1 //===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the implementation of the scalar evolution analysis
11 // engine, which is used primarily to analyze expressions involving induction
12 // variables in loops.
14 // There are several aspects to this library. First is the representation of
15 // scalar expressions, which are represented as subclasses of the SCEV class.
16 // These classes are used to represent certain types of subexpressions that we
17 // can handle. We only create one SCEV of a particular shape, so
18 // pointer-comparisons for equality are legal.
20 // One important aspect of the SCEV objects is that they are never cyclic, even
21 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
22 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
23 // recurrence) then we represent it directly as a recurrence node, otherwise we
24 // represent it as a SCEVUnknown node.
26 // In addition to being able to represent expressions of various types, we also
27 // have folders that are used to build the *canonical* representation for a
28 // particular expression. These folders are capable of using a variety of
29 // rewrite rules to simplify the expressions.
31 // Once the folders are defined, we can implement the more interesting
32 // higher-level code, such as the code that recognizes PHI nodes of various
33 // types, computes the execution count of a loop, etc.
35 // TODO: We should use these routines and value representations to implement
36 // dependence analysis!
38 //===----------------------------------------------------------------------===//
40 // There are several good references for the techniques used in this analysis.
42 // Chains of recurrences -- a method to expedite the evaluation
43 // of closed-form functions
44 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
46 // On computational properties of chains of recurrences
49 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
50 // Robert A. van Engelen
52 // Efficient Symbolic Analysis for Optimizing Compilers
53 // Robert A. van Engelen
55 // Using the chains of recurrences algebra for data dependence testing and
56 // induction variable substitution
57 // MS Thesis, Johnie Birch
59 //===----------------------------------------------------------------------===//
61 #define DEBUG_TYPE "scalar-evolution"
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/Constants.h"
64 #include "llvm/DerivedTypes.h"
65 #include "llvm/GlobalVariable.h"
66 #include "llvm/GlobalAlias.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/LLVMContext.h"
69 #include "llvm/Operator.h"
70 #include "llvm/Analysis/ConstantFolding.h"
71 #include "llvm/Analysis/Dominators.h"
72 #include "llvm/Analysis/InstructionSimplify.h"
73 #include "llvm/Analysis/LoopInfo.h"
74 #include "llvm/Analysis/ValueTracking.h"
75 #include "llvm/Assembly/Writer.h"
76 #include "llvm/Target/TargetData.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/ConstantRange.h"
79 #include "llvm/Support/Debug.h"
80 #include "llvm/Support/ErrorHandling.h"
81 #include "llvm/Support/GetElementPtrTypeIterator.h"
82 #include "llvm/Support/InstIterator.h"
83 #include "llvm/Support/MathExtras.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/ADT/Statistic.h"
86 #include "llvm/ADT/STLExtras.h"
87 #include "llvm/ADT/SmallPtrSet.h"
91 STATISTIC(NumArrayLenItCounts,
92 "Number of trip counts computed with array length");
93 STATISTIC(NumTripCountsComputed,
94 "Number of loops with predictable loop counts");
95 STATISTIC(NumTripCountsNotComputed,
96 "Number of loops without predictable loop counts");
97 STATISTIC(NumBruteForceTripCountsComputed,
98 "Number of loops with trip counts computed by force");
100 static cl::opt<unsigned>
101 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
102 cl::desc("Maximum number of iterations SCEV will "
103 "symbolically execute a constant "
107 INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
108 "Scalar Evolution Analysis", false, true)
109 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
110 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
111 INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
112 "Scalar Evolution Analysis", false, true)
113 char ScalarEvolution::ID = 0;
115 //===----------------------------------------------------------------------===//
116 // SCEV class definitions
117 //===----------------------------------------------------------------------===//
119 //===----------------------------------------------------------------------===//
120 // Implementation of the SCEV class.
123 void SCEV::dump() const {
128 void SCEV::print(raw_ostream &OS) const {
129 switch (getSCEVType()) {
131 WriteAsOperand(OS, cast<SCEVConstant>(this)->getValue(), false);
134 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(this);
135 const SCEV *Op = Trunc->getOperand();
136 OS << "(trunc " << *Op->getType() << " " << *Op << " to "
137 << *Trunc->getType() << ")";
141 const SCEVZeroExtendExpr *ZExt = cast<SCEVZeroExtendExpr>(this);
142 const SCEV *Op = ZExt->getOperand();
143 OS << "(zext " << *Op->getType() << " " << *Op << " to "
144 << *ZExt->getType() << ")";
148 const SCEVSignExtendExpr *SExt = cast<SCEVSignExtendExpr>(this);
149 const SCEV *Op = SExt->getOperand();
150 OS << "(sext " << *Op->getType() << " " << *Op << " to "
151 << *SExt->getType() << ")";
155 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(this);
156 OS << "{" << *AR->getOperand(0);
157 for (unsigned i = 1, e = AR->getNumOperands(); i != e; ++i)
158 OS << ",+," << *AR->getOperand(i);
160 if (AR->getNoWrapFlags(FlagNUW))
162 if (AR->getNoWrapFlags(FlagNSW))
164 if (AR->getNoWrapFlags(FlagNW) &&
165 !AR->getNoWrapFlags((NoWrapFlags)(FlagNUW | FlagNSW)))
167 WriteAsOperand(OS, AR->getLoop()->getHeader(), /*PrintType=*/false);
175 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(this);
176 const char *OpStr = 0;
177 switch (NAry->getSCEVType()) {
178 case scAddExpr: OpStr = " + "; break;
179 case scMulExpr: OpStr = " * "; break;
180 case scUMaxExpr: OpStr = " umax "; break;
181 case scSMaxExpr: OpStr = " smax "; break;
184 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
187 if (llvm::next(I) != E)
194 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(this);
195 OS << "(" << *UDiv->getLHS() << " /u " << *UDiv->getRHS() << ")";
199 const SCEVUnknown *U = cast<SCEVUnknown>(this);
201 if (U->isSizeOf(AllocTy)) {
202 OS << "sizeof(" << *AllocTy << ")";
205 if (U->isAlignOf(AllocTy)) {
206 OS << "alignof(" << *AllocTy << ")";
212 if (U->isOffsetOf(CTy, FieldNo)) {
213 OS << "offsetof(" << *CTy << ", ";
214 WriteAsOperand(OS, FieldNo, false);
219 // Otherwise just print it normally.
220 WriteAsOperand(OS, U->getValue(), false);
223 case scCouldNotCompute:
224 OS << "***COULDNOTCOMPUTE***";
228 llvm_unreachable("Unknown SCEV kind!");
231 Type *SCEV::getType() const {
232 switch (getSCEVType()) {
234 return cast<SCEVConstant>(this)->getType();
238 return cast<SCEVCastExpr>(this)->getType();
243 return cast<SCEVNAryExpr>(this)->getType();
245 return cast<SCEVAddExpr>(this)->getType();
247 return cast<SCEVUDivExpr>(this)->getType();
249 return cast<SCEVUnknown>(this)->getType();
250 case scCouldNotCompute:
251 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
255 llvm_unreachable("Unknown SCEV kind!");
259 bool SCEV::isZero() const {
260 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
261 return SC->getValue()->isZero();
265 bool SCEV::isOne() const {
266 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
267 return SC->getValue()->isOne();
271 bool SCEV::isAllOnesValue() const {
272 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
273 return SC->getValue()->isAllOnesValue();
277 SCEVCouldNotCompute::SCEVCouldNotCompute() :
278 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
280 bool SCEVCouldNotCompute::classof(const SCEV *S) {
281 return S->getSCEVType() == scCouldNotCompute;
284 const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
286 ID.AddInteger(scConstant);
289 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
290 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V);
291 UniqueSCEVs.InsertNode(S, IP);
295 const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
296 return getConstant(ConstantInt::get(getContext(), Val));
300 ScalarEvolution::getConstant(Type *Ty, uint64_t V, bool isSigned) {
301 IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
302 return getConstant(ConstantInt::get(ITy, V, isSigned));
305 SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
306 unsigned SCEVTy, const SCEV *op, Type *ty)
307 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
309 SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
310 const SCEV *op, Type *ty)
311 : SCEVCastExpr(ID, scTruncate, op, ty) {
312 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
313 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
314 "Cannot truncate non-integer value!");
317 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
318 const SCEV *op, Type *ty)
319 : SCEVCastExpr(ID, scZeroExtend, op, ty) {
320 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
321 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
322 "Cannot zero extend non-integer value!");
325 SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
326 const SCEV *op, Type *ty)
327 : SCEVCastExpr(ID, scSignExtend, op, ty) {
328 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
329 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
330 "Cannot sign extend non-integer value!");
333 void SCEVUnknown::deleted() {
334 // Clear this SCEVUnknown from various maps.
