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 for (unsigned i = 0; i != LNumOps; ++i) {
591 long X = compare(LC->getOperand(i), RC->getOperand(i));
595 return (int)LNumOps - (int)RNumOps;
599 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
600 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
602 // Lexicographically compare udiv expressions.
603 long X = compare(LC->getLHS(), RC->getLHS());
606 return compare(LC->getRHS(), RC->getRHS());
612 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
613 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
615 // Compare cast expressions by operand.
616 return compare(LC->getOperand(), RC->getOperand());
620 llvm_unreachable("Unknown SCEV kind!");
626 /// GroupByComplexity - Given a list of SCEV objects, order them by their
627 /// complexity, and group objects of the same complexity together by value.
628 /// When this routine is finished, we know that any duplicates in the vector are
629 /// consecutive and that complexity is monotonically increasing.
631 /// Note that we go take special precautions to ensure that we get deterministic
632 /// results from this routine. In other words, we don't want the results of
633 /// this to depend on where the addresses of various SCEV objects happened to
636 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
638 if (Ops.size() < 2) return; // Noop
639 if (Ops.size() == 2) {
640 // This is the common case, which also happens to be trivially simple.
642 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
643 if (SCEVComplexityCompare(LI)(RHS, LHS))
648 // Do the rough sort by complexity.
649 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
651 // Now that we are sorted by complexity, group elements of the same
652 // complexity. Note that this is, at worst, N^2, but the vector is likely to
653 // be extremely short in practice. Note that we take this approach because we
654 // do not want to depend on the addresses of the objects we are grouping.
655 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
656 const SCEV *S = Ops[i];
657 unsigned Complexity = S->getSCEVType();
659 // If there are any objects of the same complexity and same value as this
661 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
662 if (Ops[j] == S) { // Found a duplicate.
663 // Move it to immediately after i'th element.
664 std::swap(Ops[i+1], Ops[j]);
665 ++i; // no need to rescan it.
666 if (i == e-2) return; // Done!
674 //===----------------------------------------------------------------------===//
675 // Simple SCEV method implementations
676 //===----------------------------------------------------------------------===//
678 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
680 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
683 // Handle the simplest case efficiently.
685 return SE.getTruncateOrZeroExtend(It, ResultTy);
687 // We are using the following formula for BC(It, K):
689 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
691 // Suppose, W is the bitwidth of the return value. We must be prepared for
692 // overflow. Hence, we must assure that the result of our computation is
693 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
694 // safe in modular arithmetic.
696 // However, this code doesn't use exactly that formula; the formula it uses
697 // is something like the following, where T is the number of factors of 2 in
698 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
701 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
703 // This formula is trivially equivalent to the previous formula. However,
704 // this formula can be implemented much more efficiently. The trick is that
705 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
706 // arithmetic. To do exact division in modular arithmetic, all we have
707 // to do is multiply by the inverse. Therefore, this step can be done at
710 // The next issue is how to safely do the division by 2^T. The way this
711 // is done is by doing the multiplication step at a width of at least W + T
712 // bits. This way, the bottom W+T bits of the product are accurate. Then,
713 // when we perform the division by 2^T (which is equivalent to a right shift
714 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
715 // truncated out after the division by 2^T.
717 // In comparison to just directly using the first formula, this technique
718 // is much more efficient; using the first formula requires W * K bits,
719 // but this formula less than W + K bits. Also, the first formula requires
720 // a division step, whereas this formula only requires multiplies and shifts.
722 // It doesn't matter whether the subtraction step is done in the calculation
723 // width or the input iteration count's width; if the subtraction overflows,
724 // the result must be zero anyway. We prefer here to do it in the width of
725 // the induction variable because it helps a lot for certain cases; CodeGen
726 // isn't smart enough to ignore the overflow, which leads to much less
727 // efficient code if the width of the subtraction is wider than the native
730 // (It's possible to not widen at all by pulling out factors of 2 before
731 // the multiplication; for example, K=2 can be calculated as
732 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
733 // extra arithmetic, so it's not an obvious win, and it gets
734 // much more complicated for K > 3.)
736 // Protection from insane SCEVs; this bound is conservative,
737 // but it probably doesn't matter.
739 return SE.getCouldNotCompute();
741 unsigned W = SE.getTypeSizeInBits(ResultTy);
743 // Calculate K! / 2^T and T; we divide out the factors of two before
744 // multiplying for calculating K! / 2^T to avoid overflow.
745 // Other overflow doesn't matter because we only care about the bottom
746 // W bits of the result.
747 APInt OddFactorial(W, 1);
749 for (unsigned i = 3; i <= K; ++i) {
751 unsigned TwoFactors = Mult.countTrailingZeros();
753 Mult = Mult.lshr(TwoFactors);
754 OddFactorial *= Mult;
757 // We need at least W + T bits for the multiplication step
758 unsigned CalculationBits = W + T;
760 // Calculate 2^T, at width T+W.
761 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
763 // Calculate the multiplicative inverse of K! / 2^T;
764 // this multiplication factor will perform the exact division by
766 APInt Mod = APInt::getSignedMinValue(W+1);
767 APInt MultiplyFactor = OddFactorial.zext(W+1);
768 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
769 MultiplyFactor = MultiplyFactor.trunc(W);
771 // Calculate the product, at width T+W
772 IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
774 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
775 for (unsigned i = 1; i != K; ++i) {
776 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
777 Dividend = SE.getMulExpr(Dividend,
778 SE.getTruncateOrZeroExtend(S, CalculationTy));
782 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
784 // Truncate the result, and divide by K! / 2^T.
786 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
787 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
790 /// evaluateAtIteration - Return the value of this chain of recurrences at
791 /// the specified iteration number. We can evaluate this recurrence by
792 /// multiplying each element in the chain by the binomial coefficient
793 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
795 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
797 /// where BC(It, k) stands for binomial coefficient.
799 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
800 ScalarEvolution &SE) const {
801 const SCEV *Result = getStart();
802 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
803 // The computation is correct in the face of overflow provided that the
804 // multiplication is performed _after_ the evaluation of the binomial
806 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
807 if (isa<SCEVCouldNotCompute>(Coeff))
810 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
815 //===----------------------------------------------------------------------===//
816 // SCEV Expression folder implementations
817 //===----------------------------------------------------------------------===//
819 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
821 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
822 "This is not a truncating conversion!");
823 assert(isSCEVable(Ty) &&
824 "This is not a conversion to a SCEVable type!");
825 Ty = getEffectiveSCEVType(Ty);
828 ID.AddInteger(scTruncate);
832 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
834 // Fold if the operand is constant.
835 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
837 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
839 // trunc(trunc(x)) --> trunc(x)
840 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
841 return getTruncateExpr(ST->getOperand(), Ty);
843 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
844 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
845 return getTruncateOrSignExtend(SS->getOperand(), Ty);
847 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
848 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
849 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
851 // trunc(x1+x2+...+xN) --> trunc(x1)+trunc(x2)+...+trunc(xN) if we can
852 // eliminate all the truncates.
853 if (const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Op)) {
854 SmallVector<const SCEV *, 4> Operands;
855 bool hasTrunc = false;
856 for (unsigned i = 0, e = SA->getNumOperands(); i != e && !hasTrunc; ++i) {
857 const SCEV *S = getTruncateExpr(SA->getOperand(i), Ty);
858 hasTrunc = isa<SCEVTruncateExpr>(S);
859 Operands.push_back(S);
862 return getAddExpr(Operands);
863 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
866 // trunc(x1*x2*...*xN) --> trunc(x1)*trunc(x2)*...*trunc(xN) if we can
867 // eliminate all the truncates.
868 if (const SCEVMulExpr *SM = dyn_cast<SCEVMulExpr>(Op)) {
869 SmallVector<const SCEV *, 4> Operands;
870 bool hasTrunc = false;
871 for (unsigned i = 0, e = SM->getNumOperands(); i != e && !hasTrunc; ++i) {
872 const SCEV *S = getTruncateExpr(SM->getOperand(i), Ty);
873 hasTrunc = isa<SCEVTruncateExpr>(S);
874 Operands.push_back(S);
877 return getMulExpr(Operands);
878 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
881 // If the input value is a chrec scev, truncate the chrec's operands.
882 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
883 SmallVector<const SCEV *, 4> Operands;
884 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
885 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
886 return getAddRecExpr(Operands, AddRec->getLoop(), SCEV::FlagAnyWrap);
889 // The cast wasn't folded; create an explicit cast node. We can reuse
890 // the existing insert position since if we get here, we won't have
891 // made any changes which would invalidate it.
892 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
894 UniqueSCEVs.InsertNode(S, IP);
898 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
900 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
901 "This is not an extending conversion!");
902 assert(isSCEVable(Ty) &&
903 "This is not a conversion to a SCEVable type!");
904 Ty = getEffectiveSCEVType(Ty);
906 // Fold if the operand is constant.
907 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
909 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(), Ty)));
911 // zext(zext(x)) --> zext(x)
912 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
913 return getZeroExtendExpr(SZ->getOperand(), Ty);
915 // Before doing any expensive analysis, check to see if we've already
916 // computed a SCEV for this Op and Ty.
918 ID.AddInteger(scZeroExtend);
922 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
924 // zext(trunc(x)) --> zext(x) or x or trunc(x)
925 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
926 // It's possible the bits taken off by the truncate were all zero bits. If
927 // so, we should be able to simplify this further.
928 const SCEV *X = ST->getOperand();
929 ConstantRange CR = getUnsignedRange(X);
930 unsigned TruncBits = getTypeSizeInBits(ST->getType());
931 unsigned NewBits = getTypeSizeInBits(Ty);
932 if (CR.truncate(TruncBits).zeroExtend(NewBits).contains(
933 CR.zextOrTrunc(NewBits)))
934 return getTruncateOrZeroExtend(X, Ty);
937 // If the input value is a chrec scev, and we can prove that the value
938 // did not overflow the old, smaller, value, we can zero extend all of the
939 // operands (often constants). This allows analysis of something like
940 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
941 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
942 if (AR->isAffine()) {
943 const SCEV *Start = AR->getStart();
944 const SCEV *Step = AR->getStepRecurrence(*this);
945 unsigned BitWidth = getTypeSizeInBits(AR->getType());
946 const Loop *L = AR->getLoop();
948 // If we have special knowledge that this addrec won't overflow,
949 // we don't need to do any further analysis.
950 if (AR->getNoWrapFlags(SCEV::FlagNUW))
951 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
952 getZeroExtendExpr(Step, Ty),
953 L, AR->getNoWrapFlags());
955 // Check whether the backedge-taken count is SCEVCouldNotCompute.
956 // Note that this serves two purposes: It filters out loops that are
957 // simply not analyzable, and it covers the case where this code is
958 // being called from within backedge-taken count analysis, such that
959 // attempting to ask for the backedge-taken count would likely result
960 // in infinite recursion. In the later case, the analysis code will
961 // cope with a conservative value, and it will take care to purge
962 // that value once it has finished.
963 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
964 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
965 // Manually compute the final value for AR, checking for
968 // Check whether the backedge-taken count can be losslessly casted to
969 // the addrec's type. The count is always unsigned.
970 const SCEV *CastedMaxBECount =
971 getTruncateOrZeroExtend(MaxBECount, Start->getType());
972 const SCEV *RecastedMaxBECount =
973 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
974 if (MaxBECount == RecastedMaxBECount) {
975 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
976 // Check whether Start+Step*MaxBECount has no unsigned overflow.
977 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
978 const SCEV *ZAdd = getZeroExtendExpr(getAddExpr(Start, ZMul), WideTy);
979 const SCEV *WideStart = getZeroExtendExpr(Start, WideTy);
980 const SCEV *WideMaxBECount =
981 getZeroExtendExpr(CastedMaxBECount, WideTy);
982 const SCEV *OperandExtendedAdd =
983 getAddExpr(WideStart,
984 getMulExpr(WideMaxBECount,
985 getZeroExtendExpr(Step, WideTy)));
986 if (ZAdd == OperandExtendedAdd) {
987 // Cache knowledge of AR NUW, which is propagated to this AddRec.
988 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
989 // Return the expression with the addrec on the outside.
990 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
991 getZeroExtendExpr(Step, Ty),
992 L, AR->getNoWrapFlags());
994 // Similar to above, only this time treat the step value as signed.
995 // This covers loops that count down.
997 getAddExpr(WideStart,
998 getMulExpr(WideMaxBECount,
999 getSignExtendExpr(Step, WideTy)));
1000 if (ZAdd == OperandExtendedAdd) {
1001 // Cache knowledge of AR NW, which is propagated to this AddRec.
1002 // Negative step causes unsigned wrap, but it still can't self-wrap.
1003 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1004 // Return the expression with the addrec on the outside.
1005 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1006 getSignExtendExpr(Step, Ty),
1007 L, AR->getNoWrapFlags());
1011 // If the backedge is guarded by a comparison with the pre-inc value
1012 // the addrec is safe. Also, if the entry is guarded by a comparison
1013 // with the start value and the backedge is guarded by a comparison
1014 // with the post-inc value, the addrec is safe.
1015 if (isKnownPositive(Step)) {
1016 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
1017 getUnsignedRange(Step).getUnsignedMax());
1018 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
1019 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
1020 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
1021 AR->getPostIncExpr(*this), N))) {
1022 // Cache knowledge of AR NUW, which is propagated to this AddRec.
1023 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
1024 // Return the expression with the addrec on the outside.
1025 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1026 getZeroExtendExpr(Step, Ty),
1027 L, AR->getNoWrapFlags());
1029 } else if (isKnownNegative(Step)) {
1030 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
1031 getSignedRange(Step).getSignedMin());
1032 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
1033 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
1034 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
1035 AR->getPostIncExpr(*this), N))) {
1036 // Cache knowledge of AR NW, which is propagated to this AddRec.
1037 // Negative step causes unsigned wrap, but it still can't self-wrap.
1038 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1039 // Return the expression with the addrec on the outside.
1040 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1041 getSignExtendExpr(Step, Ty),
1042 L, AR->getNoWrapFlags());
1048 // The cast wasn't folded; create an explicit cast node.
1049 // Recompute the insert position, as it may have been invalidated.
1050 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1051 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1053 UniqueSCEVs.InsertNode(S, IP);
1057 // Get the limit of a recurrence such that incrementing by Step cannot cause
1058 // signed overflow as long as the value of the recurrence within the loop does
1059 // not exceed this limit before incrementing.
1060 static const SCEV *getOverflowLimitForStep(const SCEV *Step,
1061 ICmpInst::Predicate *Pred,
1062 ScalarEvolution *SE) {
1063 unsigned BitWidth = SE->getTypeSizeInBits(Step->getType());
1064 if (SE->isKnownPositive(Step)) {
1065 *Pred = ICmpInst::ICMP_SLT;
1066 return SE->getConstant(APInt::getSignedMinValue(BitWidth) -
1067 SE->getSignedRange(Step).getSignedMax());
1069 if (SE->isKnownNegative(Step)) {
1070 *Pred = ICmpInst::ICMP_SGT;
1071 return SE->getConstant(APInt::getSignedMaxValue(BitWidth) -
1072 SE->getSignedRange(Step).getSignedMin());
1077 // The recurrence AR has been shown to have no signed wrap. Typically, if we can
1078 // prove NSW for AR, then we can just as easily prove NSW for its preincrement
1079 // or postincrement sibling. This allows normalizing a sign extended AddRec as
1080 // such: {sext(Step + Start),+,Step} => {(Step + sext(Start),+,Step} As a
1081 // result, the expression "Step + sext(PreIncAR)" is congruent with
1082 // "sext(PostIncAR)"
1083 static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR,
1085 ScalarEvolution *SE) {
1086 const Loop *L = AR->getLoop();
1087 const SCEV *Start = AR->getStart();
1088 const SCEV *Step = AR->getStepRecurrence(*SE);
1090 // Check for a simple looking step prior to loop entry.
1091 const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start);
1095 // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV
1096 // subtraction is expensive. For this purpose, perform a quick and dirty
1097 // difference, by checking for Step in the operand list.
1098 SmallVector<const SCEV *, 4> DiffOps;
1099 for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end();
1102 DiffOps.push_back(*I);
1104 if (DiffOps.size() == SA->getNumOperands())
1107 // This is a postinc AR. Check for overflow on the preinc recurrence using the
1108 // same three conditions that getSignExtendedExpr checks.
1110 // 1. NSW flags on the step increment.
1111 const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags());
1112 const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>(
1113 SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap));
1115 if (PreAR && PreAR->getNoWrapFlags(SCEV::FlagNSW))
1118 // 2. Direct overflow check on the step operation's expression.
1119 unsigned BitWidth = SE->getTypeSizeInBits(AR->getType());
1120 Type *WideTy = IntegerType::get(SE->getContext(), BitWidth * 2);
1121 const SCEV *OperandExtendedStart =
1122 SE->getAddExpr(SE->getSignExtendExpr(PreStart, WideTy),
1123 SE->getSignExtendExpr(Step, WideTy));
1124 if (SE->getSignExtendExpr(Start, WideTy) == OperandExtendedStart) {
1125 // Cache knowledge of PreAR NSW.
1127 const_cast<SCEVAddRecExpr *>(PreAR)->setNoWrapFlags(SCEV::FlagNSW);
1128 // FIXME: this optimization needs a unit test
1129 DEBUG(dbgs() << "SCEV: untested prestart overflow check\n");
1133 // 3. Loop precondition.
1134 ICmpInst::Predicate Pred;
1135 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, SE);
1137 if (OverflowLimit &&
1138 SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) {
1144 // Get the normalized sign-extended expression for this AddRec's Start.
1145 static const SCEV *getSignExtendAddRecStart(const SCEVAddRecExpr *AR,
1147 ScalarEvolution *SE) {
1148 const SCEV *PreStart = getPreStartForSignExtend(AR, Ty, SE);
1150 return SE->getSignExtendExpr(AR->getStart(), Ty);
1152 return SE->getAddExpr(SE->getSignExtendExpr(AR->getStepRecurrence(*SE), Ty),
1153 SE->getSignExtendExpr(PreStart, Ty));
1156 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1158 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1159 "This is not an extending conversion!");
1160 assert(isSCEVable(Ty) &&
1161 "This is not a conversion to a SCEVable type!");
1162 Ty = getEffectiveSCEVType(Ty);
1164 // Fold if the operand is constant.
1165 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1167 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(), Ty)));
1169 // sext(sext(x)) --> sext(x)
1170 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1171 return getSignExtendExpr(SS->getOperand(), Ty);
1173 // sext(zext(x)) --> zext(x)
1174 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
1175 return getZeroExtendExpr(SZ->getOperand(), Ty);
1177 // Before doing any expensive analysis, check to see if we've already
1178 // computed a SCEV for this Op and Ty.
1179 FoldingSetNodeID ID;
1180 ID.AddInteger(scSignExtend);
1184 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1186 // If the input value is provably positive, build a zext instead.
1187 if (isKnownNonNegative(Op))
1188 return getZeroExtendExpr(Op, Ty);
1190 // sext(trunc(x)) --> sext(x) or x or trunc(x)
1191 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
1192 // It's possible the bits taken off by the truncate were all sign bits. If
1193 // so, we should be able to simplify this further.
1194 const SCEV *X = ST->getOperand();
1195 ConstantRange CR = getSignedRange(X);
1196 unsigned TruncBits = getTypeSizeInBits(ST->getType());
1197 unsigned NewBits = getTypeSizeInBits(Ty);
1198 if (CR.truncate(TruncBits).signExtend(NewBits).contains(
1199 CR.sextOrTrunc(NewBits)))
1200 return getTruncateOrSignExtend(X, Ty);
1203 // If the input value is a chrec scev, and we can prove that the value
1204 // did not overflow the old, smaller, value, we can sign extend all of the
1205 // operands (often constants). This allows analysis of something like
1206 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1207 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1208 if (AR->isAffine()) {
1209 const SCEV *Start = AR->getStart();
1210 const SCEV *Step = AR->getStepRecurrence(*this);
1211 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1212 const Loop *L = AR->getLoop();
1214 // If we have special knowledge that this addrec won't overflow,
1215 // we don't need to do any further analysis.
1216 if (AR->getNoWrapFlags(SCEV::FlagNSW))
1217 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1218 getSignExtendExpr(Step, Ty),
1221 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1222 // Note that this serves two purposes: It filters out loops that are
1223 // simply not analyzable, and it covers the case where this code is
1224 // being called from within backedge-taken count analysis, such that
1225 // attempting to ask for the backedge-taken count would likely result
1226 // in infinite recursion. In the later case, the analysis code will
1227 // cope with a conservative value, and it will take care to purge
1228 // that value once it has finished.
1229 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1230 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1231 // Manually compute the final value for AR, checking for
1234 // Check whether the backedge-taken count can be losslessly casted to
1235 // the addrec's type. The count is always unsigned.
1236 const SCEV *CastedMaxBECount =
1237 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1238 const SCEV *RecastedMaxBECount =
1239 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1240 if (MaxBECount == RecastedMaxBECount) {
1241 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1242 // Check whether Start+Step*MaxBECount has no signed overflow.
1243 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1244 const SCEV *SAdd = getSignExtendExpr(getAddExpr(Start, SMul), WideTy);
1245 const SCEV *WideStart = getSignExtendExpr(Start, WideTy);
1246 const SCEV *WideMaxBECount =
1247 getZeroExtendExpr(CastedMaxBECount, WideTy);
1248 const SCEV *OperandExtendedAdd =
1249 getAddExpr(WideStart,
1250 getMulExpr(WideMaxBECount,
1251 getSignExtendExpr(Step, WideTy)));
1252 if (SAdd == OperandExtendedAdd) {
1253 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1254 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1255 // Return the expression with the addrec on the outside.
1256 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1257 getSignExtendExpr(Step, Ty),
1258 L, AR->getNoWrapFlags());
1260 // Similar to above, only this time treat the step value as unsigned.
1261 // This covers loops that count up with an unsigned step.
1262 OperandExtendedAdd =
1263 getAddExpr(WideStart,
1264 getMulExpr(WideMaxBECount,
1265 getZeroExtendExpr(Step, WideTy)));
1266 if (SAdd == OperandExtendedAdd) {
1267 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1268 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1269 // Return the expression with the addrec on the outside.
1270 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1271 getZeroExtendExpr(Step, Ty),
1272 L, AR->getNoWrapFlags());
1276 // If the backedge is guarded by a comparison with the pre-inc value
1277 // the addrec is safe. Also, if the entry is guarded by a comparison
1278 // with the start value and the backedge is guarded by a comparison
1279 // with the post-inc value, the addrec is safe.
1280 ICmpInst::Predicate Pred;
1281 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, this);
1282 if (OverflowLimit &&
1283 (isLoopBackedgeGuardedByCond(L, Pred, AR, OverflowLimit) ||
1284 (isLoopEntryGuardedByCond(L, Pred, Start, OverflowLimit) &&
1285 isLoopBackedgeGuardedByCond(L, Pred, AR->getPostIncExpr(*this),
1287 // Cache knowledge of AR NSW, then propagate NSW to the wide AddRec.
1288 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1289 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1290 getSignExtendExpr(Step, Ty),
1291 L, AR->getNoWrapFlags());
1296 // The cast wasn't folded; create an explicit cast node.
1297 // Recompute the insert position, as it may have been invalidated.
1298 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1299 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1301 UniqueSCEVs.InsertNode(S, IP);
1305 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1306 /// unspecified bits out to the given type.
1308 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1310 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1311 "This is not an extending conversion!");
1312 assert(isSCEVable(Ty) &&
1313 "This is not a conversion to a SCEVable type!");
1314 Ty = getEffectiveSCEVType(Ty);
1316 // Sign-extend negative constants.
1317 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1318 if (SC->getValue()->getValue().isNegative())
1319 return getSignExtendExpr(Op, Ty);
1321 // Peel off a truncate cast.
1322 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1323 const SCEV *NewOp = T->getOperand();
1324 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1325 return getAnyExtendExpr(NewOp, Ty);
1326 return getTruncateOrNoop(NewOp, Ty);
1329 // Next try a zext cast. If the cast is folded, use it.
1330 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1331 if (!isa<SCEVZeroExtendExpr>(ZExt))
1334 // Next try a sext cast. If the cast is folded, use it.
