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
3988 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
3990 return RetVal = 0; // Not derived from same PHI.
3992 // Execute the loop symbolically to determine the exit value.
3993 if (BEs.getActiveBits() >= 32)
3994 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
3996 unsigned NumIterations = BEs.getZExtValue(); // must be in range
3997 unsigned IterationNum = 0;
3998 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
3999 if (IterationNum == NumIterations)
4000 return RetVal = PHIVal; // Got exit value!
4002 // Compute the value of the PHI node for the next iteration.
4003 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4004 if (NextPHI == PHIVal)
4005 return RetVal = NextPHI; // Stopped evolving!
4007 return 0; // Couldn't evaluate!
4012 /// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
4013 /// constant number of times (the condition evolves only from constants),
4014 /// try to evaluate a few iterations of the loop until we get the exit
4015 /// condition gets a value of ExitWhen (true or false). If we cannot
4016 /// evaluate the trip count of the loop, return getCouldNotCompute().
4018 ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
4021 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4022 if (PN == 0) return getCouldNotCompute();
4024 // Since the loop is canonicalized, the PHI node must have two entries. One
4025 // entry must be a constant (coming in from outside of the loop), and the
4026 // second must be derived from the same PHI.
4027 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4028 Constant *StartCST =
4029 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4030 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
4032 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4033 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
4034 if (PN2 != PN) return getCouldNotCompute(); // Not derived from same PHI.
4036 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4037 // the loop symbolically to determine when the condition gets a value of
4039 unsigned IterationNum = 0;
4040 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4041 for (Constant *PHIVal = StartCST;
4042 IterationNum != MaxIterations; ++IterationNum) {
4043 ConstantInt *CondVal =
4044 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
4046 // Couldn't symbolically evaluate.
4047 if (!CondVal) return getCouldNotCompute();
4049 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4050 ++NumBruteForceTripCountsComputed;
4051 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4054 // Compute the value of the PHI node for the next iteration.
4055 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4056 if (NextPHI == 0 || NextPHI == PHIVal)
4057 return getCouldNotCompute();// Couldn't evaluate or not making progress...
4061 // Too many iterations were needed to evaluate.
4062 return getCouldNotCompute();
4065 /// getSCEVAtScope - Return a SCEV expression for the specified value
4066 /// at the specified scope in the program. The L value specifies a loop
4067 /// nest to evaluate the expression at, where null is the top-level or a
4068 /// specified loop is immediately inside of the loop.
4070 /// This method can be used to compute the exit value for a variable defined
4071 /// in a loop by querying what the value will hold in the parent loop.
4073 /// In the case that a relevant loop exit value cannot be computed, the
4074 /// original value V is returned.
4075 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4076 // Check to see if we've folded this expression at this loop before.
4077 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4078 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4079 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4081 return Pair.first->second ? Pair.first->second : V;
4083 // Otherwise compute it.
4084 const SCEV *C = computeSCEVAtScope(V, L);
4085 ValuesAtScopes[V][L] = C;
4089 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
4090 if (isa<SCEVConstant>(V)) return V;
4092 // If this instruction is evolved from a constant-evolving PHI, compute the
4093 // exit value from the loop without using SCEVs.
4094 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
4095 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
4096 const Loop *LI = (*this->LI)[I->getParent()];
4097 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
4098 if (PHINode *PN = dyn_cast<PHINode>(I))
4099 if (PN->getParent() == LI->getHeader()) {
4100 // Okay, there is no closed form solution for the PHI node. Check
4101 // to see if the loop that contains it has a known backedge-taken
4102 // count. If so, we may be able to force computation of the exit
4104 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
4105 if (const SCEVConstant *BTCC =
4106 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
4107 // Okay, we know how many times the containing loop executes. If
4108 // this is a constant evolving PHI node, get the final value at
4109 // the specified iteration number.
4110 Constant *RV = getConstantEvolutionLoopExitValue(PN,
4111 BTCC->getValue()->getValue(),
4113 if (RV) return getSCEV(RV);
4117 // Okay, this is an expression that we cannot symbolically evaluate
4118 // into a SCEV. Check to see if it's possible to symbolically evaluate
4119 // the arguments into constants, and if so, try to constant propagate the
4120 // result. This is particularly useful for computing loop exit values.
4121 if (CanConstantFold(I)) {
4122 std::vector<Constant*> Operands;
4123 Operands.reserve(I->getNumOperands());
4124 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4125 Value *Op = I->getOperand(i);
4126 if (Constant *C = dyn_cast<Constant>(Op)) {
4127 Operands.push_back(C);
4129 // If any of the operands is non-constant and if they are
4130 // non-integer and non-pointer, don't even try to analyze them
4131 // with scev techniques.
4132 if (!isSCEVable(Op->getType()))
4135 const SCEV *OpV = getSCEVAtScope(Op, L);
4136 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
4137 Constant *C = SC->getValue();
4138 if (C->getType() != Op->getType())
4139 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4143 Operands.push_back(C);
4144 } else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
4145 if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
4146 if (C->getType() != Op->getType())
4148 ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4152 Operands.push_back(C);
4162 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4163 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4164 Operands[0], Operands[1], TD);
4166 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4167 &Operands[0], Operands.size(), TD);
4172 // This is some other type of SCEVUnknown, just return it.