335 SE->forgetMemoizedResults(this);
337 // Remove this SCEVUnknown from the uniquing map.
338 SE->UniqueSCEVs.RemoveNode(this);
340 // Release the value.
344 void SCEVUnknown::allUsesReplacedWith(Value *New) {
345 // Clear this SCEVUnknown from various maps.
346 SE->forgetMemoizedResults(this);
348 // Remove this SCEVUnknown from the uniquing map.
349 SE->UniqueSCEVs.RemoveNode(this);
351 // Update this SCEVUnknown to point to the new value. This is needed
352 // because there may still be outstanding SCEVs which still point to
357 bool SCEVUnknown::isSizeOf(Type *&AllocTy) const {
358 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
359 if (VCE->getOpcode() == Instruction::PtrToInt)
360 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
361 if (CE->getOpcode() == Instruction::GetElementPtr &&
362 CE->getOperand(0)->isNullValue() &&
363 CE->getNumOperands() == 2)
364 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
366 AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
374 bool SCEVUnknown::isAlignOf(Type *&AllocTy) const {
375 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
376 if (VCE->getOpcode() == Instruction::PtrToInt)
377 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
378 if (CE->getOpcode() == Instruction::GetElementPtr &&
379 CE->getOperand(0)->isNullValue()) {
381 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
382 if (StructType *STy = dyn_cast<StructType>(Ty))
383 if (!STy->isPacked() &&
384 CE->getNumOperands() == 3 &&
385 CE->getOperand(1)->isNullValue()) {
386 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
388 STy->getNumElements() == 2 &&
389 STy->getElementType(0)->isIntegerTy(1)) {
390 AllocTy = STy->getElementType(1);
399 bool SCEVUnknown::isOffsetOf(Type *&CTy, Constant *&FieldNo) const {
400 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
401 if (VCE->getOpcode() == Instruction::PtrToInt)
402 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
403 if (CE->getOpcode() == Instruction::GetElementPtr &&
404 CE->getNumOperands() == 3 &&
405 CE->getOperand(0)->isNullValue() &&
406 CE->getOperand(1)->isNullValue()) {
408 cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
409 // Ignore vector types here so that ScalarEvolutionExpander doesn't
410 // emit getelementptrs that index into vectors.
411 if (Ty->isStructTy() || Ty->isArrayTy()) {
413 FieldNo = CE->getOperand(2);
421 //===----------------------------------------------------------------------===//
423 //===----------------------------------------------------------------------===//
426 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
427 /// than the complexity of the RHS. This comparator is used to canonicalize
429 class SCEVComplexityCompare {
430 const LoopInfo *const LI;
432 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
434 // Return true or false if LHS is less than, or at least RHS, respectively.
435 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
436 return compare(LHS, RHS) < 0;
439 // Return negative, zero, or positive, if LHS is less than, equal to, or
440 // greater than RHS, respectively. A three-way result allows recursive
441 // comparisons to be more efficient.
442 int compare(const SCEV *LHS, const SCEV *RHS) const {
443 // Fast-path: SCEVs are uniqued so we can do a quick equality check.
447 // Primarily, sort the SCEVs by their getSCEVType().
448 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
450 return (int)LType - (int)RType;
452 // Aside from the getSCEVType() ordering, the particular ordering
453 // isn't very important except that it's beneficial to be consistent,
454 // so that (a + b) and (b + a) don't end up as different expressions.
457 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS);
458 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
460 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
461 // not as complete as it could be.
462 const Value *LV = LU->getValue(), *RV = RU->getValue();
464 // Order pointer values after integer values. This helps SCEVExpander
466 bool LIsPointer = LV->getType()->isPointerTy(),
467 RIsPointer = RV->getType()->isPointerTy();
468 if (LIsPointer != RIsPointer)
469 return (int)LIsPointer - (int)RIsPointer;
471 // Compare getValueID values.
472 unsigned LID = LV->getValueID(),
473 RID = RV->getValueID();
475 return (int)LID - (int)RID;
477 // Sort arguments by their position.
478 if (const Argument *LA = dyn_cast<Argument>(LV)) {
479 const Argument *RA = cast<Argument>(RV);
480 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo();
481 return (int)LArgNo - (int)RArgNo;
484 // For instructions, compare their loop depth, and their operand
485 // count. This is pretty loose.
486 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) {
487 const Instruction *RInst = cast<Instruction>(RV);
489 // Compare loop depths.
490 const BasicBlock *LParent = LInst->getParent(),
491 *RParent = RInst->getParent();
492 if (LParent != RParent) {
493 unsigned LDepth = LI->getLoopDepth(LParent),
494 RDepth = LI->getLoopDepth(RParent);
495 if (LDepth != RDepth)
496 return (int)LDepth - (int)RDepth;
499 // Compare the number of operands.
500 unsigned LNumOps = LInst->getNumOperands(),
501 RNumOps = RInst->getNumOperands();
502 return (int)LNumOps - (int)RNumOps;
509 const SCEVConstant *LC = cast<SCEVConstant>(LHS);
510 const SCEVConstant *RC = cast<SCEVConstant>(RHS);
512 // Compare constant values.
513 const APInt &LA = LC->getValue()->getValue();
514 const APInt &RA = RC->getValue()->getValue();
515 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth();
516 if (LBitWidth != RBitWidth)
517 return (int)LBitWidth - (int)RBitWidth;
518 return LA.ult(RA) ? -1 : 1;
522 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS);
523 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
525 // Compare addrec loop depths.
526 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop();
527 if (LLoop != RLoop) {
528 unsigned LDepth = LLoop->getLoopDepth(),
529 RDepth = RLoop->getLoopDepth();
530 if (LDepth != RDepth)
531 return (int)LDepth - (int)RDepth;
534 // Addrec complexity grows with operand count.
535 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands();
536 if (LNumOps != RNumOps)
537 return (int)LNumOps - (int)RNumOps;
539 // Lexicographically compare.
540 for (unsigned i = 0; i != LNumOps; ++i) {
541 long X = compare(LA->getOperand(i), RA->getOperand(i));
553 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS);
554 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
556 // Lexicographically compare n-ary expressions.
557 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
558 for (unsigned i = 0; i != LNumOps; ++i) {
561 long X = compare(LC->getOperand(i), RC->getOperand(i));
565 return (int)LNumOps - (int)RNumOps;
569 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS);
570 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
572 // Lexicographically compare udiv expressions.
573 long X = compare(LC->getLHS(), RC->getLHS());
576 return compare(LC->getRHS(), RC->getRHS());
582 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS);
583 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
585 // Compare cast expressions by operand.
586 return compare(LC->getOperand(), RC->getOperand());
593 llvm_unreachable("Unknown SCEV kind!");
599 /// GroupByComplexity - Given a list of SCEV objects, order them by their
600 /// complexity, and group objects of the same complexity together by value.
601 /// When this routine is finished, we know that any duplicates in the vector are
602 /// consecutive and that complexity is monotonically increasing.
604 /// Note that we go take special precautions to ensure that we get deterministic
605 /// results from this routine. In other words, we don't want the results of
606 /// this to depend on where the addresses of various SCEV objects happened to
609 static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
611 if (Ops.size() < 2) return; // Noop
612 if (Ops.size() == 2) {
613 // This is the common case, which also happens to be trivially simple.
615 const SCEV *&LHS = Ops[0], *&RHS = Ops[1];
616 if (SCEVComplexityCompare(LI)(RHS, LHS))
621 // Do the rough sort by complexity.
622 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
624 // Now that we are sorted by complexity, group elements of the same
625 // complexity. Note that this is, at worst, N^2, but the vector is likely to
626 // be extremely short in practice. Note that we take this approach because we
627 // do not want to depend on the addresses of the objects we are grouping.
628 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
629 const SCEV *S = Ops[i];
630 unsigned Complexity = S->getSCEVType();
632 // If there are any objects of the same complexity and same value as this
634 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
635 if (Ops[j] == S) { // Found a duplicate.
636 // Move it to immediately after i'th element.
637 std::swap(Ops[i+1], Ops[j]);
638 ++i; // no need to rescan it.
639 if (i == e-2) return; // Done!
647 //===----------------------------------------------------------------------===//
648 // Simple SCEV method implementations
649 //===----------------------------------------------------------------------===//
651 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
653 static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
656 // Handle the simplest case efficiently.
658 return SE.getTruncateOrZeroExtend(It, ResultTy);
660 // We are using the following formula for BC(It, K):
662 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
664 // Suppose, W is the bitwidth of the return value. We must be prepared for
665 // overflow. Hence, we must assure that the result of our computation is
666 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
667 // safe in modular arithmetic.
