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/ScalarEvolution.h"
63 #include "llvm/ADT/STLExtras.h"
64 #include "llvm/ADT/SmallPtrSet.h"
65 #include "llvm/ADT/Statistic.h"
66 #include "llvm/Analysis/ConstantFolding.h"
67 #include "llvm/Analysis/Dominators.h"
68 #include "llvm/Analysis/InstructionSimplify.h"
69 #include "llvm/Analysis/LoopInfo.h"
70 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
71 #include "llvm/Analysis/ValueTracking.h"
72 #include "llvm/Assembly/Writer.h"
73 #include "llvm/IR/Constants.h"
74 #include "llvm/IR/DataLayout.h"
75 #include "llvm/IR/DerivedTypes.h"
76 #include "llvm/IR/GlobalAlias.h"
77 #include "llvm/IR/GlobalVariable.h"
78 #include "llvm/IR/Instructions.h"
79 #include "llvm/IR/LLVMContext.h"
80 #include "llvm/IR/Operator.h"
81 #include "llvm/Support/CommandLine.h"
82 #include "llvm/Support/ConstantRange.h"
83 #include "llvm/Support/Debug.h"
84 #include "llvm/Support/ErrorHandling.h"
85 #include "llvm/Support/GetElementPtrTypeIterator.h"
86 #include "llvm/Support/InstIterator.h"
87 #include "llvm/Support/MathExtras.h"
88 #include "llvm/Support/raw_ostream.h"
89 #include "llvm/Target/TargetLibraryInfo.h"
93 STATISTIC(NumArrayLenItCounts,
94 "Number of trip counts computed with array length");
95 STATISTIC(NumTripCountsComputed,
96 "Number of loops with predictable loop counts");
97 STATISTIC(NumTripCountsNotComputed,
98 "Number of loops without predictable loop counts");
99 STATISTIC(NumBruteForceTripCountsComputed,
100 "Number of loops with trip counts computed by force");
102 static cl::opt<unsigned>
103 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
104 cl::desc("Maximum number of iterations SCEV will "
105 "symbolically execute a constant "
109 // FIXME: Enable this with XDEBUG when the test suite is clean.
111 VerifySCEV("verify-scev",
112 cl::desc("Verify ScalarEvolution's backedge taken counts (slow)"));
114 INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
115 "Scalar Evolution Analysis", false, true)
116 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
117 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
118 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
119 INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
120 "Scalar Evolution Analysis", false, true)
121 char ScalarEvolution::ID = 0;
123 //===----------------------------------------------------------------------===//
124 // SCEV class definitions
125 //===----------------------------------------------------------------------===//
127 //===----------------------------------------------------------------------===//
128 // Implementation of the SCEV class.
131 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
132 void SCEV::dump() const {
138 void SCEV::print(raw_ostream &OS) const {
139 switch (getSCEVType()) {
141 WriteAsOperand(OS, cast<SCEVConstant>(this)->getValue(), false);
144 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(this);
145 const SCEV *Op = Trunc->getOperand();
146 OS << "(trunc " << *Op->getType() << " " << *Op << " to "
147 << *Trunc->getType() << ")";
151 const SCEVZeroExtendExpr *ZExt = cast<SCEVZeroExtendExpr>(this);
152 const SCEV *Op = ZExt->getOperand();
153 OS << "(zext " << *Op->getType() << " " << *Op << " to "
154 << *ZExt->getType() << ")";
158 const SCEVSignExtendExpr *SExt = cast<SCEVSignExtendExpr>(this);
159 const SCEV *Op = SExt->getOperand();
160 OS << "(sext " << *Op->getType() << " " << *Op << " to "
161 << *SExt->getType() << ")";
165 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(this);
166 OS << "{" << *AR->getOperand(0);
167 for (unsigned i = 1, e = AR->getNumOperands(); i != e; ++i)
168 OS << ",+," << *AR->getOperand(i);
170 if (AR->getNoWrapFlags(FlagNUW))
172 if (AR->getNoWrapFlags(FlagNSW))
174 if (AR->getNoWrapFlags(FlagNW) &&
175 !AR->getNoWrapFlags((NoWrapFlags)(FlagNUW | FlagNSW)))
177 WriteAsOperand(OS, AR->getLoop()->getHeader(), /*PrintType=*/false);
185 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(this);
186 const char *OpStr = 0;
187 switch (NAry->getSCEVType()) {
188 case scAddExpr: OpStr = " + "; break;
189 case scMulExpr: OpStr = " * "; break;
190 case scUMaxExpr: OpStr = " umax "; break;
191 case scSMaxExpr: OpStr = " smax "; break;
194 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
197 if (llvm::next(I) != E)
201 switch (NAry->getSCEVType()) {
204 if (NAry->getNoWrapFlags(FlagNUW))
206 if (NAry->getNoWrapFlags(FlagNSW))
212 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(this);
213 OS << "(" << *UDiv->getLHS() << " /u " << *UDiv->getRHS() << ")";
217 const SCEVUnknown *U = cast<SCEVUnknown>(this);
219 if (U->isSizeOf(AllocTy)) {
220 OS << "sizeof(" << *AllocTy << ")";
223 if (U->isAlignOf(AllocTy)) {
224 OS << "alignof(" << *AllocTy << ")";
230 if (U->isOffsetOf(CTy, FieldNo)) {
231 OS << "offsetof(" << *CTy << ", ";
232 WriteAsOperand(OS, FieldNo, false);
237 // Otherwise just print it normally.
238 WriteAsOperand(OS, U->getValue(), false);
241 case scCouldNotCompute:
242 OS << "***COULDNOTCOMPUTE***";
246 llvm_unreachable("Unknown SCEV kind!");
249 Type *SCEV::getType() const {
250 switch (getSCEVType()) {
252 return cast<SCEVConstant>(this)->getType();
256 return cast<SCEVCastExpr>(this)->getType();
261 return cast<SCEVNAryExpr>(this)->getType();
263 return cast<SCEVAddExpr>(this)->getType();
265 return cast<SCEVUDivExpr>(this)->getType();
267 return cast<SCEVUnknown>(this)->getType();
268 case scCouldNotCompute:
269 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
271 llvm_unreachable("Unknown SCEV kind!");
275 bool SCEV::isZero() const {
276 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
277 return SC->getValue()->isZero();
281 bool SCEV::isOne() const {
282 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
283 return SC->getValue()->isOne();
287 bool SCEV::isAllOnesValue() const {
288 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
289 return SC->getValue()->isAllOnesValue();
293 /// isNonConstantNegative - Return true if the specified scev is negated, but
295 bool SCEV::isNonConstantNegative() const {
296 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(this);
297 if (!Mul) return false;
299 // If there is a constant factor, it will be first.
300 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
301 if (!SC) return false;
303 // Return true if the value is negative, this matches things like (-42 * V).
304 return SC->getValue()->getValue().isNegative();
307 SCEVCouldNotCompute::SCEVCouldNotCompute() :
308 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
310 bool SCEVCouldNotCompute::classof(const SCEV *S) {
311 return S->getSCEVType() == scCouldNotCompute;
314 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
316 ID.AddInteger(scConstant);
319 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
320 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V);
321 UniqueSCEVs.InsertNode(S, IP);
325 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
326 return getConstant(ConstantInt::get(getContext(), Val));
330 ScalarEvolution::getConstant(Type *Ty, uint64_t V, bool isSigned) {
331 IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
332 return getConstant(ConstantInt::get(ITy, V, isSigned));
335 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
336 unsigned SCEVTy, const SCEV *op, Type *ty)
337 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
339 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
340 const SCEV *op, Type *ty)
341 : SCEVCastExpr(ID, scTruncate, op, ty) {
342 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
343 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
344 "Cannot truncate non-integer value!");
347 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
348 const SCEV *op, Type *ty)
349 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
350 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
351 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
352 "Cannot zero extend non-integer value!");
355 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
356 const SCEV *op, Type *ty)
357 : SCEVCastExpr(ID, scSignExtend, op, ty) {
358 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
359 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
360 "Cannot sign extend non-integer value!");
363 void SCEVUnknown::deleted() {
364 // Clear this SCEVUnknown from various maps.
365 SE->forgetMemoizedResults(this);
367 // Remove this SCEVUnknown from the uniquing map.
368 SE->UniqueSCEVs.RemoveNode(this);
370 // Release the value.
374 void SCEVUnknown::allUsesReplacedWith(Value *New) {
375 // Clear this SCEVUnknown from various maps.
376 SE->forgetMemoizedResults(this);
378 // Remove this SCEVUnknown from the uniquing map.
379 SE->UniqueSCEVs.RemoveNode(this);
381 // Update this SCEVUnknown to point to the new value. This is needed
382 // because there may still be outstanding SCEVs which still point to
387 bool SCEVUnknown::isSizeOf(Type *&AllocTy) const {
388 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
389 if (VCE->getOpcode() == Instruction::PtrToInt)
390 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
391 if (CE->getOpcode() == Instruction::GetElementPtr &&
392 CE->getOperand(0)->isNullValue() &&
393 CE->getNumOperands() == 2)
394 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
396 AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
404 bool SCEVUnknown::isAlignOf(Type *&AllocTy) const {
405 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
406 if (VCE->getOpcode() == Instruction::PtrToInt)
407 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
408 if (CE->getOpcode() == Instruction::GetElementPtr &&
409 CE->getOperand(0)->isNullValue()) {
411 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
412 if (StructType *STy = dyn_cast<StructType>(Ty))
413 if (!STy->isPacked() &&
414 CE->getNumOperands() == 3 &&
415 CE->getOperand(1)->isNullValue()) {
416 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
418 STy->getNumElements() == 2 &&
419 STy->getElementType(0)->isIntegerTy(1)) {
420 AllocTy = STy->getElementType(1);
429 bool SCEVUnknown::isOffsetOf(Type *&CTy, Constant *&FieldNo) const {
430 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
431 if (VCE->getOpcode() == Instruction::PtrToInt)
432 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
433 if (CE->getOpcode() == Instruction::GetElementPtr &&
434 CE->getNumOperands() == 3 &&
435 CE->getOperand(0)->isNullValue() &&
436 CE->getOperand(1)->isNullValue()) {
438 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
439 // Ignore vector types here so that ScalarEvolutionExpander doesn't
440 // emit getelementptrs that index into vectors.
441 if (Ty->isStructTy() || Ty->isArrayTy()) {
443 FieldNo = CE->getOperand(2);
451 //===----------------------------------------------------------------------===//
453 //===----------------------------------------------------------------------===//
456 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
457 /// than the complexity of the RHS. This comparator is used to canonicalize
459 class SCEVComplexityCompare {
460 const LoopInfo *const LI;
462 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
464 // Return true or false if LHS is less than, or at least RHS, respectively.
465 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
466 return compare(LHS, RHS) < 0;
469 // Return negative, zero, or positive, if LHS is less than, equal to, or
470 // greater than RHS, respectively. A three-way result allows recursive
471 // comparisons to be more efficient.
472 int compare(const SCEV *LHS, const SCEV *RHS) const {
473 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
477 // Primarily, sort the SCEVs by their getSCEVType().
478 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
480 return (int)LType - (int)RType;
482 // Aside from the getSCEVType() ordering, the particular ordering
483 // isn't very important except that it's beneficial to be consistent,
484 // so that (a + b) and (b + a) don't end up as different expressions.
487 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS);
488 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
490 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
491 // not as complete as it could be.
492 const Value *LV = LU->getValue(), *RV = RU->getValue();
494 // Order pointer values after integer values. This helps SCEVExpander
496 bool LIsPointer = LV->getType()->isPointerTy(),
497 RIsPointer = RV->getType()->isPointerTy();
498 if (LIsPointer != RIsPointer)
499 return (int)LIsPointer - (int)RIsPointer;
501 // Compare getValueID values.
502 unsigned LID = LV->getValueID(),
503 RID = RV->getValueID();
505 return (int)LID - (int)RID;
507 // Sort arguments by their position.
508 if (const Argument *LA = dyn_cast<Argument>(LV)) {
509 const Argument *RA = cast<Argument>(RV);
510 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo();
511 return (int)LArgNo - (int)RArgNo;
514 // For instructions, compare their loop depth, and their operand
515 // count. This is pretty loose.
516 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) {
517 const Instruction *RInst = cast<Instruction>(RV);
519 // Compare loop depths.
520 const BasicBlock *LParent = LInst->getParent(),
521 *RParent = RInst->getParent();
522 if (LParent != RParent) {
523 unsigned LDepth = LI->getLoopDepth(LParent),
524 RDepth = LI->getLoopDepth(RParent);
525 if (LDepth != RDepth)
526 return (int)LDepth - (int)RDepth;
529 // Compare the number of operands.
530 unsigned LNumOps = LInst->getNumOperands(),
531 RNumOps = RInst->getNumOperands();
532 return (int)LNumOps - (int)RNumOps;
539 const SCEVConstant *LC = cast<SCEVConstant>(LHS);
540 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
542 // Compare constant values.
543 const APInt &LA = LC->getValue()->getValue();
544 const APInt &RA = RC->getValue()->getValue();
545 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth();
546 if (LBitWidth != RBitWidth)
547 return (int)LBitWidth - (int)RBitWidth;
548 return LA.ult(RA) ? -1 : 1;
552 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS);
553 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
555 // Compare addrec loop depths.
556 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop();
557 if (LLoop != RLoop) {
558 unsigned LDepth = LLoop->getLoopDepth(),
559 RDepth = RLoop->getLoopDepth();
560 if (LDepth != RDepth)
561 return (int)LDepth - (int)RDepth;
564 // Addrec complexity grows with operand count.
565 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands();
566 if (LNumOps != RNumOps)
567 return (int)LNumOps - (int)RNumOps;
569 // Lexicographically compare.
570 for (unsigned i = 0; i != LNumOps; ++i) {
571 long X = compare(LA->getOperand(i), RA->getOperand(i));
583 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS);
584 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
586 // Lexicographically compare n-ary expressions.
587 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
588 if (LNumOps != RNumOps)
589 return (int)LNumOps - (int)RNumOps;
591 for (unsigned i = 0; i != LNumOps; ++i) {
594 long X = compare(LC->getOperand(i), RC->getOperand(i));
598 return (int)LNumOps - (int)RNumOps;
602 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
603 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
605 // Lexicographically compare udiv expressions.
606 long X = compare(LC->getLHS(), RC->getLHS());
609 return compare(LC->getRHS(), RC->getRHS());
615 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
616 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
618 // Compare cast expressions by operand.
619 return compare(LC->getOperand(), RC->getOperand());
623 llvm_unreachable("Unknown SCEV kind!");
629 /// GroupByComplexity - Given a list of SCEV objects, order them by their
630 /// complexity, and group objects of the same complexity together by value.
631 /// When this routine is finished, we know that any duplicates in the vector are
632 /// consecutive and that complexity is monotonically increasing.
634 /// Note that we go take special precautions to ensure that we get deterministic
635 /// results from this routine. In other words, we don't want the results of
636 /// this to depend on where the addresses of various SCEV objects happened to
639 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
641 if (Ops.size() < 2) return; // Noop
642 if (Ops.size() == 2) {
643 // This is the common case, which also happens to be trivially simple.
645 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
646 if (SCEVComplexityCompare(LI)(RHS, LHS))
651 // Do the rough sort by complexity.
652 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
654 // Now that we are sorted by complexity, group elements of the same
655 // complexity. Note that this is, at worst, N^2, but the vector is likely to
656 // be extremely short in practice. Note that we take this approach because we
657 // do not want to depend on the addresses of the objects we are grouping.
658 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
659 const SCEV *S = Ops[i];
660 unsigned Complexity = S->getSCEVType();
662 // If there are any objects of the same complexity and same value as this
664 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
665 if (Ops[j] == S) { // Found a duplicate.
666 // Move it to immediately after i'th element.
667 std::swap(Ops[i+1], Ops[j]);
668 ++i; // no need to rescan it.
669 if (i == e-2) return; // Done!
677 //===----------------------------------------------------------------------===//
678 // Simple SCEV method implementations
679 //===----------------------------------------------------------------------===//
681 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
683 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
686 // Handle the simplest case efficiently.
688 return SE.getTruncateOrZeroExtend(It, ResultTy);
690 // We are using the following formula for BC(It, K):
692 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
694 // Suppose, W is the bitwidth of the return value. We must be prepared for
695 // overflow. Hence, we must assure that the result of our computation is
696 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
697 // safe in modular arithmetic.
699 // However, this code doesn't use exactly that formula; the formula it uses
700 // is something like the following, where T is the number of factors of 2 in
701 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
704 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
706 // This formula is trivially equivalent to the previous formula. However,
707 // this formula can be implemented much more efficiently. The trick is that
708 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
709 // arithmetic. To do exact division in modular arithmetic, all we have
710 // to do is multiply by the inverse. Therefore, this step can be done at
713 // The next issue is how to safely do the division by 2^T. The way this
714 // is done is by doing the multiplication step at a width of at least W + T
715 // bits. This way, the bottom W+T bits of the product are accurate. Then,
716 // when we perform the division by 2^T (which is equivalent to a right shift
717 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
718 // truncated out after the division by 2^T.
720 // In comparison to just directly using the first formula, this technique
721 // is much more efficient; using the first formula requires W * K bits,
722 // but this formula less than W + K bits. Also, the first formula requires
723 // a division step, whereas this formula only requires multiplies and shifts.
725 // It doesn't matter whether the subtraction step is done in the calculation
726 // width or the input iteration count's width; if the subtraction overflows,
727 // the result must be zero anyway. We prefer here to do it in the width of
728 // the induction variable because it helps a lot for certain cases; CodeGen
729 // isn't smart enough to ignore the overflow, which leads to much less
730 // efficient code if the width of the subtraction is wider than the native
733 // (It's possible to not widen at all by pulling out factors of 2 before
734 // the multiplication; for example, K=2 can be calculated as
735 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
736 // extra arithmetic, so it's not an obvious win, and it gets
737 // much more complicated for K > 3.)
739 // Protection from insane SCEVs; this bound is conservative,
740 // but it probably doesn't matter.
742 return SE.getCouldNotCompute();
744 unsigned W = SE.getTypeSizeInBits(ResultTy);
746 // Calculate K! / 2^T and T; we divide out the factors of two before
747 // multiplying for calculating K! / 2^T to avoid overflow.
748 // Other overflow doesn't matter because we only care about the bottom
749 // W bits of the result.
750 APInt OddFactorial(W, 1);
752 for (unsigned i = 3; i <= K; ++i) {
754 unsigned TwoFactors = Mult.countTrailingZeros();
756 Mult = Mult.lshr(TwoFactors);
757 OddFactorial *= Mult;
760 // We need at least W + T bits for the multiplication step
761 unsigned CalculationBits = W + T;
763 // Calculate 2^T, at width T+W.
764 APInt DivFactor = APInt::getOneBitSet(CalculationBits, T);
766 // Calculate the multiplicative inverse of K! / 2^T;
767 // this multiplication factor will perform the exact division by
769 APInt Mod = APInt::getSignedMinValue(W+1);
770 APInt MultiplyFactor = OddFactorial.zext(W+1);
771 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
772 MultiplyFactor = MultiplyFactor.trunc(W);
774 // Calculate the product, at width T+W
775 IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
777 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
778 for (unsigned i = 1; i != K; ++i) {
779 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
780 Dividend = SE.getMulExpr(Dividend,
781 SE.getTruncateOrZeroExtend(S, CalculationTy));
785 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
787 // Truncate the result, and divide by K! / 2^T.
789 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
790 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
793 /// evaluateAtIteration - Return the value of this chain of recurrences at
794 /// the specified iteration number. We can evaluate this recurrence by
795 /// multiplying each element in the chain by the binomial coefficient
796 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
798 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
800 /// where BC(It, k) stands for binomial coefficient.
802 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
803 ScalarEvolution &SE) const {
804 const SCEV *Result = getStart();
805 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
806 // The computation is correct in the face of overflow provided that the
807 // multiplication is performed _after_ the evaluation of the binomial
809 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
810 if (isa<SCEVCouldNotCompute>(Coeff))
813 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
818 //===----------------------------------------------------------------------===//
819 // SCEV Expression folder implementations
820 //===----------------------------------------------------------------------===//
822 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
824 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
825 "This is not a truncating conversion!");
826 assert(isSCEVable(Ty) &&
827 "This is not a conversion to a SCEVable type!");
828 Ty = getEffectiveSCEVType(Ty);
831 ID.AddInteger(scTruncate);
835 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
837 // Fold if the operand is constant.
838 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
840 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
842 // trunc(trunc(x)) --> trunc(x)
843 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
844 return getTruncateExpr(ST->getOperand(), Ty);
846 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
847 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
848 return getTruncateOrSignExtend(SS->getOperand(), Ty);
850 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
851 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
852 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
854 // trunc(x1+x2+...+xN) --> trunc(x1)+trunc(x2)+...+trunc(xN) if we can
855 // eliminate all the truncates.
856 if (const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Op)) {
857 SmallVector<const SCEV *, 4> Operands;
858 bool hasTrunc = false;
859 for (unsigned i = 0, e = SA->getNumOperands(); i != e && !hasTrunc; ++i) {
860 const SCEV *S = getTruncateExpr(SA->getOperand(i), Ty);
861 hasTrunc = isa<SCEVTruncateExpr>(S);
862 Operands.push_back(S);
865 return getAddExpr(Operands);
866 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
869 // trunc(x1*x2*...*xN) --> trunc(x1)*trunc(x2)*...*trunc(xN) if we can
870 // eliminate all the truncates.
871 if (const SCEVMulExpr *SM = dyn_cast<SCEVMulExpr>(Op)) {
872 SmallVector<const SCEV *, 4> Operands;
873 bool hasTrunc = false;
874 for (unsigned i = 0, e = SM->getNumOperands(); i != e && !hasTrunc; ++i) {
875 const SCEV *S = getTruncateExpr(SM->getOperand(i), Ty);
876 hasTrunc = isa<SCEVTruncateExpr>(S);
877 Operands.push_back(S);
880 return getMulExpr(Operands);
881 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
884 // If the input value is a chrec scev, truncate the chrec's operands.
885 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
886 SmallVector<const SCEV *, 4> Operands;
887 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
888 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
889 return getAddRecExpr(Operands, AddRec->getLoop(), SCEV::FlagAnyWrap);
892 // The cast wasn't folded; create an explicit cast node. We can reuse
893 // the existing insert position since if we get here, we won't have
894 // made any changes which would invalidate it.
895 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
897 UniqueSCEVs.InsertNode(S, IP);
901 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
903 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
904 "This is not an extending conversion!");
905 assert(isSCEVable(Ty) &&
906 "This is not a conversion to a SCEVable type!");
907 Ty = getEffectiveSCEVType(Ty);
909 // Fold if the operand is constant.
910 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
912 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(), Ty)));
914 // zext(zext(x)) --> zext(x)
915 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
916 return getZeroExtendExpr(SZ->getOperand(), Ty);
918 // Before doing any expensive analysis, check to see if we've already
919 // computed a SCEV for this Op and Ty.
921 ID.AddInteger(scZeroExtend);
925 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
927 // zext(trunc(x)) --> zext(x) or x or trunc(x)
928 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
929 // It's possible the bits taken off by the truncate were all zero bits. If
930 // so, we should be able to simplify this further.
931 const SCEV *X = ST->getOperand();
932 ConstantRange CR = getUnsignedRange(X);
933 unsigned TruncBits = getTypeSizeInBits(ST->getType());
934 unsigned NewBits = getTypeSizeInBits(Ty);
935 if (CR.truncate(TruncBits).zeroExtend(NewBits).contains(
936 CR.zextOrTrunc(NewBits)))
937 return getTruncateOrZeroExtend(X, Ty);
940 // If the input value is a chrec scev, and we can prove that the value
941 // did not overflow the old, smaller, value, we can zero extend all of the
942 // operands (often constants). This allows analysis of something like
943 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
944 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
945 if (AR->isAffine()) {
946 const SCEV *Start = AR->getStart();
947 const SCEV *Step = AR->getStepRecurrence(*this);
948 unsigned BitWidth = getTypeSizeInBits(AR->getType());
949 const Loop *L = AR->getLoop();
951 // If we have special knowledge that this addrec won't overflow,
952 // we don't need to do any further analysis.
953 if (AR->getNoWrapFlags(SCEV::FlagNUW))
954 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
955 getZeroExtendExpr(Step, Ty),
956 L, AR->getNoWrapFlags());
958 // Check whether the backedge-taken count is SCEVCouldNotCompute.
959 // Note that this serves two purposes: It filters out loops that are
960 // simply not analyzable, and it covers the case where this code is
961 // being called from within backedge-taken count analysis, such that
962 // attempting to ask for the backedge-taken count would likely result
963 // in infinite recursion. In the later case, the analysis code will
964 // cope with a conservative value, and it will take care to purge
965 // that value once it has finished.
966 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
967 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
968 // Manually compute the final value for AR, checking for
971 // Check whether the backedge-taken count can be losslessly casted to
972 // the addrec's type. The count is always unsigned.
973 const SCEV *CastedMaxBECount =
974 getTruncateOrZeroExtend(MaxBECount, Start->getType());
975 const SCEV *RecastedMaxBECount =
976 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
977 if (MaxBECount == RecastedMaxBECount) {
978 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
979 // Check whether Start+Step*MaxBECount has no unsigned overflow.
980 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
981 const SCEV *ZAdd = getZeroExtendExpr(getAddExpr(Start, ZMul), WideTy);
982 const SCEV *WideStart = getZeroExtendExpr(Start, WideTy);
983 const SCEV *WideMaxBECount =
984 getZeroExtendExpr(CastedMaxBECount, WideTy);
985 const SCEV *OperandExtendedAdd =
986 getAddExpr(WideStart,
987 getMulExpr(WideMaxBECount,
988 getZeroExtendExpr(Step, WideTy)));
989 if (ZAdd == OperandExtendedAdd) {
990 // Cache knowledge of AR NUW, which is propagated to this AddRec.
991 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
992 // Return the expression with the addrec on the outside.
993 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
994 getZeroExtendExpr(Step, Ty),
995 L, AR->getNoWrapFlags());
997 // Similar to above, only this time treat the step value as signed.