4176 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4177 // Avoid performing the look-up in the common case where the specified
4178 // expression has no loop-variant portions.
4179 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4180 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4181 if (OpAtScope != Comm->getOperand(i)) {
4182 // Okay, at least one of these operands is loop variant but might be
4183 // foldable. Build a new instance of the folded commutative expression.
4184 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4185 Comm->op_begin()+i);
4186 NewOps.push_back(OpAtScope);
4188 for (++i; i != e; ++i) {
4189 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4190 NewOps.push_back(OpAtScope);
4192 if (isa<SCEVAddExpr>(Comm))
4193 return getAddExpr(NewOps);
4194 if (isa<SCEVMulExpr>(Comm))
4195 return getMulExpr(NewOps);
4196 if (isa<SCEVSMaxExpr>(Comm))
4197 return getSMaxExpr(NewOps);
4198 if (isa<SCEVUMaxExpr>(Comm))
4199 return getUMaxExpr(NewOps);
4200 llvm_unreachable("Unknown commutative SCEV type!");
4203 // If we got here, all operands are loop invariant.
4207 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4208 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4209 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4210 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4211 return Div; // must be loop invariant
4212 return getUDivExpr(LHS, RHS);
4215 // If this is a loop recurrence for a loop that does not contain L, then we
4216 // are dealing with the final value computed by the loop.
4217 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4218 if (!L || !AddRec->getLoop()->contains(L)) {
4219 // To evaluate this recurrence, we need to know how many times the AddRec
4220 // loop iterates. Compute this now.
4221 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4222 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4224 // Then, evaluate the AddRec.
4225 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4230 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4231 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4232 if (Op == Cast->getOperand())
4233 return Cast; // must be loop invariant
4234 return getZeroExtendExpr(Op, Cast->getType());
4237 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
4238 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4239 if (Op == Cast->getOperand())
4240 return Cast; // must be loop invariant
4241 return getSignExtendExpr(Op, Cast->getType());
4244 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
4245 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4246 if (Op == Cast->getOperand())
4247 return Cast; // must be loop invariant
4248 return getTruncateExpr(Op, Cast->getType());
4251 llvm_unreachable("Unknown SCEV type!");
4255 /// getSCEVAtScope - This is a convenience function which does
4256 /// getSCEVAtScope(getSCEV(V), L).
4257 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
4258 return getSCEVAtScope(getSCEV(V), L);
4261 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
4262 /// following equation:
4264 /// A * X = B (mod N)
4266 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
4267 /// A and B isn't important.
4269 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
4270 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
4271 ScalarEvolution &SE) {
4272 uint32_t BW = A.getBitWidth();
4273 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
4274 assert(A != 0 && "A must be non-zero.");
4278 // The gcd of A and N may have only one prime factor: 2. The number of
4279 // trailing zeros in A is its multiplicity
4280 uint32_t Mult2 = A.countTrailingZeros();
4283 // 2. Check if B is divisible by D.
4285 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
4286 // is not less than multiplicity of this prime factor for D.
4287 if (B.countTrailingZeros() < Mult2)
4288 return SE.getCouldNotCompute();
4290 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
4293 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
4294 // bit width during computations.
4295 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
4296 APInt Mod(BW + 1, 0);
4297 Mod.set(BW - Mult2); // Mod = N / D
4298 APInt I = AD.multiplicativeInverse(Mod);
4300 // 4. Compute the minimum unsigned root of the equation:
4301 // I * (B / D) mod (N / D)
4302 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
4304 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
4306 return SE.getConstant(Result.trunc(BW));
4309 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
4310 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
4311 /// might be the same) or two SCEVCouldNotCompute objects.
4313 static std::pair<const SCEV *,const SCEV *>
4314 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
4315 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
4316 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
4317 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
4318 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
4320 // We currently can only solve this if the coefficients are constants.
4321 if (!LC || !MC || !NC) {
4322 const SCEV *CNC = SE.getCouldNotCompute();
4323 return std::make_pair(CNC, CNC);
4326 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
4327 const APInt &L = LC->getValue()->getValue();
4328 const APInt &M = MC->getValue()->getValue();
4329 const APInt &N = NC->getValue()->getValue();
4330 APInt Two(BitWidth, 2);
4331 APInt Four(BitWidth, 4);
4334 using namespace APIntOps;
4336 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
4337 // The B coefficient is M-N/2
4341 // The A coefficient is N/2
4342 APInt A(N.sdiv(Two));
4344 // Compute the B^2-4ac term.
4347 SqrtTerm -= Four * (A * C);
4349 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
4350 // integer value or else APInt::sqrt() will assert.
4351 APInt SqrtVal(SqrtTerm.sqrt());
4353 // Compute the two solutions for the quadratic formula.
4354 // The divisions must be performed as signed divisions.
4356 APInt TwoA( A << 1 );
4357 if (TwoA.isMinValue()) {
4358 const SCEV *CNC = SE.getCouldNotCompute();
4359 return std::make_pair(CNC, CNC);
4362 LLVMContext &Context = SE.getContext();
4364 ConstantInt *Solution1 =
4365 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
4366 ConstantInt *Solution2 =
4367 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
4369 return std::make_pair(SE.getConstant(Solution1),
4370 SE.getConstant(Solution2));
4371 } // end APIntOps namespace
4374 /// HowFarToZero - Return the number of times a backedge comparing the specified
4375 /// value to zero will execute. If not computable, return CouldNotCompute.