669 // However, this code doesn't use exactly that formula; the formula it uses
670 // is something like the following, where T is the number of factors of 2 in
671 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
674 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
676 // This formula is trivially equivalent to the previous formula. However,
677 // this formula can be implemented much more efficiently. The trick is that
678 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
679 // arithmetic. To do exact division in modular arithmetic, all we have
680 // to do is multiply by the inverse. Therefore, this step can be done at
683 // The next issue is how to safely do the division by 2^T. The way this
684 // is done is by doing the multiplication step at a width of at least W + T
685 // bits. This way, the bottom W+T bits of the product are accurate. Then,
686 // when we perform the division by 2^T (which is equivalent to a right shift
687 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
688 // truncated out after the division by 2^T.
690 // In comparison to just directly using the first formula, this technique
691 // is much more efficient; using the first formula requires W * K bits,
692 // but this formula less than W + K bits. Also, the first formula requires
693 // a division step, whereas this formula only requires multiplies and shifts.
695 // It doesn't matter whether the subtraction step is done in the calculation
696 // width or the input iteration count's width; if the subtraction overflows,
697 // the result must be zero anyway. We prefer here to do it in the width of
698 // the induction variable because it helps a lot for certain cases; CodeGen
699 // isn't smart enough to ignore the overflow, which leads to much less
700 // efficient code if the width of the subtraction is wider than the native
703 // (It's possible to not widen at all by pulling out factors of 2 before
704 // the multiplication; for example, K=2 can be calculated as
705 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
706 // extra arithmetic, so it's not an obvious win, and it gets
707 // much more complicated for K > 3.)
709 // Protection from insane SCEVs; this bound is conservative,
710 // but it probably doesn't matter.
712 return SE.getCouldNotCompute();
714 unsigned W = SE.getTypeSizeInBits(ResultTy);
716 // Calculate K! / 2^T and T; we divide out the factors of two before
717 // multiplying for calculating K! / 2^T to avoid overflow.
718 // Other overflow doesn't matter because we only care about the bottom
719 // W bits of the result.
720 APInt OddFactorial(W, 1);
722 for (unsigned i = 3; i <= K; ++i) {
724 unsigned TwoFactors = Mult.countTrailingZeros();
726 Mult = Mult.lshr(TwoFactors);
727 OddFactorial *= Mult;
730 // We need at least W + T bits for the multiplication step
731 unsigned CalculationBits = W + T;
733 // Calculate 2^T, at width T+W.
734 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
736 // Calculate the multiplicative inverse of K! / 2^T;
737 // this multiplication factor will perform the exact division by
739 APInt Mod = APInt::getSignedMinValue(W+1);
740 APInt MultiplyFactor = OddFactorial.zext(W+1);
741 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
742 MultiplyFactor = MultiplyFactor.trunc(W);
744 // Calculate the product, at width T+W
745 IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
747 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
748 for (unsigned i = 1; i != K; ++i) {
749 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
750 Dividend = SE.getMulExpr(Dividend,
751 SE.getTruncateOrZeroExtend(S, CalculationTy));
755 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
757 // Truncate the result, and divide by K! / 2^T.
759 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
760 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
763 /// evaluateAtIteration - Return the value of this chain of recurrences at
764 /// the specified iteration number. We can evaluate this recurrence by
765 /// multiplying each element in the chain by the binomial coefficient
766 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
768 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
770 /// where BC(It, k) stands for binomial coefficient.
772 const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
773 ScalarEvolution &SE) const {
774 const SCEV *Result = getStart();
775 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
776 // The computation is correct in the face of overflow provided that the
777 // multiplication is performed _after_ the evaluation of the binomial
779 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
780 if (isa<SCEVCouldNotCompute>(Coeff))
783 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
788 //===----------------------------------------------------------------------===//
789 // SCEV Expression folder implementations
790 //===----------------------------------------------------------------------===//
792 const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
794 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
795 "This is not a truncating conversion!");
796 assert(isSCEVable(Ty) &&
797 "This is not a conversion to a SCEVable type!");
798 Ty = getEffectiveSCEVType(Ty);
801 ID.AddInteger(scTruncate);
805 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
807 // Fold if the operand is constant.
808 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
810 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(),
811 getEffectiveSCEVType(Ty))));
813 // trunc(trunc(x)) --> trunc(x)
814 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
815 return getTruncateExpr(ST->getOperand(), Ty);
817 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
818 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
819 return getTruncateOrSignExtend(SS->getOperand(), Ty);
821 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
822 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
823 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
825 // trunc(x1+x2+...+xN) --> trunc(x1)+trunc(x2)+...+trunc(xN) if we can
826 // eliminate all the truncates.
827 if (const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Op)) {
828 SmallVector<const SCEV *, 4> Operands;
829 bool hasTrunc = false;
830 for (unsigned i = 0, e = SA->getNumOperands(); i != e && !hasTrunc; ++i) {
831 const SCEV *S = getTruncateExpr(SA->getOperand(i), Ty);
832 hasTrunc = isa<SCEVTruncateExpr>(S);
833 Operands.push_back(S);
836 return getAddExpr(Operands);
837 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
840 // trunc(x1*x2*...*xN) --> trunc(x1)*trunc(x2)*...*trunc(xN) if we can
841 // eliminate all the truncates.
842 if (const SCEVMulExpr *SM = dyn_cast<SCEVMulExpr>(Op)) {
843 SmallVector<const SCEV *, 4> Operands;
844 bool hasTrunc = false;
845 for (unsigned i = 0, e = SM->getNumOperands(); i != e && !hasTrunc; ++i) {
846 const SCEV *S = getTruncateExpr(SM->getOperand(i), Ty);
847 hasTrunc = isa<SCEVTruncateExpr>(S);
848 Operands.push_back(S);
851 return getMulExpr(Operands);
852 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
855 // If the input value is a chrec scev, truncate the chrec's operands.
856 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
857 SmallVector<const SCEV *, 4> Operands;
858 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
859 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
860 return getAddRecExpr(Operands, AddRec->getLoop(), SCEV::FlagAnyWrap);
863 // As a special case, fold trunc(undef) to undef. We don't want to
864 // know too much about SCEVUnknowns, but this special case is handy
866 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
867 if (isa<UndefValue>(U->getValue()))
868 return getSCEV(UndefValue::get(Ty));
870 // The cast wasn't folded; create an explicit cast node. We can reuse
871 // the existing insert position since if we get here, we won't have
872 // made any changes which would invalidate it.
873 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
875 UniqueSCEVs.InsertNode(S, IP);
879 const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
881 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
882 "This is not an extending conversion!");
883 assert(isSCEVable(Ty) &&
884 "This is not a conversion to a SCEVable type!");
885 Ty = getEffectiveSCEVType(Ty);
887 // Fold if the operand is constant.
888 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
890 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(),
891 getEffectiveSCEVType(Ty))));
893 // zext(zext(x)) --> zext(x)
894 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
895 return getZeroExtendExpr(SZ->getOperand(), Ty);
897 // Before doing any expensive analysis, check to see if we've already
898 // computed a SCEV for this Op and Ty.
900 ID.AddInteger(scZeroExtend);
904 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
906 // zext(trunc(x)) --> zext(x) or x or trunc(x)
907 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
908 // It's possible the bits taken off by the truncate were all zero bits. If
909 // so, we should be able to simplify this further.
910 const SCEV *X = ST->getOperand();
911 ConstantRange CR = getUnsignedRange(X);
912 unsigned TruncBits = getTypeSizeInBits(ST->getType());
913 unsigned NewBits = getTypeSizeInBits(Ty);
914 if (CR.truncate(TruncBits).zeroExtend(NewBits).contains(
915 CR.zextOrTrunc(NewBits)))
916 return getTruncateOrZeroExtend(X, Ty);
919 // If the input value is a chrec scev, and we can prove that the value
920 // did not overflow the old, smaller, value, we can zero extend all of the
921 // operands (often constants). This allows analysis of something like
922 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
923 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
924 if (AR->isAffine()) {
925 const SCEV *Start = AR->getStart();
926 const SCEV *Step = AR->getStepRecurrence(*this);
927 unsigned BitWidth = getTypeSizeInBits(AR->getType());
928 const Loop *L = AR->getLoop();
930 // If we have special knowledge that this addrec won't overflow,
931 // we don't need to do any further analysis.
932 if (AR->getNoWrapFlags(SCEV::FlagNUW))
933 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
934 getZeroExtendExpr(Step, Ty),
935 L, AR->getNoWrapFlags());
937 // Check whether the backedge-taken count is SCEVCouldNotCompute.
938 // Note that this serves two purposes: It filters out loops that are
939 // simply not analyzable, and it covers the case where this code is
940 // being called from within backedge-taken count analysis, such that
941 // attempting to ask for the backedge-taken count would likely result
942 // in infinite recursion. In the later case, the analysis code will
943 // cope with a conservative value, and it will take care to purge
944 // that value once it has finished.