1335 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1336 if (!isa<SCEVSignExtendExpr>(SExt))
1339 // Force the cast to be folded into the operands of an addrec.
1340 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1341 SmallVector<const SCEV *, 4> Ops;
1342 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1344 Ops.push_back(getAnyExtendExpr(*I, Ty));
1345 return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW);
1348 // If the expression is obviously signed, use the sext cast value.
1349 if (isa<SCEVSMaxExpr>(Op))
1352 // Absent any other information, use the zext cast value.
1356 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1357 /// a list of operands to be added under the given scale, update the given
1358 /// map. This is a helper function for getAddRecExpr. As an example of
1359 /// what it does, given a sequence of operands that would form an add
1360 /// expression like this:
1362 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1364 /// where A and B are constants, update the map with these values:
1366 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1368 /// and add 13 + A*B*29 to AccumulatedConstant.
1369 /// This will allow getAddRecExpr to produce this:
1371 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1373 /// This form often exposes folding opportunities that are hidden in
1374 /// the original operand list.
1376 /// Return true iff it appears that any interesting folding opportunities
1377 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1378 /// the common case where no interesting opportunities are present, and
1379 /// is also used as a check to avoid infinite recursion.
1382 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1383 SmallVector<const SCEV *, 8> &NewOps,
1384 APInt &AccumulatedConstant,
1385 const SCEV *const *Ops, size_t NumOperands,
1387 ScalarEvolution &SE) {
1388 bool Interesting = false;
1390 // Iterate over the add operands. They are sorted, with constants first.
1392 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1394 // Pull a buried constant out to the outside.
1395 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1397 AccumulatedConstant += Scale * C->getValue()->getValue();
1400 // Next comes everything else. We're especially interested in multiplies
1401 // here, but they're in the middle, so just visit the rest with one loop.
1402 for (; i != NumOperands; ++i) {
1403 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1404 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1406 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1407 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1408 // A multiplication of a constant with another add; recurse.
1409 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1411 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1412 Add->op_begin(), Add->getNumOperands(),
1415 // A multiplication of a constant with some other value. Update
1417 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1418 const SCEV *Key = SE.getMulExpr(MulOps);
1419 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1420 M.insert(std::make_pair(Key, NewScale));
1422 NewOps.push_back(Pair.first->first);
1424 Pair.first->second += NewScale;
1425 // The map already had an entry for this value, which may indicate
1426 // a folding opportunity.
1431 // An ordinary operand. Update the map.
1432 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1433 M.insert(std::make_pair(Ops[i], Scale));
1435 NewOps.push_back(Pair.first->first);
1437 Pair.first->second += Scale;
1438 // The map already had an entry for this value, which may indicate
1439 // a folding opportunity.
1449 struct APIntCompare {
1450 bool operator()(const APInt &LHS, const APInt &RHS) const {
1451 return LHS.ult(RHS);
1456 /// getAddExpr - Get a canonical add expression, or something simpler if
1458 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1459 SCEV::NoWrapFlags Flags) {
1460 assert(!(Flags & ~(SCEV::FlagNUW | SCEV::FlagNSW)) &&
1461 "only nuw or nsw allowed");
1462 assert(!Ops.empty() && "Cannot get empty add!");
1463 if (Ops.size() == 1) return Ops[0];
1465 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1466 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1467 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1468 "SCEVAddExpr operand types don't match!");
1471 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1473 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1474 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1475 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1477 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1478 E = Ops.end(); I != E; ++I)
1479 if (!isKnownNonNegative(*I)) {
1483 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1486 // Sort by complexity, this groups all similar expression types together.
1487 GroupByComplexity(Ops, LI);
1489 // If there are any constants, fold them together.
1491 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1493 assert(Idx < Ops.size());
1494 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1495 // We found two constants, fold them together!
1496 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1497 RHSC->getValue()->getValue());
1498 if (Ops.size() == 2) return Ops[0];
1499 Ops.erase(Ops.begin()+1); // Erase the folded element
1500 LHSC = cast<SCEVConstant>(Ops[0]);
1503 // If we are left with a constant zero being added, strip it off.
1504 if (LHSC->getValue()->isZero()) {
1505 Ops.erase(Ops.begin());
1509 if (Ops.size() == 1) return Ops[0];
1512 // Okay, check to see if the same value occurs in the operand list more than
1513 // once. If so, merge them together into an multiply expression. Since we
1514 // sorted the list, these values are required to be adjacent.
1515 Type *Ty = Ops[0]->getType();
1516 bool FoundMatch = false;
1517 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1518 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1519 // Scan ahead to count how many equal operands there are.
1521 while (i+Count != e && Ops[i+Count] == Ops[i])
1523 // Merge the values into a multiply.
1524 const SCEV *Scale = getConstant(Ty, Count);
1525 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1526 if (Ops.size() == Count)
1529 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1530 --i; e -= Count - 1;
1534 return getAddExpr(Ops, Flags);
1536 // Check for truncates. If all the operands are truncated from the same
1537 // type, see if factoring out the truncate would permit the result to be
1538 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1539 // if the contents of the resulting outer trunc fold to something simple.
1540 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1541 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1542 Type *DstType = Trunc->getType();
1543 Type *SrcType = Trunc->getOperand()->getType();
1544 SmallVector<const SCEV *, 8> LargeOps;
1546 // Check all the operands to see if they can be represented in the
1547 // source type of the truncate.
1548 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1549 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1550 if (T->getOperand()->getType() != SrcType) {
1554 LargeOps.push_back(T->getOperand());
1555 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1556 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1557 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1558 SmallVector<const SCEV *, 8> LargeMulOps;
1559 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1560 if (const SCEVTruncateExpr *T =
1561 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1562 if (T->getOperand()->getType() != SrcType) {
1566 LargeMulOps.push_back(T->getOperand());
1567 } else if (const SCEVConstant *C =
1568 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1569 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1576 LargeOps.push_back(getMulExpr(LargeMulOps));
1583 // Evaluate the expression in the larger type.
1584 const SCEV *Fold = getAddExpr(LargeOps, Flags);
1585 // If it folds to something simple, use it. Otherwise, don't.
1586 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1587 return getTruncateExpr(Fold, DstType);
1591 // Skip past any other cast SCEVs.
1592 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1595 // If there are add operands they would be next.
1596 if (Idx < Ops.size()) {
1597 bool DeletedAdd = false;
1598 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1599 // If we have an add, expand the add operands onto the end of the operands
1601 Ops.erase(Ops.begin()+Idx);
1602 Ops.append(Add->op_begin(), Add->op_end());
1606 // If we deleted at least one add, we added operands to the end of the list,
1607 // and they are not necessarily sorted. Recurse to resort and resimplify
1608 // any operands we just acquired.
1610 return getAddExpr(Ops);
1613 // Skip over the add expression until we get to a multiply.
1614 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1617 // Check to see if there are any folding opportunities present with
1618 // operands multiplied by constant values.
1619 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1620 uint64_t BitWidth = getTypeSizeInBits(Ty);
1621 DenseMap<const SCEV *, APInt> M;
1622 SmallVector<const SCEV *, 8> NewOps;
1623 APInt AccumulatedConstant(BitWidth, 0);
1624 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1625 Ops.data(), Ops.size(),
1626 APInt(BitWidth, 1), *this)) {
1627 // Some interesting folding opportunity is present, so its worthwhile to
1628 // re-generate the operands list. Group the operands by constant scale,
1629 // to avoid multiplying by the same constant scale multiple times.
1630 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1631 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1632 E = NewOps.end(); I != E; ++I)
1633 MulOpLists[M.find(*I)->second].push_back(*I);
1634 // Re-generate the operands list.
1636 if (AccumulatedConstant != 0)
1637 Ops.push_back(getConstant(AccumulatedConstant));
1638 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1639 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1641 Ops.push_back(getMulExpr(getConstant(I->first),
1642 getAddExpr(I->second)));
1644 return getConstant(Ty, 0);
1645 if (Ops.size() == 1)
1647 return getAddExpr(Ops);
1651 // If we are adding something to a multiply expression, make sure the
1652 // something is not already an operand of the multiply. If so, merge it into
1654 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1655 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1656 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1657 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1658 if (isa<SCEVConstant>(MulOpSCEV))
1660 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1661 if (MulOpSCEV == Ops[AddOp]) {
1662 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1663 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1664 if (Mul->getNumOperands() != 2) {
1665 // If the multiply has more than two operands, we must get the
1667 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1668 Mul->op_begin()+MulOp);
1669 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1670 InnerMul = getMulExpr(MulOps);
1672 const SCEV *One = getConstant(Ty, 1);
1673 const SCEV *AddOne = getAddExpr(One, InnerMul);
1674 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1675 if (Ops.size() == 2) return OuterMul;
1677 Ops.erase(Ops.begin()+AddOp);
1678 Ops.erase(Ops.begin()+Idx-1);
1680 Ops.erase(Ops.begin()+Idx);
1681 Ops.erase(Ops.begin()+AddOp-1);
1683 Ops.push_back(OuterMul);
1684 return getAddExpr(Ops);
1687 // Check this multiply against other multiplies being added together.
1688 for (unsigned OtherMulIdx = Idx+1;
1689 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1691 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1692 // If MulOp occurs in OtherMul, we can fold the two multiplies
1694 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1695 OMulOp != e; ++OMulOp)
1696 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1697 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1698 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1699 if (Mul->getNumOperands() != 2) {
1700 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1701 Mul->op_begin()+MulOp);
1702 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1703 InnerMul1 = getMulExpr(MulOps);
1705 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1706 if (OtherMul->getNumOperands() != 2) {
1707 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1708 OtherMul->op_begin()+OMulOp);
1709 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1710 InnerMul2 = getMulExpr(MulOps);
1712 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1713 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1714 if (Ops.size() == 2) return OuterMul;
1715 Ops.erase(Ops.begin()+Idx);
1716 Ops.erase(Ops.begin()+OtherMulIdx-1);
1717 Ops.push_back(OuterMul);
1718 return getAddExpr(Ops);
1724 // If there are any add recurrences in the operands list, see if any other
1725 // added values are loop invariant. If so, we can fold them into the
1727 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1730 // Scan over all recurrences, trying to fold loop invariants into them.
1731 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1732 // Scan all of the other operands to this add and add them to the vector if
1733 // they are loop invariant w.r.t. the recurrence.
1734 SmallVector<const SCEV *, 8> LIOps;
1735 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1736 const Loop *AddRecLoop = AddRec->getLoop();
1737 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1738 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1739 LIOps.push_back(Ops[i]);
1740 Ops.erase(Ops.begin()+i);
1744 // If we found some loop invariants, fold them into the recurrence.
1745 if (!LIOps.empty()) {
1746 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1747 LIOps.push_back(AddRec->getStart());
1749 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1751 AddRecOps[0] = getAddExpr(LIOps);
1753 // Build the new addrec. Propagate the NUW and NSW flags if both the
1754 // outer add and the inner addrec are guaranteed to have no overflow.
1755 // Always propagate NW.
1756 Flags = AddRec->getNoWrapFlags(setFlags(Flags, SCEV::FlagNW));
1757 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop, Flags);
1759 // If all of the other operands were loop invariant, we are done.
1760 if (Ops.size() == 1) return NewRec;
1762 // Otherwise, add the folded AddRec by the non-invariant parts.
1763 for (unsigned i = 0;; ++i)
1764 if (Ops[i] == AddRec) {
1768 return getAddExpr(Ops);
1771 // Okay, if there weren't any loop invariants to be folded, check to see if
1772 // there are multiple AddRec's with the same loop induction variable being
1773 // added together. If so, we can fold them.
1774 for (unsigned OtherIdx = Idx+1;
1775 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1777 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1778 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1779 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1781 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1783 if (const SCEVAddRecExpr *OtherAddRec =
1784 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1785 if (OtherAddRec->getLoop() == AddRecLoop) {
1786 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1788 if (i >= AddRecOps.size()) {
1789 AddRecOps.append(OtherAddRec->op_begin()+i,
1790 OtherAddRec->op_end());
1793 AddRecOps[i] = getAddExpr(AddRecOps[i],
1794 OtherAddRec->getOperand(i));
1796 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1798 // Step size has changed, so we cannot guarantee no self-wraparound.
1799 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop, SCEV::FlagAnyWrap);
1800 return getAddExpr(Ops);
1803 // Otherwise couldn't fold anything into this recurrence. Move onto the
1807 // Okay, it looks like we really DO need an add expr. Check to see if we
1808 // already have one, otherwise create a new one.
1809 FoldingSetNodeID ID;
1810 ID.AddInteger(scAddExpr);
1811 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1812 ID.AddPointer(Ops[i]);
1815 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1817 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1818 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1819 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1821 UniqueSCEVs.InsertNode(S, IP);
1823 S->setNoWrapFlags(Flags);
1827 static uint64_t umul_ov(uint64_t i, uint64_t j, bool &Overflow) {
1829 if (j > 1 && k / j != i) Overflow = true;
1833 /// Compute the result of "n choose k", the binomial coefficient. If an
1834 /// intermediate computation overflows, Overflow will be set and the return will
1835 /// be garbage. Overflow is not cleared on absence of overflow.
1836 static uint64_t Choose(uint64_t n, uint64_t k, bool &Overflow) {
1837 // We use the multiplicative formula:
1838 // n(n-1)(n-2)...(n-(k-1)) / k(k-1)(k-2)...1 .
1839 // At each iteration, we take the n-th term of the numeral and divide by the
1840 // (k-n)th term of the denominator. This division will always produce an
1841 // integral result, and helps reduce the chance of overflow in the
1842 // intermediate computations. However, we can still overflow even when the
1843 // final result would fit.
1845 if (n == 0 || n == k) return 1;
1846 if (k > n) return 0;
1852 for (uint64_t i = 1; i <= k; ++i) {
1853 r = umul_ov(r, n-(i-1), Overflow);
1859 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1861 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1862 SCEV::NoWrapFlags Flags) {
1863 assert(Flags == maskFlags(Flags, SCEV::FlagNUW | SCEV::FlagNSW) &&
1864 "only nuw or nsw allowed");
1865 assert(!Ops.empty() && "Cannot get empty mul!");
1866 if (Ops.size() == 1) return Ops[0];
1868 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1869 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1870 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1871 "SCEVMulExpr operand types don't match!");
1874 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1876 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1877 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1878 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1880 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1881 E = Ops.end(); I != E; ++I)
1882 if (!isKnownNonNegative(*I)) {
1886 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1889 // Sort by complexity, this groups all similar expression types together.
1890 GroupByComplexity(Ops, LI);
1892 // If there are any constants, fold them together.
1894 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1896 // C1*(C2+V) -> C1*C2 + C1*V
1897 if (Ops.size() == 2)
1898 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1899 if (Add->getNumOperands() == 2 &&
1900 isa<SCEVConstant>(Add->getOperand(0)))
1901 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1902 getMulExpr(LHSC, Add->getOperand(1)));
1905 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1906 // We found two constants, fold them together!
1907 ConstantInt *Fold = ConstantInt::get(getContext(),
1908 LHSC->getValue()->getValue() *
1909 RHSC->getValue()->getValue());
1910 Ops[0] = getConstant(Fold);
1911 Ops.erase(Ops.begin()+1); // Erase the folded element
1912 if (Ops.size() == 1) return Ops[0];
1913 LHSC = cast<SCEVConstant>(Ops[0]);
1916 // If we are left with a constant one being multiplied, strip it off.
1917 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1918 Ops.erase(Ops.begin());
1920 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1921 // If we have a multiply of zero, it will always be zero.
1923 } else if (Ops[0]->isAllOnesValue()) {
1924 // If we have a mul by -1 of an add, try distributing the -1 among the
1926 if (Ops.size() == 2) {
1927 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1928 SmallVector<const SCEV *, 4> NewOps;
1929 bool AnyFolded = false;
1930 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(),
1931 E = Add->op_end(); I != E; ++I) {
1932 const SCEV *Mul = getMulExpr(Ops[0], *I);
1933 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1934 NewOps.push_back(Mul);
1937 return getAddExpr(NewOps);
1939 else if (const SCEVAddRecExpr *
1940 AddRec = dyn_cast<SCEVAddRecExpr>(Ops[1])) {
1941 // Negation preserves a recurrence's no self-wrap property.
1942 SmallVector<const SCEV *, 4> Operands;
1943 for (SCEVAddRecExpr::op_iterator I = AddRec->op_begin(),
1944 E = AddRec->op_end(); I != E; ++I) {
1945 Operands.push_back(getMulExpr(Ops[0], *I));
1947 return getAddRecExpr(Operands, AddRec->getLoop(),
1948 AddRec->getNoWrapFlags(SCEV::FlagNW));
1953 if (Ops.size() == 1)
1957 // Skip over the add expression until we get to a multiply.
1958 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1961 // If there are mul operands inline them all into this expression.
1962 if (Idx < Ops.size()) {
1963 bool DeletedMul = false;
1964 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1965 // If we have an mul, expand the mul operands onto the end of the operands
1967 Ops.erase(Ops.begin()+Idx);
1968 Ops.append(Mul->op_begin(), Mul->op_end());
1972 // If we deleted at least one mul, we added operands to the end of the list,
1973 // and they are not necessarily sorted. Recurse to resort and resimplify
1974 // any operands we just acquired.
1976 return getMulExpr(Ops);
1979 // If there are any add recurrences in the operands list, see if any other
1980 // added values are loop invariant. If so, we can fold them into the
1982 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1985 // Scan over all recurrences, trying to fold loop invariants into them.
1986 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1987 // Scan all of the other operands to this mul and add them to the vector if
1988 // they are loop invariant w.r.t. the recurrence.
1989 SmallVector<const SCEV *, 8> LIOps;
1990 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1991 const Loop *AddRecLoop = AddRec->getLoop();
1992 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1993 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1994 LIOps.push_back(Ops[i]);
1995 Ops.erase(Ops.begin()+i);
1999 // If we found some loop invariants, fold them into the recurrence.
2000 if (!LIOps.empty()) {
2001 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
2002 SmallVector<const SCEV *, 4> NewOps;
2003 NewOps.reserve(AddRec->getNumOperands());
2004 const SCEV *Scale = getMulExpr(LIOps);
2005 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
2006 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
2008 // Build the new addrec. Propagate the NUW and NSW flags if both the
2009 // outer mul and the inner addrec are guaranteed to have no overflow.
2011 // No self-wrap cannot be guaranteed after changing the step size, but
2012 // will be inferred if either NUW or NSW is true.
2013 Flags = AddRec->getNoWrapFlags(clearFlags(Flags, SCEV::FlagNW));
2014 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop, Flags);
2016 // If all of the other operands were loop invariant, we are done.
2017 if (Ops.size() == 1) return NewRec;
2019 // Otherwise, multiply the folded AddRec by the non-invariant parts.
2020 for (unsigned i = 0;; ++i)
2021 if (Ops[i] == AddRec) {
2025 return getMulExpr(Ops);
2028 // Okay, if there weren't any loop invariants to be folded, check to see if
2029 // there are multiple AddRec's with the same loop induction variable being
2030 // multiplied together. If so, we can fold them.
2031 for (unsigned OtherIdx = Idx+1;
2032 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2034 if (AddRecLoop != cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop())
2037 // {A1,+,A2,+,...,+,An}<L> * {B1,+,B2,+,...,+,Bn}<L>
2038 // = {x=1 in [ sum y=x..2x [ sum z=max(y-x, y-n)..min(x,n) [
2039 // choose(x, 2x)*choose(2x-y, x-z)*A_{y-z}*B_z
2040 // ]]],+,...up to x=2n}.
2041 // Note that the arguments to choose() are always integers with values
2042 // known at compile time, never SCEV objects.
2044 // The implementation avoids pointless extra computations when the two
2045 // addrec's are of different length (mathematically, it's equivalent to
2046 // an infinite stream of zeros on the right).
2047 bool OpsModified = false;
2048 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2050 const SCEVAddRecExpr *OtherAddRec =
2051 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]);
2052 if (!OtherAddRec || OtherAddRec->getLoop() != AddRecLoop)
2055 bool Overflow = false;
2056 Type *Ty = AddRec->getType();
2057 bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64;
2058 SmallVector<const SCEV*, 7> AddRecOps;
2059 for (int x = 0, xe = AddRec->getNumOperands() +
2060 OtherAddRec->getNumOperands() - 1; x != xe && !Overflow; ++x) {
2061 const SCEV *Term = getConstant(Ty, 0);
2062 for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) {
2063 uint64_t Coeff1 = Choose(x, 2*x - y, Overflow);
2064 for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1),
2065 ze = std::min(x+1, (int)OtherAddRec->getNumOperands());
2066 z < ze && !Overflow; ++z) {
2067 uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow);
2069 if (LargerThan64Bits)
2070 Coeff = umul_ov(Coeff1, Coeff2, Overflow);
2072 Coeff = Coeff1*Coeff2;
2073 const SCEV *CoeffTerm = getConstant(Ty, Coeff);
2074 const SCEV *Term1 = AddRec->getOperand(y-z);
2075 const SCEV *Term2 = OtherAddRec->getOperand(z);
2076 Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2));
2079 AddRecOps.push_back(Term);
2082 const SCEV *NewAddRec = getAddRecExpr(AddRecOps, AddRec->getLoop(),
2084 if (Ops.size() == 2) return NewAddRec;
2085 Ops[Idx] = NewAddRec;
2086 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
2088 AddRec = dyn_cast<SCEVAddRecExpr>(NewAddRec);
2094 return getMulExpr(Ops);
2097 // Otherwise couldn't fold anything into this recurrence. Move onto the
2101 // Okay, it looks like we really DO need an mul expr. Check to see if we
2102 // already have one, otherwise create a new one.
2103 FoldingSetNodeID ID;
2104 ID.AddInteger(scMulExpr);
2105 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2106 ID.AddPointer(Ops[i]);
2109 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2111 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2112 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2113 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
2115 UniqueSCEVs.InsertNode(S, IP);
2117 S->setNoWrapFlags(Flags);
2121 /// getUDivExpr - Get a canonical unsigned division expression, or something
2122 /// simpler if possible.
2123 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
2125 assert(getEffectiveSCEVType(LHS->getType()) ==
2126 getEffectiveSCEVType(RHS->getType()) &&
2127 "SCEVUDivExpr operand types don't match!");
2129 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
2130 if (RHSC->getValue()->equalsInt(1))
2131 return LHS; // X udiv 1 --> x
2132 // If the denominator is zero, the result of the udiv is undefined. Don't
2133 // try to analyze it, because the resolution chosen here may differ from
2134 // the resolution chosen in other parts of the compiler.
2135 if (!RHSC->getValue()->isZero()) {
2136 // Determine if the division can be folded into the operands of
2138 // TODO: Generalize this to non-constants by using known-bits information.
2139 Type *Ty = LHS->getType();
2140 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
2141 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
2142 // For non-power-of-two values, effectively round the value up to the
2143 // nearest power of two.
2144 if (!RHSC->getValue()->getValue().isPowerOf2())
2146 IntegerType *ExtTy =
2147 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
2148 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
2149 if (const SCEVConstant *Step =
2150 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this))) {
2151 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
2152 const APInt &StepInt = Step->getValue()->getValue();
2153 const APInt &DivInt = RHSC->getValue()->getValue();
2154 if (!StepInt.urem(DivInt) &&
2155 getZeroExtendExpr(AR, ExtTy) ==
2156 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2157 getZeroExtendExpr(Step, ExtTy),
2158 AR->getLoop(), SCEV::FlagAnyWrap)) {
2159 SmallVector<const SCEV *, 4> Operands;
2160 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
2161 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
2162 return getAddRecExpr(Operands, AR->getLoop(),
2165 /// Get a canonical UDivExpr for a recurrence.
2166 /// {X,+,N}/C => {Y,+,N}/C where Y=X-(X%N). Safe when C%N=0.
2167 // We can currently only fold X%N if X is constant.