4376 const SCEV *ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
4377 // If the value is a constant
4378 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4379 // If the value is already zero, the branch will execute zero times.
4380 if (C->getValue()->isZero()) return C;
4381 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4384 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
4385 if (!AddRec || AddRec->getLoop() != L)
4386 return getCouldNotCompute();
4388 if (AddRec->isAffine()) {
4389 // If this is an affine expression, the execution count of this branch is
4390 // the minimum unsigned root of the following equation:
4392 // Start + Step*N = 0 (mod 2^BW)
4396 // Step*N = -Start (mod 2^BW)
4398 // where BW is the common bit width of Start and Step.
4400 // Get the initial value for the loop.
4401 const SCEV *Start = getSCEVAtScope(AddRec->getStart(),
4402 L->getParentLoop());
4403 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1),
4404 L->getParentLoop());
4406 if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
4407 // For now we handle only constant steps.
4409 // First, handle unitary steps.
4410 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
4411 return getNegativeSCEV(Start); // N = -Start (as unsigned)
4412 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
4413 return Start; // N = Start (as unsigned)
4415 // Then, try to solve the above equation provided that Start is constant.
4416 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
4417 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
4418 -StartC->getValue()->getValue(),
4421 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
4422 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
4423 // the quadratic equation to solve it.
4424 std::pair<const SCEV *,const SCEV *> Roots = SolveQuadraticEquation(AddRec,
4426 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
4427 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
4430 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
4431 << " sol#2: " << *R2 << "\n";
4433 // Pick the smallest positive root value.
4434 if (ConstantInt *CB =
4435 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
4436 R1->getValue(), R2->getValue()))) {
4437 if (CB->getZExtValue() == false)
4438 std::swap(R1, R2); // R1 is the minimum root now.
4440 // We can only use this value if the chrec ends up with an exact zero
4441 // value at this index. When solving for "X*X != 5", for example, we
4442 // should not accept a root of 2.
4443 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
4445 return R1; // We found a quadratic root!
4450 return getCouldNotCompute();
4453 /// HowFarToNonZero - Return the number of times a backedge checking the
4454 /// specified value for nonzero will execute. If not computable, return
4456 const SCEV *ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
4457 // Loops that look like: while (X == 0) are very strange indeed. We don't
4458 // handle them yet except for the trivial case. This could be expanded in the
4459 // future as needed.
4461 // If the value is a constant, check to see if it is known to be non-zero
4462 // already. If so, the backedge will execute zero times.
4463 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
4464 if (!C->getValue()->isNullValue())
4465 return getIntegerSCEV(0, C->getType());
4466 return getCouldNotCompute(); // Otherwise it will loop infinitely.
4469 // We could implement others, but I really doubt anyone writes loops like
4470 // this, and if they did, they would already be constant folded.
4471 return getCouldNotCompute();
4474 /// getLoopPredecessor - If the given loop's header has exactly one unique
4475 /// predecessor outside the loop, return it. Otherwise return null.
4477 BasicBlock *ScalarEvolution::getLoopPredecessor(const Loop *L) {
4478 BasicBlock *Header = L->getHeader();
4479 BasicBlock *Pred = 0;
4480 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header);
4482 if (!L->contains(*PI)) {
4483 if (Pred && Pred != *PI) return 0; // Multiple predecessors.
4489 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
4490 /// (which may not be an immediate predecessor) which has exactly one
4491 /// successor from which BB is reachable, or null if no such block is
4495 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
4496 // If the block has a unique predecessor, then there is no path from the
4497 // predecessor to the block that does not go through the direct edge
4498 // from the predecessor to the block.
4499 if (BasicBlock *Pred = BB->getSinglePredecessor())
4502 // A loop's header is defined to be a block that dominates the loop.
4503 // If the header has a unique predecessor outside the loop, it must be
4504 // a block that has exactly one successor that can reach the loop.
4505 if (Loop *L = LI->getLoopFor(BB))
4506 return getLoopPredecessor(L);
4511 /// HasSameValue - SCEV structural equivalence is usually sufficient for
4512 /// testing whether two expressions are equal, however for the purposes of
4513 /// looking for a condition guarding a loop, it can be useful to be a little
4514 /// more general, since a front-end may have replicated the controlling
4517 static bool HasSameValue(const SCEV *A, const SCEV *B) {
4518 // Quick check to see if they are the same SCEV.
4519 if (A == B) return true;
4521 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
4522 // two different instructions with the same value. Check for this case.
4523 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
4524 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
4525 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
4526 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
4527 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
4530 // Otherwise assume they may have a different value.