998 // This covers loops that count down.
1000 getAddExpr(WideStart,
1001 getMulExpr(WideMaxBECount,
1002 getSignExtendExpr(Step, WideTy)));
1003 if (ZAdd == OperandExtendedAdd) {
1004 // Cache knowledge of AR NW, which is propagated to this AddRec.
1005 // Negative step causes unsigned wrap, but it still can't self-wrap.
1006 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1007 // Return the expression with the addrec on the outside.
1008 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1009 getSignExtendExpr(Step, Ty),
1010 L, AR->getNoWrapFlags());
1014 // If the backedge is guarded by a comparison with the pre-inc value
1015 // the addrec is safe. Also, if the entry is guarded by a comparison
1016 // with the start value and the backedge is guarded by a comparison
1017 // with the post-inc value, the addrec is safe.
1018 if (isKnownPositive(Step)) {
1019 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
1020 getUnsignedRange(Step).getUnsignedMax());
1021 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
1022 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
1023 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
1024 AR->getPostIncExpr(*this), N))) {
1025 // Cache knowledge of AR NUW, which is propagated to this AddRec.
1026 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
1027 // Return the expression with the addrec on the outside.
1028 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1029 getZeroExtendExpr(Step, Ty),
1030 L, AR->getNoWrapFlags());
1032 } else if (isKnownNegative(Step)) {
1033 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
1034 getSignedRange(Step).getSignedMin());
1035 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
1036 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
1037 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
1038 AR->getPostIncExpr(*this), N))) {
1039 // Cache knowledge of AR NW, which is propagated to this AddRec.
1040 // Negative step causes unsigned wrap, but it still can't self-wrap.
1041 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1042 // Return the expression with the addrec on the outside.
1043 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1044 getSignExtendExpr(Step, Ty),
1045 L, AR->getNoWrapFlags());
1051 // The cast wasn't folded; create an explicit cast node.
1052 // Recompute the insert position, as it may have been invalidated.
1053 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1054 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1056 UniqueSCEVs.InsertNode(S, IP);
1060 // Get the limit of a recurrence such that incrementing by Step cannot cause
1061 // signed overflow as long as the value of the recurrence within the loop does
1062 // not exceed this limit before incrementing.
1063 static const SCEV *getOverflowLimitForStep(const SCEV *Step,
1064 ICmpInst::Predicate *Pred,
1065 ScalarEvolution *SE) {
1066 unsigned BitWidth = SE->getTypeSizeInBits(Step->getType());
1067 if (SE->isKnownPositive(Step)) {
1068 *Pred = ICmpInst::ICMP_SLT;
1069 return SE->getConstant(APInt::getSignedMinValue(BitWidth) -
1070 SE->getSignedRange(Step).getSignedMax());
1072 if (SE->isKnownNegative(Step)) {
1073 *Pred = ICmpInst::ICMP_SGT;
1074 return SE->getConstant(APInt::getSignedMaxValue(BitWidth) -
1075 SE->getSignedRange(Step).getSignedMin());
1080 // The recurrence AR has been shown to have no signed wrap. Typically, if we can
1081 // prove NSW for AR, then we can just as easily prove NSW for its preincrement
1082 // or postincrement sibling. This allows normalizing a sign extended AddRec as
1083 // such: {sext(Step + Start),+,Step} => {(Step + sext(Start),+,Step} As a
1084 // result, the expression "Step + sext(PreIncAR)" is congruent with
1085 // "sext(PostIncAR)"
1086 static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR,
1088 ScalarEvolution *SE) {
1089 const Loop *L = AR->getLoop();
1090 const SCEV *Start = AR->getStart();
1091 const SCEV *Step = AR->getStepRecurrence(*SE);
1093 // Check for a simple looking step prior to loop entry.
1094 const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start);
1098 // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV
1099 // subtraction is expensive. For this purpose, perform a quick and dirty
1100 // difference, by checking for Step in the operand list.
1101 SmallVector<const SCEV *, 4> DiffOps;
1102 for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end();
1105 DiffOps.push_back(*I);
1107 if (DiffOps.size() == SA->getNumOperands())
1110 // This is a postinc AR. Check for overflow on the preinc recurrence using the
1111 // same three conditions that getSignExtendedExpr checks.
1113 // 1. NSW flags on the step increment.
1114 const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags());
1115 const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>(
1116 SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap));
1118 if (PreAR && PreAR->getNoWrapFlags(SCEV::FlagNSW))
1121 // 2. Direct overflow check on the step operation's expression.
1122 unsigned BitWidth = SE->getTypeSizeInBits(AR->getType());
1123 Type *WideTy = IntegerType::get(SE->getContext(), BitWidth * 2);
1124 const SCEV *OperandExtendedStart =
1125 SE->getAddExpr(SE->getSignExtendExpr(PreStart, WideTy),
1126 SE->getSignExtendExpr(Step, WideTy));
1127 if (SE->getSignExtendExpr(Start, WideTy) == OperandExtendedStart) {
1128 // Cache knowledge of PreAR NSW.
1130 const_cast<SCEVAddRecExpr *>(PreAR)->setNoWrapFlags(SCEV::FlagNSW);
1131 // FIXME: this optimization needs a unit test
1132 DEBUG(dbgs() << "SCEV: untested prestart overflow check\n");
1136 // 3. Loop precondition.
1137 ICmpInst::Predicate Pred;
1138 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, SE);
1140 if (OverflowLimit &&
1141 SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) {
1147 // Get the normalized sign-extended expression for this AddRec's Start.
1148 static const SCEV *getSignExtendAddRecStart(const SCEVAddRecExpr *AR,
1150 ScalarEvolution *SE) {
1151 const SCEV *PreStart = getPreStartForSignExtend(AR, Ty, SE);
1153 return SE->getSignExtendExpr(AR->getStart(), Ty);
1155 return SE->getAddExpr(SE->getSignExtendExpr(AR->getStepRecurrence(*SE), Ty),
1156 SE->getSignExtendExpr(PreStart, Ty));
1159 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1161 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1162 "This is not an extending conversion!");
1163 assert(isSCEVable(Ty) &&
1164 "This is not a conversion to a SCEVable type!");
1165 Ty = getEffectiveSCEVType(Ty);
1167 // Fold if the operand is constant.
1168 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1170 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(), Ty)));
1172 // sext(sext(x)) --> sext(x)
1173 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1174 return getSignExtendExpr(SS->getOperand(), Ty);
1176 // sext(zext(x)) --> zext(x)
1177 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
1178 return getZeroExtendExpr(SZ->getOperand(), Ty);
1180 // Before doing any expensive analysis, check to see if we've already
1181 // computed a SCEV for this Op and Ty.
1182 FoldingSetNodeID ID;
1183 ID.AddInteger(scSignExtend);
1187 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1189 // If the input value is provably positive, build a zext instead.
1190 if (isKnownNonNegative(Op))
1191 return getZeroExtendExpr(Op, Ty);
1193 // sext(trunc(x)) --> sext(x) or x or trunc(x)
1194 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
1195 // It's possible the bits taken off by the truncate were all sign bits. If
1196 // so, we should be able to simplify this further.
1197 const SCEV *X = ST->getOperand();
1198 ConstantRange CR = getSignedRange(X);
1199 unsigned TruncBits = getTypeSizeInBits(ST->getType());
1200 unsigned NewBits = getTypeSizeInBits(Ty);
1201 if (CR.truncate(TruncBits).signExtend(NewBits).contains(
1202 CR.sextOrTrunc(NewBits)))
1203 return getTruncateOrSignExtend(X, Ty);
1206 // If the input value is a chrec scev, and we can prove that the value
1207 // did not overflow the old, smaller, value, we can sign extend all of the
1208 // operands (often constants). This allows analysis of something like
1209 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1210 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1211 if (AR->isAffine()) {
1212 const SCEV *Start = AR->getStart();
1213 const SCEV *Step = AR->getStepRecurrence(*this);
1214 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1215 const Loop *L = AR->getLoop();
1217 // If we have special knowledge that this addrec won't overflow,
1218 // we don't need to do any further analysis.
1219 if (AR->getNoWrapFlags(SCEV::FlagNSW))
1220 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1221 getSignExtendExpr(Step, Ty),
1224 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1225 // Note that this serves two purposes: It filters out loops that are
1226 // simply not analyzable, and it covers the case where this code is
1227 // being called from within backedge-taken count analysis, such that
1228 // attempting to ask for the backedge-taken count would likely result
1229 // in infinite recursion. In the later case, the analysis code will
1230 // cope with a conservative value, and it will take care to purge
1231 // that value once it has finished.
1232 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1233 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1234 // Manually compute the final value for AR, checking for
1237 // Check whether the backedge-taken count can be losslessly casted to
1238 // the addrec's type. The count is always unsigned.
1239 const SCEV *CastedMaxBECount =
1240 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1241 const SCEV *RecastedMaxBECount =
1242 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1243 if (MaxBECount == RecastedMaxBECount) {
1244 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1245 // Check whether Start+Step*MaxBECount has no signed overflow.
1246 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1247 const SCEV *SAdd = getSignExtendExpr(getAddExpr(Start, SMul), WideTy);
1248 const SCEV *WideStart = getSignExtendExpr(Start, WideTy);
1249 const SCEV *WideMaxBECount =
1250 getZeroExtendExpr(CastedMaxBECount, WideTy);
1251 const SCEV *OperandExtendedAdd =
1252 getAddExpr(WideStart,
1253 getMulExpr(WideMaxBECount,
1254 getSignExtendExpr(Step, WideTy)));
1255 if (SAdd == OperandExtendedAdd) {
1256 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1257 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1258 // Return the expression with the addrec on the outside.
1259 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1260 getSignExtendExpr(Step, Ty),
1261 L, AR->getNoWrapFlags());
1263 // Similar to above, only this time treat the step value as unsigned.
1264 // This covers loops that count up with an unsigned step.
1265 OperandExtendedAdd =
1266 getAddExpr(WideStart,
1267 getMulExpr(WideMaxBECount,
1268 getZeroExtendExpr(Step, WideTy)));
1269 if (SAdd == OperandExtendedAdd) {
1270 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1271 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1272 // Return the expression with the addrec on the outside.
1273 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1274 getZeroExtendExpr(Step, Ty),
1275 L, AR->getNoWrapFlags());
1279 // If the backedge is guarded by a comparison with the pre-inc value
1280 // the addrec is safe. Also, if the entry is guarded by a comparison
1281 // with the start value and the backedge is guarded by a comparison
1282 // with the post-inc value, the addrec is safe.
1283 ICmpInst::Predicate Pred;
1284 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, this);
1285 if (OverflowLimit &&
1286 (isLoopBackedgeGuardedByCond(L, Pred, AR, OverflowLimit) ||
1287 (isLoopEntryGuardedByCond(L, Pred, Start, OverflowLimit) &&
1288 isLoopBackedgeGuardedByCond(L, Pred, AR->getPostIncExpr(*this),
1290 // Cache knowledge of AR NSW, then propagate NSW to the wide AddRec.
1291 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1292 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1293 getSignExtendExpr(Step, Ty),
1294 L, AR->getNoWrapFlags());
1299 // The cast wasn't folded; create an explicit cast node.
1300 // Recompute the insert position, as it may have been invalidated.
1301 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1302 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1304 UniqueSCEVs.InsertNode(S, IP);
1308 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1309 /// unspecified bits out to the given type.
1311 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1313 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1314 "This is not an extending conversion!");
1315 assert(isSCEVable(Ty) &&
1316 "This is not a conversion to a SCEVable type!");
1317 Ty = getEffectiveSCEVType(Ty);
1319 // Sign-extend negative constants.
1320 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1321 if (SC->getValue()->getValue().isNegative())
1322 return getSignExtendExpr(Op, Ty);
1324 // Peel off a truncate cast.
1325 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1326 const SCEV *NewOp = T->getOperand();
1327 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1328 return getAnyExtendExpr(NewOp, Ty);
1329 return getTruncateOrNoop(NewOp, Ty);
1332 // Next try a zext cast. If the cast is folded, use it.
1333 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1334 if (!isa<SCEVZeroExtendExpr>(ZExt))
1337 // Next try a sext cast. If the cast is folded, use it.
1338 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1339 if (!isa<SCEVSignExtendExpr>(SExt))
1342 // Force the cast to be folded into the operands of an addrec.
1343 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1344 SmallVector<const SCEV *, 4> Ops;
1345 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1347 Ops.push_back(getAnyExtendExpr(*I, Ty));
1348 return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW);
1351 // If the expression is obviously signed, use the sext cast value.
1352 if (isa<SCEVSMaxExpr>(Op))
1355 // Absent any other information, use the zext cast value.
1359 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1360 /// a list of operands to be added under the given scale, update the given
1361 /// map. This is a helper function for getAddRecExpr. As an example of
1362 /// what it does, given a sequence of operands that would form an add
1363 /// expression like this:
1365 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1367 /// where A and B are constants, update the map with these values:
1369 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1371 /// and add 13 + A*B*29 to AccumulatedConstant.
1372 /// This will allow getAddRecExpr to produce this:
1374 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1376 /// This form often exposes folding opportunities that are hidden in
1377 /// the original operand list.
1379 /// Return true iff it appears that any interesting folding opportunities
1380 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1381 /// the common case where no interesting opportunities are present, and
1382 /// is also used as a check to avoid infinite recursion.
1385 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1386 SmallVectorImpl<const SCEV *> &NewOps,
1387 APInt &AccumulatedConstant,
1388 const SCEV *const *Ops, size_t NumOperands,
1390 ScalarEvolution &SE) {
1391 bool Interesting = false;
1393 // Iterate over the add operands. They are sorted, with constants first.
1395 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1397 // Pull a buried constant out to the outside.
1398 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1400 AccumulatedConstant += Scale * C->getValue()->getValue();
1403 // Next comes everything else. We're especially interested in multiplies
1404 // here, but they're in the middle, so just visit the rest with one loop.
1405 for (; i != NumOperands; ++i) {
1406 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1407 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1409 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1410 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1411 // A multiplication of a constant with another add; recurse.
1412 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1414 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1415 Add->op_begin(), Add->getNumOperands(),
1418 // A multiplication of a constant with some other value. Update
1420 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1421 const SCEV *Key = SE.getMulExpr(MulOps);
1422 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1423 M.insert(std::make_pair(Key, NewScale));
1425 NewOps.push_back(Pair.first->first);
1427 Pair.first->second += NewScale;
1428 // The map already had an entry for this value, which may indicate
1429 // a folding opportunity.
1434 // An ordinary operand. Update the map.
1435 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1436 M.insert(std::make_pair(Ops[i], Scale));
1438 NewOps.push_back(Pair.first->first);
1440 Pair.first->second += Scale;
1441 // The map already had an entry for this value, which may indicate
1442 // a folding opportunity.
1452 struct APIntCompare {
1453 bool operator()(const APInt &LHS, const APInt &RHS) const {
1454 return LHS.ult(RHS);
1459 /// getAddExpr - Get a canonical add expression, or something simpler if
1461 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1462 SCEV::NoWrapFlags Flags) {
1463 assert(!(Flags & ~(SCEV::FlagNUW | SCEV::FlagNSW)) &&
1464 "only nuw or nsw allowed");
1465 assert(!Ops.empty() && "Cannot get empty add!");
1466 if (Ops.size() == 1) return Ops[0];
1468 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1469 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1470 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1471 "SCEVAddExpr operand types don't match!");
1474 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1476 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1477 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1478 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1480 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1481 E = Ops.end(); I != E; ++I)
1482 if (!isKnownNonNegative(*I)) {
1486 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1489 // Sort by complexity, this groups all similar expression types together.
1490 GroupByComplexity(Ops, LI);
1492 // If there are any constants, fold them together.
1494 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1496 assert(Idx < Ops.size());
1497 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1498 // We found two constants, fold them together!
1499 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1500 RHSC->getValue()->getValue());
1501 if (Ops.size() == 2) return Ops[0];
1502 Ops.erase(Ops.begin()+1); // Erase the folded element
1503 LHSC = cast<SCEVConstant>(Ops[0]);
1506 // If we are left with a constant zero being added, strip it off.
1507 if (LHSC->getValue()->isZero()) {
1508 Ops.erase(Ops.begin());
1512 if (Ops.size() == 1) return Ops[0];
1515 // Okay, check to see if the same value occurs in the operand list more than
1516 // once. If so, merge them together into an multiply expression. Since we
1517 // sorted the list, these values are required to be adjacent.
1518 Type *Ty = Ops[0]->getType();
1519 bool FoundMatch = false;
1520 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1521 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1522 // Scan ahead to count how many equal operands there are.
1524 while (i+Count != e && Ops[i+Count] == Ops[i])
1526 // Merge the values into a multiply.
1527 const SCEV *Scale = getConstant(Ty, Count);
1528 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1529 if (Ops.size() == Count)
1532 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1533 --i; e -= Count - 1;
1537 return getAddExpr(Ops, Flags);
1539 // Check for truncates. If all the operands are truncated from the same
1540 // type, see if factoring out the truncate would permit the result to be
1541 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1542 // if the contents of the resulting outer trunc fold to something simple.
1543 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1544 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1545 Type *DstType = Trunc->getType();
1546 Type *SrcType = Trunc->getOperand()->getType();
1547 SmallVector<const SCEV *, 8> LargeOps;
1549 // Check all the operands to see if they can be represented in the
1550 // source type of the truncate.
1551 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1552 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1553 if (T->getOperand()->getType() != SrcType) {
1557 LargeOps.push_back(T->getOperand());
1558 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1559 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1560 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1561 SmallVector<const SCEV *, 8> LargeMulOps;
1562 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1563 if (const SCEVTruncateExpr *T =
1564 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1565 if (T->getOperand()->getType() != SrcType) {
1569 LargeMulOps.push_back(T->getOperand());
1570 } else if (const SCEVConstant *C =
1571 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1572 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1579 LargeOps.push_back(getMulExpr(LargeMulOps));
1586 // Evaluate the expression in the larger type.
1587 const SCEV *Fold = getAddExpr(LargeOps, Flags);
1588 // If it folds to something simple, use it. Otherwise, don't.
1589 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1590 return getTruncateExpr(Fold, DstType);
1594 // Skip past any other cast SCEVs.
1595 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1598 // If there are add operands they would be next.
1599 if (Idx < Ops.size()) {
1600 bool DeletedAdd = false;
1601 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1602 // If we have an add, expand the add operands onto the end of the operands
1604 Ops.erase(Ops.begin()+Idx);
1605 Ops.append(Add->op_begin(), Add->op_end());
1609 // If we deleted at least one add, we added operands to the end of the list,
1610 // and they are not necessarily sorted. Recurse to resort and resimplify
1611 // any operands we just acquired.
1613 return getAddExpr(Ops);
1616 // Skip over the add expression until we get to a multiply.
1617 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1620 // Check to see if there are any folding opportunities present with
1621 // operands multiplied by constant values.
1622 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1623 uint64_t BitWidth = getTypeSizeInBits(Ty);
1624 DenseMap<const SCEV *, APInt> M;
1625 SmallVector<const SCEV *, 8> NewOps;
1626 APInt AccumulatedConstant(BitWidth, 0);
1627 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1628 Ops.data(), Ops.size(),
1629 APInt(BitWidth, 1), *this)) {
1630 // Some interesting folding opportunity is present, so its worthwhile to
1631 // re-generate the operands list. Group the operands by constant scale,
1632 // to avoid multiplying by the same constant scale multiple times.
1633 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1634 for (SmallVectorImpl<const SCEV *>::const_iterator I = NewOps.begin(),
1635 E = NewOps.end(); I != E; ++I)
1636 MulOpLists[M.find(*I)->second].push_back(*I);
1637 // Re-generate the operands list.
1639 if (AccumulatedConstant != 0)
1640 Ops.push_back(getConstant(AccumulatedConstant));
1641 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1642 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1644 Ops.push_back(getMulExpr(getConstant(I->first),
1645 getAddExpr(I->second)));
1647 return getConstant(Ty, 0);
1648 if (Ops.size() == 1)
1650 return getAddExpr(Ops);
1654 // If we are adding something to a multiply expression, make sure the
1655 // something is not already an operand of the multiply. If so, merge it into
1657 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1658 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1659 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1660 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1661 if (isa<SCEVConstant>(MulOpSCEV))
1663 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1664 if (MulOpSCEV == Ops[AddOp]) {
1665 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1666 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1667 if (Mul->getNumOperands() != 2) {
1668 // If the multiply has more than two operands, we must get the
1670 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1671 Mul->op_begin()+MulOp);
1672 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1673 InnerMul = getMulExpr(MulOps);
1675 const SCEV *One = getConstant(Ty, 1);
1676 const SCEV *AddOne = getAddExpr(One, InnerMul);
1677 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1678 if (Ops.size() == 2) return OuterMul;
1680 Ops.erase(Ops.begin()+AddOp);
1681 Ops.erase(Ops.begin()+Idx-1);
1683 Ops.erase(Ops.begin()+Idx);
1684 Ops.erase(Ops.begin()+AddOp-1);
1686 Ops.push_back(OuterMul);
1687 return getAddExpr(Ops);
1690 // Check this multiply against other multiplies being added together.
1691 for (unsigned OtherMulIdx = Idx+1;
1692 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1694 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1695 // If MulOp occurs in OtherMul, we can fold the two multiplies
1697 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1698 OMulOp != e; ++OMulOp)
1699 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1700 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1701 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1702 if (Mul->getNumOperands() != 2) {
1703 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1704 Mul->op_begin()+MulOp);
1705 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1706 InnerMul1 = getMulExpr(MulOps);
1708 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1709 if (OtherMul->getNumOperands() != 2) {
1710 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1711 OtherMul->op_begin()+OMulOp);
1712 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1713 InnerMul2 = getMulExpr(MulOps);
1715 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1716 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1717 if (Ops.size() == 2) return OuterMul;
1718 Ops.erase(Ops.begin()+Idx);
1719 Ops.erase(Ops.begin()+OtherMulIdx-1);
1720 Ops.push_back(OuterMul);
1721 return getAddExpr(Ops);
1727 // If there are any add recurrences in the operands list, see if any other
1728 // added values are loop invariant. If so, we can fold them into the
1730 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1733 // Scan over all recurrences, trying to fold loop invariants into them.
1734 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1735 // Scan all of the other operands to this add and add them to the vector if
1736 // they are loop invariant w.r.t. the recurrence.
1737 SmallVector<const SCEV *, 8> LIOps;
1738 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1739 const Loop *AddRecLoop = AddRec->getLoop();
1740 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1741 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1742 LIOps.push_back(Ops[i]);
1743 Ops.erase(Ops.begin()+i);
1747 // If we found some loop invariants, fold them into the recurrence.
1748 if (!LIOps.empty()) {
1749 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1750 LIOps.push_back(AddRec->getStart());
1752 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1754 AddRecOps[0] = getAddExpr(LIOps);
1756 // Build the new addrec. Propagate the NUW and NSW flags if both the
1757 // outer add and the inner addrec are guaranteed to have no overflow.
1758 // Always propagate NW.
1759 Flags = AddRec->getNoWrapFlags(setFlags(Flags, SCEV::FlagNW));
1760 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop, Flags);
1762 // If all of the other operands were loop invariant, we are done.
1763 if (Ops.size() == 1) return NewRec;
1765 // Otherwise, add the folded AddRec by the non-invariant parts.
1766 for (unsigned i = 0;; ++i)
1767 if (Ops[i] == AddRec) {
1771 return getAddExpr(Ops);
1774 // Okay, if there weren't any loop invariants to be folded, check to see if
1775 // there are multiple AddRec's with the same loop induction variable being
1776 // added together. If so, we can fold them.
1777 for (unsigned OtherIdx = Idx+1;
1778 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1780 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1781 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1782 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1784 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1786 if (const SCEVAddRecExpr *OtherAddRec =
1787 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1788 if (OtherAddRec->getLoop() == AddRecLoop) {
1789 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1791 if (i >= AddRecOps.size()) {
1792 AddRecOps.append(OtherAddRec->op_begin()+i,
1793 OtherAddRec->op_end());
1796 AddRecOps[i] = getAddExpr(AddRecOps[i],
1797 OtherAddRec->getOperand(i));
1799 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1801 // Step size has changed, so we cannot guarantee no self-wraparound.
1802 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop, SCEV::FlagAnyWrap);
1803 return getAddExpr(Ops);
1806 // Otherwise couldn't fold anything into this recurrence. Move onto the
1810 // Okay, it looks like we really DO need an add expr. Check to see if we
1811 // already have one, otherwise create a new one.
1812 FoldingSetNodeID ID;
1813 ID.AddInteger(scAddExpr);
1814 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1815 ID.AddPointer(Ops[i]);
1818 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1820 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1821 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1822 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1824 UniqueSCEVs.InsertNode(S, IP);
1826 S->setNoWrapFlags(Flags);
1830 static uint64_t umul_ov(uint64_t i, uint64_t j, bool &Overflow) {
1832 if (j > 1 && k / j != i) Overflow = true;
1836 /// Compute the result of "n choose k", the binomial coefficient. If an
1837 /// intermediate computation overflows, Overflow will be set and the return will
1838 /// be garbage. Overflow is not cleared on absence of overflow.
1839 static uint64_t Choose(uint64_t n, uint64_t k, bool &Overflow) {
1840 // We use the multiplicative formula:
1841 // n(n-1)(n-2)...(n-(k-1)) / k(k-1)(k-2)...1 .
1842 // At each iteration, we take the n-th term of the numeral and divide by the
1843 // (k-n)th term of the denominator. This division will always produce an
1844 // integral result, and helps reduce the chance of overflow in the
1845 // intermediate computations. However, we can still overflow even when the
1846 // final result would fit.
1848 if (n == 0 || n == k) return 1;
1849 if (k > n) return 0;
1855 for (uint64_t i = 1; i <= k; ++i) {
1856 r = umul_ov(r, n-(i-1), Overflow);
1862 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1864 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1865 SCEV::NoWrapFlags Flags) {
1866 assert(Flags == maskFlags(Flags, SCEV::FlagNUW | SCEV::FlagNSW) &&
1867 "only nuw or nsw allowed");
1868 assert(!Ops.empty() && "Cannot get empty mul!");
1869 if (Ops.size() == 1) return Ops[0];
1871 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1872 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1873 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1874 "SCEVMulExpr operand types don't match!");
1877 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1879 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1880 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1881 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1883 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1884 E = Ops.end(); I != E; ++I)
1885 if (!isKnownNonNegative(*I)) {
1889 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1892 // Sort by complexity, this groups all similar expression types together.