2168 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
2169 if (StartC && !DivInt.urem(StepInt) &&
2170 getZeroExtendExpr(AR, ExtTy) ==
2171 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2172 getZeroExtendExpr(Step, ExtTy),
2173 AR->getLoop(), SCEV::FlagAnyWrap)) {
2174 const APInt &StartInt = StartC->getValue()->getValue();
2175 const APInt &StartRem = StartInt.urem(StepInt);
2177 LHS = getAddRecExpr(getConstant(StartInt - StartRem), Step,
2178 AR->getLoop(), SCEV::FlagNW);
2181 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
2182 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
2183 SmallVector<const SCEV *, 4> Operands;
2184 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
2185 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
2186 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
2187 // Find an operand that's safely divisible.
2188 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
2189 const SCEV *Op = M->getOperand(i);
2190 const SCEV *Div = getUDivExpr(Op, RHSC);
2191 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
2192 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
2195 return getMulExpr(Operands);
2199 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
2200 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(LHS)) {
2201 SmallVector<const SCEV *, 4> Operands;
2202 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
2203 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
2204 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
2206 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2207 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
2208 if (isa<SCEVUDivExpr>(Op) ||
2209 getMulExpr(Op, RHS) != A->getOperand(i))
2211 Operands.push_back(Op);
2213 if (Operands.size() == A->getNumOperands())
2214 return getAddExpr(Operands);
2218 // Fold if both operands are constant.
2219 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
2220 Constant *LHSCV = LHSC->getValue();
2221 Constant *RHSCV = RHSC->getValue();
2222 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
2228 FoldingSetNodeID ID;
2229 ID.AddInteger(scUDivExpr);
2233 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2234 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
2236 UniqueSCEVs.InsertNode(S, IP);
2241 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2242 /// Simplify the expression as much as possible.
2243 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start, const SCEV *Step,
2245 SCEV::NoWrapFlags Flags) {
2246 SmallVector<const SCEV *, 4> Operands;
2247 Operands.push_back(Start);
2248 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
2249 if (StepChrec->getLoop() == L) {
2250 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
2251 return getAddRecExpr(Operands, L, maskFlags(Flags, SCEV::FlagNW));
2254 Operands.push_back(Step);
2255 return getAddRecExpr(Operands, L, Flags);
2258 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2259 /// Simplify the expression as much as possible.
2261 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2262 const Loop *L, SCEV::NoWrapFlags Flags) {
2263 if (Operands.size() == 1) return Operands[0];
2265 Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2266 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2267 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2268 "SCEVAddRecExpr operand types don't match!");
2269 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2270 assert(isLoopInvariant(Operands[i], L) &&
2271 "SCEVAddRecExpr operand is not loop-invariant!");
2274 if (Operands.back()->isZero()) {
2275 Operands.pop_back();
2276 return getAddRecExpr(Operands, L, SCEV::FlagAnyWrap); // {X,+,0} --> X
2279 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2280 // use that information to infer NUW and NSW flags. However, computing a
2281 // BE count requires calling getAddRecExpr, so we may not yet have a
2282 // meaningful BE count at this point (and if we don't, we'd be stuck
2283 // with a SCEVCouldNotCompute as the cached BE count).
2285 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
2287 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
2288 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
2289 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
2291 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2292 E = Operands.end(); I != E; ++I)
2293 if (!isKnownNonNegative(*I)) {
2297 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
2300 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2301 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2302 const Loop *NestedLoop = NestedAR->getLoop();
2303 if (L->contains(NestedLoop) ?
2304 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2305 (!NestedLoop->contains(L) &&
2306 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2307 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2308 NestedAR->op_end());
2309 Operands[0] = NestedAR->getStart();
2310 // AddRecs require their operands be loop-invariant with respect to their
2311 // loops. Don't perform this transformation if it would break this
2313 bool AllInvariant = true;
2314 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2315 if (!isLoopInvariant(Operands[i], L)) {
2316 AllInvariant = false;
2320 // Create a recurrence for the outer loop with the same step size.
2322 // The outer recurrence keeps its NW flag but only keeps NUW/NSW if the
2323 // inner recurrence has the same property.
2324 SCEV::NoWrapFlags OuterFlags =
2325 maskFlags(Flags, SCEV::FlagNW | NestedAR->getNoWrapFlags());
2327 NestedOperands[0] = getAddRecExpr(Operands, L, OuterFlags);
2328 AllInvariant = true;
2329 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2330 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2331 AllInvariant = false;
2335 // Ok, both add recurrences are valid after the transformation.
2337 // The inner recurrence keeps its NW flag but only keeps NUW/NSW if
2338 // the outer recurrence has the same property.
2339 SCEV::NoWrapFlags InnerFlags =
2340 maskFlags(NestedAR->getNoWrapFlags(), SCEV::FlagNW | Flags);
2341 return getAddRecExpr(NestedOperands, NestedLoop, InnerFlags);
2344 // Reset Operands to its original state.
2345 Operands[0] = NestedAR;
2349 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2350 // already have one, otherwise create a new one.
2351 FoldingSetNodeID ID;
2352 ID.AddInteger(scAddRecExpr);
2353 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2354 ID.AddPointer(Operands[i]);
2358 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2360 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2361 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2362 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2363 O, Operands.size(), L);
2364 UniqueSCEVs.InsertNode(S, IP);
2366 S->setNoWrapFlags(Flags);
2370 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2372 SmallVector<const SCEV *, 2> Ops;
2375 return getSMaxExpr(Ops);
2379 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2380 assert(!Ops.empty() && "Cannot get empty smax!");
2381 if (Ops.size() == 1) return Ops[0];
2383 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2384 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2385 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2386 "SCEVSMaxExpr operand types don't match!");
2389 // Sort by complexity, this groups all similar expression types together.
2390 GroupByComplexity(Ops, LI);
2392 // If there are any constants, fold them together.
2394 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2396 assert(Idx < Ops.size());
2397 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2398 // We found two constants, fold them together!
2399 ConstantInt *Fold = ConstantInt::get(getContext(),
2400 APIntOps::smax(LHSC->getValue()->getValue(),
2401 RHSC->getValue()->getValue()));
2402 Ops[0] = getConstant(Fold);
2403 Ops.erase(Ops.begin()+1); // Erase the folded element
2404 if (Ops.size() == 1) return Ops[0];
2405 LHSC = cast<SCEVConstant>(Ops[0]);
2408 // If we are left with a constant minimum-int, strip it off.
2409 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2410 Ops.erase(Ops.begin());
2412 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2413 // If we have an smax with a constant maximum-int, it will always be
2418 if (Ops.size() == 1) return Ops[0];
2421 // Find the first SMax
2422 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2425 // Check to see if one of the operands is an SMax. If so, expand its operands
2426 // onto our operand list, and recurse to simplify.
2427 if (Idx < Ops.size()) {
2428 bool DeletedSMax = false;
2429 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2430 Ops.erase(Ops.begin()+Idx);
2431 Ops.append(SMax->op_begin(), SMax->op_end());
2436 return getSMaxExpr(Ops);
2439 // Okay, check to see if the same value occurs in the operand list twice. If
2440 // so, delete one. Since we sorted the list, these values are required to
2442 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2443 // X smax Y smax Y --> X smax Y
2444 // X smax Y --> X, if X is always greater than Y
2445 if (Ops[i] == Ops[i+1] ||
2446 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2447 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2449 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2450 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2454 if (Ops.size() == 1) return Ops[0];
2456 assert(!Ops.empty() && "Reduced smax down to nothing!");
2458 // Okay, it looks like we really DO need an smax expr. Check to see if we
2459 // already have one, otherwise create a new one.
2460 FoldingSetNodeID ID;
2461 ID.AddInteger(scSMaxExpr);
2462 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2463 ID.AddPointer(Ops[i]);
2465 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2466 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2467 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2468 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2470 UniqueSCEVs.InsertNode(S, IP);
2474 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2476 SmallVector<const SCEV *, 2> Ops;
2479 return getUMaxExpr(Ops);
2483 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2484 assert(!Ops.empty() && "Cannot get empty umax!");
2485 if (Ops.size() == 1) return Ops[0];
2487 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2488 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2489 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2490 "SCEVUMaxExpr operand types don't match!");
2493 // Sort by complexity, this groups all similar expression types together.
2494 GroupByComplexity(Ops, LI);
2496 // If there are any constants, fold them together.
2498 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2500 assert(Idx < Ops.size());
2501 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2502 // We found two constants, fold them together!
2503 ConstantInt *Fold = ConstantInt::get(getContext(),
2504 APIntOps::umax(LHSC->getValue()->getValue(),
2505 RHSC->getValue()->getValue()));
2506 Ops[0] = getConstant(Fold);
2507 Ops.erase(Ops.begin()+1); // Erase the folded element
2508 if (Ops.size() == 1) return Ops[0];
2509 LHSC = cast<SCEVConstant>(Ops[0]);
2512 // If we are left with a constant minimum-int, strip it off.
2513 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2514 Ops.erase(Ops.begin());
2516 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2517 // If we have an umax with a constant maximum-int, it will always be
2522 if (Ops.size() == 1) return Ops[0];
2525 // Find the first UMax
2526 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2529 // Check to see if one of the operands is a UMax. If so, expand its operands
2530 // onto our operand list, and recurse to simplify.
2531 if (Idx < Ops.size()) {
2532 bool DeletedUMax = false;
2533 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2534 Ops.erase(Ops.begin()+Idx);
2535 Ops.append(UMax->op_begin(), UMax->op_end());
2540 return getUMaxExpr(Ops);
2543 // Okay, check to see if the same value occurs in the operand list twice. If
2544 // so, delete one. Since we sorted the list, these values are required to
2546 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2547 // X umax Y umax Y --> X umax Y
2548 // X umax Y --> X, if X is always greater than Y
2549 if (Ops[i] == Ops[i+1] ||
2550 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2551 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2553 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2554 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2558 if (Ops.size() == 1) return Ops[0];
2560 assert(!Ops.empty() && "Reduced umax down to nothing!");
2562 // Okay, it looks like we really DO need a umax expr. Check to see if we
2563 // already have one, otherwise create a new one.
2564 FoldingSetNodeID ID;
2565 ID.AddInteger(scUMaxExpr);
2566 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2567 ID.AddPointer(Ops[i]);
2569 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2570 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2571 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2572 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2574 UniqueSCEVs.InsertNode(S, IP);
2578 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2580 // ~smax(~x, ~y) == smin(x, y).
2581 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2584 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2586 // ~umax(~x, ~y) == umin(x, y)
2587 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2590 const SCEV *ScalarEvolution::getSizeOfExpr(Type *AllocTy) {
2591 // If we have DataLayout, we can bypass creating a target-independent
2592 // constant expression and then folding it back into a ConstantInt.
2593 // This is just a compile-time optimization.
2595 return getConstant(TD->getIntPtrType(getContext()),
2596 TD->getTypeAllocSize(AllocTy));
2598 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2599 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2600 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2602 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2603 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2606 const SCEV *ScalarEvolution::getAlignOfExpr(Type *AllocTy) {
2607 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2608 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2609 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2611 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2612 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2615 const SCEV *ScalarEvolution::getOffsetOfExpr(StructType *STy,
2617 // If we have DataLayout, we can bypass creating a target-independent
2618 // constant expression and then folding it back into a ConstantInt.
2619 // This is just a compile-time optimization.
2621 return getConstant(TD->getIntPtrType(getContext()),
2622 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2624 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2625 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2626 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2628 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2629 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2632 const SCEV *ScalarEvolution::getOffsetOfExpr(Type *CTy,
2633 Constant *FieldNo) {
2634 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2635 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2636 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
2638 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2639 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2642 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2643 // Don't attempt to do anything other than create a SCEVUnknown object
2644 // here. createSCEV only calls getUnknown after checking for all other
2645 // interesting possibilities, and any other code that calls getUnknown
2646 // is doing so in order to hide a value from SCEV canonicalization.
2648 FoldingSetNodeID ID;
2649 ID.AddInteger(scUnknown);
2652 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2653 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2654 "Stale SCEVUnknown in uniquing map!");
2657 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2659 FirstUnknown = cast<SCEVUnknown>(S);
2660 UniqueSCEVs.InsertNode(S, IP);
2664 //===----------------------------------------------------------------------===//
2665 // Basic SCEV Analysis and PHI Idiom Recognition Code
2668 /// isSCEVable - Test if values of the given type are analyzable within
2669 /// the SCEV framework. This primarily includes integer types, and it
2670 /// can optionally include pointer types if the ScalarEvolution class
2671 /// has access to target-specific information.
2672 bool ScalarEvolution::isSCEVable(Type *Ty) const {
2673 // Integers and pointers are always SCEVable.
2674 return Ty->isIntegerTy() || Ty->isPointerTy();
2677 /// getTypeSizeInBits - Return the size in bits of the specified type,
2678 /// for which isSCEVable must return true.
2679 uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const {
2680 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2682 // If we have a DataLayout, use it!
2684 return TD->getTypeSizeInBits(Ty);
2686 // Integer types have fixed sizes.
2687 if (Ty->isIntegerTy())
2688 return Ty->getPrimitiveSizeInBits();
2690 // The only other support type is pointer. Without DataLayout, conservatively
2691 // assume pointers are 64-bit.
2692 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2696 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2697 /// the given type and which represents how SCEV will treat the given
2698 /// type, for which isSCEVable must return true. For pointer types,
2699 /// this is the pointer-sized integer type.
2700 Type *ScalarEvolution::getEffectiveSCEVType(Type *Ty) const {
2701 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2703 if (Ty->isIntegerTy())
2706 // The only other support type is pointer.
2707 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2708 if (TD) return TD->getIntPtrType(getContext());
2710 // Without DataLayout, conservatively assume pointers are 64-bit.
2711 return Type::getInt64Ty(getContext());
2714 const SCEV *ScalarEvolution::getCouldNotCompute() {
2715 return &CouldNotCompute;
2718 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2719 /// expression and create a new one.
2720 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2721 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2723 ValueExprMapType::const_iterator I = ValueExprMap.find_as(V);
2724 if (I != ValueExprMap.end()) return I->second;
2725 const SCEV *S = createSCEV(V);
2727 // The process of creating a SCEV for V may have caused other SCEVs
2728 // to have been created, so it's necessary to insert the new entry
2729 // from scratch, rather than trying to remember the insert position
2731 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2735 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2737 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2738 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2740 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2742 Type *Ty = V->getType();
2743 Ty = getEffectiveSCEVType(Ty);
2744 return getMulExpr(V,
2745 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2748 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2749 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2750 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2752 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2754 Type *Ty = V->getType();
2755 Ty = getEffectiveSCEVType(Ty);
2756 const SCEV *AllOnes =
2757 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2758 return getMinusSCEV(AllOnes, V);
2761 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1.
2762 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2763 SCEV::NoWrapFlags Flags) {
2764 assert(!maskFlags(Flags, SCEV::FlagNUW) && "subtraction does not have NUW");
2766 // Fast path: X - X --> 0.
2768 return getConstant(LHS->getType(), 0);
2771 return getAddExpr(LHS, getNegativeSCEV(RHS), Flags);
2774 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2775 /// input value to the specified type. If the type must be extended, it is zero
2778 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, Type *Ty) {
2779 Type *SrcTy = V->getType();
2780 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2781 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2782 "Cannot truncate or zero extend with non-integer arguments!");
2783 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2784 return V; // No conversion
2785 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2786 return getTruncateExpr(V, Ty);
2787 return getZeroExtendExpr(V, Ty);
2790 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2791 /// input value to the specified type. If the type must be extended, it is sign
2794 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2796 Type *SrcTy = V->getType();
2797 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2798 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2799 "Cannot truncate or zero extend with non-integer arguments!");
2800 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2801 return V; // No conversion
2802 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2803 return getTruncateExpr(V, Ty);
2804 return getSignExtendExpr(V, Ty);
2807 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2808 /// input value to the specified type. If the type must be extended, it is zero
2809 /// extended. The conversion must not be narrowing.
2811 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, Type *Ty) {
2812 Type *SrcTy = V->getType();
2813 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2814 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2815 "Cannot noop or zero extend with non-integer arguments!");
2816 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2817 "getNoopOrZeroExtend cannot truncate!");
2818 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2819 return V; // No conversion
2820 return getZeroExtendExpr(V, Ty);
2823 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2824 /// input value to the specified type. If the type must be extended, it is sign
2825 /// extended. The conversion must not be narrowing.
2827 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, Type *Ty) {
2828 Type *SrcTy = V->getType();
2829 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2830 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2831 "Cannot noop or sign extend with non-integer arguments!");
2832 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2833 "getNoopOrSignExtend cannot truncate!");
2834 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2835 return V; // No conversion
2836 return getSignExtendExpr(V, Ty);
2839 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2840 /// the input value to the specified type. If the type must be extended,
2841 /// it is extended with unspecified bits. The conversion must not be
2844 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, Type *Ty) {
2845 Type *SrcTy = V->getType();
2846 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2847 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2848 "Cannot noop or any extend with non-integer arguments!");
2849 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2850 "getNoopOrAnyExtend cannot truncate!");
2851 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2852 return V; // No conversion
2853 return getAnyExtendExpr(V, Ty);
2856 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2857 /// input value to the specified type. The conversion must not be widening.
2859 ScalarEvolution::getTruncateOrNoop(const SCEV *V, Type *Ty) {
2860 Type *SrcTy = V->getType();
2861 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2862 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2863 "Cannot truncate or noop with non-integer arguments!");
2864 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2865 "getTruncateOrNoop cannot extend!");
2866 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2867 return V; // No conversion
2868 return getTruncateExpr(V, Ty);
2871 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2872 /// the types using zero-extension, and then perform a umax operation
2874 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2876 const SCEV *PromotedLHS = LHS;
2877 const SCEV *PromotedRHS = RHS;
2879 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2880 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2882 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2884 return getUMaxExpr(PromotedLHS, PromotedRHS);
2887 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2888 /// the types using zero-extension, and then perform a umin operation
2890 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2892 const SCEV *PromotedLHS = LHS;
2893 const SCEV *PromotedRHS = RHS;
2895 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2896 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2898 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2900 return getUMinExpr(PromotedLHS, PromotedRHS);
2903 /// getPointerBase - Transitively follow the chain of pointer-type operands
2904 /// until reaching a SCEV that does not have a single pointer operand. This
2905 /// returns a SCEVUnknown pointer for well-formed pointer-type expressions,
2906 /// but corner cases do exist.
2907 const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) {
2908 // A pointer operand may evaluate to a nonpointer expression, such as null.
2909 if (!V->getType()->isPointerTy())
2912 if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(V)) {
2913 return getPointerBase(Cast->getOperand());
2915 else if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(V)) {
2916 const SCEV *PtrOp = 0;
2917 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
2919 if ((*I)->getType()->isPointerTy()) {
2920 // Cannot find the base of an expression with multiple pointer operands.
2928 return getPointerBase(PtrOp);
2933 /// PushDefUseChildren - Push users of the given Instruction
2934 /// onto the given Worklist.
2936 PushDefUseChildren(Instruction *I,
2937 SmallVectorImpl<Instruction *> &Worklist) {
2938 // Push the def-use children onto the Worklist stack.
2939 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2941 Worklist.push_back(cast<Instruction>(*UI));
2944 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2945 /// instructions that depend on the given instruction and removes them from
2946 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2949 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2950 SmallVector<Instruction *, 16> Worklist;
2951 PushDefUseChildren(PN, Worklist);
2953 SmallPtrSet<Instruction *, 8> Visited;
2955 while (!Worklist.empty()) {
2956 Instruction *I = Worklist.pop_back_val();
2957 if (!Visited.insert(I)) continue;
2959 ValueExprMapType::iterator It =
2960 ValueExprMap.find_as(static_cast<Value *>(I));
2961 if (It != ValueExprMap.end()) {
2962 const SCEV *Old = It->second;
2964 // Short-circuit the def-use traversal if the symbolic name
2965 // ceases to appear in expressions.
2966 if (Old != SymName && !hasOperand(Old, SymName))
2969 // SCEVUnknown for a PHI either means that it has an unrecognized
2970 // structure, it's a PHI that's in the progress of being computed
2971 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2972 // additional loop trip count information isn't going to change anything.
2973 // In the second case, createNodeForPHI will perform the necessary
2974 // updates on its own when it gets to that point. In the third, we do
2975 // want to forget the SCEVUnknown.
2976 if (!isa<PHINode>(I) ||
2977 !isa<SCEVUnknown>(Old) ||
2978 (I != PN && Old == SymName)) {
2979 forgetMemoizedResults(Old);
2980 ValueExprMap.erase(It);
2984 PushDefUseChildren(I, Worklist);
2988 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2989 /// a loop header, making it a potential recurrence, or it doesn't.
2991 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2992 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2993 if (L->getHeader() == PN->getParent()) {
2994 // The loop may have multiple entrances or multiple exits; we can analyze
2995 // this phi as an addrec if it has a unique entry value and a unique
2997 Value *BEValueV = 0, *StartValueV = 0;
2998 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2999 Value *V = PN->getIncomingValue(i);
3000 if (L->contains(PN->getIncomingBlock(i))) {
3003 } else if (BEValueV != V) {
3007 } else if (!StartValueV) {
3009 } else if (StartValueV != V) {
3014 if (BEValueV && StartValueV) {
3015 // While we are analyzing this PHI node, handle its value symbolically.
3016 const SCEV *SymbolicName = getUnknown(PN);
3017 assert(ValueExprMap.find_as(PN) == ValueExprMap.end() &&
3018 "PHI node already processed?");
3019 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
3021 // Using this symbolic name for the PHI, analyze the value coming around
3023 const SCEV *BEValue = getSCEV(BEValueV);
3025 // NOTE: If BEValue is loop invariant, we know that the PHI node just
3026 // has a special value for the first iteration of the loop.
3028 // If the value coming around the backedge is an add with the symbolic
3029 // value we just inserted, then we found a simple induction variable!
3030 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
3031 // If there is a single occurrence of the symbolic value, replace it
3032 // with a recurrence.
3033 unsigned FoundIndex = Add->getNumOperands();
3034 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3035 if (Add->getOperand(i) == SymbolicName)
3036 if (FoundIndex == e) {
3041 if (FoundIndex != Add->getNumOperands()) {
3042 // Create an add with everything but the specified operand.
3043 SmallVector<const SCEV *, 8> Ops;
3044 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
3045 if (i != FoundIndex)
3046 Ops.push_back(Add->getOperand(i));
3047 const SCEV *Accum = getAddExpr(Ops);
3049 // This is not a valid addrec if the step amount is varying each
3050 // loop iteration, but is not itself an addrec in this loop.
3051 if (isLoopInvariant(Accum, L) ||
3052 (isa<SCEVAddRecExpr>(Accum) &&
3053 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
3054 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap;
3056 // If the increment doesn't overflow, then neither the addrec nor
3057 // the post-increment will overflow.
3058 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
3059 if (OBO->hasNoUnsignedWrap())
3060 Flags = setFlags(Flags, SCEV::FlagNUW);
3061 if (OBO->hasNoSignedWrap())
3062 Flags = setFlags(Flags, SCEV::FlagNSW);
3063 } else if (const GEPOperator *GEP =
3064 dyn_cast<GEPOperator>(BEValueV)) {
3065 // If the increment is an inbounds GEP, then we know the address
3066 // space cannot be wrapped around. We cannot make any guarantee
3067 // about signed or unsigned overflow because pointers are
3068 // unsigned but we may have a negative index from the base
3070 if (GEP->isInBounds())
3071 Flags = setFlags(Flags, SCEV::FlagNW);
3074 const SCEV *StartVal = getSCEV(StartValueV);
3075 const SCEV *PHISCEV = getAddRecExpr(StartVal, Accum, L, Flags);
3077 // Since the no-wrap flags are on the increment, they apply to the
3078 // post-incremented value as well.
3079 if (isLoopInvariant(Accum, L))
3080 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
3083 // Okay, for the entire analysis of this edge we assumed the PHI
3084 // to be symbolic. We now need to go back and purge all of the
3085 // entries for the scalars that use the symbolic expression.
3086 ForgetSymbolicName(PN, SymbolicName);
3087 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3091 } else if (const SCEVAddRecExpr *AddRec =
3092 dyn_cast<SCEVAddRecExpr>(BEValue)) {
3093 // Otherwise, this could be a loop like this:
3094 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
3095 // In this case, j = {1,+,1} and BEValue is j.