4534 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
4535 return getSignedRange(S).getSignedMax().isNegative();
4538 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
4539 return getSignedRange(S).getSignedMin().isStrictlyPositive();
4542 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
4543 return !getSignedRange(S).getSignedMin().isNegative();
4546 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
4547 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
4550 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
4551 return isKnownNegative(S) || isKnownPositive(S);
4554 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
4555 const SCEV *LHS, const SCEV *RHS) {
4557 if (HasSameValue(LHS, RHS))
4558 return ICmpInst::isTrueWhenEqual(Pred);
4562 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4564 case ICmpInst::ICMP_SGT:
4565 Pred = ICmpInst::ICMP_SLT;
4566 std::swap(LHS, RHS);
4567 case ICmpInst::ICMP_SLT: {
4568 ConstantRange LHSRange = getSignedRange(LHS);
4569 ConstantRange RHSRange = getSignedRange(RHS);
4570 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
4572 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
4576 case ICmpInst::ICMP_SGE:
4577 Pred = ICmpInst::ICMP_SLE;
4578 std::swap(LHS, RHS);
4579 case ICmpInst::ICMP_SLE: {
4580 ConstantRange LHSRange = getSignedRange(LHS);
4581 ConstantRange RHSRange = getSignedRange(RHS);
4582 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
4584 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
4588 case ICmpInst::ICMP_UGT:
4589 Pred = ICmpInst::ICMP_ULT;
4590 std::swap(LHS, RHS);
4591 case ICmpInst::ICMP_ULT: {
4592 ConstantRange LHSRange = getUnsignedRange(LHS);
4593 ConstantRange RHSRange = getUnsignedRange(RHS);
4594 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
4596 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
4600 case ICmpInst::ICMP_UGE:
4601 Pred = ICmpInst::ICMP_ULE;
4602 std::swap(LHS, RHS);
4603 case ICmpInst::ICMP_ULE: {
4604 ConstantRange LHSRange = getUnsignedRange(LHS);
4605 ConstantRange RHSRange = getUnsignedRange(RHS);
4606 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
4608 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
4612 case ICmpInst::ICMP_NE: {
4613 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
4615 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
4618 const SCEV *Diff = getMinusSCEV(LHS, RHS);
4619 if (isKnownNonZero(Diff))
4623 case ICmpInst::ICMP_EQ:
4624 // The check at the top of the function catches the case where
4625 // the values are known to be equal.
4631 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
4632 /// protected by a conditional between LHS and RHS. This is used to
4633 /// to eliminate casts.
4635 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
4636 ICmpInst::Predicate Pred,
4637 const SCEV *LHS, const SCEV *RHS) {
4638 // Interpret a null as meaning no loop, where there is obviously no guard
4639 // (interprocedural conditions notwithstanding).
4640 if (!L) return true;
4642 BasicBlock *Latch = L->getLoopLatch();
4646 BranchInst *LoopContinuePredicate =
4647 dyn_cast<BranchInst>(Latch->getTerminator());
4648 if (!LoopContinuePredicate ||
4649 LoopContinuePredicate->isUnconditional())
4652 return isImpliedCond(LoopContinuePredicate->getCondition(), Pred, LHS, RHS,
4653 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
4656 /// isLoopGuardedByCond - Test whether entry to the loop is protected
4657 /// by a conditional between LHS and RHS. This is used to help avoid max
4658 /// expressions in loop trip counts, and to eliminate casts.
4660 ScalarEvolution::isLoopGuardedByCond(const Loop *L,
4661 ICmpInst::Predicate Pred,
4662 const SCEV *LHS, const SCEV *RHS) {
4663 // Interpret a null as meaning no loop, where there is obviously no guard
4664 // (interprocedural conditions notwithstanding).
4665 if (!L) return false;
4667 BasicBlock *Predecessor = getLoopPredecessor(L);
4668 BasicBlock *PredecessorDest = L->getHeader();
4670 // Starting at the loop predecessor, climb up the predecessor chain, as long
4671 // as there are predecessors that can be found that have unique successors
4672 // leading to the original header.
4674 PredecessorDest = Predecessor,
4675 Predecessor = getPredecessorWithUniqueSuccessorForBB(Predecessor)) {
4677 BranchInst *LoopEntryPredicate =
4678 dyn_cast<BranchInst>(Predecessor->getTerminator());
4679 if (!LoopEntryPredicate ||
4680 LoopEntryPredicate->isUnconditional())
4683 if (isImpliedCond(LoopEntryPredicate->getCondition(), Pred, LHS, RHS,
4684 LoopEntryPredicate->getSuccessor(0) != PredecessorDest))
4691 /// isImpliedCond - Test whether the condition described by Pred, LHS,
4692 /// and RHS is true whenever the given Cond value evaluates to true.
4693 bool ScalarEvolution::isImpliedCond(Value *CondValue,
4694 ICmpInst::Predicate Pred,
4695 const SCEV *LHS, const SCEV *RHS,
4697 // Recursivly handle And and Or conditions.
4698 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CondValue)) {
4699 if (BO->getOpcode() == Instruction::And) {
4701 return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
4702 isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
4703 } else if (BO->getOpcode() == Instruction::Or) {
4705 return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
4706 isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
4710 ICmpInst *ICI = dyn_cast<ICmpInst>(CondValue);
4711 if (!ICI) return false;
4713 // Bail if the ICmp's operands' types are wider than the needed type
4714 // before attempting to call getSCEV on them. This avoids infinite
4715 // recursion, since the analysis of widening casts can require loop
4716 // exit condition information for overflow checking, which would
4718 if (getTypeSizeInBits(LHS->getType()) <
4719 getTypeSizeInBits(ICI->getOperand(0)->getType()))
4722 // Now that we found a conditional branch that dominates the loop, check to
4723 // see if it is the comparison we are looking for.