1893 GroupByComplexity(Ops, LI);
1895 // If there are any constants, fold them together.
1897 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1899 // C1*(C2+V) -> C1*C2 + C1*V
1900 if (Ops.size() == 2)
1901 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1902 if (Add->getNumOperands() == 2 &&
1903 isa<SCEVConstant>(Add->getOperand(0)))
1904 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1905 getMulExpr(LHSC, Add->getOperand(1)));
1908 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1909 // We found two constants, fold them together!
1910 ConstantInt *Fold = ConstantInt::get(getContext(),
1911 LHSC->getValue()->getValue() *
1912 RHSC->getValue()->getValue());
1913 Ops[0] = getConstant(Fold);
1914 Ops.erase(Ops.begin()+1); // Erase the folded element
1915 if (Ops.size() == 1) return Ops[0];
1916 LHSC = cast<SCEVConstant>(Ops[0]);
1919 // If we are left with a constant one being multiplied, strip it off.
1920 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1921 Ops.erase(Ops.begin());
1923 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1924 // If we have a multiply of zero, it will always be zero.
1926 } else if (Ops[0]->isAllOnesValue()) {
1927 // If we have a mul by -1 of an add, try distributing the -1 among the
1929 if (Ops.size() == 2) {
1930 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1931 SmallVector<const SCEV *, 4> NewOps;
1932 bool AnyFolded = false;
1933 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(),
1934 E = Add->op_end(); I != E; ++I) {
1935 const SCEV *Mul = getMulExpr(Ops[0], *I);
1936 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1937 NewOps.push_back(Mul);
1940 return getAddExpr(NewOps);
1942 else if (const SCEVAddRecExpr *
1943 AddRec = dyn_cast<SCEVAddRecExpr>(Ops[1])) {
1944 // Negation preserves a recurrence's no self-wrap property.
1945 SmallVector<const SCEV *, 4> Operands;
1946 for (SCEVAddRecExpr::op_iterator I = AddRec->op_begin(),
1947 E = AddRec->op_end(); I != E; ++I) {
1948 Operands.push_back(getMulExpr(Ops[0], *I));
1950 return getAddRecExpr(Operands, AddRec->getLoop(),
1951 AddRec->getNoWrapFlags(SCEV::FlagNW));
1956 if (Ops.size() == 1)
1960 // Skip over the add expression until we get to a multiply.
1961 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1964 // If there are mul operands inline them all into this expression.
1965 if (Idx < Ops.size()) {
1966 bool DeletedMul = false;
1967 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1968 // If we have an mul, expand the mul operands onto the end of the operands
1970 Ops.erase(Ops.begin()+Idx);
1971 Ops.append(Mul->op_begin(), Mul->op_end());
1975 // If we deleted at least one mul, we added operands to the end of the list,
1976 // and they are not necessarily sorted. Recurse to resort and resimplify
1977 // any operands we just acquired.
1979 return getMulExpr(Ops);
1982 // If there are any add recurrences in the operands list, see if any other
1983 // added values are loop invariant. If so, we can fold them into the
1985 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1988 // Scan over all recurrences, trying to fold loop invariants into them.
1989 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1990 // Scan all of the other operands to this mul and add them to the vector if
1991 // they are loop invariant w.r.t. the recurrence.
1992 SmallVector<const SCEV *, 8> LIOps;
1993 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1994 const Loop *AddRecLoop = AddRec->getLoop();
1995 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1996 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1997 LIOps.push_back(Ops[i]);
1998 Ops.erase(Ops.begin()+i);
2002 // If we found some loop invariants, fold them into the recurrence.
2003 if (!LIOps.empty()) {
2004 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
2005 SmallVector<const SCEV *, 4> NewOps;
2006 NewOps.reserve(AddRec->getNumOperands());
2007 const SCEV *Scale = getMulExpr(LIOps);
2008 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
2009 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
2011 // Build the new addrec. Propagate the NUW and NSW flags if both the
2012 // outer mul and the inner addrec are guaranteed to have no overflow.
2014 // No self-wrap cannot be guaranteed after changing the step size, but
2015 // will be inferred if either NUW or NSW is true.
2016 Flags = AddRec->getNoWrapFlags(clearFlags(Flags, SCEV::FlagNW));
2017 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop, Flags);
2019 // If all of the other operands were loop invariant, we are done.
2020 if (Ops.size() == 1) return NewRec;
2022 // Otherwise, multiply the folded AddRec by the non-invariant parts.
2023 for (unsigned i = 0;; ++i)
2024 if (Ops[i] == AddRec) {
2028 return getMulExpr(Ops);
2031 // Okay, if there weren't any loop invariants to be folded, check to see if
2032 // there are multiple AddRec's with the same loop induction variable being
2033 // multiplied together. If so, we can fold them.
2034 for (unsigned OtherIdx = Idx+1;
2035 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2037 if (AddRecLoop != cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop())
2040 // {A1,+,A2,+,...,+,An}<L> * {B1,+,B2,+,...,+,Bn}<L>
2041 // = {x=1 in [ sum y=x..2x [ sum z=max(y-x, y-n)..min(x,n) [
2042 // choose(x, 2x)*choose(2x-y, x-z)*A_{y-z}*B_z
2043 // ]]],+,...up to x=2n}.
2044 // Note that the arguments to choose() are always integers with values
2045 // known at compile time, never SCEV objects.
2047 // The implementation avoids pointless extra computations when the two
2048 // addrec's are of different length (mathematically, it's equivalent to
2049 // an infinite stream of zeros on the right).
2050 bool OpsModified = false;
2051 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2053 const SCEVAddRecExpr *OtherAddRec =
2054 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]);
2055 if (!OtherAddRec || OtherAddRec->getLoop() != AddRecLoop)
2058 bool Overflow = false;
2059 Type *Ty = AddRec->getType();
2060 bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64;
2061 SmallVector<const SCEV*, 7> AddRecOps;
2062 for (int x = 0, xe = AddRec->getNumOperands() +
2063 OtherAddRec->getNumOperands() - 1; x != xe && !Overflow; ++x) {
2064 const SCEV *Term = getConstant(Ty, 0);
2065 for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) {
2066 uint64_t Coeff1 = Choose(x, 2*x - y, Overflow);
2067 for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1),
2068 ze = std::min(x+1, (int)OtherAddRec->getNumOperands());
2069 z < ze && !Overflow; ++z) {
2070 uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow);
2072 if (LargerThan64Bits)
2073 Coeff = umul_ov(Coeff1, Coeff2, Overflow);
2075 Coeff = Coeff1*Coeff2;
2076 const SCEV *CoeffTerm = getConstant(Ty, Coeff);
2077 const SCEV *Term1 = AddRec->getOperand(y-z);
2078 const SCEV *Term2 = OtherAddRec->getOperand(z);
2079 Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2));
2082 AddRecOps.push_back(Term);
2085 const SCEV *NewAddRec = getAddRecExpr(AddRecOps, AddRec->getLoop(),
2087 if (Ops.size() == 2) return NewAddRec;
2088 Ops[Idx] = NewAddRec;
2089 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
2091 AddRec = dyn_cast<SCEVAddRecExpr>(NewAddRec);
2097 return getMulExpr(Ops);
2100 // Otherwise couldn't fold anything into this recurrence. Move onto the
2104 // Okay, it looks like we really DO need an mul expr. Check to see if we
2105 // already have one, otherwise create a new one.
2106 FoldingSetNodeID ID;
2107 ID.AddInteger(scMulExpr);
2108 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2109 ID.AddPointer(Ops[i]);
2112 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2114 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2115 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2116 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
2118 UniqueSCEVs.InsertNode(S, IP);
2120 S->setNoWrapFlags(Flags);
2124 /// getUDivExpr - Get a canonical unsigned division expression, or something
2125 /// simpler if possible.
2126 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
2128 assert(getEffectiveSCEVType(LHS->getType()) ==
2129 getEffectiveSCEVType(RHS->getType()) &&
2130 "SCEVUDivExpr operand types don't match!");
2132 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
2133 if (RHSC->getValue()->equalsInt(1))
2134 return LHS; // X udiv 1 --> x
2135 // If the denominator is zero, the result of the udiv is undefined. Don't
2136 // try to analyze it, because the resolution chosen here may differ from
2137 // the resolution chosen in other parts of the compiler.
2138 if (!RHSC->getValue()->isZero()) {
2139 // Determine if the division can be folded into the operands of
2141 // TODO: Generalize this to non-constants by using known-bits information.
2142 Type *Ty = LHS->getType();
2143 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
2144 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
2145 // For non-power-of-two values, effectively round the value up to the
2146 // nearest power of two.
2147 if (!RHSC->getValue()->getValue().isPowerOf2())
2149 IntegerType *ExtTy =
2150 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
2151 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
2152 if (const SCEVConstant *Step =
2153 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this))) {
2154 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
2155 const APInt &StepInt = Step->getValue()->getValue();
2156 const APInt &DivInt = RHSC->getValue()->getValue();
2157 if (!StepInt.urem(DivInt) &&
2158 getZeroExtendExpr(AR, ExtTy) ==
2159 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2160 getZeroExtendExpr(Step, ExtTy),
2161 AR->getLoop(), SCEV::FlagAnyWrap)) {
2162 SmallVector<const SCEV *, 4> Operands;
2163 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
2164 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
2165 return getAddRecExpr(Operands, AR->getLoop(),
2168 /// Get a canonical UDivExpr for a recurrence.
2169 /// {X,+,N}/C => {Y,+,N}/C where Y=X-(X%N). Safe when C%N=0.
2170 // We can currently only fold X%N if X is constant.
2171 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
2172 if (StartC && !DivInt.urem(StepInt) &&
2173 getZeroExtendExpr(AR, ExtTy) ==
2174 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2175 getZeroExtendExpr(Step, ExtTy),
2176 AR->getLoop(), SCEV::FlagAnyWrap)) {
2177 const APInt &StartInt = StartC->getValue()->getValue();
2178 const APInt &StartRem = StartInt.urem(StepInt);
2180 LHS = getAddRecExpr(getConstant(StartInt - StartRem), Step,
2181 AR->getLoop(), SCEV::FlagNW);
2184 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
2185 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
2186 SmallVector<const SCEV *, 4> Operands;
2187 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
2188 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
2189 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
2190 // Find an operand that's safely divisible.
2191 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
2192 const SCEV *Op = M->getOperand(i);
2193 const SCEV *Div = getUDivExpr(Op, RHSC);
2194 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
2195 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
2198 return getMulExpr(Operands);
2202 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
2203 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(LHS)) {
2204 SmallVector<const SCEV *, 4> Operands;
2205 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
2206 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
2207 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
2209 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2210 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
2211 if (isa<SCEVUDivExpr>(Op) ||
2212 getMulExpr(Op, RHS) != A->getOperand(i))
2214 Operands.push_back(Op);
2216 if (Operands.size() == A->getNumOperands())
2217 return getAddExpr(Operands);
2221 // Fold if both operands are constant.
2222 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
2223 Constant *LHSCV = LHSC->getValue();
2224 Constant *RHSCV = RHSC->getValue();
2225 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
2231 FoldingSetNodeID ID;
2232 ID.AddInteger(scUDivExpr);
2236 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2237 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
2239 UniqueSCEVs.InsertNode(S, IP);
2244 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2245 /// Simplify the expression as much as possible.
2246 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start, const SCEV *Step,
2248 SCEV::NoWrapFlags Flags) {
2249 SmallVector<const SCEV *, 4> Operands;
2250 Operands.push_back(Start);
2251 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
2252 if (StepChrec->getLoop() == L) {
2253 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
2254 return getAddRecExpr(Operands, L, maskFlags(Flags, SCEV::FlagNW));
2257 Operands.push_back(Step);
2258 return getAddRecExpr(Operands, L, Flags);
2261 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2262 /// Simplify the expression as much as possible.
2264 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2265 const Loop *L, SCEV::NoWrapFlags Flags) {
2266 if (Operands.size() == 1) return Operands[0];
2268 Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2269 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2270 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2271 "SCEVAddRecExpr operand types don't match!");
2272 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2273 assert(isLoopInvariant(Operands[i], L) &&
2274 "SCEVAddRecExpr operand is not loop-invariant!");
2277 if (Operands.back()->isZero()) {
2278 Operands.pop_back();
2279 return getAddRecExpr(Operands, L, SCEV::FlagAnyWrap); // {X,+,0} --> X
2282 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2283 // use that information to infer NUW and NSW flags. However, computing a
2284 // BE count requires calling getAddRecExpr, so we may not yet have a
2285 // meaningful BE count at this point (and if we don't, we'd be stuck
2286 // with a SCEVCouldNotCompute as the cached BE count).
2288 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
2290 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
2291 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
2292 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
2294 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2295 E = Operands.end(); I != E; ++I)
2296 if (!isKnownNonNegative(*I)) {
2300 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
2303 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2304 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2305 const Loop *NestedLoop = NestedAR->getLoop();
2306 if (L->contains(NestedLoop) ?
2307 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2308 (!NestedLoop->contains(L) &&
2309 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2310 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2311 NestedAR->op_end());
2312 Operands[0] = NestedAR->getStart();
2313 // AddRecs require their operands be loop-invariant with respect to their
2314 // loops. Don't perform this transformation if it would break this
2316 bool AllInvariant = true;
2317 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2318 if (!isLoopInvariant(Operands[i], L)) {
2319 AllInvariant = false;
2323 // Create a recurrence for the outer loop with the same step size.
2325 // The outer recurrence keeps its NW flag but only keeps NUW/NSW if the
2326 // inner recurrence has the same property.
2327 SCEV::NoWrapFlags OuterFlags =
2328 maskFlags(Flags, SCEV::FlagNW | NestedAR->getNoWrapFlags());
2330 NestedOperands[0] = getAddRecExpr(Operands, L, OuterFlags);
2331 AllInvariant = true;
2332 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2333 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2334 AllInvariant = false;
2338 // Ok, both add recurrences are valid after the transformation.
2340 // The inner recurrence keeps its NW flag but only keeps NUW/NSW if
2341 // the outer recurrence has the same property.
2342 SCEV::NoWrapFlags InnerFlags =
2343 maskFlags(NestedAR->getNoWrapFlags(), SCEV::FlagNW | Flags);
2344 return getAddRecExpr(NestedOperands, NestedLoop, InnerFlags);
2347 // Reset Operands to its original state.
2348 Operands[0] = NestedAR;
2352 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2353 // already have one, otherwise create a new one.
2354 FoldingSetNodeID ID;
2355 ID.AddInteger(scAddRecExpr);
2356 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2357 ID.AddPointer(Operands[i]);
2361 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2363 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2364 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2365 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2366 O, Operands.size(), L);
2367 UniqueSCEVs.InsertNode(S, IP);
2369 S->setNoWrapFlags(Flags);
2373 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2375 SmallVector<const SCEV *, 2> Ops;
2378 return getSMaxExpr(Ops);
2382 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2383 assert(!Ops.empty() && "Cannot get empty smax!");
2384 if (Ops.size() == 1) return Ops[0];
2386 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2387 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2388 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2389 "SCEVSMaxExpr operand types don't match!");
2392 // Sort by complexity, this groups all similar expression types together.
2393 GroupByComplexity(Ops, LI);
2395 // If there are any constants, fold them together.
2397 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2399 assert(Idx < Ops.size());
2400 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2401 // We found two constants, fold them together!
2402 ConstantInt *Fold = ConstantInt::get(getContext(),
2403 APIntOps::smax(LHSC->getValue()->getValue(),
2404 RHSC->getValue()->getValue()));
2405 Ops[0] = getConstant(Fold);
2406 Ops.erase(Ops.begin()+1); // Erase the folded element
2407 if (Ops.size() == 1) return Ops[0];
2408 LHSC = cast<SCEVConstant>(Ops[0]);
2411 // If we are left with a constant minimum-int, strip it off.
2412 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2413 Ops.erase(Ops.begin());
2415 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2416 // If we have an smax with a constant maximum-int, it will always be
2421 if (Ops.size() == 1) return Ops[0];
2424 // Find the first SMax
2425 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2428 // Check to see if one of the operands is an SMax. If so, expand its operands
2429 // onto our operand list, and recurse to simplify.
2430 if (Idx < Ops.size()) {
2431 bool DeletedSMax = false;
2432 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2433 Ops.erase(Ops.begin()+Idx);
2434 Ops.append(SMax->op_begin(), SMax->op_end());
2439 return getSMaxExpr(Ops);
2442 // Okay, check to see if the same value occurs in the operand list twice. If
2443 // so, delete one. Since we sorted the list, these values are required to
2445 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2446 // X smax Y smax Y --> X smax Y
2447 // X smax Y --> X, if X is always greater than Y
2448 if (Ops[i] == Ops[i+1] ||
2449 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2450 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2452 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2453 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2457 if (Ops.size() == 1) return Ops[0];
2459 assert(!Ops.empty() && "Reduced smax down to nothing!");
2461 // Okay, it looks like we really DO need an smax expr. Check to see if we
2462 // already have one, otherwise create a new one.
2463 FoldingSetNodeID ID;
2464 ID.AddInteger(scSMaxExpr);
2465 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2466 ID.AddPointer(Ops[i]);
2468 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2469 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2470 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2471 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2473 UniqueSCEVs.InsertNode(S, IP);
2477 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2479 SmallVector<const SCEV *, 2> Ops;
2482 return getUMaxExpr(Ops);
2486 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2487 assert(!Ops.empty() && "Cannot get empty umax!");
2488 if (Ops.size() == 1) return Ops[0];
2490 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2491 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2492 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2493 "SCEVUMaxExpr operand types don't match!");
2496 // Sort by complexity, this groups all similar expression types together.
2497 GroupByComplexity(Ops, LI);
2499 // If there are any constants, fold them together.
2501 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2503 assert(Idx < Ops.size());
2504 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2505 // We found two constants, fold them together!
2506 ConstantInt *Fold = ConstantInt::get(getContext(),
2507 APIntOps::umax(LHSC->getValue()->getValue(),
2508 RHSC->getValue()->getValue()));
2509 Ops[0] = getConstant(Fold);
2510 Ops.erase(Ops.begin()+1); // Erase the folded element
2511 if (Ops.size() == 1) return Ops[0];
2512 LHSC = cast<SCEVConstant>(Ops[0]);
2515 // If we are left with a constant minimum-int, strip it off.
2516 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2517 Ops.erase(Ops.begin());
2519 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2520 // If we have an umax with a constant maximum-int, it will always be
2525 if (Ops.size() == 1) return Ops[0];
2528 // Find the first UMax
2529 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2532 // Check to see if one of the operands is a UMax. If so, expand its operands
2533 // onto our operand list, and recurse to simplify.
2534 if (Idx < Ops.size()) {
2535 bool DeletedUMax = false;
2536 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2537 Ops.erase(Ops.begin()+Idx);
2538 Ops.append(UMax->op_begin(), UMax->op_end());
2543 return getUMaxExpr(Ops);
2546 // Okay, check to see if the same value occurs in the operand list twice. If
2547 // so, delete one. Since we sorted the list, these values are required to
2549 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2550 // X umax Y umax Y --> X umax Y
2551 // X umax Y --> X, if X is always greater than Y
2552 if (Ops[i] == Ops[i+1] ||
2553 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2554 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2556 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2557 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2561 if (Ops.size() == 1) return Ops[0];
2563 assert(!Ops.empty() && "Reduced umax down to nothing!");
2565 // Okay, it looks like we really DO need a umax expr. Check to see if we
2566 // already have one, otherwise create a new one.
2567 FoldingSetNodeID ID;
2568 ID.AddInteger(scUMaxExpr);
2569 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2570 ID.AddPointer(Ops[i]);
2572 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2573 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2574 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2575 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2577 UniqueSCEVs.InsertNode(S, IP);
2581 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2583 // ~smax(~x, ~y) == smin(x, y).
2584 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2587 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2589 // ~umax(~x, ~y) == umin(x, y)
2590 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2593 const SCEV *ScalarEvolution::getSizeOfExpr(Type *IntTy, Type *AllocTy) {
2594 // If we have DataLayout, we can bypass creating a target-independent
2595 // constant expression and then folding it back into a ConstantInt.
2596 // This is just a compile-time optimization.
2598 return getConstant(IntTy, TD->getTypeAllocSize(AllocTy));
2600 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2601 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2602 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2604 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2605 assert(Ty == IntTy && "Effective SCEV type doesn't match");
2606 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2609 const SCEV *ScalarEvolution::getAlignOfExpr(Type *AllocTy) {
2610 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2611 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2612 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2614 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2615 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2618 const SCEV *ScalarEvolution::getOffsetOfExpr(Type *IntTy,
2621 // If we have DataLayout, we can bypass creating a target-independent
2622 // constant expression and then folding it back into a ConstantInt.
2623 // This is just a compile-time optimization.
2625 return getConstant(IntTy,
2626 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2629 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2630 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2631 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2633 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2634 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2637 const SCEV *ScalarEvolution::getOffsetOfExpr(Type *IntTy,
2639 Constant *FieldNo) {
2640 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2641 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2642 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2644 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2645 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2648 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2649 // Don't attempt to do anything other than create a SCEVUnknown object
2650 // here. createSCEV only calls getUnknown after checking for all other
2651 // interesting possibilities, and any other code that calls getUnknown
2652 // is doing so in order to hide a value from SCEV canonicalization.
2654 FoldingSetNodeID ID;
2655 ID.AddInteger(scUnknown);
2658 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2659 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2660 "Stale SCEVUnknown in uniquing map!");
2663 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2665 FirstUnknown = cast<SCEVUnknown>(S);
2666 UniqueSCEVs.InsertNode(S, IP);
2670 //===----------------------------------------------------------------------===//
2671 // Basic SCEV Analysis and PHI Idiom Recognition Code
2674 /// isSCEVable - Test if values of the given type are analyzable within
2675 /// the SCEV framework. This primarily includes integer types, and it
2676 /// can optionally include pointer types if the ScalarEvolution class
2677 /// has access to target-specific information.
2678 bool ScalarEvolution::isSCEVable(Type *Ty) const {
2679 // Integers and pointers are always SCEVable.
2680 return Ty->isIntegerTy() || Ty->isPointerTy();
2683 /// getTypeSizeInBits - Return the size in bits of the specified type,
2684 /// for which isSCEVable must return true.
2685 uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const {
2686 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2688 // If we have a DataLayout, use it!
2690 return TD->getTypeSizeInBits(Ty);
2692 // Integer types have fixed sizes.
2693 if (Ty->isIntegerTy())
2694 return Ty->getPrimitiveSizeInBits();
2696 // The only other support type is pointer. Without DataLayout, conservatively
2697 // assume pointers are 64-bit.
2698 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2702 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2703 /// the given type and which represents how SCEV will treat the given
2704 /// type, for which isSCEVable must return true. For pointer types,
2705 /// this is the pointer-sized integer type.
2706 Type *ScalarEvolution::getEffectiveSCEVType(Type *Ty) const {
2707 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2709 if (Ty->isIntegerTy()) {
2713 // The only other support type is pointer.
2714 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2717 return TD->getIntPtrType(Ty);
2719 // Without DataLayout, conservatively assume pointers are 64-bit.
2720 return Type::getInt64Ty(getContext());
2723 const SCEV *ScalarEvolution::getCouldNotCompute() {
2724 return &CouldNotCompute;
2728 // Helper class working with SCEVTraversal to figure out if a SCEV contains
2729 // a SCEVUnknown with null value-pointer. FindInvalidSCEVUnknown::FindOne
2730 // is set iff if find such SCEVUnknown.
2732 struct FindInvalidSCEVUnknown {
2734 FindInvalidSCEVUnknown() { FindOne = false; }
2735 bool follow(const SCEV *S) {
2736 switch (S->getSCEVType()) {
2740 if (!cast<SCEVUnknown>(S)->getValue())
2747 bool isDone() const { return FindOne; }
2751 bool ScalarEvolution::checkValidity(const SCEV *S) const {
2752 FindInvalidSCEVUnknown F;
2753 SCEVTraversal<FindInvalidSCEVUnknown> ST(F);
2759 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2760 /// expression and create a new one.
2761 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2762 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2764 ValueExprMapType::iterator I = ValueExprMap.find_as(V);
2765 if (I != ValueExprMap.end()) {
2766 const SCEV *S = I->second;
2767 if (checkValidity(S))
2770 ValueExprMap.erase(I);
2772 const SCEV *S = createSCEV(V);
2774 // The process of creating a SCEV for V may have caused other SCEVs
2775 // to have been created, so it's necessary to insert the new entry
2776 // from scratch, rather than trying to remember the insert position
2778 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2782 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2784 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2785 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2787 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2789 Type *Ty = V->getType();
2790 Ty = getEffectiveSCEVType(Ty);
2791 return getMulExpr(V,
2792 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2795 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2796 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2797 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2799 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2801 Type *Ty = V->getType();
2802 Ty = getEffectiveSCEVType(Ty);
2803 const SCEV *AllOnes =
2804 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2805 return getMinusSCEV(AllOnes, V);
2808 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1.
2809 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2810 SCEV::NoWrapFlags Flags) {
2811 assert(!maskFlags(Flags, SCEV::FlagNUW) && "subtraction does not have NUW");
2813 // Fast path: X - X --> 0.
2815 return getConstant(LHS->getType(), 0);
2818 return getAddExpr(LHS, getNegativeSCEV(RHS), Flags);
2821 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2822 /// input value to the specified type. If the type must be extended, it is zero
2825 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, Type *Ty) {
2826 Type *SrcTy = V->getType();
2827 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2828 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2829 "Cannot truncate or zero extend with non-integer arguments!");
2830 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2831 return V; // No conversion
2832 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2833 return getTruncateExpr(V, Ty);
2834 return getZeroExtendExpr(V, Ty);
2837 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2838 /// input value to the specified type. If the type must be extended, it is sign
2841 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2843 Type *SrcTy = V->getType();
2844 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2845 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2846 "Cannot truncate or zero extend with non-integer arguments!");
2847 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2848 return V; // No conversion
2849 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2850 return getTruncateExpr(V, Ty);
2851 return getSignExtendExpr(V, Ty);
2854 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2855 /// input value to the specified type. If the type must be extended, it is zero
2856 /// extended. The conversion must not be narrowing.