945 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
946 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
947 // Manually compute the final value for AR, checking for
950 // Check whether the backedge-taken count can be losslessly casted to
951 // the addrec's type. The count is always unsigned.
952 const SCEV *CastedMaxBECount =
953 getTruncateOrZeroExtend(MaxBECount, Start->getType());
954 const SCEV *RecastedMaxBECount =
955 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
956 if (MaxBECount == RecastedMaxBECount) {
957 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
958 // Check whether Start+Step*MaxBECount has no unsigned overflow.
959 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
960 const SCEV *Add = getAddExpr(Start, ZMul);
961 const SCEV *OperandExtendedAdd =
962 getAddExpr(getZeroExtendExpr(Start, WideTy),
963 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
964 getZeroExtendExpr(Step, WideTy)));
965 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) {
966 // Cache knowledge of AR NUW, which is propagated to this AddRec.
967 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
968 // Return the expression with the addrec on the outside.
969 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
970 getZeroExtendExpr(Step, Ty),
971 L, AR->getNoWrapFlags());
973 // Similar to above, only this time treat the step value as signed.
974 // This covers loops that count down.
975 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
976 Add = getAddExpr(Start, SMul);
978 getAddExpr(getZeroExtendExpr(Start, WideTy),
979 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
980 getSignExtendExpr(Step, WideTy)));
981 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) {
982 // Cache knowledge of AR NW, which is propagated to this AddRec.
983 // Negative step causes unsigned wrap, but it still can't self-wrap.
984 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
985 // Return the expression with the addrec on the outside.
986 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
987 getSignExtendExpr(Step, Ty),
988 L, AR->getNoWrapFlags());
992 // If the backedge is guarded by a comparison with the pre-inc value
993 // the addrec is safe. Also, if the entry is guarded by a comparison
994 // with the start value and the backedge is guarded by a comparison
995 // with the post-inc value, the addrec is safe.
996 if (isKnownPositive(Step)) {
997 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
998 getUnsignedRange(Step).getUnsignedMax());
999 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
1000 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
1001 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
1002 AR->getPostIncExpr(*this), N))) {
1003 // Cache knowledge of AR NUW, which is propagated to this AddRec.
1004 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
1005 // Return the expression with the addrec on the outside.
1006 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1007 getZeroExtendExpr(Step, Ty),
1008 L, AR->getNoWrapFlags());
1010 } else if (isKnownNegative(Step)) {
1011 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
1012 getSignedRange(Step).getSignedMin());
1013 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
1014 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
1015 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
1016 AR->getPostIncExpr(*this), N))) {
1017 // Cache knowledge of AR NW, which is propagated to this AddRec.
1018 // Negative step causes unsigned wrap, but it still can't self-wrap.
1019 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
1020 // Return the expression with the addrec on the outside.
1021 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
1022 getSignExtendExpr(Step, Ty),
1023 L, AR->getNoWrapFlags());
1029 // The cast wasn't folded; create an explicit cast node.
1030 // Recompute the insert position, as it may have been invalidated.
1031 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1032 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator),
1034 UniqueSCEVs.InsertNode(S, IP);
1038 // Get the limit of a recurrence such that incrementing by Step cannot cause
1039 // signed overflow as long as the value of the recurrence within the loop does
1040 // not exceed this limit before incrementing.
1041 static const SCEV *getOverflowLimitForStep(const SCEV *Step,
1042 ICmpInst::Predicate *Pred,
1043 ScalarEvolution *SE) {
1044 unsigned BitWidth = SE->getTypeSizeInBits(Step->getType());
1045 if (SE->isKnownPositive(Step)) {
1046 *Pred = ICmpInst::ICMP_SLT;
1047 return SE->getConstant(APInt::getSignedMinValue(BitWidth) -
1048 SE->getSignedRange(Step).getSignedMax());
1050 if (SE->isKnownNegative(Step)) {
1051 *Pred = ICmpInst::ICMP_SGT;
1052 return SE->getConstant(APInt::getSignedMaxValue(BitWidth) -
1053 SE->getSignedRange(Step).getSignedMin());
1058 // The recurrence AR has been shown to have no signed wrap. Typically, if we can
1059 // prove NSW for AR, then we can just as easily prove NSW for its preincrement
1060 // or postincrement sibling. This allows normalizing a sign extended AddRec as
1061 // such: {sext(Step + Start),+,Step} => {(Step + sext(Start),+,Step} As a
1062 // result, the expression "Step + sext(PreIncAR)" is congruent with
1063 // "sext(PostIncAR)"
1064 static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR,
1066 ScalarEvolution *SE) {
1067 const Loop *L = AR->getLoop();
1068 const SCEV *Start = AR->getStart();
1069 const SCEV *Step = AR->getStepRecurrence(*SE);
1071 // Check for a simple looking step prior to loop entry.
1072 const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start);
1076 // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV
1077 // subtraction is expensive. For this purpose, perform a quick and dirty
1078 // difference, by checking for Step in the operand list.
1079 SmallVector<const SCEV *, 4> DiffOps;
1080 for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end();
1083 DiffOps.push_back(*I);
1085 if (DiffOps.size() == SA->getNumOperands())
1088 // This is a postinc AR. Check for overflow on the preinc recurrence using the
1089 // same three conditions that getSignExtendedExpr checks.
1091 // 1. NSW flags on the step increment.
1092 const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags());
1093 const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>(
1094 SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap));
1096 if (PreAR && PreAR->getNoWrapFlags(SCEV::FlagNSW))
1099 // 2. Direct overflow check on the step operation's expression.
1100 unsigned BitWidth = SE->getTypeSizeInBits(AR->getType());
1101 Type *WideTy = IntegerType::get(SE->getContext(), BitWidth * 2);
1102 const SCEV *OperandExtendedStart =
1103 SE->getAddExpr(SE->getSignExtendExpr(PreStart, WideTy),
1104 SE->getSignExtendExpr(Step, WideTy));
1105 if (SE->getSignExtendExpr(Start, WideTy) == OperandExtendedStart) {
1106 // Cache knowledge of PreAR NSW.
1108 const_cast<SCEVAddRecExpr *>(PreAR)->setNoWrapFlags(SCEV::FlagNSW);
1109 // FIXME: this optimization needs a unit test
1110 DEBUG(dbgs() << "SCEV: untested prestart overflow check\n");
1114 // 3. Loop precondition.
1115 ICmpInst::Predicate Pred;
1116 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, SE);
1118 if (OverflowLimit &&
1119 SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) {
1125 // Get the normalized sign-extended expression for this AddRec's Start.
1126 static const SCEV *getSignExtendAddRecStart(const SCEVAddRecExpr *AR,
1128 ScalarEvolution *SE) {
1129 const SCEV *PreStart = getPreStartForSignExtend(AR, Ty, SE);
1131 return SE->getSignExtendExpr(AR->getStart(), Ty);
1133 return SE->getAddExpr(SE->getSignExtendExpr(AR->getStepRecurrence(*SE), Ty),
1134 SE->getSignExtendExpr(PreStart, Ty));
1137 const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
1139 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1140 "This is not an extending conversion!");
1141 assert(isSCEVable(Ty) &&
1142 "This is not a conversion to a SCEVable type!");
1143 Ty = getEffectiveSCEVType(Ty);
1145 // Fold if the operand is constant.
1146 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1148 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(),
1149 getEffectiveSCEVType(Ty))));
1151 // sext(sext(x)) --> sext(x)
1152 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
1153 return getSignExtendExpr(SS->getOperand(), Ty);
1155 // sext(zext(x)) --> zext(x)
1156 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
1157 return getZeroExtendExpr(SZ->getOperand(), Ty);
1159 // Before doing any expensive analysis, check to see if we've already
1160 // computed a SCEV for this Op and Ty.
1161 FoldingSetNodeID ID;
1162 ID.AddInteger(scSignExtend);
1166 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1168 // If the input value is provably positive, build a zext instead.
1169 if (isKnownNonNegative(Op))
1170 return getZeroExtendExpr(Op, Ty);
1172 // sext(trunc(x)) --> sext(x) or x or trunc(x)
1173 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) {
1174 // It's possible the bits taken off by the truncate were all sign bits. If
1175 // so, we should be able to simplify this further.
1176 const SCEV *X = ST->getOperand();
1177 ConstantRange CR = getSignedRange(X);
1178 unsigned TruncBits = getTypeSizeInBits(ST->getType());
1179 unsigned NewBits = getTypeSizeInBits(Ty);
1180 if (CR.truncate(TruncBits).signExtend(NewBits).contains(
1181 CR.sextOrTrunc(NewBits)))
1182 return getTruncateOrSignExtend(X, Ty);
1185 // If the input value is a chrec scev, and we can prove that the value
1186 // did not overflow the old, smaller, value, we can sign extend all of the
1187 // operands (often constants). This allows analysis of something like
1188 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
1189 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
1190 if (AR->isAffine()) {
1191 const SCEV *Start = AR->getStart();
1192 const SCEV *Step = AR->getStepRecurrence(*this);
1193 unsigned BitWidth = getTypeSizeInBits(AR->getType());
1194 const Loop *L = AR->getLoop();
1196 // If we have special knowledge that this addrec won't overflow,
1197 // we don't need to do any further analysis.