3096 // Because the other in-value of i (0) fits the evolution of BEValue
3097 // i really is an addrec evolution.
3098 if (AddRec->getLoop() == L && AddRec->isAffine()) {
3099 const SCEV *StartVal = getSCEV(StartValueV);
3101 // If StartVal = j.start - j.stride, we can use StartVal as the
3102 // initial step of the addrec evolution.
3103 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
3104 AddRec->getOperand(1))) {
3105 // FIXME: For constant StartVal, we should be able to infer
3107 const SCEV *PHISCEV =
3108 getAddRecExpr(StartVal, AddRec->getOperand(1), L,
3111 // Okay, for the entire analysis of this edge we assumed the PHI
3112 // to be symbolic. We now need to go back and purge all of the
3113 // entries for the scalars that use the symbolic expression.
3114 ForgetSymbolicName(PN, SymbolicName);
3115 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3123 // If the PHI has a single incoming value, follow that value, unless the
3124 // PHI's incoming blocks are in a different loop, in which case doing so
3125 // risks breaking LCSSA form. Instcombine would normally zap these, but
3126 // it doesn't have DominatorTree information, so it may miss cases.
3127 if (Value *V = SimplifyInstruction(PN, TD, TLI, DT))
3128 if (LI->replacementPreservesLCSSAForm(PN, V))
3131 // If it's not a loop phi, we can't handle it yet.
3132 return getUnknown(PN);
3135 /// createNodeForGEP - Expand GEP instructions into add and multiply
3136 /// operations. This allows them to be analyzed by regular SCEV code.
3138 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
3140 // Don't blindly transfer the inbounds flag from the GEP instruction to the
3141 // Add expression, because the Instruction may be guarded by control flow
3142 // and the no-overflow bits may not be valid for the expression in any
3144 bool isInBounds = GEP->isInBounds();
3146 Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
3147 Value *Base = GEP->getOperand(0);
3148 // Don't attempt to analyze GEPs over unsized objects.
3149 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
3150 return getUnknown(GEP);
3151 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
3152 gep_type_iterator GTI = gep_type_begin(GEP);
3153 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
3157 // Compute the (potentially symbolic) offset in bytes for this index.
3158 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
3159 // For a struct, add the member offset.
3160 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
3161 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
3163 // Add the field offset to the running total offset.
3164 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
3166 // For an array, add the element offset, explicitly scaled.
3167 const SCEV *ElementSize = getSizeOfExpr(*GTI);
3168 const SCEV *IndexS = getSCEV(Index);
3169 // Getelementptr indices are signed.
3170 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
3172 // Multiply the index by the element size to compute the element offset.
3173 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize,
3174 isInBounds ? SCEV::FlagNSW :
3177 // Add the element offset to the running total offset.
3178 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
3182 // Get the SCEV for the GEP base.
3183 const SCEV *BaseS = getSCEV(Base);
3185 // Add the total offset from all the GEP indices to the base.
3186 return getAddExpr(BaseS, TotalOffset,
3187 isInBounds ? SCEV::FlagNSW : SCEV::FlagAnyWrap);
3190 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
3191 /// guaranteed to end in (at every loop iteration). It is, at the same time,
3192 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
3193 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
3195 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
3196 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3197 return C->getValue()->getValue().countTrailingZeros();
3199 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
3200 return std::min(GetMinTrailingZeros(T->getOperand()),
3201 (uint32_t)getTypeSizeInBits(T->getType()));
3203 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
3204 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3205 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3206 getTypeSizeInBits(E->getType()) : OpRes;
3209 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
3210 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3211 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3212 getTypeSizeInBits(E->getType()) : OpRes;
3215 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
3216 // The result is the min of all operands results.
3217 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3218 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3219 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3223 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
3224 // The result is the sum of all operands results.
3225 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
3226 uint32_t BitWidth = getTypeSizeInBits(M->getType());
3227 for (unsigned i = 1, e = M->getNumOperands();
3228 SumOpRes != BitWidth && i != e; ++i)
3229 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
3234 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
3235 // The result is the min of all operands results.
3236 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3237 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3238 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3242 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
3243 // The result is the min of all operands results.
3244 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3245 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3246 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3250 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
3251 // The result is the min of all operands results.
3252 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3253 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3254 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3258 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3259 // For a SCEVUnknown, ask ValueTracking.
3260 unsigned BitWidth = getTypeSizeInBits(U->getType());
3261 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3262 ComputeMaskedBits(U->getValue(), Zeros, Ones);
3263 return Zeros.countTrailingOnes();
3270 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
3273 ScalarEvolution::getUnsignedRange(const SCEV *S) {
3274 // See if we've computed this range already.
3275 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
3276 if (I != UnsignedRanges.end())
3279 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3280 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
3282 unsigned BitWidth = getTypeSizeInBits(S->getType());
3283 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3285 // If the value has known zeros, the maximum unsigned value will have those
3286 // known zeros as well.
3287 uint32_t TZ = GetMinTrailingZeros(S);
3289 ConservativeResult =
3290 ConstantRange(APInt::getMinValue(BitWidth),
3291 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
3293 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3294 ConstantRange X = getUnsignedRange(Add->getOperand(0));
3295 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3296 X = X.add(getUnsignedRange(Add->getOperand(i)));
3297 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
3300 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3301 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
3302 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3303 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
3304 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
3307 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3308 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
3309 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3310 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
3311 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
3314 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3315 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3316 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3317 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3318 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
3321 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3322 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3323 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3324 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3327 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3328 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3329 return setUnsignedRange(ZExt,
3330 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3333 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3334 ConstantRange X = getUnsignedRange(SExt->getOperand());
3335 return setUnsignedRange(SExt,
3336 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3339 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3340 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3341 return setUnsignedRange(Trunc,
3342 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3345 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3346 // If there's no unsigned wrap, the value will never be less than its
3348 if (AddRec->getNoWrapFlags(SCEV::FlagNUW))
3349 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3350 if (!C->getValue()->isZero())
3351 ConservativeResult =
3352 ConservativeResult.intersectWith(
3353 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3355 // TODO: non-affine addrec
3356 if (AddRec->isAffine()) {
3357 Type *Ty = AddRec->getType();
3358 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3359 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3360 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3361 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3363 const SCEV *Start = AddRec->getStart();
3364 const SCEV *Step = AddRec->getStepRecurrence(*this);
3366 ConstantRange StartRange = getUnsignedRange(Start);
3367 ConstantRange StepRange = getSignedRange(Step);
3368 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3369 ConstantRange EndRange =
3370 StartRange.add(MaxBECountRange.multiply(StepRange));
3372 // Check for overflow. This must be done with ConstantRange arithmetic
3373 // because we could be called from within the ScalarEvolution overflow
3375 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3376 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3377 ConstantRange ExtMaxBECountRange =
3378 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3379 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3380 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3382 return setUnsignedRange(AddRec, ConservativeResult);
3384 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3385 EndRange.getUnsignedMin());
3386 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3387 EndRange.getUnsignedMax());
3388 if (Min.isMinValue() && Max.isMaxValue())
3389 return setUnsignedRange(AddRec, ConservativeResult);
3390 return setUnsignedRange(AddRec,
3391 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3395 return setUnsignedRange(AddRec, ConservativeResult);
3398 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3399 // For a SCEVUnknown, ask ValueTracking.
3400 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3401 ComputeMaskedBits(U->getValue(), Zeros, Ones, TD);
3402 if (Ones == ~Zeros + 1)
3403 return setUnsignedRange(U, ConservativeResult);
3404 return setUnsignedRange(U,
3405 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3408 return setUnsignedRange(S, ConservativeResult);
3411 /// getSignedRange - Determine the signed range for a particular SCEV.
3414 ScalarEvolution::getSignedRange(const SCEV *S) {
3415 // See if we've computed this range already.
3416 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3417 if (I != SignedRanges.end())
3420 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3421 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3423 unsigned BitWidth = getTypeSizeInBits(S->getType());
3424 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3426 // If the value has known zeros, the maximum signed value will have those
3427 // known zeros as well.
3428 uint32_t TZ = GetMinTrailingZeros(S);
3430 ConservativeResult =
3431 ConstantRange(APInt::getSignedMinValue(BitWidth),
3432 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3434 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3435 ConstantRange X = getSignedRange(Add->getOperand(0));
3436 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3437 X = X.add(getSignedRange(Add->getOperand(i)));
3438 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3441 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3442 ConstantRange X = getSignedRange(Mul->getOperand(0));
3443 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3444 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3445 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3448 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3449 ConstantRange X = getSignedRange(SMax->getOperand(0));
3450 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3451 X = X.smax(getSignedRange(SMax->getOperand(i)));
3452 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3455 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3456 ConstantRange X = getSignedRange(UMax->getOperand(0));
3457 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3458 X = X.umax(getSignedRange(UMax->getOperand(i)));
3459 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3462 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3463 ConstantRange X = getSignedRange(UDiv->getLHS());
3464 ConstantRange Y = getSignedRange(UDiv->getRHS());
3465 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3468 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3469 ConstantRange X = getSignedRange(ZExt->getOperand());
3470 return setSignedRange(ZExt,
3471 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3474 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3475 ConstantRange X = getSignedRange(SExt->getOperand());
3476 return setSignedRange(SExt,
3477 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3480 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3481 ConstantRange X = getSignedRange(Trunc->getOperand());
3482 return setSignedRange(Trunc,
3483 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3486 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3487 // If there's no signed wrap, and all the operands have the same sign or
3488 // zero, the value won't ever change sign.
3489 if (AddRec->getNoWrapFlags(SCEV::FlagNSW)) {
3490 bool AllNonNeg = true;
3491 bool AllNonPos = true;
3492 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3493 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3494 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3497 ConservativeResult = ConservativeResult.intersectWith(
3498 ConstantRange(APInt(BitWidth, 0),
3499 APInt::getSignedMinValue(BitWidth)));
3501 ConservativeResult = ConservativeResult.intersectWith(
3502 ConstantRange(APInt::getSignedMinValue(BitWidth),
3503 APInt(BitWidth, 1)));
3506 // TODO: non-affine addrec
3507 if (AddRec->isAffine()) {
3508 Type *Ty = AddRec->getType();
3509 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3510 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3511 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3512 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3514 const SCEV *Start = AddRec->getStart();
3515 const SCEV *Step = AddRec->getStepRecurrence(*this);
3517 ConstantRange StartRange = getSignedRange(Start);
3518 ConstantRange StepRange = getSignedRange(Step);
3519 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3520 ConstantRange EndRange =
3521 StartRange.add(MaxBECountRange.multiply(StepRange));
3523 // Check for overflow. This must be done with ConstantRange arithmetic
3524 // because we could be called from within the ScalarEvolution overflow
3526 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3527 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3528 ConstantRange ExtMaxBECountRange =
3529 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3530 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3531 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3533 return setSignedRange(AddRec, ConservativeResult);
3535 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3536 EndRange.getSignedMin());
3537 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3538 EndRange.getSignedMax());
3539 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3540 return setSignedRange(AddRec, ConservativeResult);
3541 return setSignedRange(AddRec,
3542 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3546 return setSignedRange(AddRec, ConservativeResult);
3549 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3550 // For a SCEVUnknown, ask ValueTracking.
3551 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3552 return setSignedRange(U, ConservativeResult);
3553 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3555 return setSignedRange(U, ConservativeResult);
3556 return setSignedRange(U, ConservativeResult.intersectWith(
3557 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3558 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3561 return setSignedRange(S, ConservativeResult);
3564 /// createSCEV - We know that there is no SCEV for the specified value.
3565 /// Analyze the expression.
3567 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3568 if (!isSCEVable(V->getType()))
3569 return getUnknown(V);
3571 unsigned Opcode = Instruction::UserOp1;
3572 if (Instruction *I = dyn_cast<Instruction>(V)) {
3573 Opcode = I->getOpcode();
3575 // Don't attempt to analyze instructions in blocks that aren't
3576 // reachable. Such instructions don't matter, and they aren't required
3577 // to obey basic rules for definitions dominating uses which this
3578 // analysis depends on.
3579 if (!DT->isReachableFromEntry(I->getParent()))
3580 return getUnknown(V);
3581 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3582 Opcode = CE->getOpcode();
3583 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3584 return getConstant(CI);
3585 else if (isa<ConstantPointerNull>(V))
3586 return getConstant(V->getType(), 0);
3587 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3588 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3590 return getUnknown(V);
3592 Operator *U = cast<Operator>(V);
3594 case Instruction::Add: {
3595 // The simple thing to do would be to just call getSCEV on both operands
3596 // and call getAddExpr with the result. However if we're looking at a
3597 // bunch of things all added together, this can be quite inefficient,
3598 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3599 // Instead, gather up all the operands and make a single getAddExpr call.
3600 // LLVM IR canonical form means we need only traverse the left operands.
3602 // Don't apply this instruction's NSW or NUW flags to the new
3603 // expression. The instruction may be guarded by control flow that the
3604 // no-wrap behavior depends on. Non-control-equivalent instructions can be
3605 // mapped to the same SCEV expression, and it would be incorrect to transfer
3606 // NSW/NUW semantics to those operations.
3607 SmallVector<const SCEV *, 4> AddOps;
3608 AddOps.push_back(getSCEV(U->getOperand(1)));
3609 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3610 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3611 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3613 U = cast<Operator>(Op);
3614 const SCEV *Op1 = getSCEV(U->getOperand(1));
3615 if (Opcode == Instruction::Sub)
3616 AddOps.push_back(getNegativeSCEV(Op1));
3618 AddOps.push_back(Op1);
3620 AddOps.push_back(getSCEV(U->getOperand(0)));
3621 return getAddExpr(AddOps);
3623 case Instruction::Mul: {
3624 // Don't transfer NSW/NUW for the same reason as AddExpr.
3625 SmallVector<const SCEV *, 4> MulOps;
3626 MulOps.push_back(getSCEV(U->getOperand(1)));
3627 for (Value *Op = U->getOperand(0);
3628 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3629 Op = U->getOperand(0)) {
3630 U = cast<Operator>(Op);
3631 MulOps.push_back(getSCEV(U->getOperand(1)));
3633 MulOps.push_back(getSCEV(U->getOperand(0)));
3634 return getMulExpr(MulOps);
3636 case Instruction::UDiv:
3637 return getUDivExpr(getSCEV(U->getOperand(0)),
3638 getSCEV(U->getOperand(1)));
3639 case Instruction::Sub:
3640 return getMinusSCEV(getSCEV(U->getOperand(0)),
3641 getSCEV(U->getOperand(1)));
3642 case Instruction::And:
3643 // For an expression like x&255 that merely masks off the high bits,
3644 // use zext(trunc(x)) as the SCEV expression.
3645 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3646 if (CI->isNullValue())
3647 return getSCEV(U->getOperand(1));
3648 if (CI->isAllOnesValue())
3649 return getSCEV(U->getOperand(0));
3650 const APInt &A = CI->getValue();
3652 // Instcombine's ShrinkDemandedConstant may strip bits out of
3653 // constants, obscuring what would otherwise be a low-bits mask.
3654 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3655 // knew about to reconstruct a low-bits mask value.
3656 unsigned LZ = A.countLeadingZeros();
3657 unsigned BitWidth = A.getBitWidth();
3658 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3659 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD);
3661 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3663 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3665 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3666 IntegerType::get(getContext(), BitWidth - LZ)),
3671 case Instruction::Or:
3672 // If the RHS of the Or is a constant, we may have something like:
3673 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3674 // optimizations will transparently handle this case.
3676 // In order for this transformation to be safe, the LHS must be of the
3677 // form X*(2^n) and the Or constant must be less than 2^n.
3678 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3679 const SCEV *LHS = getSCEV(U->getOperand(0));
3680 const APInt &CIVal = CI->getValue();
3681 if (GetMinTrailingZeros(LHS) >=
3682 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3683 // Build a plain add SCEV.
3684 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3685 // If the LHS of the add was an addrec and it has no-wrap flags,
3686 // transfer the no-wrap flags, since an or won't introduce a wrap.
3687 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3688 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3689 const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags(
3690 OldAR->getNoWrapFlags());
3696 case Instruction::Xor:
3697 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3698 // If the RHS of the xor is a signbit, then this is just an add.
3699 // Instcombine turns add of signbit into xor as a strength reduction step.
3700 if (CI->getValue().isSignBit())
3701 return getAddExpr(getSCEV(U->getOperand(0)),
3702 getSCEV(U->getOperand(1)));
3704 // If the RHS of xor is -1, then this is a not operation.
3705 if (CI->isAllOnesValue())
3706 return getNotSCEV(getSCEV(U->getOperand(0)));
3708 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3709 // This is a variant of the check for xor with -1, and it handles
3710 // the case where instcombine has trimmed non-demanded bits out
3711 // of an xor with -1.
3712 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3713 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3714 if (BO->getOpcode() == Instruction::And &&
3715 LCI->getValue() == CI->getValue())
3716 if (const SCEVZeroExtendExpr *Z =
3717 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3718 Type *UTy = U->getType();
3719 const SCEV *Z0 = Z->getOperand();
3720 Type *Z0Ty = Z0->getType();
3721 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3723 // If C is a low-bits mask, the zero extend is serving to
3724 // mask off the high bits. Complement the operand and
3725 // re-apply the zext.
3726 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3727 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3729 // If C is a single bit, it may be in the sign-bit position
3730 // before the zero-extend. In this case, represent the xor
3731 // using an add, which is equivalent, and re-apply the zext.
3732 APInt Trunc = CI->getValue().trunc(Z0TySize);
3733 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3735 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3741 case Instruction::Shl:
3742 // Turn shift left of a constant amount into a multiply.
3743 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3744 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3746 // If the shift count is not less than the bitwidth, the result of
3747 // the shift is undefined. Don't try to analyze it, because the
3748 // resolution chosen here may differ from the resolution chosen in
3749 // other parts of the compiler.
3750 if (SA->getValue().uge(BitWidth))
3753 Constant *X = ConstantInt::get(getContext(),
3754 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3755 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3759 case Instruction::LShr:
3760 // Turn logical shift right of a constant into a unsigned divide.
3761 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3762 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3764 // If the shift count is not less than the bitwidth, the result of
3765 // the shift is undefined. Don't try to analyze it, because the
3766 // resolution chosen here may differ from the resolution chosen in
3767 // other parts of the compiler.
3768 if (SA->getValue().uge(BitWidth))
3771 Constant *X = ConstantInt::get(getContext(),
3772 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3773 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3777 case Instruction::AShr:
3778 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3779 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3780 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3781 if (L->getOpcode() == Instruction::Shl &&
3782 L->getOperand(1) == U->getOperand(1)) {
3783 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3785 // If the shift count is not less than the bitwidth, the result of
3786 // the shift is undefined. Don't try to analyze it, because the
3787 // resolution chosen here may differ from the resolution chosen in
3788 // other parts of the compiler.
3789 if (CI->getValue().uge(BitWidth))
3792 uint64_t Amt = BitWidth - CI->getZExtValue();
3793 if (Amt == BitWidth)
3794 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3796 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3797 IntegerType::get(getContext(),
3803 case Instruction::Trunc:
3804 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3806 case Instruction::ZExt:
3807 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3809 case Instruction::SExt:
3810 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3812 case Instruction::BitCast:
3813 // BitCasts are no-op casts so we just eliminate the cast.
3814 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3815 return getSCEV(U->getOperand(0));
3818 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3819 // lead to pointer expressions which cannot safely be expanded to GEPs,
3820 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3821 // simplifying integer expressions.
3823 case Instruction::GetElementPtr:
3824 return createNodeForGEP(cast<GEPOperator>(U));
3826 case Instruction::PHI:
3827 return createNodeForPHI(cast<PHINode>(U));
3829 case Instruction::Select:
3830 // This could be a smax or umax that was lowered earlier.
3831 // Try to recover it.
3832 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3833 Value *LHS = ICI->getOperand(0);
3834 Value *RHS = ICI->getOperand(1);
3835 switch (ICI->getPredicate()) {
3836 case ICmpInst::ICMP_SLT:
3837 case ICmpInst::ICMP_SLE:
3838 std::swap(LHS, RHS);
3840 case ICmpInst::ICMP_SGT:
3841 case ICmpInst::ICMP_SGE:
3842 // a >s b ? a+x : b+x -> smax(a, b)+x
3843 // a >s b ? b+x : a+x -> smin(a, b)+x
3844 if (LHS->getType() == U->getType()) {
3845 const SCEV *LS = getSCEV(LHS);
3846 const SCEV *RS = getSCEV(RHS);
3847 const SCEV *LA = getSCEV(U->getOperand(1));
3848 const SCEV *RA = getSCEV(U->getOperand(2));
3849 const SCEV *LDiff = getMinusSCEV(LA, LS);
3850 const SCEV *RDiff = getMinusSCEV(RA, RS);
3852 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3853 LDiff = getMinusSCEV(LA, RS);
3854 RDiff = getMinusSCEV(RA, LS);
3856 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3859 case ICmpInst::ICMP_ULT:
3860 case ICmpInst::ICMP_ULE:
3861 std::swap(LHS, RHS);
3863 case ICmpInst::ICMP_UGT:
3864 case ICmpInst::ICMP_UGE:
3865 // a >u b ? a+x : b+x -> umax(a, b)+x
3866 // a >u b ? b+x : a+x -> umin(a, b)+x
3867 if (LHS->getType() == U->getType()) {
3868 const SCEV *LS = getSCEV(LHS);
3869 const SCEV *RS = getSCEV(RHS);
3870 const SCEV *LA = getSCEV(U->getOperand(1));
3871 const SCEV *RA = getSCEV(U->getOperand(2));
3872 const SCEV *LDiff = getMinusSCEV(LA, LS);
3873 const SCEV *RDiff = getMinusSCEV(RA, RS);
3875 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3876 LDiff = getMinusSCEV(LA, RS);
3877 RDiff = getMinusSCEV(RA, LS);
3879 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3882 case ICmpInst::ICMP_NE:
3883 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3884 if (LHS->getType() == U->getType() &&
3885 isa<ConstantInt>(RHS) &&
3886 cast<ConstantInt>(RHS)->isZero()) {
3887 const SCEV *One = getConstant(LHS->getType(), 1);
3888 const SCEV *LS = getSCEV(LHS);
3889 const SCEV *LA = getSCEV(U->getOperand(1));
3890 const SCEV *RA = getSCEV(U->getOperand(2));
3891 const SCEV *LDiff = getMinusSCEV(LA, LS);
3892 const SCEV *RDiff = getMinusSCEV(RA, One);
3894 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3897 case ICmpInst::ICMP_EQ:
3898 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3899 if (LHS->getType() == U->getType() &&
3900 isa<ConstantInt>(RHS) &&
3901 cast<ConstantInt>(RHS)->isZero()) {
3902 const SCEV *One = getConstant(LHS->getType(), 1);
3903 const SCEV *LS = getSCEV(LHS);
3904 const SCEV *LA = getSCEV(U->getOperand(1));
3905 const SCEV *RA = getSCEV(U->getOperand(2));
3906 const SCEV *LDiff = getMinusSCEV(LA, One);
3907 const SCEV *RDiff = getMinusSCEV(RA, LS);
3909 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3917 default: // We cannot analyze this expression.
3921 return getUnknown(V);
3926 //===----------------------------------------------------------------------===//
3927 // Iteration Count Computation Code
3930 /// getSmallConstantTripCount - Returns the maximum trip count of this loop as a
3931 /// normal unsigned value. Returns 0 if the trip count is unknown or not
3932 /// constant. Will also return 0 if the maximum trip count is very large (>=
3935 /// This "trip count" assumes that control exits via ExitingBlock. More
3936 /// precisely, it is the number of times that control may reach ExitingBlock
3937 /// before taking the branch. For loops with multiple exits, it may not be the
3938 /// number times that the loop header executes because the loop may exit
3939 /// prematurely via another branch.
3940 unsigned ScalarEvolution::
3941 getSmallConstantTripCount(Loop *L, BasicBlock *ExitingBlock) {
3942 const SCEVConstant *ExitCount =
3943 dyn_cast<SCEVConstant>(getExitCount(L, ExitingBlock));
3947 ConstantInt *ExitConst = ExitCount->getValue();
3949 // Guard against huge trip counts.