4724 ICmpInst::Predicate FoundPred;
4726 FoundPred = ICI->getInversePredicate();
4728 FoundPred = ICI->getPredicate();
4730 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
4731 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
4733 // Balance the types. The case where FoundLHS' type is wider than
4734 // LHS' type is checked for above.
4735 if (getTypeSizeInBits(LHS->getType()) >
4736 getTypeSizeInBits(FoundLHS->getType())) {
4737 if (CmpInst::isSigned(Pred)) {
4738 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
4739 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
4741 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
4742 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
4746 // Canonicalize the query to match the way instcombine will have
4747 // canonicalized the comparison.
4748 // First, put a constant operand on the right.
4749 if (isa<SCEVConstant>(LHS)) {
4750 std::swap(LHS, RHS);
4751 Pred = ICmpInst::getSwappedPredicate(Pred);
4753 // Then, canonicalize comparisons with boundary cases.
4754 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
4755 const APInt &RA = RC->getValue()->getValue();
4757 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4758 case ICmpInst::ICMP_EQ:
4759 case ICmpInst::ICMP_NE:
4761 case ICmpInst::ICMP_UGE:
4762 if ((RA - 1).isMinValue()) {
4763 Pred = ICmpInst::ICMP_NE;
4764 RHS = getConstant(RA - 1);
4767 if (RA.isMaxValue()) {
4768 Pred = ICmpInst::ICMP_EQ;
4771 if (RA.isMinValue()) return true;
4773 case ICmpInst::ICMP_ULE:
4774 if ((RA + 1).isMaxValue()) {
4775 Pred = ICmpInst::ICMP_NE;
4776 RHS = getConstant(RA + 1);
4779 if (RA.isMinValue()) {
4780 Pred = ICmpInst::ICMP_EQ;
4783 if (RA.isMaxValue()) return true;
4785 case ICmpInst::ICMP_SGE:
4786 if ((RA - 1).isMinSignedValue()) {
4787 Pred = ICmpInst::ICMP_NE;
4788 RHS = getConstant(RA - 1);
4791 if (RA.isMaxSignedValue()) {
4792 Pred = ICmpInst::ICMP_EQ;
4795 if (RA.isMinSignedValue()) return true;
4797 case ICmpInst::ICMP_SLE:
4798 if ((RA + 1).isMaxSignedValue()) {
4799 Pred = ICmpInst::ICMP_NE;
4800 RHS = getConstant(RA + 1);
4803 if (RA.isMinSignedValue()) {
4804 Pred = ICmpInst::ICMP_EQ;
4807 if (RA.isMaxSignedValue()) return true;
4809 case ICmpInst::ICMP_UGT:
4810 if (RA.isMinValue()) {
4811 Pred = ICmpInst::ICMP_NE;
4814 if ((RA + 1).isMaxValue()) {
4815 Pred = ICmpInst::ICMP_EQ;
4816 RHS = getConstant(RA + 1);
4819 if (RA.isMaxValue()) return false;
4821 case ICmpInst::ICMP_ULT:
4822 if (RA.isMaxValue()) {
4823 Pred = ICmpInst::ICMP_NE;
4826 if ((RA - 1).isMinValue()) {
4827 Pred = ICmpInst::ICMP_EQ;
4828 RHS = getConstant(RA - 1);
4831 if (RA.isMinValue()) return false;
4833 case ICmpInst::ICMP_SGT:
4834 if (RA.isMinSignedValue()) {
4835 Pred = ICmpInst::ICMP_NE;
4838 if ((RA + 1).isMaxSignedValue()) {
4839 Pred = ICmpInst::ICMP_EQ;
4840 RHS = getConstant(RA + 1);
4843 if (RA.isMaxSignedValue()) return false;
4845 case ICmpInst::ICMP_SLT:
4846 if (RA.isMaxSignedValue()) {
4847 Pred = ICmpInst::ICMP_NE;
4850 if ((RA - 1).isMinSignedValue()) {
4851 Pred = ICmpInst::ICMP_EQ;
4852 RHS = getConstant(RA - 1);
4855 if (RA.isMinSignedValue()) return false;
4860 // Check to see if we can make the LHS or RHS match.
4861 if (LHS == FoundRHS || RHS == FoundLHS) {
4862 if (isa<SCEVConstant>(RHS)) {
4863 std::swap(FoundLHS, FoundRHS);
4864 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
4866 std::swap(LHS, RHS);
4867 Pred = ICmpInst::getSwappedPredicate(Pred);
4871 // Check whether the found predicate is the same as the desired predicate.
4872 if (FoundPred == Pred)
4873 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
4875 // Check whether swapping the found predicate makes it the same as the
4876 // desired predicate.
4877 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
4878 if (isa<SCEVConstant>(RHS))
4879 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
4881 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
4882 RHS, LHS, FoundLHS, FoundRHS);
4885 // Check whether the actual condition is beyond sufficient.
4886 if (FoundPred == ICmpInst::ICMP_EQ)
4887 if (ICmpInst::isTrueWhenEqual(Pred))
4888 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
4890 if (Pred == ICmpInst::ICMP_NE)
4891 if (!ICmpInst::isTrueWhenEqual(FoundPred))
4892 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
4895 // Otherwise assume the worst.