1198 if (AR->getNoWrapFlags(SCEV::FlagNSW))
1199 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1200 getSignExtendExpr(Step, Ty),
1203 // Check whether the backedge-taken count is SCEVCouldNotCompute.
1204 // Note that this serves two purposes: It filters out loops that are
1205 // simply not analyzable, and it covers the case where this code is
1206 // being called from within backedge-taken count analysis, such that
1207 // attempting to ask for the backedge-taken count would likely result
1208 // in infinite recursion. In the later case, the analysis code will
1209 // cope with a conservative value, and it will take care to purge
1210 // that value once it has finished.
1211 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
1212 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
1213 // Manually compute the final value for AR, checking for
1216 // Check whether the backedge-taken count can be losslessly casted to
1217 // the addrec's type. The count is always unsigned.
1218 const SCEV *CastedMaxBECount =
1219 getTruncateOrZeroExtend(MaxBECount, Start->getType());
1220 const SCEV *RecastedMaxBECount =
1221 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
1222 if (MaxBECount == RecastedMaxBECount) {
1223 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
1224 // Check whether Start+Step*MaxBECount has no signed overflow.
1225 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
1226 const SCEV *Add = getAddExpr(Start, SMul);
1227 const SCEV *OperandExtendedAdd =
1228 getAddExpr(getSignExtendExpr(Start, WideTy),
1229 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1230 getSignExtendExpr(Step, WideTy)));
1231 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) {
1232 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1233 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1234 // Return the expression with the addrec on the outside.
1235 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1236 getSignExtendExpr(Step, Ty),
1237 L, AR->getNoWrapFlags());
1239 // Similar to above, only this time treat the step value as unsigned.
1240 // This covers loops that count up with an unsigned step.
1241 const SCEV *UMul = getMulExpr(CastedMaxBECount, Step);
1242 Add = getAddExpr(Start, UMul);
1243 OperandExtendedAdd =
1244 getAddExpr(getSignExtendExpr(Start, WideTy),
1245 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
1246 getZeroExtendExpr(Step, WideTy)));
1247 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) {
1248 // Cache knowledge of AR NSW, which is propagated to this AddRec.
1249 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1250 // Return the expression with the addrec on the outside.
1251 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1252 getZeroExtendExpr(Step, Ty),
1253 L, AR->getNoWrapFlags());
1257 // If the backedge is guarded by a comparison with the pre-inc value
1258 // the addrec is safe. Also, if the entry is guarded by a comparison
1259 // with the start value and the backedge is guarded by a comparison
1260 // with the post-inc value, the addrec is safe.
1261 ICmpInst::Predicate Pred;
1262 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, this);
1263 if (OverflowLimit &&
1264 (isLoopBackedgeGuardedByCond(L, Pred, AR, OverflowLimit) ||
1265 (isLoopEntryGuardedByCond(L, Pred, Start, OverflowLimit) &&
1266 isLoopBackedgeGuardedByCond(L, Pred, AR->getPostIncExpr(*this),
1268 // Cache knowledge of AR NSW, then propagate NSW to the wide AddRec.
1269 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
1270 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
1271 getSignExtendExpr(Step, Ty),
1272 L, AR->getNoWrapFlags());
1277 // The cast wasn't folded; create an explicit cast node.
1278 // Recompute the insert position, as it may have been invalidated.
1279 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
1280 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator),
1282 UniqueSCEVs.InsertNode(S, IP);
1286 /// getAnyExtendExpr - Return a SCEV for the given operand extended with
1287 /// unspecified bits out to the given type.
1289 const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
1291 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
1292 "This is not an extending conversion!");
1293 assert(isSCEVable(Ty) &&
1294 "This is not a conversion to a SCEVable type!");
1295 Ty = getEffectiveSCEVType(Ty);
1297 // Sign-extend negative constants.
1298 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
1299 if (SC->getValue()->getValue().isNegative())
1300 return getSignExtendExpr(Op, Ty);
1302 // Peel off a truncate cast.
1303 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
1304 const SCEV *NewOp = T->getOperand();
1305 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
1306 return getAnyExtendExpr(NewOp, Ty);
1307 return getTruncateOrNoop(NewOp, Ty);
1310 // Next try a zext cast. If the cast is folded, use it.
1311 const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
1312 if (!isa<SCEVZeroExtendExpr>(ZExt))
1315 // Next try a sext cast. If the cast is folded, use it.
1316 const SCEV *SExt = getSignExtendExpr(Op, Ty);
1317 if (!isa<SCEVSignExtendExpr>(SExt))
1320 // Force the cast to be folded into the operands of an addrec.
1321 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
1322 SmallVector<const SCEV *, 4> Ops;
1323 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
1325 Ops.push_back(getAnyExtendExpr(*I, Ty));
1326 return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW);
1329 // As a special case, fold anyext(undef) to undef. We don't want to
1330 // know too much about SCEVUnknowns, but this special case is handy
1332 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
1333 if (isa<UndefValue>(U->getValue()))
1334 return getSCEV(UndefValue::get(Ty));
1336 // If the expression is obviously signed, use the sext cast value.
1337 if (isa<SCEVSMaxExpr>(Op))
1340 // Absent any other information, use the zext cast value.
1344 /// CollectAddOperandsWithScales - Process the given Ops list, which is
1345 /// a list of operands to be added under the given scale, update the given
1346 /// map. This is a helper function for getAddRecExpr. As an example of
1347 /// what it does, given a sequence of operands that would form an add
1348 /// expression like this:
1350 /// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
1352 /// where A and B are constants, update the map with these values:
1354 /// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
1356 /// and add 13 + A*B*29 to AccumulatedConstant.
1357 /// This will allow getAddRecExpr to produce this:
1359 /// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
1361 /// This form often exposes folding opportunities that are hidden in
1362 /// the original operand list.
1364 /// Return true iff it appears that any interesting folding opportunities
1365 /// may be exposed. This helps getAddRecExpr short-circuit extra work in
1366 /// the common case where no interesting opportunities are present, and
1367 /// is also used as a check to avoid infinite recursion.
1370 CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
1371 SmallVector<const SCEV *, 8> &NewOps,
1372 APInt &AccumulatedConstant,
1373 const SCEV *const *Ops, size_t NumOperands,
1375 ScalarEvolution &SE) {
1376 bool Interesting = false;
1378 // Iterate over the add operands. They are sorted, with constants first.
1380 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1382 // Pull a buried constant out to the outside.
1383 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
1385 AccumulatedConstant += Scale * C->getValue()->getValue();
1388 // Next comes everything else. We're especially interested in multiplies
1389 // here, but they're in the middle, so just visit the rest with one loop.
1390 for (; i != NumOperands; ++i) {
1391 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
1392 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
1394 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
1395 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
1396 // A multiplication of a constant with another add; recurse.
1397 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1));
1399 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1400 Add->op_begin(), Add->getNumOperands(),
1403 // A multiplication of a constant with some other value. Update
1405 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
1406 const SCEV *Key = SE.getMulExpr(MulOps);
1407 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1408 M.insert(std::make_pair(Key, NewScale));
1410 NewOps.push_back(Pair.first->first);
1412 Pair.first->second += NewScale;
1413 // The map already had an entry for this value, which may indicate
1414 // a folding opportunity.
1419 // An ordinary operand. Update the map.
1420 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
1421 M.insert(std::make_pair(Ops[i], Scale));
1423 NewOps.push_back(Pair.first->first);
1425 Pair.first->second += Scale;
1426 // The map already had an entry for this value, which may indicate
1427 // a folding opportunity.
1437 struct APIntCompare {
1438 bool operator()(const APInt &LHS, const APInt &RHS) const {
1439 return LHS.ult(RHS);
1444 /// getAddExpr - Get a canonical add expression, or something simpler if
1446 const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
1447 SCEV::NoWrapFlags Flags) {
1448 assert(!(Flags & ~(SCEV::FlagNUW | SCEV::FlagNSW)) &&
1449 "only nuw or nsw allowed");
1450 assert(!Ops.empty() && "Cannot get empty add!");
1451 if (Ops.size() == 1) return Ops[0];
1453 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1454 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1455 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1456 "SCEVAddExpr operand types don't match!");
1459 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1461 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1462 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1463 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1465 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1466 E = Ops.end(); I != E; ++I)
1467 if (!isKnownNonNegative(*I)) {
1471 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1474 // Sort by complexity, this groups all similar expression types together.