3950 if (ExitConst->getValue().getActiveBits() > 32)
3953 // In case of integer overflow, this returns 0, which is correct.
3954 return ((unsigned)ExitConst->getZExtValue()) + 1;
3957 /// getSmallConstantTripMultiple - Returns the largest constant divisor of the
3958 /// trip count of this loop as a normal unsigned value, if possible. This
3959 /// means that the actual trip count is always a multiple of the returned
3960 /// value (don't forget the trip count could very well be zero as well!).
3962 /// Returns 1 if the trip count is unknown or not guaranteed to be the
3963 /// multiple of a constant (which is also the case if the trip count is simply
3964 /// constant, use getSmallConstantTripCount for that case), Will also return 1
3965 /// if the trip count is very large (>= 2^32).
3967 /// As explained in the comments for getSmallConstantTripCount, this assumes
3968 /// that control exits the loop via ExitingBlock.
3969 unsigned ScalarEvolution::
3970 getSmallConstantTripMultiple(Loop *L, BasicBlock *ExitingBlock) {
3971 const SCEV *ExitCount = getExitCount(L, ExitingBlock);
3972 if (ExitCount == getCouldNotCompute())
3975 // Get the trip count from the BE count by adding 1.
3976 const SCEV *TCMul = getAddExpr(ExitCount,
3977 getConstant(ExitCount->getType(), 1));
3978 // FIXME: SCEV distributes multiplication as V1*C1 + V2*C1. We could attempt
3979 // to factor simple cases.
3980 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(TCMul))
3981 TCMul = Mul->getOperand(0);
3983 const SCEVConstant *MulC = dyn_cast<SCEVConstant>(TCMul);
3987 ConstantInt *Result = MulC->getValue();
3989 // Guard against huge trip counts (this requires checking
3990 // for zero to handle the case where the trip count == -1 and the
3992 if (!Result || Result->getValue().getActiveBits() > 32 ||
3993 Result->getValue().getActiveBits() == 0)
3996 return (unsigned)Result->getZExtValue();
3999 // getExitCount - Get the expression for the number of loop iterations for which
4000 // this loop is guaranteed not to exit via ExitintBlock. Otherwise return
4001 // SCEVCouldNotCompute.
4002 const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) {
4003 return getBackedgeTakenInfo(L).getExact(ExitingBlock, this);
4006 /// getBackedgeTakenCount - If the specified loop has a predictable
4007 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
4008 /// object. The backedge-taken count is the number of times the loop header
4009 /// will be branched to from within the loop. This is one less than the
4010 /// trip count of the loop, since it doesn't count the first iteration,
4011 /// when the header is branched to from outside the loop.
4013 /// Note that it is not valid to call this method on a loop without a
4014 /// loop-invariant backedge-taken count (see
4015 /// hasLoopInvariantBackedgeTakenCount).
4017 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
4018 return getBackedgeTakenInfo(L).getExact(this);
4021 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
4022 /// return the least SCEV value that is known never to be less than the
4023 /// actual backedge taken count.
4024 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
4025 return getBackedgeTakenInfo(L).getMax(this);
4028 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
4029 /// onto the given Worklist.
4031 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
4032 BasicBlock *Header = L->getHeader();
4034 // Push all Loop-header PHIs onto the Worklist stack.
4035 for (BasicBlock::iterator I = Header->begin();
4036 PHINode *PN = dyn_cast<PHINode>(I); ++I)
4037 Worklist.push_back(PN);
4040 const ScalarEvolution::BackedgeTakenInfo &
4041 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
4042 // Initially insert an invalid entry for this loop. If the insertion
4043 // succeeds, proceed to actually compute a backedge-taken count and
4044 // update the value. The temporary CouldNotCompute value tells SCEV
4045 // code elsewhere that it shouldn't attempt to request a new
4046 // backedge-taken count, which could result in infinite recursion.
4047 std::pair<DenseMap<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
4048 BackedgeTakenCounts.insert(std::make_pair(L, BackedgeTakenInfo()));
4050 return Pair.first->second;
4052 // ComputeBackedgeTakenCount may allocate memory for its result. Inserting it
4053 // into the BackedgeTakenCounts map transfers ownership. Otherwise, the result
4054 // must be cleared in this scope.
4055 BackedgeTakenInfo Result = ComputeBackedgeTakenCount(L);
4057 if (Result.getExact(this) != getCouldNotCompute()) {
4058 assert(isLoopInvariant(Result.getExact(this), L) &&
4059 isLoopInvariant(Result.getMax(this), L) &&
4060 "Computed backedge-taken count isn't loop invariant for loop!");
4061 ++NumTripCountsComputed;
4063 else if (Result.getMax(this) == getCouldNotCompute() &&
4064 isa<PHINode>(L->getHeader()->begin())) {
4065 // Only count loops that have phi nodes as not being computable.
4066 ++NumTripCountsNotComputed;
4069 // Now that we know more about the trip count for this loop, forget any
4070 // existing SCEV values for PHI nodes in this loop since they are only
4071 // conservative estimates made without the benefit of trip count
4072 // information. This is similar to the code in forgetLoop, except that
4073 // it handles SCEVUnknown PHI nodes specially.
4074 if (Result.hasAnyInfo()) {
4075 SmallVector<Instruction *, 16> Worklist;
4076 PushLoopPHIs(L, Worklist);
4078 SmallPtrSet<Instruction *, 8> Visited;
4079 while (!Worklist.empty()) {
4080 Instruction *I = Worklist.pop_back_val();
4081 if (!Visited.insert(I)) continue;
4083 ValueExprMapType::iterator It =
4084 ValueExprMap.find_as(static_cast<Value *>(I));
4085 if (It != ValueExprMap.end()) {
4086 const SCEV *Old = It->second;
4088 // SCEVUnknown for a PHI either means that it has an unrecognized
4089 // structure, or it's a PHI that's in the progress of being computed
4090 // by createNodeForPHI. In the former case, additional loop trip
4091 // count information isn't going to change anything. In the later
4092 // case, createNodeForPHI will perform the necessary updates on its
4093 // own when it gets to that point.
4094 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
4095 forgetMemoizedResults(Old);
4096 ValueExprMap.erase(It);
4098 if (PHINode *PN = dyn_cast<PHINode>(I))
4099 ConstantEvolutionLoopExitValue.erase(PN);
4102 PushDefUseChildren(I, Worklist);
4106 // Re-lookup the insert position, since the call to
4107 // ComputeBackedgeTakenCount above could result in a
4108 // recusive call to getBackedgeTakenInfo (on a different
4109 // loop), which would invalidate the iterator computed
4111 return BackedgeTakenCounts.find(L)->second = Result;
4114 /// forgetLoop - This method should be called by the client when it has
4115 /// changed a loop in a way that may effect ScalarEvolution's ability to
4116 /// compute a trip count, or if the loop is deleted.
4117 void ScalarEvolution::forgetLoop(const Loop *L) {
4118 // Drop any stored trip count value.
4119 DenseMap<const Loop*, BackedgeTakenInfo>::iterator BTCPos =
4120 BackedgeTakenCounts.find(L);
4121 if (BTCPos != BackedgeTakenCounts.end()) {
4122 BTCPos->second.clear();
4123 BackedgeTakenCounts.erase(BTCPos);
4126 // Drop information about expressions based on loop-header PHIs.
4127 SmallVector<Instruction *, 16> Worklist;
4128 PushLoopPHIs(L, Worklist);
4130 SmallPtrSet<Instruction *, 8> Visited;
4131 while (!Worklist.empty()) {
4132 Instruction *I = Worklist.pop_back_val();
4133 if (!Visited.insert(I)) continue;
4135 ValueExprMapType::iterator It =
4136 ValueExprMap.find_as(static_cast<Value *>(I));
4137 if (It != ValueExprMap.end()) {
4138 forgetMemoizedResults(It->second);
4139 ValueExprMap.erase(It);
4140 if (PHINode *PN = dyn_cast<PHINode>(I))
4141 ConstantEvolutionLoopExitValue.erase(PN);
4144 PushDefUseChildren(I, Worklist);
4147 // Forget all contained loops too, to avoid dangling entries in the
4148 // ValuesAtScopes map.
4149 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
4153 /// forgetValue - This method should be called by the client when it has
4154 /// changed a value in a way that may effect its value, or which may
4155 /// disconnect it from a def-use chain linking it to a loop.
4156 void ScalarEvolution::forgetValue(Value *V) {
4157 Instruction *I = dyn_cast<Instruction>(V);
4160 // Drop information about expressions based on loop-header PHIs.
4161 SmallVector<Instruction *, 16> Worklist;
4162 Worklist.push_back(I);
4164 SmallPtrSet<Instruction *, 8> Visited;
4165 while (!Worklist.empty()) {
4166 I = Worklist.pop_back_val();
4167 if (!Visited.insert(I)) continue;
4169 ValueExprMapType::iterator It =
4170 ValueExprMap.find_as(static_cast<Value *>(I));
4171 if (It != ValueExprMap.end()) {
4172 forgetMemoizedResults(It->second);
4173 ValueExprMap.erase(It);
4174 if (PHINode *PN = dyn_cast<PHINode>(I))
4175 ConstantEvolutionLoopExitValue.erase(PN);
4178 PushDefUseChildren(I, Worklist);
4182 /// getExact - Get the exact loop backedge taken count considering all loop
4183 /// exits. A computable result can only be return for loops with a single exit.
4184 /// Returning the minimum taken count among all exits is incorrect because one
4185 /// of the loop's exit limit's may have been skipped. HowFarToZero assumes that
4186 /// the limit of each loop test is never skipped. This is a valid assumption as
4187 /// long as the loop exits via that test. For precise results, it is the
4188 /// caller's responsibility to specify the relevant loop exit using
4189 /// getExact(ExitingBlock, SE).
4191 ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
4192 // If any exits were not computable, the loop is not computable.
4193 if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute();
4195 // We need exactly one computable exit.
4196 if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
4197 assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
4199 const SCEV *BECount = 0;
4200 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4201 ENT != 0; ENT = ENT->getNextExit()) {
4203 assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV");
4206 BECount = ENT->ExactNotTaken;
4207 else if (BECount != ENT->ExactNotTaken)
4208 return SE->getCouldNotCompute();
4210 assert(BECount && "Invalid not taken count for loop exit");
4214 /// getExact - Get the exact not taken count for this loop exit.
4216 ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock,
4217 ScalarEvolution *SE) const {
4218 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4219 ENT != 0; ENT = ENT->getNextExit()) {
4221 if (ENT->ExitingBlock == ExitingBlock)
4222 return ENT->ExactNotTaken;
4224 return SE->getCouldNotCompute();
4227 /// getMax - Get the max backedge taken count for the loop.
4229 ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const {
4230 return Max ? Max : SE->getCouldNotCompute();
4233 bool ScalarEvolution::BackedgeTakenInfo::hasOperand(const SCEV *S,
4234 ScalarEvolution *SE) const {
4235 if (Max && Max != SE->getCouldNotCompute() && SE->hasOperand(Max, S))
4238 if (!ExitNotTaken.ExitingBlock)
4241 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4242 ENT != 0; ENT = ENT->getNextExit()) {
4244 if (ENT->ExactNotTaken != SE->getCouldNotCompute()
4245 && SE->hasOperand(ENT->ExactNotTaken, S)) {
4252 /// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each
4253 /// computable exit into a persistent ExitNotTakenInfo array.
4254 ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo(
4255 SmallVectorImpl< std::pair<BasicBlock *, const SCEV *> > &ExitCounts,
4256 bool Complete, const SCEV *MaxCount) : Max(MaxCount) {
4259 ExitNotTaken.setIncomplete();
4261 unsigned NumExits = ExitCounts.size();
4262 if (NumExits == 0) return;
4264 ExitNotTaken.ExitingBlock = ExitCounts[0].first;
4265 ExitNotTaken.ExactNotTaken = ExitCounts[0].second;
4266 if (NumExits == 1) return;
4268 // Handle the rare case of multiple computable exits.
4269 ExitNotTakenInfo *ENT = new ExitNotTakenInfo[NumExits-1];
4271 ExitNotTakenInfo *PrevENT = &ExitNotTaken;
4272 for (unsigned i = 1; i < NumExits; ++i, PrevENT = ENT, ++ENT) {
4273 PrevENT->setNextExit(ENT);
4274 ENT->ExitingBlock = ExitCounts[i].first;
4275 ENT->ExactNotTaken = ExitCounts[i].second;
4279 /// clear - Invalidate this result and free the ExitNotTakenInfo array.
4280 void ScalarEvolution::BackedgeTakenInfo::clear() {
4281 ExitNotTaken.ExitingBlock = 0;
4282 ExitNotTaken.ExactNotTaken = 0;
4283 delete[] ExitNotTaken.getNextExit();
4286 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
4287 /// of the specified loop will execute.
4288 ScalarEvolution::BackedgeTakenInfo
4289 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
4290 SmallVector<BasicBlock *, 8> ExitingBlocks;
4291 L->getExitingBlocks(ExitingBlocks);
4293 // Examine all exits and pick the most conservative values.
4294 const SCEV *MaxBECount = getCouldNotCompute();
4295 bool CouldComputeBECount = true;
4296 SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts;
4297 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
4298 ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]);
4299 if (EL.Exact == getCouldNotCompute())
4300 // We couldn't compute an exact value for this exit, so
4301 // we won't be able to compute an exact value for the loop.
4302 CouldComputeBECount = false;
4304 ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact));
4306 if (MaxBECount == getCouldNotCompute())
4307 MaxBECount = EL.Max;
4308 else if (EL.Max != getCouldNotCompute()) {
4309 // We cannot take the "min" MaxBECount, because non-unit stride loops may
4310 // skip some loop tests. Taking the max over the exits is sufficiently
4311 // conservative. TODO: We could do better taking into consideration
4312 // that (1) the loop has unit stride (2) the last loop test is
4313 // less-than/greater-than (3) any loop test is less-than/greater-than AND
4314 // falls-through some constant times less then the other tests.
4315 MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
4319 return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
4322 /// ComputeExitLimit - Compute the number of times the backedge of the specified
4323 /// loop will execute if it exits via the specified block.
4324 ScalarEvolution::ExitLimit
4325 ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) {
4327 // Okay, we've chosen an exiting block. See what condition causes us to
4328 // exit at this block.
4330 // FIXME: we should be able to handle switch instructions (with a single exit)
4331 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
4332 if (ExitBr == 0) return getCouldNotCompute();
4333 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
4335 // At this point, we know we have a conditional branch that determines whether
4336 // the loop is exited. However, we don't know if the branch is executed each
4337 // time through the loop. If not, then the execution count of the branch will
4338 // not be equal to the trip count of the loop.
4340 // Currently we check for this by checking to see if the Exit branch goes to
4341 // the loop header. If so, we know it will always execute the same number of
4342 // times as the loop. We also handle the case where the exit block *is* the
4343 // loop header. This is common for un-rotated loops.
4345 // If both of those tests fail, walk up the unique predecessor chain to the
4346 // header, stopping if there is an edge that doesn't exit the loop. If the
4347 // header is reached, the execution count of the branch will be equal to the
4348 // trip count of the loop.
4350 // More extensive analysis could be done to handle more cases here.
4352 if (ExitBr->getSuccessor(0) != L->getHeader() &&
4353 ExitBr->getSuccessor(1) != L->getHeader() &&
4354 ExitBr->getParent() != L->getHeader()) {
4355 // The simple checks failed, try climbing the unique predecessor chain
4356 // up to the header.
4358 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
4359 BasicBlock *Pred = BB->getUniquePredecessor();
4361 return getCouldNotCompute();
4362 TerminatorInst *PredTerm = Pred->getTerminator();
4363 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
4364 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
4367 // If the predecessor has a successor that isn't BB and isn't
4368 // outside the loop, assume the worst.
4369 if (L->contains(PredSucc))
4370 return getCouldNotCompute();
4372 if (Pred == L->getHeader()) {
4379 return getCouldNotCompute();
4382 // Proceed to the next level to examine the exit condition expression.
4383 return ComputeExitLimitFromCond(L, ExitBr->getCondition(),
4384 ExitBr->getSuccessor(0),
4385 ExitBr->getSuccessor(1),
4386 /*IsSubExpr=*/false);
4389 /// ComputeExitLimitFromCond - Compute the number of times the
4390 /// backedge of the specified loop will execute if its exit condition
4391 /// were a conditional branch of ExitCond, TBB, and FBB.
4393 /// @param IsSubExpr is true if ExitCond does not directly control the exit
4394 /// branch. In this case, we cannot assume that the loop only exits when the
4395 /// condition is true and cannot infer that failing to meet the condition prior
4396 /// to integer wraparound results in undefined behavior.
4397 ScalarEvolution::ExitLimit
4398 ScalarEvolution::ComputeExitLimitFromCond(const Loop *L,
4403 // Check if the controlling expression for this loop is an And or Or.
4404 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
4405 if (BO->getOpcode() == Instruction::And) {
4406 // Recurse on the operands of the and.
4407 bool EitherMayExit = L->contains(TBB);
4408 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB,
4409 IsSubExpr || EitherMayExit);
4410 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB,
4411 IsSubExpr || EitherMayExit);
4412 const SCEV *BECount = getCouldNotCompute();
4413 const SCEV *MaxBECount = getCouldNotCompute();
4414 if (EitherMayExit) {
4415 // Both conditions must be true for the loop to continue executing.
4416 // Choose the less conservative count.
4417 if (EL0.Exact == getCouldNotCompute() ||
4418 EL1.Exact == getCouldNotCompute())
4419 BECount = getCouldNotCompute();
4421 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4422 if (EL0.Max == getCouldNotCompute())
4423 MaxBECount = EL1.Max;
4424 else if (EL1.Max == getCouldNotCompute())
4425 MaxBECount = EL0.Max;
4427 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4429 // Both conditions must be true at the same time for the loop to exit.
4430 // For now, be conservative.
4431 assert(L->contains(FBB) && "Loop block has no successor in loop!");
4432 if (EL0.Max == EL1.Max)
4433 MaxBECount = EL0.Max;
4434 if (EL0.Exact == EL1.Exact)
4435 BECount = EL0.Exact;
4438 return ExitLimit(BECount, MaxBECount);
4440 if (BO->getOpcode() == Instruction::Or) {
4441 // Recurse on the operands of the or.
4442 bool EitherMayExit = L->contains(FBB);
4443 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB,
4444 IsSubExpr || EitherMayExit);
4445 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB,
4446 IsSubExpr || EitherMayExit);
4447 const SCEV *BECount = getCouldNotCompute();
4448 const SCEV *MaxBECount = getCouldNotCompute();
4449 if (EitherMayExit) {
4450 // Both conditions must be false for the loop to continue executing.
4451 // Choose the less conservative count.
4452 if (EL0.Exact == getCouldNotCompute() ||
4453 EL1.Exact == getCouldNotCompute())
4454 BECount = getCouldNotCompute();
4456 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4457 if (EL0.Max == getCouldNotCompute())
4458 MaxBECount = EL1.Max;
4459 else if (EL1.Max == getCouldNotCompute())
4460 MaxBECount = EL0.Max;
4462 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4464 // Both conditions must be false at the same time for the loop to exit.
4465 // For now, be conservative.
4466 assert(L->contains(TBB) && "Loop block has no successor in loop!");
4467 if (EL0.Max == EL1.Max)
4468 MaxBECount = EL0.Max;
4469 if (EL0.Exact == EL1.Exact)
4470 BECount = EL0.Exact;
4473 return ExitLimit(BECount, MaxBECount);
4477 // With an icmp, it may be feasible to compute an exact backedge-taken count.
4478 // Proceed to the next level to examine the icmp.
4479 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
4480 return ComputeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB, IsSubExpr);
4482 // Check for a constant condition. These are normally stripped out by
4483 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
4484 // preserve the CFG and is temporarily leaving constant conditions
4486 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
4487 if (L->contains(FBB) == !CI->getZExtValue())
4488 // The backedge is always taken.
4489 return getCouldNotCompute();
4491 // The backedge is never taken.
4492 return getConstant(CI->getType(), 0);
4495 // If it's not an integer or pointer comparison then compute it the hard way.
4496 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4499 /// ComputeExitLimitFromICmp - Compute the number of times the
4500 /// backedge of the specified loop will execute if its exit condition
4501 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4502 ScalarEvolution::ExitLimit
4503 ScalarEvolution::ComputeExitLimitFromICmp(const Loop *L,
4509 // If the condition was exit on true, convert the condition to exit on false
4510 ICmpInst::Predicate Cond;
4511 if (!L->contains(FBB))
4512 Cond = ExitCond->getPredicate();
4514 Cond = ExitCond->getInversePredicate();
4516 // Handle common loops like: for (X = "string"; *X; ++X)
4517 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4518 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4520 ComputeLoadConstantCompareExitLimit(LI, RHS, L, Cond);
4521 if (ItCnt.hasAnyInfo())
4525 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4526 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4528 // Try to evaluate any dependencies out of the loop.
4529 LHS = getSCEVAtScope(LHS, L);
4530 RHS = getSCEVAtScope(RHS, L);
4532 // At this point, we would like to compute how many iterations of the
4533 // loop the predicate will return true for these inputs.
4534 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4535 // If there is a loop-invariant, force it into the RHS.
4536 std::swap(LHS, RHS);
4537 Cond = ICmpInst::getSwappedPredicate(Cond);
4540 // Simplify the operands before analyzing them.
4541 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4543 // If we have a comparison of a chrec against a constant, try to use value
4544 // ranges to answer this query.
4545 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4546 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4547 if (AddRec->getLoop() == L) {
4548 // Form the constant range.
4549 ConstantRange CompRange(
4550 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4552 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4553 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4557 case ICmpInst::ICMP_NE: { // while (X != Y)
4558 // Convert to: while (X-Y != 0)
4559 ExitLimit EL = HowFarToZero(getMinusSCEV(LHS, RHS), L, IsSubExpr);
4560 if (EL.hasAnyInfo()) return EL;
4563 case ICmpInst::ICMP_EQ: { // while (X == Y)
4564 // Convert to: while (X-Y == 0)
4565 ExitLimit EL = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4566 if (EL.hasAnyInfo()) return EL;
4569 case ICmpInst::ICMP_SLT: {
4570 ExitLimit EL = HowManyLessThans(LHS, RHS, L, true, IsSubExpr);
4571 if (EL.hasAnyInfo()) return EL;
4574 case ICmpInst::ICMP_SGT: {
4575 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4576 getNotSCEV(RHS), L, true, IsSubExpr);
4577 if (EL.hasAnyInfo()) return EL;
4580 case ICmpInst::ICMP_ULT: {
4581 ExitLimit EL = HowManyLessThans(LHS, RHS, L, false, IsSubExpr);
4582 if (EL.hasAnyInfo()) return EL;
4585 case ICmpInst::ICMP_UGT: {
4586 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4587 getNotSCEV(RHS), L, false, IsSubExpr);
4588 if (EL.hasAnyInfo()) return EL;
4593 dbgs() << "ComputeBackedgeTakenCount ";
4594 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4595 dbgs() << "[unsigned] ";
4596 dbgs() << *LHS << " "
4597 << Instruction::getOpcodeName(Instruction::ICmp)
4598 << " " << *RHS << "\n";
4602 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4605 static ConstantInt *
4606 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4607 ScalarEvolution &SE) {
4608 const SCEV *InVal = SE.getConstant(C);
4609 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4610 assert(isa<SCEVConstant>(Val) &&
4611 "Evaluation of SCEV at constant didn't fold correctly?");
4612 return cast<SCEVConstant>(Val)->getValue();
4615 /// ComputeLoadConstantCompareExitLimit - Given an exit condition of
4616 /// 'icmp op load X, cst', try to see if we can compute the backedge
4617 /// execution count.
4618 ScalarEvolution::ExitLimit
4619 ScalarEvolution::ComputeLoadConstantCompareExitLimit(
4623 ICmpInst::Predicate predicate) {
4625 if (LI->isVolatile()) return getCouldNotCompute();
4627 // Check to see if the loaded pointer is a getelementptr of a global.
4628 // TODO: Use SCEV instead of manually grubbing with GEPs.