2858 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, Type *Ty) {
2859 Type *SrcTy = V->getType();
2860 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2861 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2862 "Cannot noop or zero extend with non-integer arguments!");
2863 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2864 "getNoopOrZeroExtend cannot truncate!");
2865 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2866 return V; // No conversion
2867 return getZeroExtendExpr(V, Ty);
2870 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2871 /// input value to the specified type. If the type must be extended, it is sign
2872 /// extended. The conversion must not be narrowing.
2874 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, Type *Ty) {
2875 Type *SrcTy = V->getType();
2876 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2877 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2878 "Cannot noop or sign extend with non-integer arguments!");
2879 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2880 "getNoopOrSignExtend cannot truncate!");
2881 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2882 return V; // No conversion
2883 return getSignExtendExpr(V, Ty);
2886 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2887 /// the input value to the specified type. If the type must be extended,
2888 /// it is extended with unspecified bits. The conversion must not be
2891 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, Type *Ty) {
2892 Type *SrcTy = V->getType();
2893 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2894 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2895 "Cannot noop or any extend with non-integer arguments!");
2896 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2897 "getNoopOrAnyExtend cannot truncate!");
2898 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2899 return V; // No conversion
2900 return getAnyExtendExpr(V, Ty);
2903 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2904 /// input value to the specified type. The conversion must not be widening.
2906 ScalarEvolution::getTruncateOrNoop(const SCEV *V, Type *Ty) {
2907 Type *SrcTy = V->getType();
2908 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2909 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2910 "Cannot truncate or noop with non-integer arguments!");
2911 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2912 "getTruncateOrNoop cannot extend!");
2913 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2914 return V; // No conversion
2915 return getTruncateExpr(V, Ty);
2918 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2919 /// the types using zero-extension, and then perform a umax operation
2921 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2923 const SCEV *PromotedLHS = LHS;
2924 const SCEV *PromotedRHS = RHS;
2926 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2927 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2929 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2931 return getUMaxExpr(PromotedLHS, PromotedRHS);
2934 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2935 /// the types using zero-extension, and then perform a umin operation
2937 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2939 const SCEV *PromotedLHS = LHS;
2940 const SCEV *PromotedRHS = RHS;
2942 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2943 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2945 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2947 return getUMinExpr(PromotedLHS, PromotedRHS);
2950 /// getPointerBase - Transitively follow the chain of pointer-type operands
2951 /// until reaching a SCEV that does not have a single pointer operand. This
2952 /// returns a SCEVUnknown pointer for well-formed pointer-type expressions,
2953 /// but corner cases do exist.
2954 const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) {
2955 // A pointer operand may evaluate to a nonpointer expression, such as null.
2956 if (!V->getType()->isPointerTy())
2959 if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(V)) {
2960 return getPointerBase(Cast->getOperand());
2962 else if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(V)) {
2963 const SCEV *PtrOp = 0;
2964 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
2966 if ((*I)->getType()->isPointerTy()) {
2967 // Cannot find the base of an expression with multiple pointer operands.
2975 return getPointerBase(PtrOp);
2980 /// PushDefUseChildren - Push users of the given Instruction
2981 /// onto the given Worklist.
2983 PushDefUseChildren(Instruction *I,
2984 SmallVectorImpl<Instruction *> &Worklist) {
2985 // Push the def-use children onto the Worklist stack.
2986 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2988 Worklist.push_back(cast<Instruction>(*UI));
2991 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2992 /// instructions that depend on the given instruction and removes them from
2993 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2996 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2997 SmallVector<Instruction *, 16> Worklist;
2998 PushDefUseChildren(PN, Worklist);
3000 SmallPtrSet<Instruction *, 8> Visited;
3002 while (!Worklist.empty()) {
3003 Instruction *I = Worklist.pop_back_val();
3004 if (!Visited.insert(I)) continue;
3006 ValueExprMapType::iterator It =
3007 ValueExprMap.find_as(static_cast<Value *>(I));
3008 if (It != ValueExprMap.end()) {
3009 const SCEV *Old = It->second;
3011 // Short-circuit the def-use traversal if the symbolic name
3012 // ceases to appear in expressions.
3013 if (Old != SymName && !hasOperand(Old, SymName))
3016 // SCEVUnknown for a PHI either means that it has an unrecognized
3017 // structure, it's a PHI that's in the progress of being computed
3018 // by createNodeForPHI, or it's a single-value PHI. In the first case,
3019 // additional loop trip count information isn't going to change anything.
3020 // In the second case, createNodeForPHI will perform the necessary
3021 // updates on its own when it gets to that point. In the third, we do
3022 // want to forget the SCEVUnknown.
3023 if (!isa<PHINode>(I) ||
3024 !isa<SCEVUnknown>(Old) ||
3025 (I != PN && Old == SymName)) {
3026 forgetMemoizedResults(Old);
3027 ValueExprMap.erase(It);
3031 PushDefUseChildren(I, Worklist);
3035 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
3036 /// a loop header, making it a potential recurrence, or it doesn't.
3038 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
3039 if (const Loop *L = LI->getLoopFor(PN->getParent()))
3040 if (L->getHeader() == PN->getParent()) {
3041 // The loop may have multiple entrances or multiple exits; we can analyze
3042 // this phi as an addrec if it has a unique entry value and a unique
3044 Value *BEValueV = 0, *StartValueV = 0;
3045 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
3046 Value *V = PN->getIncomingValue(i);
3047 if (L->contains(PN->getIncomingBlock(i))) {
3050 } else if (BEValueV != V) {
3054 } else if (!StartValueV) {
3056 } else if (StartValueV != V) {
3061 if (BEValueV && StartValueV) {
3062 // While we are analyzing this PHI node, handle its value symbolically.
3063 const SCEV *SymbolicName = getUnknown(PN);
3064 assert(ValueExprMap.find_as(PN) == ValueExprMap.end() &&
3065 "PHI node already processed?");
3066 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
3068 // Using this symbolic name for the PHI, analyze the value coming around
3070 const SCEV *BEValue = getSCEV(BEValueV);
3072 // NOTE: If BEValue is loop invariant, we know that the PHI node just
3073 // has a special value for the first iteration of the loop.
3075 // If the value coming around the backedge is an add with the symbolic
3076 // value we just inserted, then we found a simple induction variable!
3077 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
3078 // If there is a single occurrence of the symbolic value, replace it
3079 // with a recurrence.
3080 unsigned FoundIndex = Add->getNumOperands();
3081 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3082 if (Add->getOperand(i) == SymbolicName)
3083 if (FoundIndex == e) {
3088 if (FoundIndex != Add->getNumOperands()) {
3089 // Create an add with everything but the specified operand.
3090 SmallVector<const SCEV *, 8> Ops;
3091 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3092 if (i != FoundIndex)
3093 Ops.push_back(Add->getOperand(i));
3094 const SCEV *Accum = getAddExpr(Ops);
3096 // This is not a valid addrec if the step amount is varying each
3097 // loop iteration, but is not itself an addrec in this loop.
3098 if (isLoopInvariant(Accum, L) ||
3099 (isa<SCEVAddRecExpr>(Accum) &&
3100 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
3101 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap;
3103 // If the increment doesn't overflow, then neither the addrec nor
3104 // the post-increment will overflow.
3105 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
3106 if (OBO->hasNoUnsignedWrap())
3107 Flags = setFlags(Flags, SCEV::FlagNUW);
3108 if (OBO->hasNoSignedWrap())
3109 Flags = setFlags(Flags, SCEV::FlagNSW);
3110 } else if (const GEPOperator *GEP =
3111 dyn_cast<GEPOperator>(BEValueV)) {
3112 // If the increment is an inbounds GEP, then we know the address
3113 // space cannot be wrapped around. We cannot make any guarantee
3114 // about signed or unsigned overflow because pointers are
3115 // unsigned but we may have a negative index from the base
3117 if (GEP->isInBounds())
3118 Flags = setFlags(Flags, SCEV::FlagNW);
3121 const SCEV *StartVal = getSCEV(StartValueV);
3122 const SCEV *PHISCEV = getAddRecExpr(StartVal, Accum, L, Flags);
3124 // Since the no-wrap flags are on the increment, they apply to the
3125 // post-incremented value as well.
3126 if (isLoopInvariant(Accum, L))
3127 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
3130 // Okay, for the entire analysis of this edge we assumed the PHI
3131 // to be symbolic. We now need to go back and purge all of the
3132 // entries for the scalars that use the symbolic expression.
3133 ForgetSymbolicName(PN, SymbolicName);
3134 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3138 } else if (const SCEVAddRecExpr *AddRec =
3139 dyn_cast<SCEVAddRecExpr>(BEValue)) {
3140 // Otherwise, this could be a loop like this:
3141 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
3142 // In this case, j = {1,+,1} and BEValue is j.
3143 // Because the other in-value of i (0) fits the evolution of BEValue
3144 // i really is an addrec evolution.
3145 if (AddRec->getLoop() == L && AddRec->isAffine()) {
3146 const SCEV *StartVal = getSCEV(StartValueV);
3148 // If StartVal = j.start - j.stride, we can use StartVal as the
3149 // initial step of the addrec evolution.
3150 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
3151 AddRec->getOperand(1))) {
3152 // FIXME: For constant StartVal, we should be able to infer
3154 const SCEV *PHISCEV =
3155 getAddRecExpr(StartVal, AddRec->getOperand(1), L,
3158 // Okay, for the entire analysis of this edge we assumed the PHI
3159 // to be symbolic. We now need to go back and purge all of the
3160 // entries for the scalars that use the symbolic expression.
3161 ForgetSymbolicName(PN, SymbolicName);
3162 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3170 // If the PHI has a single incoming value, follow that value, unless the
3171 // PHI's incoming blocks are in a different loop, in which case doing so
3172 // risks breaking LCSSA form. Instcombine would normally zap these, but
3173 // it doesn't have DominatorTree information, so it may miss cases.
3174 if (Value *V = SimplifyInstruction(PN, TD, TLI, DT))
3175 if (LI->replacementPreservesLCSSAForm(PN, V))
3178 // If it's not a loop phi, we can't handle it yet.
3179 return getUnknown(PN);
3182 /// createNodeForGEP - Expand GEP instructions into add and multiply
3183 /// operations. This allows them to be analyzed by regular SCEV code.
3185 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
3187 // Don't blindly transfer the inbounds flag from the GEP instruction to the
3188 // Add expression, because the Instruction may be guarded by control flow
3189 // and the no-overflow bits may not be valid for the expression in any
3191 bool isInBounds = GEP->isInBounds();
3193 Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
3194 Value *Base = GEP->getOperand(0);
3195 // Don't attempt to analyze GEPs over unsized objects.
3196 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
3197 return getUnknown(GEP);
3198 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
3199 gep_type_iterator GTI = gep_type_begin(GEP);
3200 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
3204 // Compute the (potentially symbolic) offset in bytes for this index.
3205 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
3206 // For a struct, add the member offset.
3207 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
3208 const SCEV *FieldOffset = getOffsetOfExpr(IntPtrTy, STy, FieldNo);
3210 // Add the field offset to the running total offset.
3211 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
3213 // For an array, add the element offset, explicitly scaled.
3214 const SCEV *ElementSize = getSizeOfExpr(IntPtrTy, *GTI);
3215 const SCEV *IndexS = getSCEV(Index);
3216 // Getelementptr indices are signed.
3217 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
3219 // Multiply the index by the element size to compute the element offset.
3220 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize,
3221 isInBounds ? SCEV::FlagNSW :
3224 // Add the element offset to the running total offset.
3225 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
3229 // Get the SCEV for the GEP base.
3230 const SCEV *BaseS = getSCEV(Base);
3232 // Add the total offset from all the GEP indices to the base.
3233 return getAddExpr(BaseS, TotalOffset,
3234 isInBounds ? SCEV::FlagNSW : SCEV::FlagAnyWrap);
3237 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
3238 /// guaranteed to end in (at every loop iteration). It is, at the same time,
3239 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
3240 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
3242 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
3243 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3244 return C->getValue()->getValue().countTrailingZeros();
3246 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
3247 return std::min(GetMinTrailingZeros(T->getOperand()),
3248 (uint32_t)getTypeSizeInBits(T->getType()));
3250 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
3251 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3252 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3253 getTypeSizeInBits(E->getType()) : OpRes;
3256 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
3257 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3258 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3259 getTypeSizeInBits(E->getType()) : OpRes;
3262 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
3263 // The result is the min of all operands results.
3264 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3265 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3266 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3270 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
3271 // The result is the sum of all operands results.
3272 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
3273 uint32_t BitWidth = getTypeSizeInBits(M->getType());
3274 for (unsigned i = 1, e = M->getNumOperands();
3275 SumOpRes != BitWidth && i != e; ++i)
3276 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
3281 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
3282 // The result is the min of all operands results.
3283 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3284 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3285 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3289 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
3290 // The result is the min of all operands results.
3291 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3292 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3293 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3297 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
3298 // The result is the min of all operands results.
3299 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3300 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3301 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3305 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3306 // For a SCEVUnknown, ask ValueTracking.
3307 unsigned BitWidth = getTypeSizeInBits(U->getType());
3308 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3309 ComputeMaskedBits(U->getValue(), Zeros, Ones);
3310 return Zeros.countTrailingOnes();
3317 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
3320 ScalarEvolution::getUnsignedRange(const SCEV *S) {
3321 // See if we've computed this range already.
3322 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
3323 if (I != UnsignedRanges.end())
3326 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3327 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
3329 unsigned BitWidth = getTypeSizeInBits(S->getType());
3330 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3332 // If the value has known zeros, the maximum unsigned value will have those
3333 // known zeros as well.
3334 uint32_t TZ = GetMinTrailingZeros(S);
3336 ConservativeResult =
3337 ConstantRange(APInt::getMinValue(BitWidth),
3338 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
3340 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3341 ConstantRange X = getUnsignedRange(Add->getOperand(0));
3342 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3343 X = X.add(getUnsignedRange(Add->getOperand(i)));
3344 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
3347 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3348 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
3349 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3350 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
3351 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
3354 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3355 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
3356 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3357 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
3358 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
3361 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3362 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3363 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3364 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3365 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
3368 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3369 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3370 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3371 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3374 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3375 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3376 return setUnsignedRange(ZExt,
3377 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3380 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3381 ConstantRange X = getUnsignedRange(SExt->getOperand());
3382 return setUnsignedRange(SExt,
3383 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3386 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3387 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3388 return setUnsignedRange(Trunc,
3389 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3392 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3393 // If there's no unsigned wrap, the value will never be less than its
3395 if (AddRec->getNoWrapFlags(SCEV::FlagNUW))
3396 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3397 if (!C->getValue()->isZero())
3398 ConservativeResult =
3399 ConservativeResult.intersectWith(
3400 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3402 // TODO: non-affine addrec
3403 if (AddRec->isAffine()) {
3404 Type *Ty = AddRec->getType();
3405 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3406 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3407 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3408 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3410 const SCEV *Start = AddRec->getStart();
3411 const SCEV *Step = AddRec->getStepRecurrence(*this);
3413 ConstantRange StartRange = getUnsignedRange(Start);
3414 ConstantRange StepRange = getSignedRange(Step);
3415 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3416 ConstantRange EndRange =
3417 StartRange.add(MaxBECountRange.multiply(StepRange));
3419 // Check for overflow. This must be done with ConstantRange arithmetic
3420 // because we could be called from within the ScalarEvolution overflow
3422 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3423 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3424 ConstantRange ExtMaxBECountRange =
3425 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3426 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3427 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3429 return setUnsignedRange(AddRec, ConservativeResult);
3431 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3432 EndRange.getUnsignedMin());
3433 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3434 EndRange.getUnsignedMax());
3435 if (Min.isMinValue() && Max.isMaxValue())
3436 return setUnsignedRange(AddRec, ConservativeResult);
3437 return setUnsignedRange(AddRec,
3438 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3442 return setUnsignedRange(AddRec, ConservativeResult);
3445 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3446 // For a SCEVUnknown, ask ValueTracking.
3447 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3448 ComputeMaskedBits(U->getValue(), Zeros, Ones, TD);
3449 if (Ones == ~Zeros + 1)
3450 return setUnsignedRange(U, ConservativeResult);
3451 return setUnsignedRange(U,
3452 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3455 return setUnsignedRange(S, ConservativeResult);
3458 /// getSignedRange - Determine the signed range for a particular SCEV.
3461 ScalarEvolution::getSignedRange(const SCEV *S) {
3462 // See if we've computed this range already.
3463 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3464 if (I != SignedRanges.end())
3467 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3468 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3470 unsigned BitWidth = getTypeSizeInBits(S->getType());
3471 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3473 // If the value has known zeros, the maximum signed value will have those
3474 // known zeros as well.
3475 uint32_t TZ = GetMinTrailingZeros(S);
3477 ConservativeResult =
3478 ConstantRange(APInt::getSignedMinValue(BitWidth),
3479 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3481 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3482 ConstantRange X = getSignedRange(Add->getOperand(0));
3483 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3484 X = X.add(getSignedRange(Add->getOperand(i)));
3485 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3488 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3489 ConstantRange X = getSignedRange(Mul->getOperand(0));
3490 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3491 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3492 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3495 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3496 ConstantRange X = getSignedRange(SMax->getOperand(0));
3497 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3498 X = X.smax(getSignedRange(SMax->getOperand(i)));
3499 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3502 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3503 ConstantRange X = getSignedRange(UMax->getOperand(0));
3504 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3505 X = X.umax(getSignedRange(UMax->getOperand(i)));
3506 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3509 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3510 ConstantRange X = getSignedRange(UDiv->getLHS());
3511 ConstantRange Y = getSignedRange(UDiv->getRHS());
3512 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3515 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3516 ConstantRange X = getSignedRange(ZExt->getOperand());
3517 return setSignedRange(ZExt,
3518 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3521 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3522 ConstantRange X = getSignedRange(SExt->getOperand());
3523 return setSignedRange(SExt,
3524 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3527 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3528 ConstantRange X = getSignedRange(Trunc->getOperand());
3529 return setSignedRange(Trunc,
3530 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3533 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3534 // If there's no signed wrap, and all the operands have the same sign or
3535 // zero, the value won't ever change sign.
3536 if (AddRec->getNoWrapFlags(SCEV::FlagNSW)) {
3537 bool AllNonNeg = true;
3538 bool AllNonPos = true;
3539 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3540 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3541 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3544 ConservativeResult = ConservativeResult.intersectWith(
3545 ConstantRange(APInt(BitWidth, 0),
3546 APInt::getSignedMinValue(BitWidth)));
3548 ConservativeResult = ConservativeResult.intersectWith(
3549 ConstantRange(APInt::getSignedMinValue(BitWidth),
3550 APInt(BitWidth, 1)));
3553 // TODO: non-affine addrec
3554 if (AddRec->isAffine()) {
3555 Type *Ty = AddRec->getType();
3556 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3557 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3558 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3559 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3561 const SCEV *Start = AddRec->getStart();
3562 const SCEV *Step = AddRec->getStepRecurrence(*this);
3564 ConstantRange StartRange = getSignedRange(Start);
3565 ConstantRange StepRange = getSignedRange(Step);
3566 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3567 ConstantRange EndRange =
3568 StartRange.add(MaxBECountRange.multiply(StepRange));
3570 // Check for overflow. This must be done with ConstantRange arithmetic
3571 // because we could be called from within the ScalarEvolution overflow
3573 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3574 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3575 ConstantRange ExtMaxBECountRange =
3576 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3577 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3578 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3580 return setSignedRange(AddRec, ConservativeResult);
3582 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3583 EndRange.getSignedMin());
3584 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3585 EndRange.getSignedMax());
3586 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3587 return setSignedRange(AddRec, ConservativeResult);
3588 return setSignedRange(AddRec,
3589 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3593 return setSignedRange(AddRec, ConservativeResult);
3596 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3597 // For a SCEVUnknown, ask ValueTracking.
3598 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3599 return setSignedRange(U, ConservativeResult);
3600 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3602 return setSignedRange(U, ConservativeResult);
3603 return setSignedRange(U, ConservativeResult.intersectWith(
3604 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3605 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3608 return setSignedRange(S, ConservativeResult);
3611 /// createSCEV - We know that there is no SCEV for the specified value.
3612 /// Analyze the expression.
3614 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3615 if (!isSCEVable(V->getType()))
3616 return getUnknown(V);
3618 unsigned Opcode = Instruction::UserOp1;
3619 if (Instruction *I = dyn_cast<Instruction>(V)) {
3620 Opcode = I->getOpcode();
3622 // Don't attempt to analyze instructions in blocks that aren't
3623 // reachable. Such instructions don't matter, and they aren't required
3624 // to obey basic rules for definitions dominating uses which this
3625 // analysis depends on.
3626 if (!DT->isReachableFromEntry(I->getParent()))
3627 return getUnknown(V);
3628 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3629 Opcode = CE->getOpcode();
3630 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3631 return getConstant(CI);
3632 else if (isa<ConstantPointerNull>(V))
3633 return getConstant(V->getType(), 0);
3634 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3635 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3637 return getUnknown(V);
3639 Operator *U = cast<Operator>(V);
3641 case Instruction::Add: {
3642 // The simple thing to do would be to just call getSCEV on both operands
3643 // and call getAddExpr with the result. However if we're looking at a
3644 // bunch of things all added together, this can be quite inefficient,
3645 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3646 // Instead, gather up all the operands and make a single getAddExpr call.
3647 // LLVM IR canonical form means we need only traverse the left operands.
3649 // Don't apply this instruction's NSW or NUW flags to the new
3650 // expression. The instruction may be guarded by control flow that the
3651 // no-wrap behavior depends on. Non-control-equivalent instructions can be
3652 // mapped to the same SCEV expression, and it would be incorrect to transfer
3653 // NSW/NUW semantics to those operations.
3654 SmallVector<const SCEV *, 4> AddOps;
3655 AddOps.push_back(getSCEV(U->getOperand(1)));
3656 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3657 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3658 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3660 U = cast<Operator>(Op);
3661 const SCEV *Op1 = getSCEV(U->getOperand(1));
3662 if (Opcode == Instruction::Sub)
3663 AddOps.push_back(getNegativeSCEV(Op1));
3665 AddOps.push_back(Op1);
3667 AddOps.push_back(getSCEV(U->getOperand(0)));
3668 return getAddExpr(AddOps);
3670 case Instruction::Mul: {
3671 // Don't transfer NSW/NUW for the same reason as AddExpr.
3672 SmallVector<const SCEV *, 4> MulOps;
3673 MulOps.push_back(getSCEV(U->getOperand(1)));
3674 for (Value *Op = U->getOperand(0);
3675 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3676 Op = U->getOperand(0)) {
3677 U = cast<Operator>(Op);
3678 MulOps.push_back(getSCEV(U->getOperand(1)));
3680 MulOps.push_back(getSCEV(U->getOperand(0)));
3681 return getMulExpr(MulOps);
3683 case Instruction::UDiv:
3684 return getUDivExpr(getSCEV(U->getOperand(0)),
3685 getSCEV(U->getOperand(1)));
3686 case Instruction::Sub:
3687 return getMinusSCEV(getSCEV(U->getOperand(0)),
3688 getSCEV(U->getOperand(1)));
3689 case Instruction::And:
3690 // For an expression like x&255 that merely masks off the high bits,
3691 // use zext(trunc(x)) as the SCEV expression.
3692 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3693 if (CI->isNullValue())
3694 return getSCEV(U->getOperand(1));
3695 if (CI->isAllOnesValue())
3696 return getSCEV(U->getOperand(0));
3697 const APInt &A = CI->getValue();
3699 // Instcombine's ShrinkDemandedConstant may strip bits out of
3700 // constants, obscuring what would otherwise be a low-bits mask.
3701 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3702 // knew about to reconstruct a low-bits mask value.
3703 unsigned LZ = A.countLeadingZeros();
3704 unsigned BitWidth = A.getBitWidth();
3705 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3706 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD);
3708 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3710 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3712 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3713 IntegerType::get(getContext(), BitWidth - LZ)),
3718 case Instruction::Or:
3719 // If the RHS of the Or is a constant, we may have something like:
3720 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3721 // optimizations will transparently handle this case.
3723 // In order for this transformation to be safe, the LHS must be of the
3724 // form X*(2^n) and the Or constant must be less than 2^n.
3725 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3726 const SCEV *LHS = getSCEV(U->getOperand(0));
3727 const APInt &CIVal = CI->getValue();
3728 if (GetMinTrailingZeros(LHS) >=
3729 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3730 // Build a plain add SCEV.
3731 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3732 // If the LHS of the add was an addrec and it has no-wrap flags,
3733 // transfer the no-wrap flags, since an or won't introduce a wrap.
3734 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3735 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3736 const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags(
3737 OldAR->getNoWrapFlags());
3743 case Instruction::Xor:
3744 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3745 // If the RHS of the xor is a signbit, then this is just an add.
3746 // Instcombine turns add of signbit into xor as a strength reduction step.
3747 if (CI->getValue().isSignBit())
3748 return getAddExpr(getSCEV(U->getOperand(0)),
3749 getSCEV(U->getOperand(1)));
3751 // If the RHS of xor is -1, then this is a not operation.
3752 if (CI->isAllOnesValue())
3753 return getNotSCEV(getSCEV(U->getOperand(0)));
3755 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3756 // This is a variant of the check for xor with -1, and it handles
3757 // the case where instcombine has trimmed non-demanded bits out
3758 // of an xor with -1.
3759 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3760 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3761 if (BO->getOpcode() == Instruction::And &&
3762 LCI->getValue() == CI->getValue())
3763 if (const SCEVZeroExtendExpr *Z =
3764 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3765 Type *UTy = U->getType();
3766 const SCEV *Z0 = Z->getOperand();
3767 Type *Z0Ty = Z0->getType();
3768 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3770 // If C is a low-bits mask, the zero extend is serving to
3771 // mask off the high bits. Complement the operand and
3772 // re-apply the zext.