4899 /// isImpliedCondOperands - Test whether the condition described by Pred,
4900 /// LHS, and RHS is true whenever the condition desribed by Pred, FoundLHS,
4901 /// and FoundRHS is true.
4902 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
4903 const SCEV *LHS, const SCEV *RHS,
4904 const SCEV *FoundLHS,
4905 const SCEV *FoundRHS) {
4906 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
4907 FoundLHS, FoundRHS) ||
4908 // ~x < ~y --> x > y
4909 isImpliedCondOperandsHelper(Pred, LHS, RHS,
4910 getNotSCEV(FoundRHS),
4911 getNotSCEV(FoundLHS));
4914 /// isImpliedCondOperandsHelper - Test whether the condition described by
4915 /// Pred, LHS, and RHS is true whenever the condition desribed by Pred,
4916 /// FoundLHS, and FoundRHS is true.
4918 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
4919 const SCEV *LHS, const SCEV *RHS,
4920 const SCEV *FoundLHS,
4921 const SCEV *FoundRHS) {
4923 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
4924 case ICmpInst::ICMP_EQ:
4925 case ICmpInst::ICMP_NE:
4926 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
4929 case ICmpInst::ICMP_SLT:
4930 case ICmpInst::ICMP_SLE:
4931 if (isKnownPredicate(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
4932 isKnownPredicate(ICmpInst::ICMP_SGE, RHS, FoundRHS))
4935 case ICmpInst::ICMP_SGT:
4936 case ICmpInst::ICMP_SGE:
4937 if (isKnownPredicate(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
4938 isKnownPredicate(ICmpInst::ICMP_SLE, RHS, FoundRHS))
4941 case ICmpInst::ICMP_ULT:
4942 case ICmpInst::ICMP_ULE:
4943 if (isKnownPredicate(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
4944 isKnownPredicate(ICmpInst::ICMP_UGE, RHS, FoundRHS))
4947 case ICmpInst::ICMP_UGT:
4948 case ICmpInst::ICMP_UGE:
4949 if (isKnownPredicate(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
4950 isKnownPredicate(ICmpInst::ICMP_ULE, RHS, FoundRHS))
4958 /// getBECount - Subtract the end and start values and divide by the step,
4959 /// rounding up, to get the number of times the backedge is executed. Return
4960 /// CouldNotCompute if an intermediate computation overflows.
4961 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
4965 assert(!isKnownNegative(Step) &&
4966 "This code doesn't handle negative strides yet!");
4968 const Type *Ty = Start->getType();
4969 const SCEV *NegOne = getIntegerSCEV(-1, Ty);
4970 const SCEV *Diff = getMinusSCEV(End, Start);
4971 const SCEV *RoundUp = getAddExpr(Step, NegOne);
4973 // Add an adjustment to the difference between End and Start so that
4974 // the division will effectively round up.
4975 const SCEV *Add = getAddExpr(Diff, RoundUp);
4978 // Check Add for unsigned overflow.
4979 // TODO: More sophisticated things could be done here.
4980 const Type *WideTy = IntegerType::get(getContext(),
4981 getTypeSizeInBits(Ty) + 1);
4982 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
4983 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
4984 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
4985 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
4986 return getCouldNotCompute();
4989 return getUDivExpr(Add, Step);
4992 /// HowManyLessThans - Return the number of times a backedge containing the
4993 /// specified less-than comparison will execute. If not computable, return
4994 /// CouldNotCompute.
4995 ScalarEvolution::BackedgeTakenInfo
4996 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
4997 const Loop *L, bool isSigned) {
4998 // Only handle: "ADDREC < LoopInvariant".
4999 if (!RHS->isLoopInvariant(L)) return getCouldNotCompute();
5001 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
5002 if (!AddRec || AddRec->getLoop() != L)
5003 return getCouldNotCompute();
5005 // Check to see if we have a flag which makes analysis easy.
5006 bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
5007 AddRec->hasNoUnsignedWrap();
5009 if (AddRec->isAffine()) {
5010 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
5011 const SCEV *Step = AddRec->getStepRecurrence(*this);
5014 return getCouldNotCompute();
5015 if (Step->isOne()) {
5016 // With unit stride, the iteration never steps past the limit value.
5017 } else if (isKnownPositive(Step)) {
5018 // Test whether a positive iteration iteration can step past the limit
5019 // value and past the maximum value for its type in a single step.
5020 // Note that it's not sufficient to check NoWrap here, because even
5021 // though the value after a wrap is undefined, it's not undefined
5022 // behavior, so if wrap does occur, the loop could either terminate or
5023 // loop infinitely, but in either case, the loop is guaranteed to
5024 // iterate at least until the iteration where the wrapping occurs.
5025 const SCEV *One = getIntegerSCEV(1, Step->getType());
5027 APInt Max = APInt::getSignedMaxValue(BitWidth);
5028 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
5029 .slt(getSignedRange(RHS).getSignedMax()))
5030 return getCouldNotCompute();
5032 APInt Max = APInt::getMaxValue(BitWidth);
5033 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
5034 .ult(getUnsignedRange(RHS).getUnsignedMax()))
5035 return getCouldNotCompute();
5038 // TODO: Handle negative strides here and below.
5039 return getCouldNotCompute();
5041 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
5042 // m. So, we count the number of iterations in which {n,+,s} < m is true.