4629 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4630 if (!GEP) return getCouldNotCompute();
4632 // Make sure that it is really a constant global we are gepping, with an
4633 // initializer, and make sure the first IDX is really 0.
4634 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4635 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4636 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4637 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4638 return getCouldNotCompute();
4640 // Okay, we allow one non-constant index into the GEP instruction.
4642 std::vector<Constant*> Indexes;
4643 unsigned VarIdxNum = 0;
4644 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4645 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4646 Indexes.push_back(CI);
4647 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4648 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4649 VarIdx = GEP->getOperand(i);
4651 Indexes.push_back(0);
4654 // Loop-invariant loads may be a byproduct of loop optimization. Skip them.
4656 return getCouldNotCompute();
4658 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4659 // Check to see if X is a loop variant variable value now.
4660 const SCEV *Idx = getSCEV(VarIdx);
4661 Idx = getSCEVAtScope(Idx, L);
4663 // We can only recognize very limited forms of loop index expressions, in
4664 // particular, only affine AddRec's like {C1,+,C2}.
4665 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4666 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4667 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4668 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4669 return getCouldNotCompute();
4671 unsigned MaxSteps = MaxBruteForceIterations;
4672 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4673 ConstantInt *ItCst = ConstantInt::get(
4674 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4675 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4677 // Form the GEP offset.
4678 Indexes[VarIdxNum] = Val;
4680 Constant *Result = ConstantFoldLoadThroughGEPIndices(GV->getInitializer(),
4682 if (Result == 0) break; // Cannot compute!
4684 // Evaluate the condition for this iteration.
4685 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4686 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4687 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4689 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4690 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4693 ++NumArrayLenItCounts;
4694 return getConstant(ItCst); // Found terminating iteration!
4697 return getCouldNotCompute();
4701 /// CanConstantFold - Return true if we can constant fold an instruction of the
4702 /// specified type, assuming that all operands were constants.
4703 static bool CanConstantFold(const Instruction *I) {
4704 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4705 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) ||
4709 if (const CallInst *CI = dyn_cast<CallInst>(I))
4710 if (const Function *F = CI->getCalledFunction())
4711 return canConstantFoldCallTo(F);
4715 /// Determine whether this instruction can constant evolve within this loop
4716 /// assuming its operands can all constant evolve.
4717 static bool canConstantEvolve(Instruction *I, const Loop *L) {
4718 // An instruction outside of the loop can't be derived from a loop PHI.
4719 if (!L->contains(I)) return false;
4721 if (isa<PHINode>(I)) {
4722 if (L->getHeader() == I->getParent())
4725 // We don't currently keep track of the control flow needed to evaluate
4726 // PHIs, so we cannot handle PHIs inside of loops.
4730 // If we won't be able to constant fold this expression even if the operands
4731 // are constants, bail early.
4732 return CanConstantFold(I);
4735 /// getConstantEvolvingPHIOperands - Implement getConstantEvolvingPHI by
4736 /// recursing through each instruction operand until reaching a loop header phi.
4738 getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L,
4739 DenseMap<Instruction *, PHINode *> &PHIMap) {
4741 // Otherwise, we can evaluate this instruction if all of its operands are
4742 // constant or derived from a PHI node themselves.
4744 for (Instruction::op_iterator OpI = UseInst->op_begin(),
4745 OpE = UseInst->op_end(); OpI != OpE; ++OpI) {
4747 if (isa<Constant>(*OpI)) continue;
4749 Instruction *OpInst = dyn_cast<Instruction>(*OpI);
4750 if (!OpInst || !canConstantEvolve(OpInst, L)) return 0;
4752 PHINode *P = dyn_cast<PHINode>(OpInst);
4754 // If this operand is already visited, reuse the prior result.
4755 // We may have P != PHI if this is the deepest point at which the
4756 // inconsistent paths meet.
4757 P = PHIMap.lookup(OpInst);
4759 // Recurse and memoize the results, whether a phi is found or not.
4760 // This recursive call invalidates pointers into PHIMap.
4761 P = getConstantEvolvingPHIOperands(OpInst, L, PHIMap);
4764 if (P == 0) return 0; // Not evolving from PHI
4765 if (PHI && PHI != P) return 0; // Evolving from multiple different PHIs.
4768 // This is a expression evolving from a constant PHI!
4772 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4773 /// in the loop that V is derived from. We allow arbitrary operations along the
4774 /// way, but the operands of an operation must either be constants or a value
4775 /// derived from a constant PHI. If this expression does not fit with these
4776 /// constraints, return null.
4777 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4778 Instruction *I = dyn_cast<Instruction>(V);
4779 if (I == 0 || !canConstantEvolve(I, L)) return 0;
4781 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4785 // Record non-constant instructions contained by the loop.
4786 DenseMap<Instruction *, PHINode *> PHIMap;
4787 return getConstantEvolvingPHIOperands(I, L, PHIMap);
4790 /// EvaluateExpression - Given an expression that passes the
4791 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4792 /// in the loop has the value PHIVal. If we can't fold this expression for some
4793 /// reason, return null.
4794 static Constant *EvaluateExpression(Value *V, const Loop *L,
4795 DenseMap<Instruction *, Constant *> &Vals,
4796 const DataLayout *TD,
4797 const TargetLibraryInfo *TLI) {
4798 // Convenient constant check, but redundant for recursive calls.
4799 if (Constant *C = dyn_cast<Constant>(V)) return C;
4800 Instruction *I = dyn_cast<Instruction>(V);
4803 if (Constant *C = Vals.lookup(I)) return C;
4805 // An instruction inside the loop depends on a value outside the loop that we
4806 // weren't given a mapping for, or a value such as a call inside the loop.
4807 if (!canConstantEvolve(I, L)) return 0;
4809 // An unmapped PHI can be due to a branch or another loop inside this loop,
4810 // or due to this not being the initial iteration through a loop where we
4811 // couldn't compute the evolution of this particular PHI last time.
4812 if (isa<PHINode>(I)) return 0;
4814 std::vector<Constant*> Operands(I->getNumOperands());
4816 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4817 Instruction *Operand = dyn_cast<Instruction>(I->getOperand(i));
4819 Operands[i] = dyn_cast<Constant>(I->getOperand(i));
4820 if (!Operands[i]) return 0;
4823 Constant *C = EvaluateExpression(Operand, L, Vals, TD, TLI);
4829 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4830 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4831 Operands[1], TD, TLI);
4832 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4833 if (!LI->isVolatile())
4834 return ConstantFoldLoadFromConstPtr(Operands[0], TD);
4836 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Operands, TD,
4840 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4841 /// in the header of its containing loop, we know the loop executes a
4842 /// constant number of times, and the PHI node is just a recurrence
4843 /// involving constants, fold it.
4845 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4848 DenseMap<PHINode*, Constant*>::const_iterator I =
4849 ConstantEvolutionLoopExitValue.find(PN);
4850 if (I != ConstantEvolutionLoopExitValue.end())
4853 if (BEs.ugt(MaxBruteForceIterations))
4854 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4856 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4858 DenseMap<Instruction *, Constant *> CurrentIterVals;
4859 BasicBlock *Header = L->getHeader();
4860 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
4862 // Since the loop is canonicalized, the PHI node must have two entries. One
4863 // entry must be a constant (coming in from outside of the loop), and the
4864 // second must be derived from the same PHI.
4865 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4867 for (BasicBlock::iterator I = Header->begin();
4868 (PHI = dyn_cast<PHINode>(I)); ++I) {
4869 Constant *StartCST =
4870 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
4871 if (StartCST == 0) continue;
4872 CurrentIterVals[PHI] = StartCST;
4874 if (!CurrentIterVals.count(PN))
4877 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4879 // Execute the loop symbolically to determine the exit value.
4880 if (BEs.getActiveBits() >= 32)
4881 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4883 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4884 unsigned IterationNum = 0;
4885 for (; ; ++IterationNum) {
4886 if (IterationNum == NumIterations)
4887 return RetVal = CurrentIterVals[PN]; // Got exit value!
4889 // Compute the value of the PHIs for the next iteration.
4890 // EvaluateExpression adds non-phi values to the CurrentIterVals map.
4891 DenseMap<Instruction *, Constant *> NextIterVals;
4892 Constant *NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD,
4895 return 0; // Couldn't evaluate!
4896 NextIterVals[PN] = NextPHI;
4898 bool StoppedEvolving = NextPHI == CurrentIterVals[PN];
4900 // Also evaluate the other PHI nodes. However, we don't get to stop if we
4901 // cease to be able to evaluate one of them or if they stop evolving,
4902 // because that doesn't necessarily prevent us from computing PN.
4903 SmallVector<std::pair<PHINode *, Constant *>, 8> PHIsToCompute;
4904 for (DenseMap<Instruction *, Constant *>::const_iterator
4905 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
4906 PHINode *PHI = dyn_cast<PHINode>(I->first);
4907 if (!PHI || PHI == PN || PHI->getParent() != Header) continue;
4908 PHIsToCompute.push_back(std::make_pair(PHI, I->second));
4910 // We use two distinct loops because EvaluateExpression may invalidate any
4911 // iterators into CurrentIterVals.
4912 for (SmallVectorImpl<std::pair<PHINode *, Constant*> >::const_iterator
4913 I = PHIsToCompute.begin(), E = PHIsToCompute.end(); I != E; ++I) {
4914 PHINode *PHI = I->first;
4915 Constant *&NextPHI = NextIterVals[PHI];
4916 if (!NextPHI) { // Not already computed.
4917 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
4918 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
4920 if (NextPHI != I->second)
4921 StoppedEvolving = false;
4924 // If all entries in CurrentIterVals == NextIterVals then we can stop
4925 // iterating, the loop can't continue to change.
4926 if (StoppedEvolving)
4927 return RetVal = CurrentIterVals[PN];
4929 CurrentIterVals.swap(NextIterVals);
4933 /// ComputeExitCountExhaustively - If the loop is known to execute a
4934 /// constant number of times (the condition evolves only from constants),
4935 /// try to evaluate a few iterations of the loop until we get the exit
4936 /// condition gets a value of ExitWhen (true or false). If we cannot
4937 /// evaluate the trip count of the loop, return getCouldNotCompute().
4938 const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L,
4941 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4942 if (PN == 0) return getCouldNotCompute();
4944 // If the loop is canonicalized, the PHI will have exactly two entries.
4945 // That's the only form we support here.
4946 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4948 DenseMap<Instruction *, Constant *> CurrentIterVals;
4949 BasicBlock *Header = L->getHeader();
4950 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
4952 // One entry must be a constant (coming in from outside of the loop), and the
4953 // second must be derived from the same PHI.
4954 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4956 for (BasicBlock::iterator I = Header->begin();
4957 (PHI = dyn_cast<PHINode>(I)); ++I) {
4958 Constant *StartCST =
4959 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
4960 if (StartCST == 0) continue;
4961 CurrentIterVals[PHI] = StartCST;
4963 if (!CurrentIterVals.count(PN))
4964 return getCouldNotCompute();
4966 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4967 // the loop symbolically to determine when the condition gets a value of
4970 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4971 for (unsigned IterationNum = 0; IterationNum != MaxIterations;++IterationNum){
4972 ConstantInt *CondVal =
4973 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, L, CurrentIterVals,
4976 // Couldn't symbolically evaluate.
4977 if (!CondVal) return getCouldNotCompute();
4979 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4980 ++NumBruteForceTripCountsComputed;
4981 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4984 // Update all the PHI nodes for the next iteration.
4985 DenseMap<Instruction *, Constant *> NextIterVals;
4987 // Create a list of which PHIs we need to compute. We want to do this before
4988 // calling EvaluateExpression on them because that may invalidate iterators
4989 // into CurrentIterVals.
4990 SmallVector<PHINode *, 8> PHIsToCompute;
4991 for (DenseMap<Instruction *, Constant *>::const_iterator
4992 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
4993 PHINode *PHI = dyn_cast<PHINode>(I->first);
4994 if (!PHI || PHI->getParent() != Header) continue;
4995 PHIsToCompute.push_back(PHI);
4997 for (SmallVectorImpl<PHINode *>::const_iterator I = PHIsToCompute.begin(),
4998 E = PHIsToCompute.end(); I != E; ++I) {
5000 Constant *&NextPHI = NextIterVals[PHI];
5001 if (NextPHI) continue; // Already computed!
5003 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
5004 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
5006 CurrentIterVals.swap(NextIterVals);
5009 // Too many iterations were needed to evaluate.
5010 return getCouldNotCompute();
5013 /// getSCEVAtScope - Return a SCEV expression for the specified value
5014 /// at the specified scope in the program. The L value specifies a loop
5015 /// nest to evaluate the expression at, where null is the top-level or a
5016 /// specified loop is immediately inside of the loop.
5018 /// This method can be used to compute the exit value for a variable defined
5019 /// in a loop by querying what the value will hold in the parent loop.
5021 /// In the case that a relevant loop exit value cannot be computed, the
5022 /// original value V is returned.
5023 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
5024 // Check to see if we've folded this expression at this loop before.
5025 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
5026 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
5027 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
5029 return Pair.first->second ? Pair.first->second : V;
5031 // Otherwise compute it.
5032 const SCEV *C = computeSCEVAtScope(V, L);
5033 ValuesAtScopes[V][L] = C;
5037 /// This builds up a Constant using the ConstantExpr interface. That way, we
5038 /// will return Constants for objects which aren't represented by a
5039 /// SCEVConstant, because SCEVConstant is restricted to ConstantInt.
5040 /// Returns NULL if the SCEV isn't representable as a Constant.
5041 static Constant *BuildConstantFromSCEV(const SCEV *V) {
5042 switch (V->getSCEVType()) {
5043 default: // TODO: smax, umax.
5044 case scCouldNotCompute:
5048 return cast<SCEVConstant>(V)->getValue();
5050 return dyn_cast<Constant>(cast<SCEVUnknown>(V)->getValue());
5051 case scSignExtend: {
5052 const SCEVSignExtendExpr *SS = cast<SCEVSignExtendExpr>(V);
5053 if (Constant *CastOp = BuildConstantFromSCEV(SS->getOperand()))
5054 return ConstantExpr::getSExt(CastOp, SS->getType());
5057 case scZeroExtend: {
5058 const SCEVZeroExtendExpr *SZ = cast<SCEVZeroExtendExpr>(V);
5059 if (Constant *CastOp = BuildConstantFromSCEV(SZ->getOperand()))
5060 return ConstantExpr::getZExt(CastOp, SZ->getType());
5064 const SCEVTruncateExpr *ST = cast<SCEVTruncateExpr>(V);
5065 if (Constant *CastOp = BuildConstantFromSCEV(ST->getOperand()))
5066 return ConstantExpr::getTrunc(CastOp, ST->getType());
5070 const SCEVAddExpr *SA = cast<SCEVAddExpr>(V);
5071 if (Constant *C = BuildConstantFromSCEV(SA->getOperand(0))) {
5072 if (C->getType()->isPointerTy())
5073 C = ConstantExpr::getBitCast(C, Type::getInt8PtrTy(C->getContext()));
5074 for (unsigned i = 1, e = SA->getNumOperands(); i != e; ++i) {
5075 Constant *C2 = BuildConstantFromSCEV(SA->getOperand(i));
5079 if (!C->getType()->isPointerTy() && C2->getType()->isPointerTy()) {
5081 // The offsets have been converted to bytes. We can add bytes to an
5082 // i8* by GEP with the byte count in the first index.
5083 C = ConstantExpr::getBitCast(C,Type::getInt8PtrTy(C->getContext()));
5086 // Don't bother trying to sum two pointers. We probably can't
5087 // statically compute a load that results from it anyway.
5088 if (C2->getType()->isPointerTy())
5091 if (C->getType()->isPointerTy()) {
5092 if (cast<PointerType>(C->getType())->getElementType()->isStructTy())
5093 C2 = ConstantExpr::getIntegerCast(
5094 C2, Type::getInt32Ty(C->getContext()), true);
5095 C = ConstantExpr::getGetElementPtr(C, C2);
5097 C = ConstantExpr::getAdd(C, C2);
5104 const SCEVMulExpr *SM = cast<SCEVMulExpr>(V);
5105 if (Constant *C = BuildConstantFromSCEV(SM->getOperand(0))) {
5106 // Don't bother with pointers at all.
5107 if (C->getType()->isPointerTy()) return 0;
5108 for (unsigned i = 1, e = SM->getNumOperands(); i != e; ++i) {
5109 Constant *C2 = BuildConstantFromSCEV(SM->getOperand(i));
5110 if (!C2 || C2->getType()->isPointerTy()) return 0;
5111 C = ConstantExpr::getMul(C, C2);
5118 const SCEVUDivExpr *SU = cast<SCEVUDivExpr>(V);
5119 if (Constant *LHS = BuildConstantFromSCEV(SU->getLHS()))
5120 if (Constant *RHS = BuildConstantFromSCEV(SU->getRHS()))
5121 if (LHS->getType() == RHS->getType())
5122 return ConstantExpr::getUDiv(LHS, RHS);
5129 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
5130 if (isa<SCEVConstant>(V)) return V;
5132 // If this instruction is evolved from a constant-evolving PHI, compute the
5133 // exit value from the loop without using SCEVs.
5134 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
5135 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
5136 const Loop *LI = (*this->LI)[I->getParent()];
5137 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
5138 if (PHINode *PN = dyn_cast<PHINode>(I))
5139 if (PN->getParent() == LI->getHeader()) {
5140 // Okay, there is no closed form solution for the PHI node. Check
5141 // to see if the loop that contains it has a known backedge-taken
5142 // count. If so, we may be able to force computation of the exit
5144 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
5145 if (const SCEVConstant *BTCC =
5146 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
5147 // Okay, we know how many times the containing loop executes. If
5148 // this is a constant evolving PHI node, get the final value at
5149 // the specified iteration number.
5150 Constant *RV = getConstantEvolutionLoopExitValue(PN,
5151 BTCC->getValue()->getValue(),
5153 if (RV) return getSCEV(RV);
5157 // Okay, this is an expression that we cannot symbolically evaluate
5158 // into a SCEV. Check to see if it's possible to symbolically evaluate
5159 // the arguments into constants, and if so, try to constant propagate the
5160 // result. This is particularly useful for computing loop exit values.
5161 if (CanConstantFold(I)) {
5162 SmallVector<Constant *, 4> Operands;
5163 bool MadeImprovement = false;
5164 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
5165 Value *Op = I->getOperand(i);
5166 if (Constant *C = dyn_cast<Constant>(Op)) {
5167 Operands.push_back(C);
5171 // If any of the operands is non-constant and if they are
5172 // non-integer and non-pointer, don't even try to analyze them
5173 // with scev techniques.
5174 if (!isSCEVable(Op->getType()))
5177 const SCEV *OrigV = getSCEV(Op);
5178 const SCEV *OpV = getSCEVAtScope(OrigV, L);
5179 MadeImprovement |= OrigV != OpV;
5181 Constant *C = BuildConstantFromSCEV(OpV);
5183 if (C->getType() != Op->getType())
5184 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5188 Operands.push_back(C);
5191 // Check to see if getSCEVAtScope actually made an improvement.
5192 if (MadeImprovement) {
5194 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
5195 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
5196 Operands[0], Operands[1], TD,
5198 else if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
5199 if (!LI->isVolatile())
5200 C = ConstantFoldLoadFromConstPtr(Operands[0], TD);
5202 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
5210 // This is some other type of SCEVUnknown, just return it.
5214 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
5215 // Avoid performing the look-up in the common case where the specified
5216 // expression has no loop-variant portions.
5217 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
5218 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5219 if (OpAtScope != Comm->getOperand(i)) {
5220 // Okay, at least one of these operands is loop variant but might be
5221 // foldable. Build a new instance of the folded commutative expression.
5222 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
5223 Comm->op_begin()+i);
5224 NewOps.push_back(OpAtScope);
5226 for (++i; i != e; ++i) {
5227 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
5228 NewOps.push_back(OpAtScope);
5230 if (isa<SCEVAddExpr>(Comm))
5231 return getAddExpr(NewOps);
5232 if (isa<SCEVMulExpr>(Comm))
5233 return getMulExpr(NewOps);
5234 if (isa<SCEVSMaxExpr>(Comm))
5235 return getSMaxExpr(NewOps);
5236 if (isa<SCEVUMaxExpr>(Comm))
5237 return getUMaxExpr(NewOps);
5238 llvm_unreachable("Unknown commutative SCEV type!");
5241 // If we got here, all operands are loop invariant.
5245 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
5246 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
5247 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
5248 if (LHS == Div->getLHS() && RHS == Div->getRHS())
5249 return Div; // must be loop invariant
5250 return getUDivExpr(LHS, RHS);
5253 // If this is a loop recurrence for a loop that does not contain L, then we
5254 // are dealing with the final value computed by the loop.
5255 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
5256 // First, attempt to evaluate each operand.
5257 // Avoid performing the look-up in the common case where the specified
5258 // expression has no loop-variant portions.
5259 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
5260 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
5261 if (OpAtScope == AddRec->getOperand(i))
5264 // Okay, at least one of these operands is loop variant but might be
5265 // foldable. Build a new instance of the folded commutative expression.
5266 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
5267 AddRec->op_begin()+i);
5268 NewOps.push_back(OpAtScope);
5269 for (++i; i != e; ++i)
5270 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
5272 const SCEV *FoldedRec =
5273 getAddRecExpr(NewOps, AddRec->getLoop(),
5274 AddRec->getNoWrapFlags(SCEV::FlagNW));
5275 AddRec = dyn_cast<SCEVAddRecExpr>(FoldedRec);
5276 // The addrec may be folded to a nonrecurrence, for example, if the
5277 // induction variable is multiplied by zero after constant folding. Go
5278 // ahead and return the folded value.
5284 // If the scope is outside the addrec's loop, evaluate it by using the
5285 // loop exit value of the addrec.
5286 if (!AddRec->getLoop()->contains(L)) {
5287 // To evaluate this recurrence, we need to know how many times the AddRec
5288 // loop iterates. Compute this now.
5289 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
5290 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
5292 // Then, evaluate the AddRec.
5293 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
5299 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
5300 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5301 if (Op == Cast->getOperand())
5302 return Cast; // must be loop invariant
5303 return getZeroExtendExpr(Op, Cast->getType());
5306 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
5307 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5308 if (Op == Cast->getOperand())
5309 return Cast; // must be loop invariant
5310 return getSignExtendExpr(Op, Cast->getType());
5313 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
5314 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5315 if (Op == Cast->getOperand())
5316 return Cast; // must be loop invariant
5317 return getTruncateExpr(Op, Cast->getType());
5320 llvm_unreachable("Unknown SCEV type!");
5323 /// getSCEVAtScope - This is a convenience function which does
5324 /// getSCEVAtScope(getSCEV(V), L).
5325 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
5326 return getSCEVAtScope(getSCEV(V), L);
5329 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
5330 /// following equation:
5332 /// A * X = B (mod N)
5334 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
5335 /// A and B isn't important.
5337 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
5338 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
5339 ScalarEvolution &SE) {
5340 uint32_t BW = A.getBitWidth();
5341 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
5342 assert(A != 0 && "A must be non-zero.");
5346 // The gcd of A and N may have only one prime factor: 2. The number of
5347 // trailing zeros in A is its multiplicity
5348 uint32_t Mult2 = A.countTrailingZeros();
5351 // 2. Check if B is divisible by D.
5353 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
5354 // is not less than multiplicity of this prime factor for D.
5355 if (B.countTrailingZeros() < Mult2)
5356 return SE.getCouldNotCompute();
5358 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
5361 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
5362 // bit width during computations.
5363 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
5364 APInt Mod(BW + 1, 0);
5365 Mod.setBit(BW - Mult2); // Mod = N / D
5366 APInt I = AD.multiplicativeInverse(Mod);
5368 // 4. Compute the minimum unsigned root of the equation:
5369 // I * (B / D) mod (N / D)
5370 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
5372 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
5374 return SE.getConstant(Result.trunc(BW));
5377 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
5378 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
5379 /// might be the same) or two SCEVCouldNotCompute objects.
5381 static std::pair<const SCEV *,const SCEV *>
5382 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
5383 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
5384 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
5385 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
5386 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
5388 // We currently can only solve this if the coefficients are constants.