1475 GroupByComplexity(Ops, LI);
1477 // If there are any constants, fold them together.
1479 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1481 assert(Idx < Ops.size());
1482 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1483 // We found two constants, fold them together!
1484 Ops[0] = getConstant(LHSC->getValue()->getValue() +
1485 RHSC->getValue()->getValue());
1486 if (Ops.size() == 2) return Ops[0];
1487 Ops.erase(Ops.begin()+1); // Erase the folded element
1488 LHSC = cast<SCEVConstant>(Ops[0]);
1491 // If we are left with a constant zero being added, strip it off.
1492 if (LHSC->getValue()->isZero()) {
1493 Ops.erase(Ops.begin());
1497 if (Ops.size() == 1) return Ops[0];
1500 // Okay, check to see if the same value occurs in the operand list more than
1501 // once. If so, merge them together into an multiply expression. Since we
1502 // sorted the list, these values are required to be adjacent.
1503 Type *Ty = Ops[0]->getType();
1504 bool FoundMatch = false;
1505 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
1506 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
1507 // Scan ahead to count how many equal operands there are.
1509 while (i+Count != e && Ops[i+Count] == Ops[i])
1511 // Merge the values into a multiply.
1512 const SCEV *Scale = getConstant(Ty, Count);
1513 const SCEV *Mul = getMulExpr(Scale, Ops[i]);
1514 if (Ops.size() == Count)
1517 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count);
1518 --i; e -= Count - 1;
1522 return getAddExpr(Ops, Flags);
1524 // Check for truncates. If all the operands are truncated from the same
1525 // type, see if factoring out the truncate would permit the result to be
1526 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
1527 // if the contents of the resulting outer trunc fold to something simple.
1528 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1529 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1530 Type *DstType = Trunc->getType();
1531 Type *SrcType = Trunc->getOperand()->getType();
1532 SmallVector<const SCEV *, 8> LargeOps;
1534 // Check all the operands to see if they can be represented in the
1535 // source type of the truncate.
1536 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1537 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1538 if (T->getOperand()->getType() != SrcType) {
1542 LargeOps.push_back(T->getOperand());
1543 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1544 LargeOps.push_back(getAnyExtendExpr(C, SrcType));
1545 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1546 SmallVector<const SCEV *, 8> LargeMulOps;
1547 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1548 if (const SCEVTruncateExpr *T =
1549 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1550 if (T->getOperand()->getType() != SrcType) {
1554 LargeMulOps.push_back(T->getOperand());
1555 } else if (const SCEVConstant *C =
1556 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1557 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType));
1564 LargeOps.push_back(getMulExpr(LargeMulOps));
1571 // Evaluate the expression in the larger type.
1572 const SCEV *Fold = getAddExpr(LargeOps, Flags);
1573 // If it folds to something simple, use it. Otherwise, don't.
1574 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1575 return getTruncateExpr(Fold, DstType);
1579 // Skip past any other cast SCEVs.
1580 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1583 // If there are add operands they would be next.
1584 if (Idx < Ops.size()) {
1585 bool DeletedAdd = false;
1586 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1587 // If we have an add, expand the add operands onto the end of the operands
1589 Ops.erase(Ops.begin()+Idx);
1590 Ops.append(Add->op_begin(), Add->op_end());
1594 // If we deleted at least one add, we added operands to the end of the list,
1595 // and they are not necessarily sorted. Recurse to resort and resimplify
1596 // any operands we just acquired.
1598 return getAddExpr(Ops);
1601 // Skip over the add expression until we get to a multiply.
1602 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1605 // Check to see if there are any folding opportunities present with
1606 // operands multiplied by constant values.
1607 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
1608 uint64_t BitWidth = getTypeSizeInBits(Ty);
1609 DenseMap<const SCEV *, APInt> M;
1610 SmallVector<const SCEV *, 8> NewOps;
1611 APInt AccumulatedConstant(BitWidth, 0);
1612 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
1613 Ops.data(), Ops.size(),
1614 APInt(BitWidth, 1), *this)) {
1615 // Some interesting folding opportunity is present, so its worthwhile to
1616 // re-generate the operands list. Group the operands by constant scale,
1617 // to avoid multiplying by the same constant scale multiple times.
1618 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
1619 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(),
1620 E = NewOps.end(); I != E; ++I)
1621 MulOpLists[M.find(*I)->second].push_back(*I);
1622 // Re-generate the operands list.
1624 if (AccumulatedConstant != 0)
1625 Ops.push_back(getConstant(AccumulatedConstant));
1626 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
1627 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
1629 Ops.push_back(getMulExpr(getConstant(I->first),
1630 getAddExpr(I->second)));
1632 return getConstant(Ty, 0);
1633 if (Ops.size() == 1)
1635 return getAddExpr(Ops);
1639 // If we are adding something to a multiply expression, make sure the
1640 // something is not already an operand of the multiply. If so, merge it into
1642 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1643 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1644 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1645 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1646 if (isa<SCEVConstant>(MulOpSCEV))
1648 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1649 if (MulOpSCEV == Ops[AddOp]) {
1650 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1651 const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
1652 if (Mul->getNumOperands() != 2) {
1653 // If the multiply has more than two operands, we must get the
1655 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1656 Mul->op_begin()+MulOp);
1657 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1658 InnerMul = getMulExpr(MulOps);
1660 const SCEV *One = getConstant(Ty, 1);
1661 const SCEV *AddOne = getAddExpr(One, InnerMul);
1662 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV);
1663 if (Ops.size() == 2) return OuterMul;
1665 Ops.erase(Ops.begin()+AddOp);
1666 Ops.erase(Ops.begin()+Idx-1);
1668 Ops.erase(Ops.begin()+Idx);
1669 Ops.erase(Ops.begin()+AddOp-1);
1671 Ops.push_back(OuterMul);
1672 return getAddExpr(Ops);
1675 // Check this multiply against other multiplies being added together.
1676 for (unsigned OtherMulIdx = Idx+1;
1677 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1679 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1680 // If MulOp occurs in OtherMul, we can fold the two multiplies
1682 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1683 OMulOp != e; ++OMulOp)
1684 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1685 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1686 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
1687 if (Mul->getNumOperands() != 2) {
1688 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
1689 Mul->op_begin()+MulOp);
1690 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end());
1691 InnerMul1 = getMulExpr(MulOps);
1693 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1694 if (OtherMul->getNumOperands() != 2) {
1695 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
1696 OtherMul->op_begin()+OMulOp);
1697 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end());
1698 InnerMul2 = getMulExpr(MulOps);
1700 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1701 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1702 if (Ops.size() == 2) return OuterMul;
1703 Ops.erase(Ops.begin()+Idx);
1704 Ops.erase(Ops.begin()+OtherMulIdx-1);
1705 Ops.push_back(OuterMul);
1706 return getAddExpr(Ops);
1712 // If there are any add recurrences in the operands list, see if any other
1713 // added values are loop invariant. If so, we can fold them into the
1715 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1718 // Scan over all recurrences, trying to fold loop invariants into them.
1719 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1720 // Scan all of the other operands to this add and add them to the vector if
1721 // they are loop invariant w.r.t. the recurrence.
1722 SmallVector<const SCEV *, 8> LIOps;
1723 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1724 const Loop *AddRecLoop = AddRec->getLoop();
1725 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1726 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1727 LIOps.push_back(Ops[i]);
1728 Ops.erase(Ops.begin()+i);
1732 // If we found some loop invariants, fold them into the recurrence.
1733 if (!LIOps.empty()) {
1734 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1735 LIOps.push_back(AddRec->getStart());
1737 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1739 AddRecOps[0] = getAddExpr(LIOps);
1741 // Build the new addrec. Propagate the NUW and NSW flags if both the
1742 // outer add and the inner addrec are guaranteed to have no overflow.
1743 // Always propagate NW.
1744 Flags = AddRec->getNoWrapFlags(setFlags(Flags, SCEV::FlagNW));
1745 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop, Flags);
1747 // If all of the other operands were loop invariant, we are done.
1748 if (Ops.size() == 1) return NewRec;
1750 // Otherwise, add the folded AddRec by the non-invariant parts.
1751 for (unsigned i = 0;; ++i)
1752 if (Ops[i] == AddRec) {
1756 return getAddExpr(Ops);
1759 // Okay, if there weren't any loop invariants to be folded, check to see if
1760 // there are multiple AddRec's with the same loop induction variable being
1761 // added together. If so, we can fold them.