5043 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
5044 // treat m-n as signed nor unsigned due to overflow possibility.
5046 // First, we get the value of the LHS in the first iteration: n
5047 const SCEV *Start = AddRec->getOperand(0);
5049 // Determine the minimum constant start value.
5050 const SCEV *MinStart = getConstant(isSigned ?
5051 getSignedRange(Start).getSignedMin() :
5052 getUnsignedRange(Start).getUnsignedMin());
5054 // If we know that the condition is true in order to enter the loop,
5055 // then we know that it will run exactly (m-n)/s times. Otherwise, we
5056 // only know that it will execute (max(m,n)-n)/s times. In both cases,
5057 // the division must round up.
5058 const SCEV *End = RHS;
5059 if (!isLoopGuardedByCond(L,
5060 isSigned ? ICmpInst::ICMP_SLT :
5062 getMinusSCEV(Start, Step), RHS))
5063 End = isSigned ? getSMaxExpr(RHS, Start)
5064 : getUMaxExpr(RHS, Start);
5066 // Determine the maximum constant end value.
5067 const SCEV *MaxEnd = getConstant(isSigned ?
5068 getSignedRange(End).getSignedMax() :
5069 getUnsignedRange(End).getUnsignedMax());
5071 // If MaxEnd is within a step of the maximum integer value in its type,
5072 // adjust it down to the minimum value which would produce the same effect.
5073 // This allows the subsequent ceiling divison of (N+(step-1))/step to
5074 // compute the correct value.
5075 const SCEV *StepMinusOne = getMinusSCEV(Step,
5076 getIntegerSCEV(1, Step->getType()));
5079 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
5082 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
5085 // Finally, we subtract these two values and divide, rounding up, to get
5086 // the number of times the backedge is executed.
5087 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
5089 // The maximum backedge count is similar, except using the minimum start
5090 // value and the maximum end value.
5091 const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
5093 return BackedgeTakenInfo(BECount, MaxBECount);
5096 return getCouldNotCompute();
5099 /// getNumIterationsInRange - Return the number of iterations of this loop that
5100 /// produce values in the specified constant range. Another way of looking at
5101 /// this is that it returns the first iteration number where the value is not in
5102 /// the condition, thus computing the exit count. If the iteration count can't
5103 /// be computed, an instance of SCEVCouldNotCompute is returned.
5104 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
5105 ScalarEvolution &SE) const {
5106 if (Range.isFullSet()) // Infinite loop.
5107 return SE.getCouldNotCompute();
5109 // If the start is a non-zero constant, shift the range to simplify things.
5110 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
5111 if (!SC->getValue()->isZero()) {
5112 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
5113 Operands[0] = SE.getIntegerSCEV(0, SC->getType());
5114 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
5115 if (const SCEVAddRecExpr *ShiftedAddRec =
5116 dyn_cast<SCEVAddRecExpr>(Shifted))
5117 return ShiftedAddRec->getNumIterationsInRange(
5118 Range.subtract(SC->getValue()->getValue()), SE);
5119 // This is strange and shouldn't happen.
5120 return SE.getCouldNotCompute();
5123 // The only time we can solve this is when we have all constant indices.
5124 // Otherwise, we cannot determine the overflow conditions.
5125 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
5126 if (!isa<SCEVConstant>(getOperand(i)))
5127 return SE.getCouldNotCompute();
5130 // Okay at this point we know that all elements of the chrec are constants and
5131 // that the start element is zero.
5133 // First check to see if the range contains zero. If not, the first
5135 unsigned BitWidth = SE.getTypeSizeInBits(getType());
5136 if (!Range.contains(APInt(BitWidth, 0)))
5137 return SE.getIntegerSCEV(0, getType());
5140 // If this is an affine expression then we have this situation:
5141 // Solve {0,+,A} in Range === Ax in Range
5143 // We know that zero is in the range. If A is positive then we know that
5144 // the upper value of the range must be the first possible exit value.
5145 // If A is negative then the lower of the range is the last possible loop
5146 // value. Also note that we already checked for a full range.
5147 APInt One(BitWidth,1);
5148 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
5149 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
5151 // The exit value should be (End+A)/A.
5152 APInt ExitVal = (End + A).udiv(A);
5153 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
5155 // Evaluate at the exit value. If we really did fall out of the valid
5156 // range, then we computed our trip count, otherwise wrap around or other
5157 // things must have happened.
5158 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
5159 if (Range.contains(Val->getValue()))
5160 return SE.getCouldNotCompute(); // Something strange happened
5162 // Ensure that the previous value is in the range. This is a sanity check.
5163 assert(Range.contains(
5164 EvaluateConstantChrecAtConstant(this,
5165 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
5166 "Linear scev computation is off in a bad way!");
5167 return SE.getConstant(ExitValue);
5168 } else if (isQuadratic()) {
5169 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
5170 // quadratic equation to solve it. To do this, we must frame our problem in
5171 // terms of figuring out when zero is crossed, instead of when
5172 // Range.getUpper() is crossed.
5173 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
5174 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
5175 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
5177 // Next, solve the constructed addrec
5178 std::pair<const SCEV *,const SCEV *> Roots =
5179 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
5180 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5181 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5183 // Pick the smallest positive root value.