3773 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3774 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3776 // If C is a single bit, it may be in the sign-bit position
3777 // before the zero-extend. In this case, represent the xor
3778 // using an add, which is equivalent, and re-apply the zext.
3779 APInt Trunc = CI->getValue().trunc(Z0TySize);
3780 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3782 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3788 case Instruction::Shl:
3789 // Turn shift left of a constant amount into a multiply.
3790 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3791 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3793 // If the shift count is not less than the bitwidth, the result of
3794 // the shift is undefined. Don't try to analyze it, because the
3795 // resolution chosen here may differ from the resolution chosen in
3796 // other parts of the compiler.
3797 if (SA->getValue().uge(BitWidth))
3800 Constant *X = ConstantInt::get(getContext(),
3801 APInt::getOneBitSet(BitWidth, SA->getZExtValue()));
3802 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3806 case Instruction::LShr:
3807 // Turn logical shift right of a constant into a unsigned divide.
3808 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3809 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3811 // If the shift count is not less than the bitwidth, the result of
3812 // the shift is undefined. Don't try to analyze it, because the
3813 // resolution chosen here may differ from the resolution chosen in
3814 // other parts of the compiler.
3815 if (SA->getValue().uge(BitWidth))
3818 Constant *X = ConstantInt::get(getContext(),
3819 APInt::getOneBitSet(BitWidth, SA->getZExtValue()));
3820 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3824 case Instruction::AShr:
3825 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3826 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3827 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3828 if (L->getOpcode() == Instruction::Shl &&
3829 L->getOperand(1) == U->getOperand(1)) {
3830 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3832 // If the shift count is not less than the bitwidth, the result of
3833 // the shift is undefined. Don't try to analyze it, because the
3834 // resolution chosen here may differ from the resolution chosen in
3835 // other parts of the compiler.
3836 if (CI->getValue().uge(BitWidth))
3839 uint64_t Amt = BitWidth - CI->getZExtValue();
3840 if (Amt == BitWidth)
3841 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3843 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3844 IntegerType::get(getContext(),
3850 case Instruction::Trunc:
3851 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3853 case Instruction::ZExt:
3854 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3856 case Instruction::SExt:
3857 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3859 case Instruction::BitCast:
3860 // BitCasts are no-op casts so we just eliminate the cast.
3861 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3862 return getSCEV(U->getOperand(0));
3865 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3866 // lead to pointer expressions which cannot safely be expanded to GEPs,
3867 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3868 // simplifying integer expressions.
3870 case Instruction::GetElementPtr:
3871 return createNodeForGEP(cast<GEPOperator>(U));
3873 case Instruction::PHI:
3874 return createNodeForPHI(cast<PHINode>(U));
3876 case Instruction::Select:
3877 // This could be a smax or umax that was lowered earlier.
3878 // Try to recover it.
3879 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3880 Value *LHS = ICI->getOperand(0);
3881 Value *RHS = ICI->getOperand(1);
3882 switch (ICI->getPredicate()) {
3883 case ICmpInst::ICMP_SLT:
3884 case ICmpInst::ICMP_SLE:
3885 std::swap(LHS, RHS);
3887 case ICmpInst::ICMP_SGT:
3888 case ICmpInst::ICMP_SGE:
3889 // a >s b ? a+x : b+x -> smax(a, b)+x
3890 // a >s b ? b+x : a+x -> smin(a, b)+x
3891 if (LHS->getType() == U->getType()) {
3892 const SCEV *LS = getSCEV(LHS);
3893 const SCEV *RS = getSCEV(RHS);
3894 const SCEV *LA = getSCEV(U->getOperand(1));
3895 const SCEV *RA = getSCEV(U->getOperand(2));
3896 const SCEV *LDiff = getMinusSCEV(LA, LS);
3897 const SCEV *RDiff = getMinusSCEV(RA, RS);
3899 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3900 LDiff = getMinusSCEV(LA, RS);
3901 RDiff = getMinusSCEV(RA, LS);
3903 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3906 case ICmpInst::ICMP_ULT:
3907 case ICmpInst::ICMP_ULE:
3908 std::swap(LHS, RHS);
3910 case ICmpInst::ICMP_UGT:
3911 case ICmpInst::ICMP_UGE:
3912 // a >u b ? a+x : b+x -> umax(a, b)+x
3913 // a >u b ? b+x : a+x -> umin(a, b)+x
3914 if (LHS->getType() == U->getType()) {
3915 const SCEV *LS = getSCEV(LHS);
3916 const SCEV *RS = getSCEV(RHS);
3917 const SCEV *LA = getSCEV(U->getOperand(1));
3918 const SCEV *RA = getSCEV(U->getOperand(2));
3919 const SCEV *LDiff = getMinusSCEV(LA, LS);
3920 const SCEV *RDiff = getMinusSCEV(RA, RS);
3922 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3923 LDiff = getMinusSCEV(LA, RS);
3924 RDiff = getMinusSCEV(RA, LS);
3926 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3929 case ICmpInst::ICMP_NE:
3930 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3931 if (LHS->getType() == U->getType() &&
3932 isa<ConstantInt>(RHS) &&
3933 cast<ConstantInt>(RHS)->isZero()) {
3934 const SCEV *One = getConstant(LHS->getType(), 1);
3935 const SCEV *LS = getSCEV(LHS);
3936 const SCEV *LA = getSCEV(U->getOperand(1));
3937 const SCEV *RA = getSCEV(U->getOperand(2));
3938 const SCEV *LDiff = getMinusSCEV(LA, LS);
3939 const SCEV *RDiff = getMinusSCEV(RA, One);
3941 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3944 case ICmpInst::ICMP_EQ:
3945 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3946 if (LHS->getType() == U->getType() &&
3947 isa<ConstantInt>(RHS) &&
3948 cast<ConstantInt>(RHS)->isZero()) {
3949 const SCEV *One = getConstant(LHS->getType(), 1);
3950 const SCEV *LS = getSCEV(LHS);
3951 const SCEV *LA = getSCEV(U->getOperand(1));
3952 const SCEV *RA = getSCEV(U->getOperand(2));
3953 const SCEV *LDiff = getMinusSCEV(LA, One);
3954 const SCEV *RDiff = getMinusSCEV(RA, LS);
3956 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3964 default: // We cannot analyze this expression.
3968 return getUnknown(V);
3973 //===----------------------------------------------------------------------===//
3974 // Iteration Count Computation Code
3977 /// getSmallConstantTripCount - Returns the maximum trip count of this loop as a
3978 /// normal unsigned value. Returns 0 if the trip count is unknown or not
3979 /// constant. Will also return 0 if the maximum trip count is very large (>=
3982 /// This "trip count" assumes that control exits via ExitingBlock. More
3983 /// precisely, it is the number of times that control may reach ExitingBlock
3984 /// before taking the branch. For loops with multiple exits, it may not be the
3985 /// number times that the loop header executes because the loop may exit
3986 /// prematurely via another branch.
3988 /// FIXME: We conservatively call getBackedgeTakenCount(L) instead of
3989 /// getExitCount(L, ExitingBlock) to compute a safe trip count considering all
3990 /// loop exits. getExitCount() may return an exact count for this branch
3991 /// assuming no-signed-wrap. The number of well-defined iterations may actually
3992 /// be higher than this trip count if this exit test is skipped and the loop
3993 /// exits via a different branch. Ideally, getExitCount() would know whether it
3994 /// depends on a NSW assumption, and we would only fall back to a conservative
3995 /// trip count in that case.
3996 unsigned ScalarEvolution::
3997 getSmallConstantTripCount(Loop *L, BasicBlock * /*ExitingBlock*/) {
3998 const SCEVConstant *ExitCount =
3999 dyn_cast<SCEVConstant>(getBackedgeTakenCount(L));
4003 ConstantInt *ExitConst = ExitCount->getValue();
4005 // Guard against huge trip counts.
4006 if (ExitConst->getValue().getActiveBits() > 32)
4009 // In case of integer overflow, this returns 0, which is correct.
4010 return ((unsigned)ExitConst->getZExtValue()) + 1;
4013 /// getSmallConstantTripMultiple - Returns the largest constant divisor of the
4014 /// trip count of this loop as a normal unsigned value, if possible. This
4015 /// means that the actual trip count is always a multiple of the returned
4016 /// value (don't forget the trip count could very well be zero as well!).
4018 /// Returns 1 if the trip count is unknown or not guaranteed to be the
4019 /// multiple of a constant (which is also the case if the trip count is simply
4020 /// constant, use getSmallConstantTripCount for that case), Will also return 1
4021 /// if the trip count is very large (>= 2^32).
4023 /// As explained in the comments for getSmallConstantTripCount, this assumes
4024 /// that control exits the loop via ExitingBlock.
4025 unsigned ScalarEvolution::
4026 getSmallConstantTripMultiple(Loop *L, BasicBlock * /*ExitingBlock*/) {
4027 const SCEV *ExitCount = getBackedgeTakenCount(L);
4028 if (ExitCount == getCouldNotCompute())
4031 // Get the trip count from the BE count by adding 1.
4032 const SCEV *TCMul = getAddExpr(ExitCount,
4033 getConstant(ExitCount->getType(), 1));
4034 // FIXME: SCEV distributes multiplication as V1*C1 + V2*C1. We could attempt
4035 // to factor simple cases.
4036 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(TCMul))
4037 TCMul = Mul->getOperand(0);
4039 const SCEVConstant *MulC = dyn_cast<SCEVConstant>(TCMul);
4043 ConstantInt *Result = MulC->getValue();
4045 // Guard against huge trip counts (this requires checking
4046 // for zero to handle the case where the trip count == -1 and the
4048 if (!Result || Result->getValue().getActiveBits() > 32 ||
4049 Result->getValue().getActiveBits() == 0)
4052 return (unsigned)Result->getZExtValue();
4055 // getExitCount - Get the expression for the number of loop iterations for which
4056 // this loop is guaranteed not to exit via ExitingBlock. Otherwise return
4057 // SCEVCouldNotCompute.
4058 const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) {
4059 return getBackedgeTakenInfo(L).getExact(ExitingBlock, this);
4062 /// getBackedgeTakenCount - If the specified loop has a predictable
4063 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
4064 /// object. The backedge-taken count is the number of times the loop header
4065 /// will be branched to from within the loop. This is one less than the
4066 /// trip count of the loop, since it doesn't count the first iteration,
4067 /// when the header is branched to from outside the loop.
4069 /// Note that it is not valid to call this method on a loop without a
4070 /// loop-invariant backedge-taken count (see
4071 /// hasLoopInvariantBackedgeTakenCount).
4073 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
4074 return getBackedgeTakenInfo(L).getExact(this);
4077 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
4078 /// return the least SCEV value that is known never to be less than the
4079 /// actual backedge taken count.
4080 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
4081 return getBackedgeTakenInfo(L).getMax(this);
4084 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
4085 /// onto the given Worklist.
4087 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
4088 BasicBlock *Header = L->getHeader();
4090 // Push all Loop-header PHIs onto the Worklist stack.
4091 for (BasicBlock::iterator I = Header->begin();
4092 PHINode *PN = dyn_cast<PHINode>(I); ++I)
4093 Worklist.push_back(PN);
4096 const ScalarEvolution::BackedgeTakenInfo &
4097 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
4098 // Initially insert an invalid entry for this loop. If the insertion
4099 // succeeds, proceed to actually compute a backedge-taken count and
4100 // update the value. The temporary CouldNotCompute value tells SCEV
4101 // code elsewhere that it shouldn't attempt to request a new
4102 // backedge-taken count, which could result in infinite recursion.
4103 std::pair<DenseMap<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
4104 BackedgeTakenCounts.insert(std::make_pair(L, BackedgeTakenInfo()));
4106 return Pair.first->second;
4108 // ComputeBackedgeTakenCount may allocate memory for its result. Inserting it
4109 // into the BackedgeTakenCounts map transfers ownership. Otherwise, the result
4110 // must be cleared in this scope.
4111 BackedgeTakenInfo Result = ComputeBackedgeTakenCount(L);
4113 if (Result.getExact(this) != getCouldNotCompute()) {
4114 assert(isLoopInvariant(Result.getExact(this), L) &&
4115 isLoopInvariant(Result.getMax(this), L) &&
4116 "Computed backedge-taken count isn't loop invariant for loop!");
4117 ++NumTripCountsComputed;
4119 else if (Result.getMax(this) == getCouldNotCompute() &&
4120 isa<PHINode>(L->getHeader()->begin())) {
4121 // Only count loops that have phi nodes as not being computable.
4122 ++NumTripCountsNotComputed;
4125 // Now that we know more about the trip count for this loop, forget any
4126 // existing SCEV values for PHI nodes in this loop since they are only
4127 // conservative estimates made without the benefit of trip count
4128 // information. This is similar to the code in forgetLoop, except that
4129 // it handles SCEVUnknown PHI nodes specially.
4130 if (Result.hasAnyInfo()) {
4131 SmallVector<Instruction *, 16> Worklist;
4132 PushLoopPHIs(L, Worklist);
4134 SmallPtrSet<Instruction *, 8> Visited;
4135 while (!Worklist.empty()) {
4136 Instruction *I = Worklist.pop_back_val();
4137 if (!Visited.insert(I)) continue;
4139 ValueExprMapType::iterator It =
4140 ValueExprMap.find_as(static_cast<Value *>(I));
4141 if (It != ValueExprMap.end()) {
4142 const SCEV *Old = It->second;
4144 // SCEVUnknown for a PHI either means that it has an unrecognized
4145 // structure, or it's a PHI that's in the progress of being computed
4146 // by createNodeForPHI. In the former case, additional loop trip
4147 // count information isn't going to change anything. In the later
4148 // case, createNodeForPHI will perform the necessary updates on its
4149 // own when it gets to that point.
4150 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
4151 forgetMemoizedResults(Old);
4152 ValueExprMap.erase(It);
4154 if (PHINode *PN = dyn_cast<PHINode>(I))
4155 ConstantEvolutionLoopExitValue.erase(PN);
4158 PushDefUseChildren(I, Worklist);
4162 // Re-lookup the insert position, since the call to
4163 // ComputeBackedgeTakenCount above could result in a
4164 // recusive call to getBackedgeTakenInfo (on a different
4165 // loop), which would invalidate the iterator computed
4167 return BackedgeTakenCounts.find(L)->second = Result;
4170 /// forgetLoop - This method should be called by the client when it has
4171 /// changed a loop in a way that may effect ScalarEvolution's ability to
4172 /// compute a trip count, or if the loop is deleted.
4173 void ScalarEvolution::forgetLoop(const Loop *L) {
4174 // Drop any stored trip count value.
4175 DenseMap<const Loop*, BackedgeTakenInfo>::iterator BTCPos =
4176 BackedgeTakenCounts.find(L);
4177 if (BTCPos != BackedgeTakenCounts.end()) {
4178 BTCPos->second.clear();
4179 BackedgeTakenCounts.erase(BTCPos);
4182 // Drop information about expressions based on loop-header PHIs.
4183 SmallVector<Instruction *, 16> Worklist;
4184 PushLoopPHIs(L, Worklist);
4186 SmallPtrSet<Instruction *, 8> Visited;
4187 while (!Worklist.empty()) {
4188 Instruction *I = Worklist.pop_back_val();
4189 if (!Visited.insert(I)) continue;
4191 ValueExprMapType::iterator It =
4192 ValueExprMap.find_as(static_cast<Value *>(I));
4193 if (It != ValueExprMap.end()) {
4194 forgetMemoizedResults(It->second);
4195 ValueExprMap.erase(It);
4196 if (PHINode *PN = dyn_cast<PHINode>(I))
4197 ConstantEvolutionLoopExitValue.erase(PN);
4200 PushDefUseChildren(I, Worklist);
4203 // Forget all contained loops too, to avoid dangling entries in the
4204 // ValuesAtScopes map.
4205 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
4209 /// forgetValue - This method should be called by the client when it has
4210 /// changed a value in a way that may effect its value, or which may
4211 /// disconnect it from a def-use chain linking it to a loop.
4212 void ScalarEvolution::forgetValue(Value *V) {
4213 Instruction *I = dyn_cast<Instruction>(V);
4216 // Drop information about expressions based on loop-header PHIs.
4217 SmallVector<Instruction *, 16> Worklist;
4218 Worklist.push_back(I);
4220 SmallPtrSet<Instruction *, 8> Visited;
4221 while (!Worklist.empty()) {
4222 I = Worklist.pop_back_val();
4223 if (!Visited.insert(I)) continue;
4225 ValueExprMapType::iterator It =
4226 ValueExprMap.find_as(static_cast<Value *>(I));
4227 if (It != ValueExprMap.end()) {
4228 forgetMemoizedResults(It->second);
4229 ValueExprMap.erase(It);
4230 if (PHINode *PN = dyn_cast<PHINode>(I))
4231 ConstantEvolutionLoopExitValue.erase(PN);
4234 PushDefUseChildren(I, Worklist);
4238 /// getExact - Get the exact loop backedge taken count considering all loop
4239 /// exits. A computable result can only be return for loops with a single exit.
4240 /// Returning the minimum taken count among all exits is incorrect because one
4241 /// of the loop's exit limit's may have been skipped. HowFarToZero assumes that
4242 /// the limit of each loop test is never skipped. This is a valid assumption as
4243 /// long as the loop exits via that test. For precise results, it is the
4244 /// caller's responsibility to specify the relevant loop exit using
4245 /// getExact(ExitingBlock, SE).
4247 ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
4248 // If any exits were not computable, the loop is not computable.
4249 if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute();
4251 // We need exactly one computable exit.
4252 if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
4253 assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
4255 const SCEV *BECount = 0;
4256 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4257 ENT != 0; ENT = ENT->getNextExit()) {
4259 assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV");
4262 BECount = ENT->ExactNotTaken;
4263 else if (BECount != ENT->ExactNotTaken)
4264 return SE->getCouldNotCompute();
4266 assert(BECount && "Invalid not taken count for loop exit");
4270 /// getExact - Get the exact not taken count for this loop exit.
4272 ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock,
4273 ScalarEvolution *SE) const {
4274 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4275 ENT != 0; ENT = ENT->getNextExit()) {
4277 if (ENT->ExitingBlock == ExitingBlock)
4278 return ENT->ExactNotTaken;
4280 return SE->getCouldNotCompute();
4283 /// getMax - Get the max backedge taken count for the loop.
4285 ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const {
4286 return Max ? Max : SE->getCouldNotCompute();
4289 bool ScalarEvolution::BackedgeTakenInfo::hasOperand(const SCEV *S,
4290 ScalarEvolution *SE) const {
4291 if (Max && Max != SE->getCouldNotCompute() && SE->hasOperand(Max, S))
4294 if (!ExitNotTaken.ExitingBlock)
4297 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4298 ENT != 0; ENT = ENT->getNextExit()) {
4300 if (ENT->ExactNotTaken != SE->getCouldNotCompute()
4301 && SE->hasOperand(ENT->ExactNotTaken, S)) {
4308 /// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each
4309 /// computable exit into a persistent ExitNotTakenInfo array.
4310 ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo(
4311 SmallVectorImpl< std::pair<BasicBlock *, const SCEV *> > &ExitCounts,
4312 bool Complete, const SCEV *MaxCount) : Max(MaxCount) {
4315 ExitNotTaken.setIncomplete();
4317 unsigned NumExits = ExitCounts.size();
4318 if (NumExits == 0) return;
4320 ExitNotTaken.ExitingBlock = ExitCounts[0].first;
4321 ExitNotTaken.ExactNotTaken = ExitCounts[0].second;
4322 if (NumExits == 1) return;
4324 // Handle the rare case of multiple computable exits.
4325 ExitNotTakenInfo *ENT = new ExitNotTakenInfo[NumExits-1];
4327 ExitNotTakenInfo *PrevENT = &ExitNotTaken;
4328 for (unsigned i = 1; i < NumExits; ++i, PrevENT = ENT, ++ENT) {
4329 PrevENT->setNextExit(ENT);
4330 ENT->ExitingBlock = ExitCounts[i].first;
4331 ENT->ExactNotTaken = ExitCounts[i].second;
4335 /// clear - Invalidate this result and free the ExitNotTakenInfo array.
4336 void ScalarEvolution::BackedgeTakenInfo::clear() {
4337 ExitNotTaken.ExitingBlock = 0;
4338 ExitNotTaken.ExactNotTaken = 0;
4339 delete[] ExitNotTaken.getNextExit();
4342 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
4343 /// of the specified loop will execute.
4344 ScalarEvolution::BackedgeTakenInfo
4345 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
4346 SmallVector<BasicBlock *, 8> ExitingBlocks;
4347 L->getExitingBlocks(ExitingBlocks);
4349 // Examine all exits and pick the most conservative values.
4350 const SCEV *MaxBECount = getCouldNotCompute();
4351 bool CouldComputeBECount = true;
4352 SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts;
4353 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
4354 ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]);
4355 if (EL.Exact == getCouldNotCompute())
4356 // We couldn't compute an exact value for this exit, so
4357 // we won't be able to compute an exact value for the loop.
4358 CouldComputeBECount = false;
4360 ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact));
4362 if (MaxBECount == getCouldNotCompute())
4363 MaxBECount = EL.Max;
4364 else if (EL.Max != getCouldNotCompute()) {
4365 // We cannot take the "min" MaxBECount, because non-unit stride loops may
4366 // skip some loop tests. Taking the max over the exits is sufficiently
4367 // conservative. TODO: We could do better taking into consideration
4368 // that (1) the loop has unit stride (2) the last loop test is
4369 // less-than/greater-than (3) any loop test is less-than/greater-than AND
4370 // falls-through some constant times less then the other tests.
4371 MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
4375 return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
4378 /// ComputeExitLimit - Compute the number of times the backedge of the specified
4379 /// loop will execute if it exits via the specified block.
4380 ScalarEvolution::ExitLimit
4381 ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) {
4383 // Okay, we've chosen an exiting block. See what condition causes us to
4384 // exit at this block.
4386 // FIXME: we should be able to handle switch instructions (with a single exit)
4387 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
4388 if (ExitBr == 0) return getCouldNotCompute();
4389 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
4391 // At this point, we know we have a conditional branch that determines whether
4392 // the loop is exited. However, we don't know if the branch is executed each
4393 // time through the loop. If not, then the execution count of the branch will
4394 // not be equal to the trip count of the loop.
4396 // Currently we check for this by checking to see if the Exit branch goes to
4397 // the loop header. If so, we know it will always execute the same number of
4398 // times as the loop. We also handle the case where the exit block *is* the
4399 // loop header. This is common for un-rotated loops.
4401 // If both of those tests fail, walk up the unique predecessor chain to the
4402 // header, stopping if there is an edge that doesn't exit the loop. If the
4403 // header is reached, the execution count of the branch will be equal to the
4404 // trip count of the loop.
4406 // More extensive analysis could be done to handle more cases here.
4408 if (ExitBr->getSuccessor(0) != L->getHeader() &&
4409 ExitBr->getSuccessor(1) != L->getHeader() &&
4410 ExitBr->getParent() != L->getHeader()) {
4411 // The simple checks failed, try climbing the unique predecessor chain
4412 // up to the header.
4414 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
4415 BasicBlock *Pred = BB->getUniquePredecessor();
4417 return getCouldNotCompute();
4418 TerminatorInst *PredTerm = Pred->getTerminator();
4419 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
4420 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
4423 // If the predecessor has a successor that isn't BB and isn't
4424 // outside the loop, assume the worst.
4425 if (L->contains(PredSucc))
4426 return getCouldNotCompute();
4428 if (Pred == L->getHeader()) {
4435 return getCouldNotCompute();
4438 // Proceed to the next level to examine the exit condition expression.
4439 return ComputeExitLimitFromCond(L, ExitBr->getCondition(),
4440 ExitBr->getSuccessor(0),
4441 ExitBr->getSuccessor(1),
4442 /*IsSubExpr=*/false);
4445 /// ComputeExitLimitFromCond - Compute the number of times the
4446 /// backedge of the specified loop will execute if its exit condition
4447 /// were a conditional branch of ExitCond, TBB, and FBB.
4449 /// @param IsSubExpr is true if ExitCond does not directly control the exit
4450 /// branch. In this case, we cannot assume that the loop only exits when the
4451 /// condition is true and cannot infer that failing to meet the condition prior
4452 /// to integer wraparound results in undefined behavior.
4453 ScalarEvolution::ExitLimit
4454 ScalarEvolution::ComputeExitLimitFromCond(const Loop *L,
4459 // Check if the controlling expression for this loop is an And or Or.
4460 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
4461 if (BO->getOpcode() == Instruction::And) {
4462 // Recurse on the operands of the and.
4463 bool EitherMayExit = L->contains(TBB);
4464 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB,
4465 IsSubExpr || EitherMayExit);
4466 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB,
4467 IsSubExpr || EitherMayExit);
4468 const SCEV *BECount = getCouldNotCompute();
4469 const SCEV *MaxBECount = getCouldNotCompute();
4470 if (EitherMayExit) {
4471 // Both conditions must be true for the loop to continue executing.
4472 // Choose the less conservative count.
4473 if (EL0.Exact == getCouldNotCompute() ||
4474 EL1.Exact == getCouldNotCompute())
4475 BECount = getCouldNotCompute();
4477 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4478 if (EL0.Max == getCouldNotCompute())
4479 MaxBECount = EL1.Max;
4480 else if (EL1.Max == getCouldNotCompute())
4481 MaxBECount = EL0.Max;
4483 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4485 // Both conditions must be true at the same time for the loop to exit.
4486 // For now, be conservative.
4487 assert(L->contains(FBB) && "Loop block has no successor in loop!");
4488 if (EL0.Max == EL1.Max)
4489 MaxBECount = EL0.Max;
4490 if (EL0.Exact == EL1.Exact)
4491 BECount = EL0.Exact;
4494 return ExitLimit(BECount, MaxBECount);
4496 if (BO->getOpcode() == Instruction::Or) {
4497 // Recurse on the operands of the or.