1762 for (unsigned OtherIdx = Idx+1;
1763 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1765 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1766 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L>
1767 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
1769 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1771 if (const SCEVAddRecExpr *OtherAddRec =
1772 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
1773 if (OtherAddRec->getLoop() == AddRecLoop) {
1774 for (unsigned i = 0, e = OtherAddRec->getNumOperands();
1776 if (i >= AddRecOps.size()) {
1777 AddRecOps.append(OtherAddRec->op_begin()+i,
1778 OtherAddRec->op_end());
1781 AddRecOps[i] = getAddExpr(AddRecOps[i],
1782 OtherAddRec->getOperand(i));
1784 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
1786 // Step size has changed, so we cannot guarantee no self-wraparound.
1787 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop, SCEV::FlagAnyWrap);
1788 return getAddExpr(Ops);
1791 // Otherwise couldn't fold anything into this recurrence. Move onto the
1795 // Okay, it looks like we really DO need an add expr. Check to see if we
1796 // already have one, otherwise create a new one.
1797 FoldingSetNodeID ID;
1798 ID.AddInteger(scAddExpr);
1799 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1800 ID.AddPointer(Ops[i]);
1803 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
1805 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
1806 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
1807 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator),
1809 UniqueSCEVs.InsertNode(S, IP);
1811 S->setNoWrapFlags(Flags);
1815 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1817 const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
1818 SCEV::NoWrapFlags Flags) {
1819 assert(Flags == maskFlags(Flags, SCEV::FlagNUW | SCEV::FlagNSW) &&
1820 "only nuw or nsw allowed");
1821 assert(!Ops.empty() && "Cannot get empty mul!");
1822 if (Ops.size() == 1) return Ops[0];
1824 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
1825 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1826 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
1827 "SCEVMulExpr operand types don't match!");
1830 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
1832 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
1833 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
1834 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
1836 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
1837 E = Ops.end(); I != E; ++I)
1838 if (!isKnownNonNegative(*I)) {
1842 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
1845 // Sort by complexity, this groups all similar expression types together.
1846 GroupByComplexity(Ops, LI);
1848 // If there are any constants, fold them together.
1850 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1852 // C1*(C2+V) -> C1*C2 + C1*V
1853 if (Ops.size() == 2)
1854 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1855 if (Add->getNumOperands() == 2 &&
1856 isa<SCEVConstant>(Add->getOperand(0)))
1857 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1858 getMulExpr(LHSC, Add->getOperand(1)));
1861 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1862 // We found two constants, fold them together!
1863 ConstantInt *Fold = ConstantInt::get(getContext(),
1864 LHSC->getValue()->getValue() *
1865 RHSC->getValue()->getValue());
1866 Ops[0] = getConstant(Fold);
1867 Ops.erase(Ops.begin()+1); // Erase the folded element
1868 if (Ops.size() == 1) return Ops[0];
1869 LHSC = cast<SCEVConstant>(Ops[0]);
1872 // If we are left with a constant one being multiplied, strip it off.
1873 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1874 Ops.erase(Ops.begin());
1876 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1877 // If we have a multiply of zero, it will always be zero.
1879 } else if (Ops[0]->isAllOnesValue()) {
1880 // If we have a mul by -1 of an add, try distributing the -1 among the
1882 if (Ops.size() == 2) {
1883 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
1884 SmallVector<const SCEV *, 4> NewOps;
1885 bool AnyFolded = false;
1886 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(),
1887 E = Add->op_end(); I != E; ++I) {
1888 const SCEV *Mul = getMulExpr(Ops[0], *I);
1889 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
1890 NewOps.push_back(Mul);
1893 return getAddExpr(NewOps);
1895 else if (const SCEVAddRecExpr *
1896 AddRec = dyn_cast<SCEVAddRecExpr>(Ops[1])) {
1897 // Negation preserves a recurrence's no self-wrap property.
1898 SmallVector<const SCEV *, 4> Operands;
1899 for (SCEVAddRecExpr::op_iterator I = AddRec->op_begin(),
1900 E = AddRec->op_end(); I != E; ++I) {
1901 Operands.push_back(getMulExpr(Ops[0], *I));
1903 return getAddRecExpr(Operands, AddRec->getLoop(),
1904 AddRec->getNoWrapFlags(SCEV::FlagNW));
1909 if (Ops.size() == 1)
1913 // Skip over the add expression until we get to a multiply.
1914 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1917 // If there are mul operands inline them all into this expression.
1918 if (Idx < Ops.size()) {
1919 bool DeletedMul = false;
1920 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1921 // If we have an mul, expand the mul operands onto the end of the operands
1923 Ops.erase(Ops.begin()+Idx);
1924 Ops.append(Mul->op_begin(), Mul->op_end());
1928 // If we deleted at least one mul, we added operands to the end of the list,
1929 // and they are not necessarily sorted. Recurse to resort and resimplify
1930 // any operands we just acquired.
1932 return getMulExpr(Ops);
1935 // If there are any add recurrences in the operands list, see if any other
1936 // added values are loop invariant. If so, we can fold them into the
1938 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1941 // Scan over all recurrences, trying to fold loop invariants into them.
1942 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1943 // Scan all of the other operands to this mul and add them to the vector if
1944 // they are loop invariant w.r.t. the recurrence.
1945 SmallVector<const SCEV *, 8> LIOps;
1946 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1947 const Loop *AddRecLoop = AddRec->getLoop();
1948 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1949 if (isLoopInvariant(Ops[i], AddRecLoop)) {
1950 LIOps.push_back(Ops[i]);
1951 Ops.erase(Ops.begin()+i);
1955 // If we found some loop invariants, fold them into the recurrence.
1956 if (!LIOps.empty()) {
1957 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1958 SmallVector<const SCEV *, 4> NewOps;
1959 NewOps.reserve(AddRec->getNumOperands());
1960 const SCEV *Scale = getMulExpr(LIOps);
1961 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1962 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1964 // Build the new addrec. Propagate the NUW and NSW flags if both the
1965 // outer mul and the inner addrec are guaranteed to have no overflow.
1967 // No self-wrap cannot be guaranteed after changing the step size, but
1968 // will be inferred if either NUW or NSW is true.
1969 Flags = AddRec->getNoWrapFlags(clearFlags(Flags, SCEV::FlagNW));
1970 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop, Flags);
1972 // If all of the other operands were loop invariant, we are done.
1973 if (Ops.size() == 1) return NewRec;
1975 // Otherwise, multiply the folded AddRec by the non-invariant parts.
1976 for (unsigned i = 0;; ++i)
1977 if (Ops[i] == AddRec) {
1981 return getMulExpr(Ops);
1984 // Okay, if there weren't any loop invariants to be folded, check to see if
1985 // there are multiple AddRec's with the same loop induction variable being
1986 // multiplied together. If so, we can fold them.
1987 for (unsigned OtherIdx = Idx+1;
1988 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
1991 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
1992 // {A,+,B}<L> * {C,+,D}<L> --> {A*C,+,A*D + B*C + B*D,+,2*B*D}<L>
1994 // {A,+,B} * {C,+,D} = A+It*B * C+It*D = A*C + (A*D + B*C)*It + B*D*It^2
1995 // Given an equation of the form x + y*It + z*It^2 (above), we want to
1996 // express it in terms of {X,+,Y,+,Z}.
1997 // {X,+,Y,+,Z} = X + Y*It + Z*(It^2 - It)/2.
1998 // Rearranging, X = x, Y = y+z, Z = 2z.
2000 // x = A*C, y = (A*D + B*C), z = B*D.
2001 // Therefore X = A*C, Y = A*D + B*C + B*D and Z = 2*B*D.
2002 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
2004 if (const SCEVAddRecExpr *OtherAddRec =
2005 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
2006 if (OtherAddRec->getLoop() == AddRecLoop) {
2007 const SCEV *A = AddRec->getStart();
2008 const SCEV *B = AddRec->getStepRecurrence(*this);
2009 const SCEV *C = OtherAddRec->getStart();
2010 const SCEV *D = OtherAddRec->getStepRecurrence(*this);
2011 const SCEV *NewStart = getMulExpr(A, C);
2012 const SCEV *BD = getMulExpr(B, D);
2013 const SCEV *NewStep = getAddExpr(getMulExpr(A, D),
2014 getMulExpr(B, C), BD);
2015 const SCEV *NewSecondOrderStep =
2016 getMulExpr(BD, getConstant(BD->getType(), 2));
2018 // This can happen when AddRec or OtherAddRec have >3 operands.
2019 // TODO: support these add-recs.
2020 if (isLoopInvariant(NewStart, AddRecLoop) &&
2021 isLoopInvariant(NewStep, AddRecLoop) &&
2022 isLoopInvariant(NewSecondOrderStep, AddRecLoop)) {
2023 SmallVector<const SCEV *, 3> AddRecOps;
2024 AddRecOps.push_back(NewStart);
2025 AddRecOps.push_back(NewStep);
2026 AddRecOps.push_back(NewSecondOrderStep);
2027 const SCEV *NewAddRec = getAddRecExpr(AddRecOps,
2030 if (Ops.size() == 2) return NewAddRec;
2031 Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
2032 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
2037 return getMulExpr(Ops);
2041 // Otherwise couldn't fold anything into this recurrence. Move onto the
2045 // Okay, it looks like we really DO need an mul expr. Check to see if we
2046 // already have one, otherwise create a new one.