5389 if (!LC || !MC || !NC) {
5390 const SCEV *CNC = SE.getCouldNotCompute();
5391 return std::make_pair(CNC, CNC);
5394 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
5395 const APInt &L = LC->getValue()->getValue();
5396 const APInt &M = MC->getValue()->getValue();
5397 const APInt &N = NC->getValue()->getValue();
5398 APInt Two(BitWidth, 2);
5399 APInt Four(BitWidth, 4);
5402 using namespace APIntOps;
5404 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
5405 // The B coefficient is M-N/2
5409 // The A coefficient is N/2
5410 APInt A(N.sdiv(Two));
5412 // Compute the B^2-4ac term.
5415 SqrtTerm -= Four * (A * C);
5417 if (SqrtTerm.isNegative()) {
5418 // The loop is provably infinite.
5419 const SCEV *CNC = SE.getCouldNotCompute();
5420 return std::make_pair(CNC, CNC);
5423 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
5424 // integer value or else APInt::sqrt() will assert.
5425 APInt SqrtVal(SqrtTerm.sqrt());
5427 // Compute the two solutions for the quadratic formula.
5428 // The divisions must be performed as signed divisions.
5431 if (TwoA.isMinValue()) {
5432 const SCEV *CNC = SE.getCouldNotCompute();
5433 return std::make_pair(CNC, CNC);
5436 LLVMContext &Context = SE.getContext();
5438 ConstantInt *Solution1 =
5439 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
5440 ConstantInt *Solution2 =
5441 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
5443 return std::make_pair(SE.getConstant(Solution1),
5444 SE.getConstant(Solution2));
5445 } // end APIntOps namespace
5448 /// HowFarToZero - Return the number of times a backedge comparing the specified
5449 /// value to zero will execute. If not computable, return CouldNotCompute.
5451 /// This is only used for loops with a "x != y" exit test. The exit condition is
5452 /// now expressed as a single expression, V = x-y. So the exit test is
5453 /// effectively V != 0. We know and take advantage of the fact that this
5454 /// expression only being used in a comparison by zero context.
5455 ScalarEvolution::ExitLimit
5456 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L, bool IsSubExpr) {
5457 // If the value is a constant
5458 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5459 // If the value is already zero, the branch will execute zero times.
5460 if (C->getValue()->isZero()) return C;
5461 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5464 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
5465 if (!AddRec || AddRec->getLoop() != L)
5466 return getCouldNotCompute();
5468 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
5469 // the quadratic equation to solve it.
5470 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
5471 std::pair<const SCEV *,const SCEV *> Roots =
5472 SolveQuadraticEquation(AddRec, *this);
5473 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5474 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5477 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
5478 << " sol#2: " << *R2 << "\n";
5480 // Pick the smallest positive root value.
5481 if (ConstantInt *CB =
5482 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
5485 if (CB->getZExtValue() == false)
5486 std::swap(R1, R2); // R1 is the minimum root now.
5488 // We can only use this value if the chrec ends up with an exact zero
5489 // value at this index. When solving for "X*X != 5", for example, we
5490 // should not accept a root of 2.
5491 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
5493 return R1; // We found a quadratic root!
5496 return getCouldNotCompute();
5499 // Otherwise we can only handle this if it is affine.
5500 if (!AddRec->isAffine())
5501 return getCouldNotCompute();
5503 // If this is an affine expression, the execution count of this branch is
5504 // the minimum unsigned root of the following equation:
5506 // Start + Step*N = 0 (mod 2^BW)
5510 // Step*N = -Start (mod 2^BW)
5512 // where BW is the common bit width of Start and Step.
5514 // Get the initial value for the loop.
5515 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
5516 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
5518 // For now we handle only constant steps.
5520 // TODO: Handle a nonconstant Step given AddRec<NUW>. If the
5521 // AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap
5522 // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step.
5523 // We have not yet seen any such cases.
5524 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
5525 if (StepC == 0 || StepC->getValue()->equalsInt(0))
5526 return getCouldNotCompute();
5528 // For positive steps (counting up until unsigned overflow):
5529 // N = -Start/Step (as unsigned)
5530 // For negative steps (counting down to zero):
5532 // First compute the unsigned distance from zero in the direction of Step.
5533 bool CountDown = StepC->getValue()->getValue().isNegative();
5534 const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
5536 // Handle unitary steps, which cannot wraparound.
5537 // 1*N = -Start; -1*N = Start (mod 2^BW), so:
5538 // N = Distance (as unsigned)
5539 if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) {
5540 ConstantRange CR = getUnsignedRange(Start);
5541 const SCEV *MaxBECount;
5542 if (!CountDown && CR.getUnsignedMin().isMinValue())
5543 // When counting up, the worst starting value is 1, not 0.
5544 MaxBECount = CR.getUnsignedMax().isMinValue()
5545 ? getConstant(APInt::getMinValue(CR.getBitWidth()))
5546 : getConstant(APInt::getMaxValue(CR.getBitWidth()));
5548 MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
5549 : -CR.getUnsignedMin());
5550 return ExitLimit(Distance, MaxBECount);
5553 // If the recurrence is known not to wraparound, unsigned divide computes the
5554 // back edge count. (Ideally we would have an "isexact" bit for udiv). We know
5555 // that the value will either become zero (and thus the loop terminates), that
5556 // the loop will terminate through some other exit condition first, or that
5557 // the loop has undefined behavior. This means we can't "miss" the exit
5558 // value, even with nonunit stride.
5560 // This is only valid for expressions that directly compute the loop exit. It
5561 // is invalid for subexpressions in which the loop may exit through this
5562 // branch even if this subexpression is false. In that case, the trip count
5563 // computed by this udiv could be smaller than the number of well-defined
5565 if (!IsSubExpr && AddRec->getNoWrapFlags(SCEV::FlagNW))
5566 return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
5568 // Then, try to solve the above equation provided that Start is constant.
5569 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
5570 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
5571 -StartC->getValue()->getValue(),
5573 return getCouldNotCompute();
5576 /// HowFarToNonZero - Return the number of times a backedge checking the
5577 /// specified value for nonzero will execute. If not computable, return
5579 ScalarEvolution::ExitLimit
5580 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
5581 // Loops that look like: while (X == 0) are very strange indeed. We don't
5582 // handle them yet except for the trivial case. This could be expanded in the
5583 // future as needed.
5585 // If the value is a constant, check to see if it is known to be non-zero
5586 // already. If so, the backedge will execute zero times.
5587 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5588 if (!C->getValue()->isNullValue())
5589 return getConstant(C->getType(), 0);
5590 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5593 // We could implement others, but I really doubt anyone writes loops like
5594 // this, and if they did, they would already be constant folded.
5595 return getCouldNotCompute();
5598 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
5599 /// (which may not be an immediate predecessor) which has exactly one
5600 /// successor from which BB is reachable, or null if no such block is
5603 std::pair<BasicBlock *, BasicBlock *>
5604 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
5605 // If the block has a unique predecessor, then there is no path from the
5606 // predecessor to the block that does not go through the direct edge
5607 // from the predecessor to the block.
5608 if (BasicBlock *Pred = BB->getSinglePredecessor())
5609 return std::make_pair(Pred, BB);
5611 // A loop's header is defined to be a block that dominates the loop.
5612 // If the header has a unique predecessor outside the loop, it must be
5613 // a block that has exactly one successor that can reach the loop.
5614 if (Loop *L = LI->getLoopFor(BB))
5615 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
5617 return std::pair<BasicBlock *, BasicBlock *>();
5620 /// HasSameValue - SCEV structural equivalence is usually sufficient for
5621 /// testing whether two expressions are equal, however for the purposes of
5622 /// looking for a condition guarding a loop, it can be useful to be a little
5623 /// more general, since a front-end may have replicated the controlling
5626 static bool HasSameValue(const SCEV *A, const SCEV *B) {
5627 // Quick check to see if they are the same SCEV.
5628 if (A == B) return true;
5630 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
5631 // two different instructions with the same value. Check for this case.
5632 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
5633 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
5634 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
5635 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
5636 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
5639 // Otherwise assume they may have a different value.
5643 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
5644 /// predicate Pred. Return true iff any changes were made.
5646 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
5647 const SCEV *&LHS, const SCEV *&RHS,
5649 bool Changed = false;
5651 // If we hit the max recursion limit bail out.
5655 // Canonicalize a constant to the right side.
5656 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
5657 // Check for both operands constant.
5658 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
5659 if (ConstantExpr::getICmp(Pred,
5661 RHSC->getValue())->isNullValue())
5662 goto trivially_false;
5664 goto trivially_true;
5666 // Otherwise swap the operands to put the constant on the right.
5667 std::swap(LHS, RHS);
5668 Pred = ICmpInst::getSwappedPredicate(Pred);
5672 // If we're comparing an addrec with a value which is loop-invariant in the
5673 // addrec's loop, put the addrec on the left. Also make a dominance check,
5674 // as both operands could be addrecs loop-invariant in each other's loop.
5675 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
5676 const Loop *L = AR->getLoop();
5677 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
5678 std::swap(LHS, RHS);
5679 Pred = ICmpInst::getSwappedPredicate(Pred);
5684 // If there's a constant operand, canonicalize comparisons with boundary
5685 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
5686 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
5687 const APInt &RA = RC->getValue()->getValue();
5689 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5690 case ICmpInst::ICMP_EQ:
5691 case ICmpInst::ICMP_NE:
5692 // Fold ((-1) * %a) + %b == 0 (equivalent to %b-%a == 0) into %a == %b.
5694 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(LHS))
5695 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(AE->getOperand(0)))
5696 if (AE->getNumOperands() == 2 && ME->getNumOperands() == 2 &&
5697 ME->getOperand(0)->isAllOnesValue()) {
5698 RHS = AE->getOperand(1);
5699 LHS = ME->getOperand(1);
5703 case ICmpInst::ICMP_UGE:
5704 if ((RA - 1).isMinValue()) {
5705 Pred = ICmpInst::ICMP_NE;
5706 RHS = getConstant(RA - 1);
5710 if (RA.isMaxValue()) {
5711 Pred = ICmpInst::ICMP_EQ;
5715 if (RA.isMinValue()) goto trivially_true;
5717 Pred = ICmpInst::ICMP_UGT;
5718 RHS = getConstant(RA - 1);
5721 case ICmpInst::ICMP_ULE:
5722 if ((RA + 1).isMaxValue()) {
5723 Pred = ICmpInst::ICMP_NE;
5724 RHS = getConstant(RA + 1);
5728 if (RA.isMinValue()) {
5729 Pred = ICmpInst::ICMP_EQ;
5733 if (RA.isMaxValue()) goto trivially_true;
5735 Pred = ICmpInst::ICMP_ULT;
5736 RHS = getConstant(RA + 1);
5739 case ICmpInst::ICMP_SGE:
5740 if ((RA - 1).isMinSignedValue()) {
5741 Pred = ICmpInst::ICMP_NE;
5742 RHS = getConstant(RA - 1);
5746 if (RA.isMaxSignedValue()) {
5747 Pred = ICmpInst::ICMP_EQ;
5751 if (RA.isMinSignedValue()) goto trivially_true;
5753 Pred = ICmpInst::ICMP_SGT;
5754 RHS = getConstant(RA - 1);
5757 case ICmpInst::ICMP_SLE:
5758 if ((RA + 1).isMaxSignedValue()) {
5759 Pred = ICmpInst::ICMP_NE;
5760 RHS = getConstant(RA + 1);
5764 if (RA.isMinSignedValue()) {
5765 Pred = ICmpInst::ICMP_EQ;
5769 if (RA.isMaxSignedValue()) goto trivially_true;
5771 Pred = ICmpInst::ICMP_SLT;
5772 RHS = getConstant(RA + 1);
5775 case ICmpInst::ICMP_UGT:
5776 if (RA.isMinValue()) {
5777 Pred = ICmpInst::ICMP_NE;
5781 if ((RA + 1).isMaxValue()) {
5782 Pred = ICmpInst::ICMP_EQ;
5783 RHS = getConstant(RA + 1);
5787 if (RA.isMaxValue()) goto trivially_false;
5789 case ICmpInst::ICMP_ULT:
5790 if (RA.isMaxValue()) {
5791 Pred = ICmpInst::ICMP_NE;
5795 if ((RA - 1).isMinValue()) {
5796 Pred = ICmpInst::ICMP_EQ;
5797 RHS = getConstant(RA - 1);
5801 if (RA.isMinValue()) goto trivially_false;
5803 case ICmpInst::ICMP_SGT:
5804 if (RA.isMinSignedValue()) {
5805 Pred = ICmpInst::ICMP_NE;
5809 if ((RA + 1).isMaxSignedValue()) {
5810 Pred = ICmpInst::ICMP_EQ;
5811 RHS = getConstant(RA + 1);
5815 if (RA.isMaxSignedValue()) goto trivially_false;
5817 case ICmpInst::ICMP_SLT:
5818 if (RA.isMaxSignedValue()) {
5819 Pred = ICmpInst::ICMP_NE;
5823 if ((RA - 1).isMinSignedValue()) {
5824 Pred = ICmpInst::ICMP_EQ;
5825 RHS = getConstant(RA - 1);
5829 if (RA.isMinSignedValue()) goto trivially_false;
5834 // Check for obvious equality.
5835 if (HasSameValue(LHS, RHS)) {
5836 if (ICmpInst::isTrueWhenEqual(Pred))
5837 goto trivially_true;
5838 if (ICmpInst::isFalseWhenEqual(Pred))
5839 goto trivially_false;
5842 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5843 // adding or subtracting 1 from one of the operands.
5845 case ICmpInst::ICMP_SLE:
5846 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5847 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5849 Pred = ICmpInst::ICMP_SLT;
5851 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5852 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5854 Pred = ICmpInst::ICMP_SLT;
5858 case ICmpInst::ICMP_SGE:
5859 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5860 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5862 Pred = ICmpInst::ICMP_SGT;
5864 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5865 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5867 Pred = ICmpInst::ICMP_SGT;
5871 case ICmpInst::ICMP_ULE:
5872 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5873 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5875 Pred = ICmpInst::ICMP_ULT;
5877 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5878 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5880 Pred = ICmpInst::ICMP_ULT;
5884 case ICmpInst::ICMP_UGE:
5885 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5886 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5888 Pred = ICmpInst::ICMP_UGT;
5890 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5891 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5893 Pred = ICmpInst::ICMP_UGT;
5901 // TODO: More simplifications are possible here.
5903 // Recursively simplify until we either hit a recursion limit or nothing
5906 return SimplifyICmpOperands(Pred, LHS, RHS, Depth+1);
5912 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5913 Pred = ICmpInst::ICMP_EQ;
5918 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5919 Pred = ICmpInst::ICMP_NE;
5923 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5924 return getSignedRange(S).getSignedMax().isNegative();
5927 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5928 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5931 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5932 return !getSignedRange(S).getSignedMin().isNegative();
5935 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5936 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5939 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5940 return isKnownNegative(S) || isKnownPositive(S);
5943 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5944 const SCEV *LHS, const SCEV *RHS) {
5945 // Canonicalize the inputs first.
5946 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5948 // If LHS or RHS is an addrec, check to see if the condition is true in
5949 // every iteration of the loop.
5950 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5951 if (isLoopEntryGuardedByCond(
5952 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5953 isLoopBackedgeGuardedByCond(
5954 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5956 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5957 if (isLoopEntryGuardedByCond(
5958 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5959 isLoopBackedgeGuardedByCond(
5960 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5963 // Otherwise see what can be done with known constant ranges.
5964 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5968 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5969 const SCEV *LHS, const SCEV *RHS) {
5970 if (HasSameValue(LHS, RHS))
5971 return ICmpInst::isTrueWhenEqual(Pred);
5973 // This code is split out from isKnownPredicate because it is called from
5974 // within isLoopEntryGuardedByCond.
5977 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5978 case ICmpInst::ICMP_SGT:
5979 Pred = ICmpInst::ICMP_SLT;
5980 std::swap(LHS, RHS);
5981 case ICmpInst::ICMP_SLT: {
5982 ConstantRange LHSRange = getSignedRange(LHS);
5983 ConstantRange RHSRange = getSignedRange(RHS);
5984 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5986 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5990 case ICmpInst::ICMP_SGE:
5991 Pred = ICmpInst::ICMP_SLE;
5992 std::swap(LHS, RHS);
5993 case ICmpInst::ICMP_SLE: {
5994 ConstantRange LHSRange = getSignedRange(LHS);
5995 ConstantRange RHSRange = getSignedRange(RHS);
5996 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5998 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
6002 case ICmpInst::ICMP_UGT:
6003 Pred = ICmpInst::ICMP_ULT;
6004 std::swap(LHS, RHS);
6005 case ICmpInst::ICMP_ULT: {
6006 ConstantRange LHSRange = getUnsignedRange(LHS);
6007 ConstantRange RHSRange = getUnsignedRange(RHS);
6008 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
6010 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
6014 case ICmpInst::ICMP_UGE:
6015 Pred = ICmpInst::ICMP_ULE;
6016 std::swap(LHS, RHS);
6017 case ICmpInst::ICMP_ULE: {
6018 ConstantRange LHSRange = getUnsignedRange(LHS);
6019 ConstantRange RHSRange = getUnsignedRange(RHS);
6020 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
6022 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
6026 case ICmpInst::ICMP_NE: {
6027 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
6029 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
6032 const SCEV *Diff = getMinusSCEV(LHS, RHS);
6033 if (isKnownNonZero(Diff))
6037 case ICmpInst::ICMP_EQ:
6038 // The check at the top of the function catches the case where
6039 // the values are known to be equal.
6045 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
6046 /// protected by a conditional between LHS and RHS. This is used to
6047 /// to eliminate casts.
6049 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
6050 ICmpInst::Predicate Pred,
6051 const SCEV *LHS, const SCEV *RHS) {
6052 // Interpret a null as meaning no loop, where there is obviously no guard
6053 // (interprocedural conditions notwithstanding).
6054 if (!L) return true;
6056 BasicBlock *Latch = L->getLoopLatch();
6060 BranchInst *LoopContinuePredicate =
6061 dyn_cast<BranchInst>(Latch->getTerminator());
6062 if (!LoopContinuePredicate ||
6063 LoopContinuePredicate->isUnconditional())
6066 return isImpliedCond(Pred, LHS, RHS,
6067 LoopContinuePredicate->getCondition(),
6068 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
6071 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
6072 /// by a conditional between LHS and RHS. This is used to help avoid max
6073 /// expressions in loop trip counts, and to eliminate casts.
6075 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
6076 ICmpInst::Predicate Pred,
6077 const SCEV *LHS, const SCEV *RHS) {
6078 // Interpret a null as meaning no loop, where there is obviously no guard
6079 // (interprocedural conditions notwithstanding).
6080 if (!L) return false;
6082 // Starting at the loop predecessor, climb up the predecessor chain, as long
6083 // as there are predecessors that can be found that have unique successors
6084 // leading to the original header.
6085 for (std::pair<BasicBlock *, BasicBlock *>
6086 Pair(L->getLoopPredecessor(), L->getHeader());
6088 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
6090 BranchInst *LoopEntryPredicate =
6091 dyn_cast<BranchInst>(Pair.first->getTerminator());
6092 if (!LoopEntryPredicate ||
6093 LoopEntryPredicate->isUnconditional())
6096 if (isImpliedCond(Pred, LHS, RHS,
6097 LoopEntryPredicate->getCondition(),
6098 LoopEntryPredicate->getSuccessor(0) != Pair.second))
6105 /// RAII wrapper to prevent recursive application of isImpliedCond.
6106 /// ScalarEvolution's PendingLoopPredicates set must be empty unless we are
6107 /// currently evaluating isImpliedCond.
6108 struct MarkPendingLoopPredicate {
6110 DenseSet<Value*> &LoopPreds;
6113 MarkPendingLoopPredicate(Value *C, DenseSet<Value*> &LP)
6114 : Cond(C), LoopPreds(LP) {
6115 Pending = !LoopPreds.insert(Cond).second;
6117 ~MarkPendingLoopPredicate() {
6119 LoopPreds.erase(Cond);
6123 /// isImpliedCond - Test whether the condition described by Pred, LHS,
6124 /// and RHS is true whenever the given Cond value evaluates to true.
6125 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
6126 const SCEV *LHS, const SCEV *RHS,
6127 Value *FoundCondValue,
6129 MarkPendingLoopPredicate Mark(FoundCondValue, PendingLoopPredicates);
6133 // Recursively handle And and Or conditions.
6134 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
6135 if (BO->getOpcode() == Instruction::And) {
6137 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6138 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6139 } else if (BO->getOpcode() == Instruction::Or) {
6141 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
6142 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
6146 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
6147 if (!ICI) return false;
6149 // Bail if the ICmp's operands' types are wider than the needed type
6150 // before attempting to call getSCEV on them. This avoids infinite
6151 // recursion, since the analysis of widening casts can require loop
6152 // exit condition information for overflow checking, which would
6154 if (getTypeSizeInBits(LHS->getType()) <
6155 getTypeSizeInBits(ICI->getOperand(0)->getType()))
6158 // Now that we found a conditional branch that dominates the loop or controls
6159 // the loop latch. Check to see if it is the comparison we are looking for.
6160 ICmpInst::Predicate FoundPred;
6162 FoundPred = ICI->getInversePredicate();
6164 FoundPred = ICI->getPredicate();
6166 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
6167 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
6169 // Balance the types. The case where FoundLHS' type is wider than
6170 // LHS' type is checked for above.
6171 if (getTypeSizeInBits(LHS->getType()) >
6172 getTypeSizeInBits(FoundLHS->getType())) {
6173 if (CmpInst::isSigned(Pred)) {
6174 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
6175 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
6177 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
6178 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
6182 // Canonicalize the query to match the way instcombine will have
6183 // canonicalized the comparison.
6184 if (SimplifyICmpOperands(Pred, LHS, RHS))
6186 return CmpInst::isTrueWhenEqual(Pred);
6187 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
6188 if (FoundLHS == FoundRHS)
6189 return CmpInst::isFalseWhenEqual(FoundPred);
6191 // Check to see if we can make the LHS or RHS match.
6192 if (LHS == FoundRHS || RHS == FoundLHS) {
6193 if (isa<SCEVConstant>(RHS)) {
6194 std::swap(FoundLHS, FoundRHS);
6195 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
6197 std::swap(LHS, RHS);
6198 Pred = ICmpInst::getSwappedPredicate(Pred);
6202 // Check whether the found predicate is the same as the desired predicate.
6203 if (FoundPred == Pred)
6204 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
6206 // Check whether swapping the found predicate makes it the same as the
6207 // desired predicate.
6208 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
6209 if (isa<SCEVConstant>(RHS))
6210 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
6212 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
6213 RHS, LHS, FoundLHS, FoundRHS);
6216 // Check whether the actual condition is beyond sufficient.
6217 if (FoundPred == ICmpInst::ICMP_EQ)
6218 if (ICmpInst::isTrueWhenEqual(Pred))
6219 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
6221 if (Pred == ICmpInst::ICMP_NE)
6222 if (!ICmpInst::isTrueWhenEqual(FoundPred))
6223 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
6226 // Otherwise assume the worst.
6230 /// isImpliedCondOperands - Test whether the condition described by Pred,
6231 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
6232 /// and FoundRHS is true.
6233 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
6234 const SCEV *LHS, const SCEV *RHS,
6235 const SCEV *FoundLHS,
6236 const SCEV *FoundRHS) {
6237 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
6238 FoundLHS, FoundRHS) ||
6239 // ~x < ~y --> x > y
6240 isImpliedCondOperandsHelper(Pred, LHS, RHS,
6241 getNotSCEV(FoundRHS),
6242 getNotSCEV(FoundLHS));
6245 /// isImpliedCondOperandsHelper - Test whether the condition described by
6246 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
6247 /// FoundLHS, and FoundRHS is true.