4498 bool EitherMayExit = L->contains(FBB);
4499 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB,
4500 IsSubExpr || EitherMayExit);
4501 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB,
4502 IsSubExpr || EitherMayExit);
4503 const SCEV *BECount = getCouldNotCompute();
4504 const SCEV *MaxBECount = getCouldNotCompute();
4505 if (EitherMayExit) {
4506 // Both conditions must be false for the loop to continue executing.
4507 // Choose the less conservative count.
4508 if (EL0.Exact == getCouldNotCompute() ||
4509 EL1.Exact == getCouldNotCompute())
4510 BECount = getCouldNotCompute();
4512 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4513 if (EL0.Max == getCouldNotCompute())
4514 MaxBECount = EL1.Max;
4515 else if (EL1.Max == getCouldNotCompute())
4516 MaxBECount = EL0.Max;
4518 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4520 // Both conditions must be false at the same time for the loop to exit.
4521 // For now, be conservative.
4522 assert(L->contains(TBB) && "Loop block has no successor in loop!");
4523 if (EL0.Max == EL1.Max)
4524 MaxBECount = EL0.Max;
4525 if (EL0.Exact == EL1.Exact)
4526 BECount = EL0.Exact;
4529 return ExitLimit(BECount, MaxBECount);
4533 // With an icmp, it may be feasible to compute an exact backedge-taken count.
4534 // Proceed to the next level to examine the icmp.
4535 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
4536 return ComputeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB, IsSubExpr);
4538 // Check for a constant condition. These are normally stripped out by
4539 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
4540 // preserve the CFG and is temporarily leaving constant conditions
4542 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
4543 if (L->contains(FBB) == !CI->getZExtValue())
4544 // The backedge is always taken.
4545 return getCouldNotCompute();
4547 // The backedge is never taken.
4548 return getConstant(CI->getType(), 0);
4551 // If it's not an integer or pointer comparison then compute it the hard way.
4552 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4555 /// ComputeExitLimitFromICmp - Compute the number of times the
4556 /// backedge of the specified loop will execute if its exit condition
4557 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4558 ScalarEvolution::ExitLimit
4559 ScalarEvolution::ComputeExitLimitFromICmp(const Loop *L,
4565 // If the condition was exit on true, convert the condition to exit on false
4566 ICmpInst::Predicate Cond;
4567 if (!L->contains(FBB))
4568 Cond = ExitCond->getPredicate();
4570 Cond = ExitCond->getInversePredicate();
4572 // Handle common loops like: for (X = "string"; *X; ++X)
4573 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4574 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4576 ComputeLoadConstantCompareExitLimit(LI, RHS, L, Cond);
4577 if (ItCnt.hasAnyInfo())
4581 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4582 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4584 // Try to evaluate any dependencies out of the loop.
4585 LHS = getSCEVAtScope(LHS, L);
4586 RHS = getSCEVAtScope(RHS, L);
4588 // At this point, we would like to compute how many iterations of the
4589 // loop the predicate will return true for these inputs.
4590 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4591 // If there is a loop-invariant, force it into the RHS.
4592 std::swap(LHS, RHS);
4593 Cond = ICmpInst::getSwappedPredicate(Cond);
4596 // Simplify the operands before analyzing them.
4597 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4599 // If we have a comparison of a chrec against a constant, try to use value
4600 // ranges to answer this query.
4601 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4602 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4603 if (AddRec->getLoop() == L) {
4604 // Form the constant range.
4605 ConstantRange CompRange(
4606 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4608 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4609 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4613 case ICmpInst::ICMP_NE: { // while (X != Y)
4614 // Convert to: while (X-Y != 0)
4615 ExitLimit EL = HowFarToZero(getMinusSCEV(LHS, RHS), L, IsSubExpr);
4616 if (EL.hasAnyInfo()) return EL;
4619 case ICmpInst::ICMP_EQ: { // while (X == Y)
4620 // Convert to: while (X-Y == 0)
4621 ExitLimit EL = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4622 if (EL.hasAnyInfo()) return EL;
4625 case ICmpInst::ICMP_SLT: {
4626 ExitLimit EL = HowManyLessThans(LHS, RHS, L, true, IsSubExpr);
4627 if (EL.hasAnyInfo()) return EL;
4630 case ICmpInst::ICMP_SGT: {
4631 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4632 getNotSCEV(RHS), L, true, IsSubExpr);
4633 if (EL.hasAnyInfo()) return EL;
4636 case ICmpInst::ICMP_ULT: {
4637 ExitLimit EL = HowManyLessThans(LHS, RHS, L, false, IsSubExpr);
4638 if (EL.hasAnyInfo()) return EL;
4641 case ICmpInst::ICMP_UGT: {
4642 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4643 getNotSCEV(RHS), L, false, IsSubExpr);
4644 if (EL.hasAnyInfo()) return EL;
4649 dbgs() << "ComputeBackedgeTakenCount ";
4650 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4651 dbgs() << "[unsigned] ";
4652 dbgs() << *LHS << " "
4653 << Instruction::getOpcodeName(Instruction::ICmp)
4654 << " " << *RHS << "\n";
4658 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4661 static ConstantInt *
4662 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4663 ScalarEvolution &SE) {
4664 const SCEV *InVal = SE.getConstant(C);
4665 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4666 assert(isa<SCEVConstant>(Val) &&
4667 "Evaluation of SCEV at constant didn't fold correctly?");
4668 return cast<SCEVConstant>(Val)->getValue();
4671 /// ComputeLoadConstantCompareExitLimit - Given an exit condition of
4672 /// 'icmp op load X, cst', try to see if we can compute the backedge
4673 /// execution count.
4674 ScalarEvolution::ExitLimit
4675 ScalarEvolution::ComputeLoadConstantCompareExitLimit(
4679 ICmpInst::Predicate predicate) {
4681 if (LI->isVolatile()) return getCouldNotCompute();
4683 // Check to see if the loaded pointer is a getelementptr of a global.
4684 // TODO: Use SCEV instead of manually grubbing with GEPs.
4685 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4686 if (!GEP) return getCouldNotCompute();
4688 // Make sure that it is really a constant global we are gepping, with an
4689 // initializer, and make sure the first IDX is really 0.
4690 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4691 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4692 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4693 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4694 return getCouldNotCompute();
4696 // Okay, we allow one non-constant index into the GEP instruction.
4698 std::vector<Constant*> Indexes;
4699 unsigned VarIdxNum = 0;
4700 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4701 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4702 Indexes.push_back(CI);
4703 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4704 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4705 VarIdx = GEP->getOperand(i);
4707 Indexes.push_back(0);
4710 // Loop-invariant loads may be a byproduct of loop optimization. Skip them.
4712 return getCouldNotCompute();
4714 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4715 // Check to see if X is a loop variant variable value now.
4716 const SCEV *Idx = getSCEV(VarIdx);
4717 Idx = getSCEVAtScope(Idx, L);
4719 // We can only recognize very limited forms of loop index expressions, in
4720 // particular, only affine AddRec's like {C1,+,C2}.
4721 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4722 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4723 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4724 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4725 return getCouldNotCompute();
4727 unsigned MaxSteps = MaxBruteForceIterations;
4728 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4729 ConstantInt *ItCst = ConstantInt::get(
4730 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4731 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4733 // Form the GEP offset.
4734 Indexes[VarIdxNum] = Val;
4736 Constant *Result = ConstantFoldLoadThroughGEPIndices(GV->getInitializer(),
4738 if (Result == 0) break; // Cannot compute!
4740 // Evaluate the condition for this iteration.
4741 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4742 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4743 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4745 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4746 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4749 ++NumArrayLenItCounts;
4750 return getConstant(ItCst); // Found terminating iteration!
4753 return getCouldNotCompute();
4757 /// CanConstantFold - Return true if we can constant fold an instruction of the
4758 /// specified type, assuming that all operands were constants.
4759 static bool CanConstantFold(const Instruction *I) {
4760 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4761 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) ||
4765 if (const CallInst *CI = dyn_cast<CallInst>(I))
4766 if (const Function *F = CI->getCalledFunction())
4767 return canConstantFoldCallTo(F);
4771 /// Determine whether this instruction can constant evolve within this loop
4772 /// assuming its operands can all constant evolve.
4773 static bool canConstantEvolve(Instruction *I, const Loop *L) {
4774 // An instruction outside of the loop can't be derived from a loop PHI.
4775 if (!L->contains(I)) return false;
4777 if (isa<PHINode>(I)) {
4778 if (L->getHeader() == I->getParent())
4781 // We don't currently keep track of the control flow needed to evaluate
4782 // PHIs, so we cannot handle PHIs inside of loops.
4786 // If we won't be able to constant fold this expression even if the operands
4787 // are constants, bail early.
4788 return CanConstantFold(I);
4791 /// getConstantEvolvingPHIOperands - Implement getConstantEvolvingPHI by
4792 /// recursing through each instruction operand until reaching a loop header phi.
4794 getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L,
4795 DenseMap<Instruction *, PHINode *> &PHIMap) {
4797 // Otherwise, we can evaluate this instruction if all of its operands are
4798 // constant or derived from a PHI node themselves.
4800 for (Instruction::op_iterator OpI = UseInst->op_begin(),
4801 OpE = UseInst->op_end(); OpI != OpE; ++OpI) {
4803 if (isa<Constant>(*OpI)) continue;
4805 Instruction *OpInst = dyn_cast<Instruction>(*OpI);
4806 if (!OpInst || !canConstantEvolve(OpInst, L)) return 0;
4808 PHINode *P = dyn_cast<PHINode>(OpInst);
4810 // If this operand is already visited, reuse the prior result.
4811 // We may have P != PHI if this is the deepest point at which the
4812 // inconsistent paths meet.
4813 P = PHIMap.lookup(OpInst);
4815 // Recurse and memoize the results, whether a phi is found or not.
4816 // This recursive call invalidates pointers into PHIMap.
4817 P = getConstantEvolvingPHIOperands(OpInst, L, PHIMap);
4820 if (P == 0) return 0; // Not evolving from PHI
4821 if (PHI && PHI != P) return 0; // Evolving from multiple different PHIs.
4824 // This is a expression evolving from a constant PHI!
4828 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4829 /// in the loop that V is derived from. We allow arbitrary operations along the
4830 /// way, but the operands of an operation must either be constants or a value
4831 /// derived from a constant PHI. If this expression does not fit with these
4832 /// constraints, return null.
4833 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4834 Instruction *I = dyn_cast<Instruction>(V);
4835 if (I == 0 || !canConstantEvolve(I, L)) return 0;
4837 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4841 // Record non-constant instructions contained by the loop.
4842 DenseMap<Instruction *, PHINode *> PHIMap;
4843 return getConstantEvolvingPHIOperands(I, L, PHIMap);
4846 /// EvaluateExpression - Given an expression that passes the
4847 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4848 /// in the loop has the value PHIVal. If we can't fold this expression for some
4849 /// reason, return null.
4850 static Constant *EvaluateExpression(Value *V, const Loop *L,
4851 DenseMap<Instruction *, Constant *> &Vals,
4852 const DataLayout *TD,
4853 const TargetLibraryInfo *TLI) {
4854 // Convenient constant check, but redundant for recursive calls.
4855 if (Constant *C = dyn_cast<Constant>(V)) return C;
4856 Instruction *I = dyn_cast<Instruction>(V);
4859 if (Constant *C = Vals.lookup(I)) return C;
4861 // An instruction inside the loop depends on a value outside the loop that we
4862 // weren't given a mapping for, or a value such as a call inside the loop.
4863 if (!canConstantEvolve(I, L)) return 0;
4865 // An unmapped PHI can be due to a branch or another loop inside this loop,
4866 // or due to this not being the initial iteration through a loop where we
4867 // couldn't compute the evolution of this particular PHI last time.
4868 if (isa<PHINode>(I)) return 0;
4870 std::vector<Constant*> Operands(I->getNumOperands());
4872 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4873 Instruction *Operand = dyn_cast<Instruction>(I->getOperand(i));
4875 Operands[i] = dyn_cast<Constant>(I->getOperand(i));
4876 if (!Operands[i]) return 0;
4879 Constant *C = EvaluateExpression(Operand, L, Vals, TD, TLI);
4885 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4886 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4887 Operands[1], TD, TLI);
4888 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4889 if (!LI->isVolatile())
4890 return ConstantFoldLoadFromConstPtr(Operands[0], TD);
4892 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Operands, TD,
4896 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4897 /// in the header of its containing loop, we know the loop executes a
4898 /// constant number of times, and the PHI node is just a recurrence
4899 /// involving constants, fold it.
4901 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4904 DenseMap<PHINode*, Constant*>::const_iterator I =
4905 ConstantEvolutionLoopExitValue.find(PN);
4906 if (I != ConstantEvolutionLoopExitValue.end())
4909 if (BEs.ugt(MaxBruteForceIterations))
4910 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4912 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4914 DenseMap<Instruction *, Constant *> CurrentIterVals;
4915 BasicBlock *Header = L->getHeader();
4916 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
4918 // Since the loop is canonicalized, the PHI node must have two entries. One
4919 // entry must be a constant (coming in from outside of the loop), and the
4920 // second must be derived from the same PHI.
4921 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4923 for (BasicBlock::iterator I = Header->begin();
4924 (PHI = dyn_cast<PHINode>(I)); ++I) {
4925 Constant *StartCST =
4926 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
4927 if (StartCST == 0) continue;
4928 CurrentIterVals[PHI] = StartCST;
4930 if (!CurrentIterVals.count(PN))
4933 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4935 // Execute the loop symbolically to determine the exit value.
4936 if (BEs.getActiveBits() >= 32)
4937 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4939 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4940 unsigned IterationNum = 0;
4941 for (; ; ++IterationNum) {
4942 if (IterationNum == NumIterations)
4943 return RetVal = CurrentIterVals[PN]; // Got exit value!
4945 // Compute the value of the PHIs for the next iteration.
4946 // EvaluateExpression adds non-phi values to the CurrentIterVals map.
4947 DenseMap<Instruction *, Constant *> NextIterVals;
4948 Constant *NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD,
4951 return 0; // Couldn't evaluate!
4952 NextIterVals[PN] = NextPHI;
4954 bool StoppedEvolving = NextPHI == CurrentIterVals[PN];
4956 // Also evaluate the other PHI nodes. However, we don't get to stop if we
4957 // cease to be able to evaluate one of them or if they stop evolving,
4958 // because that doesn't necessarily prevent us from computing PN.
4959 SmallVector<std::pair<PHINode *, Constant *>, 8> PHIsToCompute;
4960 for (DenseMap<Instruction *, Constant *>::const_iterator
4961 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
4962 PHINode *PHI = dyn_cast<PHINode>(I->first);
4963 if (!PHI || PHI == PN || PHI->getParent() != Header) continue;
4964 PHIsToCompute.push_back(std::make_pair(PHI, I->second));
4966 // We use two distinct loops because EvaluateExpression may invalidate any
4967 // iterators into CurrentIterVals.
4968 for (SmallVectorImpl<std::pair<PHINode *, Constant*> >::const_iterator
4969 I = PHIsToCompute.begin(), E = PHIsToCompute.end(); I != E; ++I) {
4970 PHINode *PHI = I->first;
4971 Constant *&NextPHI = NextIterVals[PHI];
4972 if (!NextPHI) { // Not already computed.
4973 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
4974 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
4976 if (NextPHI != I->second)
4977 StoppedEvolving = false;
4980 // If all entries in CurrentIterVals == NextIterVals then we can stop
4981 // iterating, the loop can't continue to change.
4982 if (StoppedEvolving)
4983 return RetVal = CurrentIterVals[PN];
4985 CurrentIterVals.swap(NextIterVals);
4989 /// ComputeExitCountExhaustively - If the loop is known to execute a
4990 /// constant number of times (the condition evolves only from constants),
4991 /// try to evaluate a few iterations of the loop until we get the exit
4992 /// condition gets a value of ExitWhen (true or false). If we cannot
4993 /// evaluate the trip count of the loop, return getCouldNotCompute().
4994 const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L,
4997 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4998 if (PN == 0) return getCouldNotCompute();
5000 // If the loop is canonicalized, the PHI will have exactly two entries.
5001 // That's the only form we support here.
5002 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
5004 DenseMap<Instruction *, Constant *> CurrentIterVals;
5005 BasicBlock *Header = L->getHeader();
5006 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
5008 // One entry must be a constant (coming in from outside of the loop), and the
5009 // second must be derived from the same PHI.
5010 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
5012 for (BasicBlock::iterator I = Header->begin();
5013 (PHI = dyn_cast<PHINode>(I)); ++I) {
5014 Constant *StartCST =
5015 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
5016 if (StartCST == 0) continue;
5017 CurrentIterVals[PHI] = StartCST;
5019 if (!CurrentIterVals.count(PN))
5020 return getCouldNotCompute();
5022 // Okay, we find a PHI node that defines the trip count of this loop. Execute
5023 // the loop symbolically to determine when the condition gets a value of
5026 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
5027 for (unsigned IterationNum = 0; IterationNum != MaxIterations;++IterationNum){
5028 ConstantInt *CondVal =
5029 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, L, CurrentIterVals,
5032 // Couldn't symbolically evaluate.
5033 if (!CondVal) return getCouldNotCompute();
5035 if (CondVal->getValue() == uint64_t(ExitWhen)) {
5036 ++NumBruteForceTripCountsComputed;
5037 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
5040 // Update all the PHI nodes for the next iteration.
5041 DenseMap<Instruction *, Constant *> NextIterVals;
5043 // Create a list of which PHIs we need to compute. We want to do this before
5044 // calling EvaluateExpression on them because that may invalidate iterators
5045 // into CurrentIterVals.
5046 SmallVector<PHINode *, 8> PHIsToCompute;
5047 for (DenseMap<Instruction *, Constant *>::const_iterator
5048 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
5049 PHINode *PHI = dyn_cast<PHINode>(I->first);
5050 if (!PHI || PHI->getParent() != Header) continue;
5051 PHIsToCompute.push_back(PHI);
5053 for (SmallVectorImpl<PHINode *>::const_iterator I = PHIsToCompute.begin(),
5054 E = PHIsToCompute.end(); I != E; ++I) {
5056 Constant *&NextPHI = NextIterVals[PHI];
5057 if (NextPHI) continue; // Already computed!
5059 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
5060 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
5062 CurrentIterVals.swap(NextIterVals);
5065 // Too many iterations were needed to evaluate.
5066 return getCouldNotCompute();
5069 /// getSCEVAtScope - Return a SCEV expression for the specified value
5070 /// at the specified scope in the program. The L value specifies a loop
5071 /// nest to evaluate the expression at, where null is the top-level or a
5072 /// specified loop is immediately inside of the loop.
5074 /// This method can be used to compute the exit value for a variable defined
5075 /// in a loop by querying what the value will hold in the parent loop.
5077 /// In the case that a relevant loop exit value cannot be computed, the
5078 /// original value V is returned.
5079 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
5080 // Check to see if we've folded this expression at this loop before.
5081 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
5082 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
5083 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
5085 return Pair.first->second ? Pair.first->second : V;
5087 // Otherwise compute it.
5088 const SCEV *C = computeSCEVAtScope(V, L);
5089 ValuesAtScopes[V][L] = C;
5093 /// This builds up a Constant using the ConstantExpr interface. That way, we
5094 /// will return Constants for objects which aren't represented by a
5095 /// SCEVConstant, because SCEVConstant is restricted to ConstantInt.
5096 /// Returns NULL if the SCEV isn't representable as a Constant.
5097 static Constant *BuildConstantFromSCEV(const SCEV *V) {
5098 switch (V->getSCEVType()) {
5099 default: // TODO: smax, umax.
5100 case scCouldNotCompute:
5104 return cast<SCEVConstant>(V)->getValue();
5106 return dyn_cast<Constant>(cast<SCEVUnknown>(V)->getValue());
5107 case scSignExtend: {
5108 const SCEVSignExtendExpr *SS = cast<SCEVSignExtendExpr>(V);
5109 if (Constant *CastOp = BuildConstantFromSCEV(SS->getOperand()))
5110 return ConstantExpr::getSExt(CastOp, SS->getType());
5113 case scZeroExtend: {
5114 const SCEVZeroExtendExpr *SZ = cast<SCEVZeroExtendExpr>(V);
5115 if (Constant *CastOp = BuildConstantFromSCEV(SZ->getOperand()))
5116 return ConstantExpr::getZExt(CastOp, SZ->getType());
5120 const SCEVTruncateExpr *ST = cast<SCEVTruncateExpr>(V);
5121 if (Constant *CastOp = BuildConstantFromSCEV(ST->getOperand()))
5122 return ConstantExpr::getTrunc(CastOp, ST->getType());
5126 const SCEVAddExpr *SA = cast<SCEVAddExpr>(V);
5127 if (Constant *C = BuildConstantFromSCEV(SA->getOperand(0))) {
5128 if (C->getType()->isPointerTy())
5129 C = ConstantExpr::getBitCast(C, Type::getInt8PtrTy(C->getContext()));
5130 for (unsigned i = 1, e = SA->getNumOperands(); i != e; ++i) {
5131 Constant *C2 = BuildConstantFromSCEV(SA->getOperand(i));
5135 if (!C->getType()->isPointerTy() && C2->getType()->isPointerTy()) {
5137 // The offsets have been converted to bytes. We can add bytes to an
5138 // i8* by GEP with the byte count in the first index.
5139 C = ConstantExpr::getBitCast(C,Type::getInt8PtrTy(C->getContext()));
5142 // Don't bother trying to sum two pointers. We probably can't
5143 // statically compute a load that results from it anyway.
5144 if (C2->getType()->isPointerTy())
5147 if (C->getType()->isPointerTy()) {
5148 if (cast<PointerType>(C->getType())->getElementType()->isStructTy())
5149 C2 = ConstantExpr::getIntegerCast(
5150 C2, Type::getInt32Ty(C->getContext()), true);
5151 C = ConstantExpr::getGetElementPtr(C, C2);
5153 C = ConstantExpr::getAdd(C, C2);
5160 const SCEVMulExpr *SM = cast<SCEVMulExpr>(V);
5161 if (Constant *C = BuildConstantFromSCEV(SM->getOperand(0))) {
5162 // Don't bother with pointers at all.
5163 if (C->getType()->isPointerTy()) return 0;
5164 for (unsigned i = 1, e = SM->getNumOperands(); i != e; ++i) {
5165 Constant *C2 = BuildConstantFromSCEV(SM->getOperand(i));
5166 if (!C2 || C2->getType()->isPointerTy()) return 0;
5167 C = ConstantExpr::getMul(C, C2);
5174 const SCEVUDivExpr *SU = cast<SCEVUDivExpr>(V);
5175 if (Constant *LHS = BuildConstantFromSCEV(SU->getLHS()))
5176 if (Constant *RHS = BuildConstantFromSCEV(SU->getRHS()))
5177 if (LHS->getType() == RHS->getType())
5178 return ConstantExpr::getUDiv(LHS, RHS);
5185 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
5186 if (isa<SCEVConstant>(V)) return V;
5188 // If this instruction is evolved from a constant-evolving PHI, compute the
5189 // exit value from the loop without using SCEVs.
5190 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
5191 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
5192 const Loop *LI = (*this->LI)[I->getParent()];
5193 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
5194 if (PHINode *PN = dyn_cast<PHINode>(I))
5195 if (PN->getParent() == LI->getHeader()) {
5196 // Okay, there is no closed form solution for the PHI node. Check
5197 // to see if the loop that contains it has a known backedge-taken
5198 // count. If so, we may be able to force computation of the exit
5200 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
5201 if (const SCEVConstant *BTCC =
5202 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
5203 // Okay, we know how many times the containing loop executes. If
5204 // this is a constant evolving PHI node, get the final value at
5205 // the specified iteration number.
5206 Constant *RV = getConstantEvolutionLoopExitValue(PN,
5207 BTCC->getValue()->getValue(),
5209 if (RV) return getSCEV(RV);
5213 // Okay, this is an expression that we cannot symbolically evaluate
5214 // into a SCEV. Check to see if it's possible to symbolically evaluate
5215 // the arguments into constants, and if so, try to constant propagate the
5216 // result. This is particularly useful for computing loop exit values.
5217 if (CanConstantFold(I)) {
5218 SmallVector<Constant *, 4> Operands;
5219 bool MadeImprovement = false;
5220 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
5221 Value *Op = I->getOperand(i);
5222 if (Constant *C = dyn_cast<Constant>(Op)) {
5223 Operands.push_back(C);
5227 // If any of the operands is non-constant and if they are
5228 // non-integer and non-pointer, don't even try to analyze them
5229 // with scev techniques.
5230 if (!isSCEVable(Op->getType()))
5233 const SCEV *OrigV = getSCEV(Op);
5234 const SCEV *OpV = getSCEVAtScope(OrigV, L);
5235 MadeImprovement |= OrigV != OpV;
5237 Constant *C = BuildConstantFromSCEV(OpV);
5239 if (C->getType() != Op->getType())
5240 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5244 Operands.push_back(C);
5247 // Check to see if getSCEVAtScope actually made an improvement.
5248 if (MadeImprovement) {
5250 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
5251 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
5252 Operands[0], Operands[1], TD,
5254 else if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
5255 if (!LI->isVolatile())
5256 C = ConstantFoldLoadFromConstPtr(Operands[0], TD);
5258 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
5266 // This is some other type of SCEVUnknown, just return it.
5270 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
5271 // Avoid performing the look-up in the common case where the specified
5272 // expression has no loop-variant portions.
5273 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
5274 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5275 if (OpAtScope != Comm->getOperand(i)) {
5276 // Okay, at least one of these operands is loop variant but might be
5277 // foldable. Build a new instance of the folded commutative expression.
5278 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
5279 Comm->op_begin()+i);
5280 NewOps.push_back(OpAtScope);
5282 for (++i; i != e; ++i) {
5283 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5284 NewOps.push_back(OpAtScope);
5286 if (isa<SCEVAddExpr>(Comm))
5287 return getAddExpr(NewOps);
5288 if (isa<SCEVMulExpr>(Comm))
5289 return getMulExpr(NewOps);
5290 if (isa<SCEVSMaxExpr>(Comm))
5291 return getSMaxExpr(NewOps);
5292 if (isa<SCEVUMaxExpr>(Comm))
5293 return getUMaxExpr(NewOps);
5294 llvm_unreachable("Unknown commutative SCEV type!");
5297 // If we got here, all operands are loop invariant.