5184 if (ConstantInt *CB =
5185 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
5186 R1->getValue(), R2->getValue()))) {
5187 if (CB->getZExtValue() == false)
5188 std::swap(R1, R2); // R1 is the minimum root now.
5190 // Make sure the root is not off by one. The returned iteration should
5191 // not be in the range, but the previous one should be. When solving
5192 // for "X*X < 5", for example, we should not return a root of 2.
5193 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
5196 if (Range.contains(R1Val->getValue())) {
5197 // The next iteration must be out of the range...
5198 ConstantInt *NextVal =
5199 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
5201 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5202 if (!Range.contains(R1Val->getValue()))
5203 return SE.getConstant(NextVal);
5204 return SE.getCouldNotCompute(); // Something strange happened
5207 // If R1 was not in the range, then it is a good return value. Make
5208 // sure that R1-1 WAS in the range though, just in case.
5209 ConstantInt *NextVal =
5210 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
5211 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
5212 if (Range.contains(R1Val->getValue()))
5214 return SE.getCouldNotCompute(); // Something strange happened
5219 return SE.getCouldNotCompute();
5224 //===----------------------------------------------------------------------===//
5225 // SCEVCallbackVH Class Implementation
5226 //===----------------------------------------------------------------------===//
5228 void ScalarEvolution::SCEVCallbackVH::deleted() {
5229 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5230 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
5231 SE->ConstantEvolutionLoopExitValue.erase(PN);
5232 SE->Scalars.erase(getValPtr());
5233 // this now dangles!
5236 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *) {
5237 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
5239 // Forget all the expressions associated with users of the old value,
5240 // so that future queries will recompute the expressions using the new
5242 SmallVector<User *, 16> Worklist;
5243 SmallPtrSet<User *, 8> Visited;
5244 Value *Old = getValPtr();
5245 bool DeleteOld = false;
5246 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
5248 Worklist.push_back(*UI);
5249 while (!Worklist.empty()) {
5250 User *U = Worklist.pop_back_val();
5251 // Deleting the Old value will cause this to dangle. Postpone
5252 // that until everything else is done.
5257 if (!Visited.insert(U))
5259 if (PHINode *PN = dyn_cast<PHINode>(U))
5260 SE->ConstantEvolutionLoopExitValue.erase(PN);
5261 SE->Scalars.erase(U);
5262 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
5264 Worklist.push_back(*UI);
5266 // Delete the Old value if it (indirectly) references itself.
5268 if (PHINode *PN = dyn_cast<PHINode>(Old))
5269 SE->ConstantEvolutionLoopExitValue.erase(PN);
5270 SE->Scalars.erase(Old);
5271 // this now dangles!
5276 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
5277 : CallbackVH(V), SE(se) {}
5279 //===----------------------------------------------------------------------===//
5280 // ScalarEvolution Class Implementation
5281 //===----------------------------------------------------------------------===//
5283 ScalarEvolution::ScalarEvolution()
5284 : FunctionPass(&ID) {
5287 bool ScalarEvolution::runOnFunction(Function &F) {
5289 LI = &getAnalysis<LoopInfo>();
5290 DT = &getAnalysis<DominatorTree>();
5291 TD = getAnalysisIfAvailable<TargetData>();
5295 void ScalarEvolution::releaseMemory() {
5297 BackedgeTakenCounts.clear();
5298 ConstantEvolutionLoopExitValue.clear();
5299 ValuesAtScopes.clear();
5300 UniqueSCEVs.clear();
5301 SCEVAllocator.Reset();
5304 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
5305 AU.setPreservesAll();
5306 AU.addRequiredTransitive<LoopInfo>();
5307 AU.addRequiredTransitive<DominatorTree>();
5310 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
5311 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
5314 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
5316 // Print all inner loops first
5317 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
5318 PrintLoopInfo(OS, SE, *I);
5321 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5324 SmallVector<BasicBlock *, 8> ExitBlocks;
5325 L->getExitBlocks(ExitBlocks);
5326 if (ExitBlocks.size() != 1)
5327 OS << "<multiple exits> ";
5329 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
5330 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
5332 OS << "Unpredictable backedge-taken count. ";
5337 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
5340 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
5341 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
5343 OS << "Unpredictable max backedge-taken count. ";
5349 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
5350 // ScalarEvolution's implementaiton of the print method is to print
5351 // out SCEV values of all instructions that are interesting. Doing
5352 // this potentially causes it to create new SCEV objects though,
5353 // which technically conflicts with the const qualifier. This isn't
5354 // observable from outside the class though, so casting away the
5355 // const isn't dangerous.
5356 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
5358 OS << "Classifying expressions for: ";
5359 WriteAsOperand(OS, F, /*PrintType=*/false);
5361 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
5362 if (isSCEVable(I->getType())) {
5365 const SCEV *SV = SE.getSCEV(&*I);
5368 const Loop *L = LI->getLoopFor((*I).getParent());
5370 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
5377 OS << "\t\t" "Exits: ";
5378 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
5379 if (!ExitValue->isLoopInvariant(L)) {
5380 OS << "<<Unknown>>";
5389 OS << "Determining loop execution counts for: ";
5390 WriteAsOperand(OS, F, /*PrintType=*/false);
5392 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
5393 PrintLoopInfo(OS, &SE, *I);