6249 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
6250 const SCEV *LHS, const SCEV *RHS,
6251 const SCEV *FoundLHS,
6252 const SCEV *FoundRHS) {
6254 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
6255 case ICmpInst::ICMP_EQ:
6256 case ICmpInst::ICMP_NE:
6257 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
6260 case ICmpInst::ICMP_SLT:
6261 case ICmpInst::ICMP_SLE:
6262 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
6263 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
6266 case ICmpInst::ICMP_SGT:
6267 case ICmpInst::ICMP_SGE:
6268 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
6269 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
6272 case ICmpInst::ICMP_ULT:
6273 case ICmpInst::ICMP_ULE:
6274 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
6275 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
6278 case ICmpInst::ICMP_UGT:
6279 case ICmpInst::ICMP_UGE:
6280 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
6281 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
6289 /// getBECount - Subtract the end and start values and divide by the step,
6290 /// rounding up, to get the number of times the backedge is executed. Return
6291 /// CouldNotCompute if an intermediate computation overflows.
6292 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
6296 assert(!isKnownNegative(Step) &&
6297 "This code doesn't handle negative strides yet!");
6299 Type *Ty = Start->getType();
6301 // When Start == End, we have an exact BECount == 0. Short-circuit this case
6302 // here because SCEV may not be able to determine that the unsigned division
6303 // after rounding is zero.
6305 return getConstant(Ty, 0);
6307 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
6308 const SCEV *Diff = getMinusSCEV(End, Start);
6309 const SCEV *RoundUp = getAddExpr(Step, NegOne);
6311 // Add an adjustment to the difference between End and Start so that
6312 // the division will effectively round up.
6313 const SCEV *Add = getAddExpr(Diff, RoundUp);
6316 // Check Add for unsigned overflow.
6317 // TODO: More sophisticated things could be done here.
6318 Type *WideTy = IntegerType::get(getContext(),
6319 getTypeSizeInBits(Ty) + 1);
6320 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
6321 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
6322 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
6323 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
6324 return getCouldNotCompute();
6327 return getUDivExpr(Add, Step);
6330 /// HowManyLessThans - Return the number of times a backedge containing the
6331 /// specified less-than comparison will execute. If not computable, return
6332 /// CouldNotCompute.
6334 /// @param IsSubExpr is true when the LHS < RHS condition does not directly
6335 /// control the branch. In this case, we can only compute an iteration count for
6336 /// a subexpression that cannot overflow before evaluating true.
6337 ScalarEvolution::ExitLimit
6338 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
6339 const Loop *L, bool isSigned,
6341 // Only handle: "ADDREC < LoopInvariant".
6342 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute();
6344 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
6345 if (!AddRec || AddRec->getLoop() != L)
6346 return getCouldNotCompute();
6348 // Check to see if we have a flag which makes analysis easy.
6349 bool NoWrap = false;
6351 NoWrap = AddRec->getNoWrapFlags(
6352 (SCEV::NoWrapFlags)(((isSigned ? SCEV::FlagNSW : SCEV::FlagNUW))
6355 if (AddRec->isAffine()) {
6356 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
6357 const SCEV *Step = AddRec->getStepRecurrence(*this);
6360 return getCouldNotCompute();
6361 if (Step->isOne()) {
6362 // With unit stride, the iteration never steps past the limit value.
6363 } else if (isKnownPositive(Step)) {
6364 // Test whether a positive iteration can step past the limit
6365 // value and past the maximum value for its type in a single step.
6366 // Note that it's not sufficient to check NoWrap here, because even
6367 // though the value after a wrap is undefined, it's not undefined
6368 // behavior, so if wrap does occur, the loop could either terminate or
6369 // loop infinitely, but in either case, the loop is guaranteed to
6370 // iterate at least until the iteration where the wrapping occurs.
6371 const SCEV *One = getConstant(Step->getType(), 1);
6373 APInt Max = APInt::getSignedMaxValue(BitWidth);
6374 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
6375 .slt(getSignedRange(RHS).getSignedMax()))
6376 return getCouldNotCompute();
6378 APInt Max = APInt::getMaxValue(BitWidth);
6379 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
6380 .ult(getUnsignedRange(RHS).getUnsignedMax()))
6381 return getCouldNotCompute();
6384 // TODO: Handle negative strides here and below.
6385 return getCouldNotCompute();
6387 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
6388 // m. So, we count the number of iterations in which {n,+,s} < m is true.
6389 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
6390 // treat m-n as signed nor unsigned due to overflow possibility.
6392 // First, we get the value of the LHS in the first iteration: n
6393 const SCEV *Start = AddRec->getOperand(0);
6395 // Determine the minimum constant start value.
6396 const SCEV *MinStart = getConstant(isSigned ?
6397 getSignedRange(Start).getSignedMin() :
6398 getUnsignedRange(Start).getUnsignedMin());
6400 // If we know that the condition is true in order to enter the loop,
6401 // then we know that it will run exactly (m-n)/s times. Otherwise, we
6402 // only know that it will execute (max(m,n)-n)/s times. In both cases,
6403 // the division must round up.
6404 const SCEV *End = RHS;
6405 if (!isLoopEntryGuardedByCond(L,
6406 isSigned ? ICmpInst::ICMP_SLT :
6408 getMinusSCEV(Start, Step), RHS))
6409 End = isSigned ? getSMaxExpr(RHS, Start)
6410 : getUMaxExpr(RHS, Start);
6412 // Determine the maximum constant end value.
6413 const SCEV *MaxEnd = getConstant(isSigned ?
6414 getSignedRange(End).getSignedMax() :
6415 getUnsignedRange(End).getUnsignedMax());
6417 // If MaxEnd is within a step of the maximum integer value in its type,
6418 // adjust it down to the minimum value which would produce the same effect.
6419 // This allows the subsequent ceiling division of (N+(step-1))/step to
6420 // compute the correct value.
6421 const SCEV *StepMinusOne = getMinusSCEV(Step,
6422 getConstant(Step->getType(), 1));
6425 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
6428 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
6431 // Finally, we subtract these two values and divide, rounding up, to get
6432 // the number of times the backedge is executed.
6433 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
6435 // The maximum backedge count is similar, except using the minimum start
6436 // value and the maximum end value.
6437 // If we already have an exact constant BECount, use it instead.
6438 const SCEV *MaxBECount = isa<SCEVConstant>(BECount) ? BECount
6439 : getBECount(MinStart, MaxEnd, Step, NoWrap);
6441 // If the stride is nonconstant, and NoWrap == true, then
6442 // getBECount(MinStart, MaxEnd) may not compute. This would result in an
6443 // exact BECount and invalid MaxBECount, which should be avoided to catch
6444 // more optimization opportunities.
6445 if (isa<SCEVCouldNotCompute>(MaxBECount))
6446 MaxBECount = BECount;
6448 return ExitLimit(BECount, MaxBECount);
6451 return getCouldNotCompute();
6454 /// getNumIterationsInRange - Return the number of iterations of this loop that
6455 /// produce values in the specified constant range. Another way of looking at
6456 /// this is that it returns the first iteration number where the value is not in
6457 /// the condition, thus computing the exit count. If the iteration count can't
6458 /// be computed, an instance of SCEVCouldNotCompute is returned.
6459 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
6460 ScalarEvolution &SE) const {
6461 if (Range.isFullSet()) // Infinite loop.
6462 return SE.getCouldNotCompute();
6464 // If the start is a non-zero constant, shift the range to simplify things.
6465 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
6466 if (!SC->getValue()->isZero()) {
6467 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
6468 Operands[0] = SE.getConstant(SC->getType(), 0);
6469 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop(),
6470 getNoWrapFlags(FlagNW));
6471 if (const SCEVAddRecExpr *ShiftedAddRec =
6472 dyn_cast<SCEVAddRecExpr>(Shifted))
6473 return ShiftedAddRec->getNumIterationsInRange(
6474 Range.subtract(SC->getValue()->getValue()), SE);
6475 // This is strange and shouldn't happen.
6476 return SE.getCouldNotCompute();
6479 // The only time we can solve this is when we have all constant indices.
6480 // Otherwise, we cannot determine the overflow conditions.
6481 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
6482 if (!isa<SCEVConstant>(getOperand(i)))
6483 return SE.getCouldNotCompute();
6486 // Okay at this point we know that all elements of the chrec are constants and
6487 // that the start element is zero.
6489 // First check to see if the range contains zero. If not, the first
6491 unsigned BitWidth = SE.getTypeSizeInBits(getType());
6492 if (!Range.contains(APInt(BitWidth, 0)))
6493 return SE.getConstant(getType(), 0);
6496 // If this is an affine expression then we have this situation:
6497 // Solve {0,+,A} in Range === Ax in Range
6499 // We know that zero is in the range. If A is positive then we know that
6500 // the upper value of the range must be the first possible exit value.
6501 // If A is negative then the lower of the range is the last possible loop
6502 // value. Also note that we already checked for a full range.
6503 APInt One(BitWidth,1);
6504 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
6505 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
6507 // The exit value should be (End+A)/A.
6508 APInt ExitVal = (End + A).udiv(A);
6509 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
6511 // Evaluate at the exit value. If we really did fall out of the valid
6512 // range, then we computed our trip count, otherwise wrap around or other
6513 // things must have happened.
6514 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
6515 if (Range.contains(Val->getValue()))
6516 return SE.getCouldNotCompute(); // Something strange happened
6518 // Ensure that the previous value is in the range. This is a sanity check.
6519 assert(Range.contains(
6520 EvaluateConstantChrecAtConstant(this,
6521 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
6522 "Linear scev computation is off in a bad way!");
6523 return SE.getConstant(ExitValue);
6524 } else if (isQuadratic()) {
6525 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
6526 // quadratic equation to solve it. To do this, we must frame our problem in
6527 // terms of figuring out when zero is crossed, instead of when
6528 // Range.getUpper() is crossed.
6529 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
6530 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
6531 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop(),
6532 // getNoWrapFlags(FlagNW)
6535 // Next, solve the constructed addrec
6536 std::pair<const SCEV *,const SCEV *> Roots =
6537 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
6538 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
6539 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
6541 // Pick the smallest positive root value.
6542 if (ConstantInt *CB =
6543 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
6544 R1->getValue(), R2->getValue()))) {
6545 if (CB->getZExtValue() == false)
6546 std::swap(R1, R2); // R1 is the minimum root now.
6548 // Make sure the root is not off by one. The returned iteration should
6549 // not be in the range, but the previous one should be. When solving
6550 // for "X*X < 5", for example, we should not return a root of 2.
6551 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
6554 if (Range.contains(R1Val->getValue())) {
6555 // The next iteration must be out of the range...
6556 ConstantInt *NextVal =
6557 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
6559 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6560 if (!Range.contains(R1Val->getValue()))
6561 return SE.getConstant(NextVal);
6562 return SE.getCouldNotCompute(); // Something strange happened
6565 // If R1 was not in the range, then it is a good return value. Make
6566 // sure that R1-1 WAS in the range though, just in case.
6567 ConstantInt *NextVal =
6568 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
6569 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6570 if (Range.contains(R1Val->getValue()))
6572 return SE.getCouldNotCompute(); // Something strange happened
6577 return SE.getCouldNotCompute();
6582 //===----------------------------------------------------------------------===//
6583 // SCEVCallbackVH Class Implementation
6584 //===----------------------------------------------------------------------===//
6586 void ScalarEvolution::SCEVCallbackVH::deleted() {
6587 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6588 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
6589 SE->ConstantEvolutionLoopExitValue.erase(PN);
6590 SE->ValueExprMap.erase(getValPtr());
6591 // this now dangles!
6594 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
6595 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6597 // Forget all the expressions associated with users of the old value,
6598 // so that future queries will recompute the expressions using the new
6600 Value *Old = getValPtr();
6601 SmallVector<User *, 16> Worklist;
6602 SmallPtrSet<User *, 8> Visited;
6603 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
6605 Worklist.push_back(*UI);
6606 while (!Worklist.empty()) {
6607 User *U = Worklist.pop_back_val();
6608 // Deleting the Old value will cause this to dangle. Postpone
6609 // that until everything else is done.
6612 if (!Visited.insert(U))
6614 if (PHINode *PN = dyn_cast<PHINode>(U))
6615 SE->ConstantEvolutionLoopExitValue.erase(PN);
6616 SE->ValueExprMap.erase(U);
6617 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
6619 Worklist.push_back(*UI);
6621 // Delete the Old value.
6622 if (PHINode *PN = dyn_cast<PHINode>(Old))
6623 SE->ConstantEvolutionLoopExitValue.erase(PN);
6624 SE->ValueExprMap.erase(Old);
6625 // this now dangles!
6628 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
6629 : CallbackVH(V), SE(se) {}
6631 //===----------------------------------------------------------------------===//
6632 // ScalarEvolution Class Implementation
6633 //===----------------------------------------------------------------------===//
6635 ScalarEvolution::ScalarEvolution()
6636 : FunctionPass(ID), FirstUnknown(0) {
6637 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
6640 bool ScalarEvolution::runOnFunction(Function &F) {
6642 LI = &getAnalysis<LoopInfo>();
6643 TD = getAnalysisIfAvailable<DataLayout>();
6644 TLI = &getAnalysis<TargetLibraryInfo>();
6645 DT = &getAnalysis<DominatorTree>();
6649 void ScalarEvolution::releaseMemory() {
6650 // Iterate through all the SCEVUnknown instances and call their
6651 // destructors, so that they release their references to their values.
6652 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
6656 ValueExprMap.clear();
6658 // Free any extra memory created for ExitNotTakenInfo in the unlikely event
6659 // that a loop had multiple computable exits.
6660 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
6661 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end();
6666 assert(PendingLoopPredicates.empty() && "isImpliedCond garbage");
6668 BackedgeTakenCounts.clear();
6669 ConstantEvolutionLoopExitValue.clear();
6670 ValuesAtScopes.clear();
6671 LoopDispositions.clear();
6672 BlockDispositions.clear();
6673 UnsignedRanges.clear();
6674 SignedRanges.clear();
6675 UniqueSCEVs.clear();
6676 SCEVAllocator.Reset();
6679 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
6680 AU.setPreservesAll();
6681 AU.addRequiredTransitive<LoopInfo>();
6682 AU.addRequiredTransitive<DominatorTree>();
6683 AU.addRequired<TargetLibraryInfo>();
6686 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
6687 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
6690 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
6692 // Print all inner loops first
6693 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
6694 PrintLoopInfo(OS, SE, *I);
6697 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6700 SmallVector<BasicBlock *, 8> ExitBlocks;
6701 L->getExitBlocks(ExitBlocks);
6702 if (ExitBlocks.size() != 1)
6703 OS << "<multiple exits> ";
6705 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
6706 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
6708 OS << "Unpredictable backedge-taken count. ";
6713 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6716 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
6717 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
6719 OS << "Unpredictable max backedge-taken count. ";
6725 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
6726 // ScalarEvolution's implementation of the print method is to print
6727 // out SCEV values of all instructions that are interesting. Doing
6728 // this potentially causes it to create new SCEV objects though,
6729 // which technically conflicts with the const qualifier. This isn't
6730 // observable from outside the class though, so casting away the
6731 // const isn't dangerous.
6732 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
6734 OS << "Classifying expressions for: ";
6735 WriteAsOperand(OS, F, /*PrintType=*/false);
6737 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
6738 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
6741 const SCEV *SV = SE.getSCEV(&*I);
6744 const Loop *L = LI->getLoopFor((*I).getParent());
6746 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
6753 OS << "\t\t" "Exits: ";
6754 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
6755 if (!SE.isLoopInvariant(ExitValue, L)) {
6756 OS << "<<Unknown>>";
6765 OS << "Determining loop execution counts for: ";
6766 WriteAsOperand(OS, F, /*PrintType=*/false);
6768 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6769 PrintLoopInfo(OS, &SE, *I);
6772 ScalarEvolution::LoopDisposition
6773 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
6774 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S];
6775 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair =
6776 Values.insert(std::make_pair(L, LoopVariant));
6778 return Pair.first->second;
6780 LoopDisposition D = computeLoopDisposition(S, L);
6781 return LoopDispositions[S][L] = D;
6784 ScalarEvolution::LoopDisposition
6785 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
6786 switch (S->getSCEVType()) {
6788 return LoopInvariant;
6792 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
6793 case scAddRecExpr: {
6794 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6796 // If L is the addrec's loop, it's computable.
6797 if (AR->getLoop() == L)
6798 return LoopComputable;
6800 // Add recurrences are never invariant in the function-body (null loop).
6804 // This recurrence is variant w.r.t. L if L contains AR's loop.
6805 if (L->contains(AR->getLoop()))
6808 // This recurrence is invariant w.r.t. L if AR's loop contains L.
6809 if (AR->getLoop()->contains(L))
6810 return LoopInvariant;
6812 // This recurrence is variant w.r.t. L if any of its operands
6814 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
6816 if (!isLoopInvariant(*I, L))
6819 // Otherwise it's loop-invariant.
6820 return LoopInvariant;
6826 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6827 bool HasVarying = false;
6828 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6830 LoopDisposition D = getLoopDisposition(*I, L);
6831 if (D == LoopVariant)
6833 if (D == LoopComputable)
6836 return HasVarying ? LoopComputable : LoopInvariant;
6839 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6840 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
6841 if (LD == LoopVariant)
6843 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
6844 if (RD == LoopVariant)
6846 return (LD == LoopInvariant && RD == LoopInvariant) ?
6847 LoopInvariant : LoopComputable;
6850 // All non-instruction values are loop invariant. All instructions are loop
6851 // invariant if they are not contained in the specified loop.
6852 // Instructions are never considered invariant in the function body
6853 // (null loop) because they are defined within the "loop".
6854 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
6855 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
6856 return LoopInvariant;
6857 case scCouldNotCompute:
6858 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6859 default: llvm_unreachable("Unknown SCEV kind!");
6863 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
6864 return getLoopDisposition(S, L) == LoopInvariant;
6867 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
6868 return getLoopDisposition(S, L) == LoopComputable;
6871 ScalarEvolution::BlockDisposition
6872 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6873 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S];
6874 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool>
6875 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock));
6877 return Pair.first->second;
6879 BlockDisposition D = computeBlockDisposition(S, BB);
6880 return BlockDispositions[S][BB] = D;
6883 ScalarEvolution::BlockDisposition
6884 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6885 switch (S->getSCEVType()) {
6887 return ProperlyDominatesBlock;
6891 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
6892 case scAddRecExpr: {
6893 // This uses a "dominates" query instead of "properly dominates" query
6894 // to test for proper dominance too, because the instruction which
6895 // produces the addrec's value is a PHI, and a PHI effectively properly
6896 // dominates its entire containing block.
6897 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6898 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
6899 return DoesNotDominateBlock;
6901 // FALL THROUGH into SCEVNAryExpr handling.
6906 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6908 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6910 BlockDisposition D = getBlockDisposition(*I, BB);
6911 if (D == DoesNotDominateBlock)
6912 return DoesNotDominateBlock;
6913 if (D == DominatesBlock)
6916 return Proper ? ProperlyDominatesBlock : DominatesBlock;
6919 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6920 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6921 BlockDisposition LD = getBlockDisposition(LHS, BB);
6922 if (LD == DoesNotDominateBlock)
6923 return DoesNotDominateBlock;
6924 BlockDisposition RD = getBlockDisposition(RHS, BB);
6925 if (RD == DoesNotDominateBlock)
6926 return DoesNotDominateBlock;
6927 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
6928 ProperlyDominatesBlock : DominatesBlock;
6931 if (Instruction *I =
6932 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
6933 if (I->getParent() == BB)
6934 return DominatesBlock;
6935 if (DT->properlyDominates(I->getParent(), BB))
6936 return ProperlyDominatesBlock;
6937 return DoesNotDominateBlock;
6939 return ProperlyDominatesBlock;
6940 case scCouldNotCompute:
6941 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6943 llvm_unreachable("Unknown SCEV kind!");
6947 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
6948 return getBlockDisposition(S, BB) >= DominatesBlock;
6951 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
6952 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
6956 // Search for a SCEV expression node within an expression tree.
6957 // Implements SCEVTraversal::Visitor.
6962 SCEVSearch(const SCEV *N): Node(N), IsFound(false) {}
6964 bool follow(const SCEV *S) {
6965 IsFound |= (S == Node);
6968 bool isDone() const { return IsFound; }
6972 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
6973 SCEVSearch Search(Op);
6974 visitAll(S, Search);
6975 return Search.IsFound;
6978 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
6979 ValuesAtScopes.erase(S);
6980 LoopDispositions.erase(S);
6981 BlockDispositions.erase(S);
6982 UnsignedRanges.erase(S);
6983 SignedRanges.erase(S);
6985 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
6986 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end(); I != E; ) {
6987 BackedgeTakenInfo &BEInfo = I->second;
6988 if (BEInfo.hasOperand(S, this)) {
6990 BackedgeTakenCounts.erase(I++);
6997 typedef DenseMap<const Loop *, std::string> VerifyMap;
6999 /// replaceSubString - Replaces all occurences of From in Str with To.
7000 static void replaceSubString(std::string &Str, StringRef From, StringRef To) {
7002 while ((Pos = Str.find(From, Pos)) != std::string::npos) {
7003 Str.replace(Pos, From.size(), To.data(), To.size());
7008 /// getLoopBackedgeTakenCounts - Helper method for verifyAnalysis.
7010 getLoopBackedgeTakenCounts(Loop *L, VerifyMap &Map, ScalarEvolution &SE) {
7011 for (Loop::reverse_iterator I = L->rbegin(), E = L->rend(); I != E; ++I) {
7012 getLoopBackedgeTakenCounts(*I, Map, SE); // recurse.
7014 std::string &S = Map[L];
7016 raw_string_ostream OS(S);
7017 SE.getBackedgeTakenCount(L)->print(OS);
7019 // false and 0 are semantically equivalent. This can happen in dead loops.
7020 replaceSubString(OS.str(), "false", "0");
7021 // Remove wrap flags, their use in SCEV is highly fragile.
7022 // FIXME: Remove this when SCEV gets smarter about them.
7023 replaceSubString(OS.str(), "<nw>", "");
7024 replaceSubString(OS.str(), "<nsw>", "");
7025 replaceSubString(OS.str(), "<nuw>", "");
7030 void ScalarEvolution::verifyAnalysis() const {
7034 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
7036 // Gather stringified backedge taken counts for all loops using SCEV's caches.
7037 // FIXME: It would be much better to store actual values instead of strings,
7038 // but SCEV pointers will change if we drop the caches.
7039 VerifyMap BackedgeDumpsOld, BackedgeDumpsNew;
7040 for (LoopInfo::reverse_iterator I = LI->rbegin(), E = LI->rend(); I != E; ++I)
7041 getLoopBackedgeTakenCounts(*I, BackedgeDumpsOld, SE);
7043 // Gather stringified backedge taken counts for all loops without using
7046 for (LoopInfo::reverse_iterator I = LI->rbegin(), E = LI->rend(); I != E; ++I)
7047 getLoopBackedgeTakenCounts(*I, BackedgeDumpsNew, SE);
7049 // Now compare whether they're the same with and without caches. This allows
7050 // verifying that no pass changed the cache.
7051 assert(BackedgeDumpsOld.size() == BackedgeDumpsNew.size() &&
7052 "New loops suddenly appeared!");
7054 for (VerifyMap::iterator OldI = BackedgeDumpsOld.begin(),
7055 OldE = BackedgeDumpsOld.end(),
7056 NewI = BackedgeDumpsNew.begin();
7057 OldI != OldE; ++OldI, ++NewI) {
7058 assert(OldI->first == NewI->first && "Loop order changed!");
7060 // Compare the stringified SCEVs. We don't care if undef backedgetaken count
7062 // FIXME: We currently ignore SCEV changes from/to CouldNotCompute. This
7063 // means that a pass is buggy or SCEV has to learn a new pattern but is
7064 // usually not harmful.
7065 if (OldI->second != NewI->second &&
7066 OldI->second.find("undef") == std::string::npos &&
7067 NewI->second.find("undef") == std::string::npos &&
7068 OldI->second != "***COULDNOTCOMPUTE***" &&
7069 NewI->second != "***COULDNOTCOMPUTE***") {
7070 dbgs() << "SCEVValidator: SCEV for loop '"
7071 << OldI->first->getHeader()->getName()
7072 << "' changed from '" << OldI->second
7073 << "' to '" << NewI->second << "'!\n";
7078 // TODO: Verify more things.