5301 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
5302 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
5303 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
5304 if (LHS == Div->getLHS() && RHS == Div->getRHS())
5305 return Div; // must be loop invariant
5306 return getUDivExpr(LHS, RHS);
5309 // If this is a loop recurrence for a loop that does not contain L, then we
5310 // are dealing with the final value computed by the loop.
5311 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
5312 // First, attempt to evaluate each operand.
5313 // Avoid performing the look-up in the common case where the specified
5314 // expression has no loop-variant portions.
5315 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
5316 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
5317 if (OpAtScope == AddRec->getOperand(i))
5320 // Okay, at least one of these operands is loop variant but might be
5321 // foldable. Build a new instance of the folded commutative expression.
5322 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
5323 AddRec->op_begin()+i);
5324 NewOps.push_back(OpAtScope);
5325 for (++i; i != e; ++i)
5326 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
5328 const SCEV *FoldedRec =
5329 getAddRecExpr(NewOps, AddRec->getLoop(),
5330 AddRec->getNoWrapFlags(SCEV::FlagNW));
5331 AddRec = dyn_cast<SCEVAddRecExpr>(FoldedRec);
5332 // The addrec may be folded to a nonrecurrence, for example, if the
5333 // induction variable is multiplied by zero after constant folding. Go
5334 // ahead and return the folded value.
5340 // If the scope is outside the addrec's loop, evaluate it by using the
5341 // loop exit value of the addrec.
5342 if (!AddRec->getLoop()->contains(L)) {
5343 // To evaluate this recurrence, we need to know how many times the AddRec
5344 // loop iterates. Compute this now.
5345 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
5346 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
5348 // Then, evaluate the AddRec.
5349 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
5355 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
5356 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5357 if (Op == Cast->getOperand())
5358 return Cast; // must be loop invariant
5359 return getZeroExtendExpr(Op, Cast->getType());
5362 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
5363 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5364 if (Op == Cast->getOperand())
5365 return Cast; // must be loop invariant
5366 return getSignExtendExpr(Op, Cast->getType());
5369 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
5370 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5371 if (Op == Cast->getOperand())
5372 return Cast; // must be loop invariant
5373 return getTruncateExpr(Op, Cast->getType());
5376 llvm_unreachable("Unknown SCEV type!");
5379 /// getSCEVAtScope - This is a convenience function which does
5380 /// getSCEVAtScope(getSCEV(V), L).
5381 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
5382 return getSCEVAtScope(getSCEV(V), L);
5385 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
5386 /// following equation:
5388 /// A * X = B (mod N)
5390 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
5391 /// A and B isn't important.
5393 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
5394 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
5395 ScalarEvolution &SE) {
5396 uint32_t BW = A.getBitWidth();
5397 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
5398 assert(A != 0 && "A must be non-zero.");
5402 // The gcd of A and N may have only one prime factor: 2. The number of
5403 // trailing zeros in A is its multiplicity
5404 uint32_t Mult2 = A.countTrailingZeros();
5407 // 2. Check if B is divisible by D.
5409 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
5410 // is not less than multiplicity of this prime factor for D.
5411 if (B.countTrailingZeros() < Mult2)
5412 return SE.getCouldNotCompute();
5414 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
5417 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
5418 // bit width during computations.
5419 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
5420 APInt Mod(BW + 1, 0);
5421 Mod.setBit(BW - Mult2); // Mod = N / D
5422 APInt I = AD.multiplicativeInverse(Mod);
5424 // 4. Compute the minimum unsigned root of the equation:
5425 // I * (B / D) mod (N / D)
5426 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
5428 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
5430 return SE.getConstant(Result.trunc(BW));
5433 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
5434 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
5435 /// might be the same) or two SCEVCouldNotCompute objects.
5437 static std::pair<const SCEV *,const SCEV *>
5438 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
5439 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
5440 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
5441 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
5442 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
5444 // We currently can only solve this if the coefficients are constants.
5445 if (!LC || !MC || !NC) {
5446 const SCEV *CNC = SE.getCouldNotCompute();
5447 return std::make_pair(CNC, CNC);
5450 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
5451 const APInt &L = LC->getValue()->getValue();
5452 const APInt &M = MC->getValue()->getValue();
5453 const APInt &N = NC->getValue()->getValue();
5454 APInt Two(BitWidth, 2);
5455 APInt Four(BitWidth, 4);
5458 using namespace APIntOps;
5460 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
5461 // The B coefficient is M-N/2
5465 // The A coefficient is N/2
5466 APInt A(N.sdiv(Two));
5468 // Compute the B^2-4ac term.
5471 SqrtTerm -= Four * (A * C);
5473 if (SqrtTerm.isNegative()) {
5474 // The loop is provably infinite.
5475 const SCEV *CNC = SE.getCouldNotCompute();
5476 return std::make_pair(CNC, CNC);
5479 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
5480 // integer value or else APInt::sqrt() will assert.
5481 APInt SqrtVal(SqrtTerm.sqrt());
5483 // Compute the two solutions for the quadratic formula.
5484 // The divisions must be performed as signed divisions.
5487 if (TwoA.isMinValue()) {
5488 const SCEV *CNC = SE.getCouldNotCompute();
5489 return std::make_pair(CNC, CNC);
5492 LLVMContext &Context = SE.getContext();
5494 ConstantInt *Solution1 =
5495 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
5496 ConstantInt *Solution2 =
5497 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
5499 return std::make_pair(SE.getConstant(Solution1),
5500 SE.getConstant(Solution2));
5501 } // end APIntOps namespace
5504 /// HowFarToZero - Return the number of times a backedge comparing the specified
5505 /// value to zero will execute. If not computable, return CouldNotCompute.
5507 /// This is only used for loops with a "x != y" exit test. The exit condition is
5508 /// now expressed as a single expression, V = x-y. So the exit test is
5509 /// effectively V != 0. We know and take advantage of the fact that this
5510 /// expression only being used in a comparison by zero context.
5511 ScalarEvolution::ExitLimit
5512 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L, bool IsSubExpr) {
5513 // If the value is a constant
5514 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5515 // If the value is already zero, the branch will execute zero times.
5516 if (C->getValue()->isZero()) return C;
5517 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5520 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
5521 if (!AddRec || AddRec->getLoop() != L)
5522 return getCouldNotCompute();
5524 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
5525 // the quadratic equation to solve it.
5526 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
5527 std::pair<const SCEV *,const SCEV *> Roots =
5528 SolveQuadraticEquation(AddRec, *this);
5529 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5530 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5533 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
5534 << " sol#2: " << *R2 << "\n";
5536 // Pick the smallest positive root value.
5537 if (ConstantInt *CB =
5538 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
5541 if (CB->getZExtValue() == false)
5542 std::swap(R1, R2); // R1 is the minimum root now.
5544 // We can only use this value if the chrec ends up with an exact zero
5545 // value at this index. When solving for "X*X != 5", for example, we
5546 // should not accept a root of 2.
5547 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
5549 return R1; // We found a quadratic root!
5552 return getCouldNotCompute();
5555 // Otherwise we can only handle this if it is affine.
5556 if (!AddRec->isAffine())
5557 return getCouldNotCompute();
5559 // If this is an affine expression, the execution count of this branch is
5560 // the minimum unsigned root of the following equation:
5562 // Start + Step*N = 0 (mod 2^BW)
5566 // Step*N = -Start (mod 2^BW)
5568 // where BW is the common bit width of Start and Step.
5570 // Get the initial value for the loop.
5571 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
5572 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
5574 // For now we handle only constant steps.
5576 // TODO: Handle a nonconstant Step given AddRec<NUW>. If the
5577 // AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap
5578 // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step.
5579 // We have not yet seen any such cases.
5580 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
5581 if (StepC == 0 || StepC->getValue()->equalsInt(0))
5582 return getCouldNotCompute();
5584 // For positive steps (counting up until unsigned overflow):
5585 // N = -Start/Step (as unsigned)
5586 // For negative steps (counting down to zero):
5588 // First compute the unsigned distance from zero in the direction of Step.
5589 bool CountDown = StepC->getValue()->getValue().isNegative();
5590 const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
5592 // Handle unitary steps, which cannot wraparound.
5593 // 1*N = -Start; -1*N = Start (mod 2^BW), so:
5594 // N = Distance (as unsigned)
5595 if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) {
5596 ConstantRange CR = getUnsignedRange(Start);
5597 const SCEV *MaxBECount;
5598 if (!CountDown && CR.getUnsignedMin().isMinValue())
5599 // When counting up, the worst starting value is 1, not 0.
5600 MaxBECount = CR.getUnsignedMax().isMinValue()
5601 ? getConstant(APInt::getMinValue(CR.getBitWidth()))
5602 : getConstant(APInt::getMaxValue(CR.getBitWidth()));
5604 MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
5605 : -CR.getUnsignedMin());
5606 return ExitLimit(Distance, MaxBECount);
5609 // If the recurrence is known not to wraparound, unsigned divide computes the
5610 // back edge count. (Ideally we would have an "isexact" bit for udiv). We know
5611 // that the value will either become zero (and thus the loop terminates), that
5612 // the loop will terminate through some other exit condition first, or that
5613 // the loop has undefined behavior. This means we can't "miss" the exit
5614 // value, even with nonunit stride.
5616 // This is only valid for expressions that directly compute the loop exit. It
5617 // is invalid for subexpressions in which the loop may exit through this
5618 // branch even if this subexpression is false. In that case, the trip count
5619 // computed by this udiv could be smaller than the number of well-defined
5621 if (!IsSubExpr && AddRec->getNoWrapFlags(SCEV::FlagNW))
5622 return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
5624 // Then, try to solve the above equation provided that Start is constant.
5625 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
5626 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
5627 -StartC->getValue()->getValue(),
5629 return getCouldNotCompute();
5632 /// HowFarToNonZero - Return the number of times a backedge checking the
5633 /// specified value for nonzero will execute. If not computable, return
5635 ScalarEvolution::ExitLimit
5636 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
5637 // Loops that look like: while (X == 0) are very strange indeed. We don't
5638 // handle them yet except for the trivial case. This could be expanded in the
5639 // future as needed.
5641 // If the value is a constant, check to see if it is known to be non-zero
5642 // already. If so, the backedge will execute zero times.
5643 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5644 if (!C->getValue()->isNullValue())
5645 return getConstant(C->getType(), 0);
5646 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5649 // We could implement others, but I really doubt anyone writes loops like
5650 // this, and if they did, they would already be constant folded.
5651 return getCouldNotCompute();
5654 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
5655 /// (which may not be an immediate predecessor) which has exactly one
5656 /// successor from which BB is reachable, or null if no such block is
5659 std::pair<BasicBlock *, BasicBlock *>
5660 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
5661 // If the block has a unique predecessor, then there is no path from the
5662 // predecessor to the block that does not go through the direct edge
5663 // from the predecessor to the block.
5664 if (BasicBlock *Pred = BB->getSinglePredecessor())
5665 return std::make_pair(Pred, BB);
5667 // A loop's header is defined to be a block that dominates the loop.
5668 // If the header has a unique predecessor outside the loop, it must be
5669 // a block that has exactly one successor that can reach the loop.
5670 if (Loop *L = LI->getLoopFor(BB))
5671 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
5673 return std::pair<BasicBlock *, BasicBlock *>();
5676 /// HasSameValue - SCEV structural equivalence is usually sufficient for
5677 /// testing whether two expressions are equal, however for the purposes of
5678 /// looking for a condition guarding a loop, it can be useful to be a little
5679 /// more general, since a front-end may have replicated the controlling
5682 static bool HasSameValue(const SCEV *A, const SCEV *B) {
5683 // Quick check to see if they are the same SCEV.
5684 if (A == B) return true;
5686 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
5687 // two different instructions with the same value. Check for this case.
5688 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
5689 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
5690 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
5691 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
5692 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
5695 // Otherwise assume they may have a different value.
5699 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
5700 /// predicate Pred. Return true iff any changes were made.
5702 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
5703 const SCEV *&LHS, const SCEV *&RHS,
5705 bool Changed = false;
5707 // If we hit the max recursion limit bail out.
5711 // Canonicalize a constant to the right side.
5712 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
5713 // Check for both operands constant.
5714 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
5715 if (ConstantExpr::getICmp(Pred,
5717 RHSC->getValue())->isNullValue())
5718 goto trivially_false;
5720 goto trivially_true;
5722 // Otherwise swap the operands to put the constant on the right.
5723 std::swap(LHS, RHS);
5724 Pred = ICmpInst::getSwappedPredicate(Pred);
5728 // If we're comparing an addrec with a value which is loop-invariant in the
5729 // addrec's loop, put the addrec on the left. Also make a dominance check,
5730 // as both operands could be addrecs loop-invariant in each other's loop.
5731 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
5732 const Loop *L = AR->getLoop();
5733 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
5734 std::swap(LHS, RHS);
5735 Pred = ICmpInst::getSwappedPredicate(Pred);
5740 // If there's a constant operand, canonicalize comparisons with boundary
5741 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
5742 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
5743 const APInt &RA = RC->getValue()->getValue();
5745 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5746 case ICmpInst::ICMP_EQ:
5747 case ICmpInst::ICMP_NE:
5748 // Fold ((-1) * %a) + %b == 0 (equivalent to %b-%a == 0) into %a == %b.
5750 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(LHS))
5751 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(AE->getOperand(0)))
5752 if (AE->getNumOperands() == 2 && ME->getNumOperands() == 2 &&
5753 ME->getOperand(0)->isAllOnesValue()) {
5754 RHS = AE->getOperand(1);
5755 LHS = ME->getOperand(1);
5759 case ICmpInst::ICMP_UGE:
5760 if ((RA - 1).isMinValue()) {
5761 Pred = ICmpInst::ICMP_NE;
5762 RHS = getConstant(RA - 1);
5766 if (RA.isMaxValue()) {
5767 Pred = ICmpInst::ICMP_EQ;
5771 if (RA.isMinValue()) goto trivially_true;
5773 Pred = ICmpInst::ICMP_UGT;
5774 RHS = getConstant(RA - 1);
5777 case ICmpInst::ICMP_ULE:
5778 if ((RA + 1).isMaxValue()) {
5779 Pred = ICmpInst::ICMP_NE;
5780 RHS = getConstant(RA + 1);
5784 if (RA.isMinValue()) {
5785 Pred = ICmpInst::ICMP_EQ;
5789 if (RA.isMaxValue()) goto trivially_true;
5791 Pred = ICmpInst::ICMP_ULT;
5792 RHS = getConstant(RA + 1);
5795 case ICmpInst::ICMP_SGE:
5796 if ((RA - 1).isMinSignedValue()) {
5797 Pred = ICmpInst::ICMP_NE;
5798 RHS = getConstant(RA - 1);
5802 if (RA.isMaxSignedValue()) {
5803 Pred = ICmpInst::ICMP_EQ;
5807 if (RA.isMinSignedValue()) goto trivially_true;
5809 Pred = ICmpInst::ICMP_SGT;
5810 RHS = getConstant(RA - 1);
5813 case ICmpInst::ICMP_SLE:
5814 if ((RA + 1).isMaxSignedValue()) {
5815 Pred = ICmpInst::ICMP_NE;
5816 RHS = getConstant(RA + 1);
5820 if (RA.isMinSignedValue()) {
5821 Pred = ICmpInst::ICMP_EQ;
5825 if (RA.isMaxSignedValue()) goto trivially_true;
5827 Pred = ICmpInst::ICMP_SLT;
5828 RHS = getConstant(RA + 1);
5831 case ICmpInst::ICMP_UGT:
5832 if (RA.isMinValue()) {
5833 Pred = ICmpInst::ICMP_NE;
5837 if ((RA + 1).isMaxValue()) {
5838 Pred = ICmpInst::ICMP_EQ;
5839 RHS = getConstant(RA + 1);
5843 if (RA.isMaxValue()) goto trivially_false;
5845 case ICmpInst::ICMP_ULT:
5846 if (RA.isMaxValue()) {
5847 Pred = ICmpInst::ICMP_NE;
5851 if ((RA - 1).isMinValue()) {
5852 Pred = ICmpInst::ICMP_EQ;
5853 RHS = getConstant(RA - 1);
5857 if (RA.isMinValue()) goto trivially_false;
5859 case ICmpInst::ICMP_SGT:
5860 if (RA.isMinSignedValue()) {
5861 Pred = ICmpInst::ICMP_NE;
5865 if ((RA + 1).isMaxSignedValue()) {
5866 Pred = ICmpInst::ICMP_EQ;
5867 RHS = getConstant(RA + 1);
5871 if (RA.isMaxSignedValue()) goto trivially_false;
5873 case ICmpInst::ICMP_SLT:
5874 if (RA.isMaxSignedValue()) {
5875 Pred = ICmpInst::ICMP_NE;
5879 if ((RA - 1).isMinSignedValue()) {
5880 Pred = ICmpInst::ICMP_EQ;
5881 RHS = getConstant(RA - 1);
5885 if (RA.isMinSignedValue()) goto trivially_false;
5890 // Check for obvious equality.
5891 if (HasSameValue(LHS, RHS)) {
5892 if (ICmpInst::isTrueWhenEqual(Pred))
5893 goto trivially_true;
5894 if (ICmpInst::isFalseWhenEqual(Pred))
5895 goto trivially_false;
5898 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5899 // adding or subtracting 1 from one of the operands.
5901 case ICmpInst::ICMP_SLE:
5902 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5903 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5905 Pred = ICmpInst::ICMP_SLT;
5907 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5908 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5910 Pred = ICmpInst::ICMP_SLT;
5914 case ICmpInst::ICMP_SGE:
5915 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5916 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5918 Pred = ICmpInst::ICMP_SGT;
5920 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5921 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5923 Pred = ICmpInst::ICMP_SGT;
5927 case ICmpInst::ICMP_ULE:
5928 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5929 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5931 Pred = ICmpInst::ICMP_ULT;
5933 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5934 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5936 Pred = ICmpInst::ICMP_ULT;
5940 case ICmpInst::ICMP_UGE:
5941 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5942 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5944 Pred = ICmpInst::ICMP_UGT;
5946 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5947 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5949 Pred = ICmpInst::ICMP_UGT;
5957 // TODO: More simplifications are possible here.
5959 // Recursively simplify until we either hit a recursion limit or nothing
5962 return SimplifyICmpOperands(Pred, LHS, RHS, Depth+1);
5968 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5969 Pred = ICmpInst::ICMP_EQ;
5974 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5975 Pred = ICmpInst::ICMP_NE;
5979 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5980 return getSignedRange(S).getSignedMax().isNegative();
5983 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5984 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5987 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5988 return !getSignedRange(S).getSignedMin().isNegative();
5991 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5992 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5995 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5996 return isKnownNegative(S) || isKnownPositive(S);
5999 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
6000 const SCEV *LHS, const SCEV *RHS) {
6001 // Canonicalize the inputs first.
6002 (void)SimplifyICmpOperands(Pred, LHS, RHS);
6004 // If LHS or RHS is an addrec, check to see if the condition is true in
6005 // every iteration of the loop.
6006 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
6007 if (isLoopEntryGuardedByCond(
6008 AR->getLoop(), Pred, AR->getStart(), RHS) &&
6009 isLoopBackedgeGuardedByCond(
6010 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
6012 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
6013 if (isLoopEntryGuardedByCond(
6014 AR->getLoop(), Pred, LHS, AR->getStart()) &&
6015 isLoopBackedgeGuardedByCond(
6016 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
6019 // Otherwise see what can be done with known constant ranges.
6020 return isKnownPredicateWithRanges(Pred, LHS, RHS);
6024 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
6025 const SCEV *LHS, const SCEV *RHS) {
6026 if (HasSameValue(LHS, RHS))
6027 return ICmpInst::isTrueWhenEqual(Pred);
6029 // This code is split out from isKnownPredicate because it is called from
6030 // within isLoopEntryGuardedByCond.
6033 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
6034 case ICmpInst::ICMP_SGT:
6035 Pred = ICmpInst::ICMP_SLT;
6036 std::swap(LHS, RHS);
6037 case ICmpInst::ICMP_SLT: {
6038 ConstantRange LHSRange = getSignedRange(LHS);
6039 ConstantRange RHSRange = getSignedRange(RHS);
6040 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
6042 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
6046 case ICmpInst::ICMP_SGE:
6047 Pred = ICmpInst::ICMP_SLE;
6048 std::swap(LHS, RHS);
6049 case ICmpInst::ICMP_SLE: {
6050 ConstantRange LHSRange = getSignedRange(LHS);
6051 ConstantRange RHSRange = getSignedRange(RHS);
6052 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
6054 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
6058 case ICmpInst::ICMP_UGT:
6059 Pred = ICmpInst::ICMP_ULT;
6060 std::swap(LHS, RHS);
6061 case ICmpInst::ICMP_ULT: {
6062 ConstantRange LHSRange = getUnsignedRange(LHS);
6063 ConstantRange RHSRange = getUnsignedRange(RHS);
6064 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
6066 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
6070 case ICmpInst::ICMP_UGE:
6071 Pred = ICmpInst::ICMP_ULE;
6072 std::swap(LHS, RHS);
6073 case ICmpInst::ICMP_ULE: {
6074 ConstantRange LHSRange = getUnsignedRange(LHS);
6075 ConstantRange RHSRange = getUnsignedRange(RHS);
6076 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
6078 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
6082 case ICmpInst::ICMP_NE: {
6083 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
6085 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
6088 const SCEV *Diff = getMinusSCEV(LHS, RHS);
6089 if (isKnownNonZero(Diff))
6093 case ICmpInst::ICMP_EQ:
6094 // The check at the top of the function catches the case where
6095 // the values are known to be equal.
6101 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
6102 /// protected by a conditional between LHS and RHS. This is used to
6103 /// to eliminate casts.
6105 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
6106 ICmpInst::Predicate Pred,
6107 const SCEV *LHS, const SCEV *RHS) {
6108 // Interpret a null as meaning no loop, where there is obviously no guard
6109 // (interprocedural conditions notwithstanding).
6110 if (!L) return true;
6112 BasicBlock *Latch = L->getLoopLatch();
6116 BranchInst *LoopContinuePredicate =
6117 dyn_cast<BranchInst>(Latch->getTerminator());
6118 if (!LoopContinuePredicate ||
6119 LoopContinuePredicate->isUnconditional())
6122 return isImpliedCond(Pred, LHS, RHS,
6123 LoopContinuePredicate->getCondition(),
6124 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
6127 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
6128 /// by a conditional between LHS and RHS. This is used to help avoid max
6129 /// expressions in loop trip counts, and to eliminate casts.
6131 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
6132 ICmpInst::Predicate Pred,
6133 const SCEV *LHS, const SCEV *RHS) {
6134 // Interpret a null as meaning no loop, where there is obviously no guard
6135 // (interprocedural conditions notwithstanding).
6136 if (!L) return false;
6138 // Starting at the loop predecessor, climb up the predecessor chain, as long
6139 // as there are predecessors that can be found that have unique successors
6140 // leading to the original header.
6141 for (std::pair<BasicBlock *, BasicBlock *>
6142 Pair(L->getLoopPredecessor(), L->getHeader());
6144 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
6146 BranchInst *LoopEntryPredicate =
6147 dyn_cast<BranchInst>(Pair.first->getTerminator());
6148 if (!LoopEntryPredicate ||
6149 LoopEntryPredicate->isUnconditional())
6152 if (isImpliedCond(Pred, LHS, RHS,
6153 LoopEntryPredicate->getCondition(),
6154 LoopEntryPredicate->getSuccessor(0) != Pair.second))
6161 /// RAII wrapper to prevent recursive application of isImpliedCond.
6162 /// ScalarEvolution's PendingLoopPredicates set must be empty unless we are
6163 /// currently evaluating isImpliedCond.
6164 struct MarkPendingLoopPredicate {
6166 DenseSet<Value*> &LoopPreds;
6169 MarkPendingLoopPredicate(Value *C, DenseSet<Value*> &LP)
6170 : Cond(C), LoopPreds(LP) {
6171 Pending = !LoopPreds.insert(Cond).second;
6173 ~MarkPendingLoopPredicate() {
6175 LoopPreds.erase(Cond);
6179 /// isImpliedCond - Test whether the condition described by Pred, LHS,
6180 /// and RHS is true whenever the given Cond value evaluates to true.
6181 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
6182 const SCEV *LHS, const SCEV *RHS,
6183 Value *FoundCondValue,
6185 MarkPendingLoopPredicate Mark(FoundCondValue, PendingLoopPredicates);
6189 // Recursively handle And and Or conditions.
6190 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
6191 if (BO->getOpcode() == Instruction::And) {
6193 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6194 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6195 } else if (BO->getOpcode() == Instruction::Or) {
6197 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6198 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6202 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
6203 if (!ICI) return false;
6205 // Bail if the ICmp's operands' types are wider than the needed type
6206 // before attempting to call getSCEV on them. This avoids infinite
6207 // recursion, since the analysis of widening casts can require loop
6208 // exit condition information for overflow checking, which would
6210 if (getTypeSizeInBits(LHS->getType()) <
6211 getTypeSizeInBits(ICI->getOperand(0)->getType()))
6214 // Now that we found a conditional branch that dominates the loop or controls
6215 // the loop latch. Check to see if it is the comparison we are looking for.
6216 ICmpInst::Predicate FoundPred;
6218 FoundPred = ICI->getInversePredicate();
6220 FoundPred = ICI->getPredicate();
6222 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
6223 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
6225 // Balance the types. The case where FoundLHS' type is wider than
6226 // LHS' type is checked for above.
6227 if (getTypeSizeInBits(LHS->getType()) >
6228 getTypeSizeInBits(FoundLHS->getType())) {
6229 if (CmpInst::isSigned(Pred)) {
6230 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
6231 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
6233 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
6234 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
6238 // Canonicalize the query to match the way instcombine will have
6239 // canonicalized the comparison.
6240 if (SimplifyICmpOperands(Pred, LHS, RHS))
6242 return CmpInst::isTrueWhenEqual(Pred);
6243 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
6244 if (FoundLHS == FoundRHS)
6245 return CmpInst::isFalseWhenEqual(FoundPred);
6247 // Check to see if we can make the LHS or RHS match.
6248 if (LHS == FoundRHS || RHS == FoundLHS) {
6249 if (isa<SCEVConstant>(RHS)) {
6250 std::swap(FoundLHS, FoundRHS);
6251 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
6253 std::swap(LHS, RHS);
6254 Pred = ICmpInst::getSwappedPredicate(Pred);
6258 // Check whether the found predicate is the same as the desired predicate.
6259 if (FoundPred == Pred)
6260 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
6262 // Check whether swapping the found predicate makes it the same as the
6263 // desired predicate.
6264 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
6265 if (isa<SCEVConstant>(RHS))
6266 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
6268 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
6269 RHS, LHS, FoundLHS, FoundRHS);
6272 // Check whether the actual condition is beyond sufficient.
6273 if (FoundPred == ICmpInst::ICMP_EQ)
6274 if (ICmpInst::isTrueWhenEqual(Pred))
6275 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
6277 if (Pred == ICmpInst::ICMP_NE)
6278 if (!ICmpInst::isTrueWhenEqual(FoundPred))
6279 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
6282 // Otherwise assume the worst.
6286 /// isImpliedCondOperands - Test whether the condition described by Pred,
6287 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
6288 /// and FoundRHS is true.
6289 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
6290 const SCEV *LHS, const SCEV *RHS,
6291 const SCEV *FoundLHS,
6292 const SCEV *FoundRHS) {
6293 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
6294 FoundLHS, FoundRHS) ||
6295 // ~x < ~y --> x > y
6296 isImpliedCondOperandsHelper(Pred, LHS, RHS,
6297 getNotSCEV(FoundRHS),
6298 getNotSCEV(FoundLHS));
6301 /// isImpliedCondOperandsHelper - Test whether the condition described by
6302 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
6303 /// FoundLHS, and FoundRHS is true.
6305 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
6306 const SCEV *LHS, const SCEV *RHS,
6307 const SCEV *FoundLHS,
6308 const SCEV *FoundRHS) {
6310 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
6311 case ICmpInst::ICMP_EQ:
6312 case ICmpInst::ICMP_NE:
6313 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
6316 case ICmpInst::ICMP_SLT:
6317 case ICmpInst::ICMP_SLE:
6318 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
6319 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
6322 case ICmpInst::ICMP_SGT:
6323 case ICmpInst::ICMP_SGE:
6324 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
6325 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
6328 case ICmpInst::ICMP_ULT:
6329 case ICmpInst::ICMP_ULE:
6330 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
6331 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
6334 case ICmpInst::ICMP_UGT:
6335 case ICmpInst::ICMP_UGE:
6336 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
6337 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
6345 /// getBECount - Subtract the end and start values and divide by the step,
6346 /// rounding up, to get the number of times the backedge is executed. Return
6347 /// CouldNotCompute if an intermediate computation overflows.
6348 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
6352 assert(!isKnownNegative(Step) &&
6353 "This code doesn't handle negative strides yet!");
6355 Type *Ty = Start->getType();
6357 // When Start == End, we have an exact BECount == 0. Short-circuit this case
6358 // here because SCEV may not be able to determine that the unsigned division
6359 // after rounding is zero.
6361 return getConstant(Ty, 0);
6363 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
6364 const SCEV *Diff = getMinusSCEV(End, Start);
6365 const SCEV *RoundUp = getAddExpr(Step, NegOne);
6367 // Add an adjustment to the difference between End and Start so that
6368 // the division will effectively round up.
6369 const SCEV *Add = getAddExpr(Diff, RoundUp);
6372 // Check Add for unsigned overflow.
6373 // TODO: More sophisticated things could be done here.
6374 Type *WideTy = IntegerType::get(getContext(),
6375 getTypeSizeInBits(Ty) + 1);
6376 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
6377 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
6378 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
6379 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
6380 return getCouldNotCompute();
6383 return getUDivExpr(Add, Step);
6386 /// HowManyLessThans - Return the number of times a backedge containing the
6387 /// specified less-than comparison will execute. If not computable, return
6388 /// CouldNotCompute.
6390 /// @param IsSubExpr is true when the LHS < RHS condition does not directly
6391 /// control the branch. In this case, we can only compute an iteration count for
6392 /// a subexpression that cannot overflow before evaluating true.
6393 ScalarEvolution::ExitLimit
6394 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
6395 const Loop *L, bool isSigned,
6397 // Only handle: "ADDREC < LoopInvariant".
6398 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute();
6400 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
6401 if (!AddRec || AddRec->getLoop() != L)
6402 return getCouldNotCompute();
6404 // Check to see if we have a flag which makes analysis easy.
6405 bool NoWrap = false;
6407 NoWrap = AddRec->getNoWrapFlags(
6408 (SCEV::NoWrapFlags)(((isSigned ? SCEV::FlagNSW : SCEV::FlagNUW))
6411 if (AddRec->isAffine()) {
6412 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
6413 const SCEV *Step = AddRec->getStepRecurrence(*this);
6416 return getCouldNotCompute();
6417 if (Step->isOne()) {
6418 // With unit stride, the iteration never steps past the limit value.
6419 } else if (isKnownPositive(Step)) {
6420 // Test whether a positive iteration can step past the limit
6421 // value and past the maximum value for its type in a single step.
6422 // Note that it's not sufficient to check NoWrap here, because even
6423 // though the value after a wrap is undefined, it's not undefined
6424 // behavior, so if wrap does occur, the loop could either terminate or
6425 // loop infinitely, but in either case, the loop is guaranteed to
6426 // iterate at least until the iteration where the wrapping occurs.
6427 const SCEV *One = getConstant(Step->getType(), 1);
6429 APInt Max = APInt::getSignedMaxValue(BitWidth);
6430 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
6431 .slt(getSignedRange(RHS).getSignedMax()))
6432 return getCouldNotCompute();
6434 APInt Max = APInt::getMaxValue(BitWidth);
6435 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
6436 .ult(getUnsignedRange(RHS).getUnsignedMax()))
6437 return getCouldNotCompute();
6440 // TODO: Handle negative strides here and below.
6441 return getCouldNotCompute();
6443 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
6444 // m. So, we count the number of iterations in which {n,+,s} < m is true.
6445 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
6446 // treat m-n as signed nor unsigned due to overflow possibility.
6448 // First, we get the value of the LHS in the first iteration: n
6449 const SCEV *Start = AddRec->getOperand(0);
6451 // Determine the minimum constant start value.
6452 const SCEV *MinStart = getConstant(isSigned ?
6453 getSignedRange(Start).getSignedMin() :
6454 getUnsignedRange(Start).getUnsignedMin());
6456 // If we know that the condition is true in order to enter the loop,
6457 // then we know that it will run exactly (m-n)/s times. Otherwise, we
6458 // only know that it will execute (max(m,n)-n)/s times. In both cases,
6459 // the division must round up.
6460 const SCEV *End = RHS;
6461 if (!isLoopEntryGuardedByCond(L,
6462 isSigned ? ICmpInst::ICMP_SLT :
6464 getMinusSCEV(Start, Step), RHS))
6465 End = isSigned ? getSMaxExpr(RHS, Start)
6466 : getUMaxExpr(RHS, Start);
6468 // Determine the maximum constant end value.
6469 const SCEV *MaxEnd = getConstant(isSigned ?
6470 getSignedRange(End).getSignedMax() :
6471 getUnsignedRange(End).getUnsignedMax());
6473 // If MaxEnd is within a step of the maximum integer value in its type,
6474 // adjust it down to the minimum value which would produce the same effect.
6475 // This allows the subsequent ceiling division of (N+(step-1))/step to
6476 // compute the correct value.
6477 const SCEV *StepMinusOne = getMinusSCEV(Step,
6478 getConstant(Step->getType(), 1));
6481 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
6484 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
6487 // Finally, we subtract these two values and divide, rounding up, to get
6488 // the number of times the backedge is executed.
6489 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
6491 // The maximum backedge count is similar, except using the minimum start
6492 // value and the maximum end value.
6493 // If we already have an exact constant BECount, use it instead.
6494 const SCEV *MaxBECount = isa<SCEVConstant>(BECount) ? BECount
6495 : getBECount(MinStart, MaxEnd, Step, NoWrap);
6497 // If the stride is nonconstant, and NoWrap == true, then
6498 // getBECount(MinStart, MaxEnd) may not compute. This would result in an
6499 // exact BECount and invalid MaxBECount, which should be avoided to catch
6500 // more optimization opportunities.
6501 if (isa<SCEVCouldNotCompute>(MaxBECount))
6502 MaxBECount = BECount;
6504 return ExitLimit(BECount, MaxBECount);
6507 return getCouldNotCompute();
6510 /// getNumIterationsInRange - Return the number of iterations of this loop that
6511 /// produce values in the specified constant range. Another way of looking at
6512 /// this is that it returns the first iteration number where the value is not in
6513 /// the condition, thus computing the exit count. If the iteration count can't
6514 /// be computed, an instance of SCEVCouldNotCompute is returned.
6515 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
6516 ScalarEvolution &SE) const {
6517 if (Range.isFullSet()) // Infinite loop.
6518 return SE.getCouldNotCompute();
6520 // If the start is a non-zero constant, shift the range to simplify things.
6521 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
6522 if (!SC->getValue()->isZero()) {
6523 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
6524 Operands[0] = SE.getConstant(SC->getType(), 0);
6525 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop(),
6526 getNoWrapFlags(FlagNW));
6527 if (const SCEVAddRecExpr *ShiftedAddRec =
6528 dyn_cast<SCEVAddRecExpr>(Shifted))
6529 return ShiftedAddRec->getNumIterationsInRange(
6530 Range.subtract(SC->getValue()->getValue()), SE);
6531 // This is strange and shouldn't happen.
6532 return SE.getCouldNotCompute();
6535 // The only time we can solve this is when we have all constant indices.
6536 // Otherwise, we cannot determine the overflow conditions.
6537 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
6538 if (!isa<SCEVConstant>(getOperand(i)))
6539 return SE.getCouldNotCompute();
6542 // Okay at this point we know that all elements of the chrec are constants and
6543 // that the start element is zero.
6545 // First check to see if the range contains zero. If not, the first
6547 unsigned BitWidth = SE.getTypeSizeInBits(getType());
6548 if (!Range.contains(APInt(BitWidth, 0)))
6549 return SE.getConstant(getType(), 0);
6552 // If this is an affine expression then we have this situation:
6553 // Solve {0,+,A} in Range === Ax in Range
6555 // We know that zero is in the range. If A is positive then we know that
6556 // the upper value of the range must be the first possible exit value.
6557 // If A is negative then the lower of the range is the last possible loop
6558 // value. Also note that we already checked for a full range.
6559 APInt One(BitWidth,1);
6560 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
6561 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
6563 // The exit value should be (End+A)/A.
6564 APInt ExitVal = (End + A).udiv(A);
6565 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
6567 // Evaluate at the exit value. If we really did fall out of the valid
6568 // range, then we computed our trip count, otherwise wrap around or other
6569 // things must have happened.
6570 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
6571 if (Range.contains(Val->getValue()))
6572 return SE.getCouldNotCompute(); // Something strange happened
6574 // Ensure that the previous value is in the range. This is a sanity check.
6575 assert(Range.contains(
6576 EvaluateConstantChrecAtConstant(this,
6577 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
6578 "Linear scev computation is off in a bad way!");
6579 return SE.getConstant(ExitValue);
6580 } else if (isQuadratic()) {
6581 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
6582 // quadratic equation to solve it. To do this, we must frame our problem in
6583 // terms of figuring out when zero is crossed, instead of when
6584 // Range.getUpper() is crossed.
6585 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
6586 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
6587 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop(),
6588 // getNoWrapFlags(FlagNW)
6591 // Next, solve the constructed addrec
6592 std::pair<const SCEV *,const SCEV *> Roots =
6593 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
6594 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
6595 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
6597 // Pick the smallest positive root value.
6598 if (ConstantInt *CB =
6599 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
6600 R1->getValue(), R2->getValue()))) {
6601 if (CB->getZExtValue() == false)
6602 std::swap(R1, R2); // R1 is the minimum root now.
6604 // Make sure the root is not off by one. The returned iteration should
6605 // not be in the range, but the previous one should be. When solving
6606 // for "X*X < 5", for example, we should not return a root of 2.
6607 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
6610 if (Range.contains(R1Val->getValue())) {
6611 // The next iteration must be out of the range...
6612 ConstantInt *NextVal =
6613 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
6615 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6616 if (!Range.contains(R1Val->getValue()))
6617 return SE.getConstant(NextVal);
6618 return SE.getCouldNotCompute(); // Something strange happened
6621 // If R1 was not in the range, then it is a good return value. Make
6622 // sure that R1-1 WAS in the range though, just in case.
6623 ConstantInt *NextVal =
6624 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
6625 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6626 if (Range.contains(R1Val->getValue()))
6628 return SE.getCouldNotCompute(); // Something strange happened
6633 return SE.getCouldNotCompute();
6638 //===----------------------------------------------------------------------===//
6639 // SCEVCallbackVH Class Implementation
6640 //===----------------------------------------------------------------------===//
6642 void ScalarEvolution::SCEVCallbackVH::deleted() {
6643 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6644 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
6645 SE->ConstantEvolutionLoopExitValue.erase(PN);
6646 SE->ValueExprMap.erase(getValPtr());
6647 // this now dangles!
6650 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
6651 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6653 // Forget all the expressions associated with users of the old value,
6654 // so that future queries will recompute the expressions using the new
6656 Value *Old = getValPtr();
6657 SmallVector<User *, 16> Worklist;
6658 SmallPtrSet<User *, 8> Visited;
6659 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
6661 Worklist.push_back(*UI);
6662 while (!Worklist.empty()) {
6663 User *U = Worklist.pop_back_val();
6664 // Deleting the Old value will cause this to dangle. Postpone
6665 // that until everything else is done.
6668 if (!Visited.insert(U))
6670 if (PHINode *PN = dyn_cast<PHINode>(U))
6671 SE->ConstantEvolutionLoopExitValue.erase(PN);
6672 SE->ValueExprMap.erase(U);
6673 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
6675 Worklist.push_back(*UI);
6677 // Delete the Old value.
6678 if (PHINode *PN = dyn_cast<PHINode>(Old))
6679 SE->ConstantEvolutionLoopExitValue.erase(PN);
6680 SE->ValueExprMap.erase(Old);
6681 // this now dangles!
6684 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
6685 : CallbackVH(V), SE(se) {}
6687 //===----------------------------------------------------------------------===//
6688 // ScalarEvolution Class Implementation
6689 //===----------------------------------------------------------------------===//
6691 ScalarEvolution::ScalarEvolution()
6692 : FunctionPass(ID), FirstUnknown(0) {
6693 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
6696 bool ScalarEvolution::runOnFunction(Function &F) {
6698 LI = &getAnalysis<LoopInfo>();
6699 TD = getAnalysisIfAvailable<DataLayout>();
6700 TLI = &getAnalysis<TargetLibraryInfo>();
6701 DT = &getAnalysis<DominatorTree>();
6705 void ScalarEvolution::releaseMemory() {
6706 // Iterate through all the SCEVUnknown instances and call their
6707 // destructors, so that they release their references to their values.
6708 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
6712 ValueExprMap.clear();
6714 // Free any extra memory created for ExitNotTakenInfo in the unlikely event
6715 // that a loop had multiple computable exits.
6716 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
6717 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end();
6722 assert(PendingLoopPredicates.empty() && "isImpliedCond garbage");
6724 BackedgeTakenCounts.clear();
6725 ConstantEvolutionLoopExitValue.clear();
6726 ValuesAtScopes.clear();
6727 LoopDispositions.clear();
6728 BlockDispositions.clear();
6729 UnsignedRanges.clear();
6730 SignedRanges.clear();
6731 UniqueSCEVs.clear();
6732 SCEVAllocator.Reset();
6735 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
6736 AU.setPreservesAll();
6737 AU.addRequiredTransitive<LoopInfo>();
6738 AU.addRequiredTransitive<DominatorTree>();
6739 AU.addRequired<TargetLibraryInfo>();
6742 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
6743 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
6746 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
6748 // Print all inner loops first
6749 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
6750 PrintLoopInfo(OS, SE, *I);
6753 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6756 SmallVector<BasicBlock *, 8> ExitBlocks;
6757 L->getExitBlocks(ExitBlocks);
6758 if (ExitBlocks.size() != 1)
6759 OS << "<multiple exits> ";
6761 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
6762 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
6764 OS << "Unpredictable backedge-taken count. ";
6769 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6772 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
6773 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
6775 OS << "Unpredictable max backedge-taken count. ";
6781 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
6782 // ScalarEvolution's implementation of the print method is to print
6783 // out SCEV values of all instructions that are interesting. Doing
6784 // this potentially causes it to create new SCEV objects though,
6785 // which technically conflicts with the const qualifier. This isn't
6786 // observable from outside the class though, so casting away the
6787 // const isn't dangerous.
6788 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
6790 OS << "Classifying expressions for: ";
6791 WriteAsOperand(OS, F, /*PrintType=*/false);
6793 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
6794 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
6797 const SCEV *SV = SE.getSCEV(&*I);
6800 const Loop *L = LI->getLoopFor((*I).getParent());
6802 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
6809 OS << "\t\t" "Exits: ";
6810 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
6811 if (!SE.isLoopInvariant(ExitValue, L)) {
6812 OS << "<<Unknown>>";
6821 OS << "Determining loop execution counts for: ";
6822 WriteAsOperand(OS, F, /*PrintType=*/false);
6824 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6825 PrintLoopInfo(OS, &SE, *I);
6828 ScalarEvolution::LoopDisposition
6829 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
6830 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S];
6831 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair =
6832 Values.insert(std::make_pair(L, LoopVariant));
6834 return Pair.first->second;
6836 LoopDisposition D = computeLoopDisposition(S, L);
6837 return LoopDispositions[S][L] = D;
6840 ScalarEvolution::LoopDisposition
6841 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
6842 switch (S->getSCEVType()) {
6844 return LoopInvariant;
6848 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
6849 case scAddRecExpr: {
6850 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6852 // If L is the addrec's loop, it's computable.
6853 if (AR->getLoop() == L)
6854 return LoopComputable;
6856 // Add recurrences are never invariant in the function-body (null loop).
6860 // This recurrence is variant w.r.t. L if L contains AR's loop.
6861 if (L->contains(AR->getLoop()))
6864 // This recurrence is invariant w.r.t. L if AR's loop contains L.
6865 if (AR->getLoop()->contains(L))
6866 return LoopInvariant;
6868 // This recurrence is variant w.r.t. L if any of its operands
6870 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
6872 if (!isLoopInvariant(*I, L))
6875 // Otherwise it's loop-invariant.
6876 return LoopInvariant;
6882 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6883 bool HasVarying = false;
6884 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6886 LoopDisposition D = getLoopDisposition(*I, L);
6887 if (D == LoopVariant)
6889 if (D == LoopComputable)
6892 return HasVarying ? LoopComputable : LoopInvariant;
6895 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6896 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
6897 if (LD == LoopVariant)
6899 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
6900 if (RD == LoopVariant)
6902 return (LD == LoopInvariant && RD == LoopInvariant) ?
6903 LoopInvariant : LoopComputable;
6906 // All non-instruction values are loop invariant. All instructions are loop
6907 // invariant if they are not contained in the specified loop.
6908 // Instructions are never considered invariant in the function body
6909 // (null loop) because they are defined within the "loop".
6910 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
6911 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
6912 return LoopInvariant;
6913 case scCouldNotCompute:
6914 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6915 default: llvm_unreachable("Unknown SCEV kind!");
6919 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
6920 return getLoopDisposition(S, L) == LoopInvariant;
6923 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
6924 return getLoopDisposition(S, L) == LoopComputable;
6927 ScalarEvolution::BlockDisposition
6928 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6929 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S];
6930 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool>
6931 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock));
6933 return Pair.first->second;
6935 BlockDisposition D = computeBlockDisposition(S, BB);
6936 return BlockDispositions[S][BB] = D;
6939 ScalarEvolution::BlockDisposition
6940 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6941 switch (S->getSCEVType()) {
6943 return ProperlyDominatesBlock;
6947 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
6948 case scAddRecExpr: {
6949 // This uses a "dominates" query instead of "properly dominates" query
6950 // to test for proper dominance too, because the instruction which
6951 // produces the addrec's value is a PHI, and a PHI effectively properly
6952 // dominates its entire containing block.
6953 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6954 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
6955 return DoesNotDominateBlock;
6957 // FALL THROUGH into SCEVNAryExpr handling.
6962 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6964 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6966 BlockDisposition D = getBlockDisposition(*I, BB);
6967 if (D == DoesNotDominateBlock)
6968 return DoesNotDominateBlock;
6969 if (D == DominatesBlock)
6972 return Proper ? ProperlyDominatesBlock : DominatesBlock;
6975 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6976 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6977 BlockDisposition LD = getBlockDisposition(LHS, BB);
6978 if (LD == DoesNotDominateBlock)
6979 return DoesNotDominateBlock;
6980 BlockDisposition RD = getBlockDisposition(RHS, BB);
6981 if (RD == DoesNotDominateBlock)
6982 return DoesNotDominateBlock;
6983 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
6984 ProperlyDominatesBlock : DominatesBlock;
6987 if (Instruction *I =
6988 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
6989 if (I->getParent() == BB)
6990 return DominatesBlock;
6991 if (DT->properlyDominates(I->getParent(), BB))
6992 return ProperlyDominatesBlock;
6993 return DoesNotDominateBlock;
6995 return ProperlyDominatesBlock;
6996 case scCouldNotCompute:
6997 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6999 llvm_unreachable("Unknown SCEV kind!");
7003 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
7004 return getBlockDisposition(S, BB) >= DominatesBlock;
7007 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
7008 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
7012 // Search for a SCEV expression node within an expression tree.
7013 // Implements SCEVTraversal::Visitor.
7018 SCEVSearch(const SCEV *N): Node(N), IsFound(false) {}
7020 bool follow(const SCEV *S) {
7021 IsFound |= (S == Node);
7024 bool isDone() const { return IsFound; }
7028 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
7029 SCEVSearch Search(Op);
7030 visitAll(S, Search);
7031 return Search.IsFound;
7034 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
7035 ValuesAtScopes.erase(S);
7036 LoopDispositions.erase(S);
7037 BlockDispositions.erase(S);
7038 UnsignedRanges.erase(S);
7039 SignedRanges.erase(S);
7041 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
7042 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end(); I != E; ) {
7043 BackedgeTakenInfo &BEInfo = I->second;
7044 if (BEInfo.hasOperand(S, this)) {
7046 BackedgeTakenCounts.erase(I++);
7053 typedef DenseMap<const Loop *, std::string> VerifyMap;
7055 /// replaceSubString - Replaces all occurences of From in Str with To.
7056 static void replaceSubString(std::string &Str, StringRef From, StringRef To) {
7058 while ((Pos = Str.find(From, Pos)) != std::string::npos) {
7059 Str.replace(Pos, From.size(), To.data(), To.size());
7064 /// getLoopBackedgeTakenCounts - Helper method for verifyAnalysis.
7066 getLoopBackedgeTakenCounts(Loop *L, VerifyMap &Map, ScalarEvolution &SE) {
7067 for (Loop::reverse_iterator I = L->rbegin(), E = L->rend(); I != E; ++I) {
7068 getLoopBackedgeTakenCounts(*I, Map, SE); // recurse.
7070 std::string &S = Map[L];
7072 raw_string_ostream OS(S);
7073 SE.getBackedgeTakenCount(L)->print(OS);
7075 // false and 0 are semantically equivalent. This can happen in dead loops.
7076 replaceSubString(OS.str(), "false", "0");
7077 // Remove wrap flags, their use in SCEV is highly fragile.
7078 // FIXME: Remove this when SCEV gets smarter about them.
7079 replaceSubString(OS.str(), "<nw>", "");
7080 replaceSubString(OS.str(), "<nsw>", "");
7081 replaceSubString(OS.str(), "<nuw>", "");
7086 void ScalarEvolution::verifyAnalysis() const {
7090 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
7092 // Gather stringified backedge taken counts for all loops using SCEV's caches.
7093 // FIXME: It would be much better to store actual values instead of strings,
7094 // but SCEV pointers will change if we drop the caches.
7095 VerifyMap BackedgeDumpsOld, BackedgeDumpsNew;
7096 for (LoopInfo::reverse_iterator I = LI->rbegin(), E = LI->rend(); I != E; ++I)
7097 getLoopBackedgeTakenCounts(*I, BackedgeDumpsOld, SE);
7099 // Gather stringified backedge taken counts for all loops without using
7102 for (LoopInfo::reverse_iterator I = LI->rbegin(), E = LI->rend(); I != E; ++I)
7103 getLoopBackedgeTakenCounts(*I, BackedgeDumpsNew, SE);
7105 // Now compare whether they're the same with and without caches. This allows
7106 // verifying that no pass changed the cache.
7107 assert(BackedgeDumpsOld.size() == BackedgeDumpsNew.size() &&
7108 "New loops suddenly appeared!");
7110 for (VerifyMap::iterator OldI = BackedgeDumpsOld.begin(),
7111 OldE = BackedgeDumpsOld.end(),
7112 NewI = BackedgeDumpsNew.begin();
7113 OldI != OldE; ++OldI, ++NewI) {
7114 assert(OldI->first == NewI->first && "Loop order changed!");
7116 // Compare the stringified SCEVs. We don't care if undef backedgetaken count
7118 // FIXME: We currently ignore SCEV changes from/to CouldNotCompute. This
7119 // means that a pass is buggy or SCEV has to learn a new pattern but is
7120 // usually not harmful.
7121 if (OldI->second != NewI->second &&
7122 OldI->second.find("undef") == std::string::npos &&
7123 NewI->second.find("undef") == std::string::npos &&
7124 OldI->second != "***COULDNOTCOMPUTE***" &&
7125 NewI->second != "***COULDNOTCOMPUTE***") {
7126 dbgs() << "SCEVValidator: SCEV for loop '"
7127 << OldI->first->getHeader()->getName()
7128 << "' changed from '" << OldI->second
7129 << "' to '" << NewI->second << "'!\n";
7134 // TODO: Verify more things.