2047 FoldingSetNodeID ID;
2048 ID.AddInteger(scMulExpr);
2049 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2050 ID.AddPointer(Ops[i]);
2053 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2055 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2056 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2057 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator),
2059 UniqueSCEVs.InsertNode(S, IP);
2061 S->setNoWrapFlags(Flags);
2065 /// getUDivExpr - Get a canonical unsigned division expression, or something
2066 /// simpler if possible.
2067 const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
2069 assert(getEffectiveSCEVType(LHS->getType()) ==
2070 getEffectiveSCEVType(RHS->getType()) &&
2071 "SCEVUDivExpr operand types don't match!");
2073 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
2074 if (RHSC->getValue()->equalsInt(1))
2075 return LHS; // X udiv 1 --> x
2076 // If the denominator is zero, the result of the udiv is undefined. Don't
2077 // try to analyze it, because the resolution chosen here may differ from
2078 // the resolution chosen in other parts of the compiler.
2079 if (!RHSC->getValue()->isZero()) {
2080 // Determine if the division can be folded into the operands of
2082 // TODO: Generalize this to non-constants by using known-bits information.
2083 Type *Ty = LHS->getType();
2084 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
2085 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
2086 // For non-power-of-two values, effectively round the value up to the
2087 // nearest power of two.
2088 if (!RHSC->getValue()->getValue().isPowerOf2())
2090 IntegerType *ExtTy =
2091 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
2092 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
2093 if (const SCEVConstant *Step =
2094 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this))) {
2095 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
2096 const APInt &StepInt = Step->getValue()->getValue();
2097 const APInt &DivInt = RHSC->getValue()->getValue();
2098 if (!StepInt.urem(DivInt) &&
2099 getZeroExtendExpr(AR, ExtTy) ==
2100 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2101 getZeroExtendExpr(Step, ExtTy),
2102 AR->getLoop(), SCEV::FlagAnyWrap)) {
2103 SmallVector<const SCEV *, 4> Operands;
2104 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
2105 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
2106 return getAddRecExpr(Operands, AR->getLoop(),
2109 /// Get a canonical UDivExpr for a recurrence.
2110 /// {X,+,N}/C => {Y,+,N}/C where Y=X-(X%N). Safe when C%N=0.
2111 // We can currently only fold X%N if X is constant.
2112 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
2113 if (StartC && !DivInt.urem(StepInt) &&
2114 getZeroExtendExpr(AR, ExtTy) ==
2115 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
2116 getZeroExtendExpr(Step, ExtTy),
2117 AR->getLoop(), SCEV::FlagAnyWrap)) {
2118 const APInt &StartInt = StartC->getValue()->getValue();
2119 const APInt &StartRem = StartInt.urem(StepInt);
2121 LHS = getAddRecExpr(getConstant(StartInt - StartRem), Step,
2122 AR->getLoop(), SCEV::FlagNW);
2125 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
2126 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
2127 SmallVector<const SCEV *, 4> Operands;
2128 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
2129 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
2130 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
2131 // Find an operand that's safely divisible.
2132 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
2133 const SCEV *Op = M->getOperand(i);
2134 const SCEV *Div = getUDivExpr(Op, RHSC);
2135 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
2136 Operands = SmallVector<const SCEV *, 4>(M->op_begin(),
2139 return getMulExpr(Operands);
2143 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
2144 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(LHS)) {
2145 SmallVector<const SCEV *, 4> Operands;
2146 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
2147 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
2148 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
2150 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
2151 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
2152 if (isa<SCEVUDivExpr>(Op) ||
2153 getMulExpr(Op, RHS) != A->getOperand(i))
2155 Operands.push_back(Op);
2157 if (Operands.size() == A->getNumOperands())
2158 return getAddExpr(Operands);
2162 // Fold if both operands are constant.
2163 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
2164 Constant *LHSCV = LHSC->getValue();
2165 Constant *RHSCV = RHSC->getValue();
2166 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
2172 FoldingSetNodeID ID;
2173 ID.AddInteger(scUDivExpr);
2177 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2178 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator),
2180 UniqueSCEVs.InsertNode(S, IP);
2185 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2186 /// Simplify the expression as much as possible.
2187 const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start, const SCEV *Step,
2189 SCEV::NoWrapFlags Flags) {
2190 SmallVector<const SCEV *, 4> Operands;
2191 Operands.push_back(Start);
2192 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
2193 if (StepChrec->getLoop() == L) {
2194 Operands.append(StepChrec->op_begin(), StepChrec->op_end());
2195 return getAddRecExpr(Operands, L, maskFlags(Flags, SCEV::FlagNW));
2198 Operands.push_back(Step);
2199 return getAddRecExpr(Operands, L, Flags);
2202 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
2203 /// Simplify the expression as much as possible.
2205 ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
2206 const Loop *L, SCEV::NoWrapFlags Flags) {
2207 if (Operands.size() == 1) return Operands[0];
2209 Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
2210 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
2211 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
2212 "SCEVAddRecExpr operand types don't match!");
2213 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2214 assert(isLoopInvariant(Operands[i], L) &&
2215 "SCEVAddRecExpr operand is not loop-invariant!");
2218 if (Operands.back()->isZero()) {
2219 Operands.pop_back();
2220 return getAddRecExpr(Operands, L, SCEV::FlagAnyWrap); // {X,+,0} --> X
2223 // It's tempting to want to call getMaxBackedgeTakenCount count here and
2224 // use that information to infer NUW and NSW flags. However, computing a
2225 // BE count requires calling getAddRecExpr, so we may not yet have a
2226 // meaningful BE count at this point (and if we don't, we'd be stuck
2227 // with a SCEVCouldNotCompute as the cached BE count).
2229 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
2231 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
2232 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
2233 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
2235 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
2236 E = Operands.end(); I != E; ++I)
2237 if (!isKnownNonNegative(*I)) {
2241 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
2244 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
2245 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
2246 const Loop *NestedLoop = NestedAR->getLoop();
2247 if (L->contains(NestedLoop) ?
2248 (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
2249 (!NestedLoop->contains(L) &&
2250 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
2251 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
2252 NestedAR->op_end());
2253 Operands[0] = NestedAR->getStart();
2254 // AddRecs require their operands be loop-invariant with respect to their
2255 // loops. Don't perform this transformation if it would break this
2257 bool AllInvariant = true;
2258 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2259 if (!isLoopInvariant(Operands[i], L)) {
2260 AllInvariant = false;
2264 // Create a recurrence for the outer loop with the same step size.
2266 // The outer recurrence keeps its NW flag but only keeps NUW/NSW if the
2267 // inner recurrence has the same property.
2268 SCEV::NoWrapFlags OuterFlags =
2269 maskFlags(Flags, SCEV::FlagNW | NestedAR->getNoWrapFlags());
2271 NestedOperands[0] = getAddRecExpr(Operands, L, OuterFlags);
2272 AllInvariant = true;
2273 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
2274 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
2275 AllInvariant = false;
2279 // Ok, both add recurrences are valid after the transformation.
2281 // The inner recurrence keeps its NW flag but only keeps NUW/NSW if
2282 // the outer recurrence has the same property.
2283 SCEV::NoWrapFlags InnerFlags =
2284 maskFlags(NestedAR->getNoWrapFlags(), SCEV::FlagNW | Flags);
2285 return getAddRecExpr(NestedOperands, NestedLoop, InnerFlags);
2288 // Reset Operands to its original state.
2289 Operands[0] = NestedAR;
2293 // Okay, it looks like we really DO need an addrec expr. Check to see if we
2294 // already have one, otherwise create a new one.
2295 FoldingSetNodeID ID;
2296 ID.AddInteger(scAddRecExpr);
2297 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
2298 ID.AddPointer(Operands[i]);
2302 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
2304 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size());
2305 std::uninitialized_copy(Operands.begin(), Operands.end(), O);
2306 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator),
2307 O, Operands.size(), L);
2308 UniqueSCEVs.InsertNode(S, IP);
2310 S->setNoWrapFlags(Flags);
2314 const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
2316 SmallVector<const SCEV *, 2> Ops;
2319 return getSMaxExpr(Ops);
2323 ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2324 assert(!Ops.empty() && "Cannot get empty smax!");
2325 if (Ops.size() == 1) return Ops[0];
2327 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2328 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2329 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2330 "SCEVSMaxExpr operand types don't match!");
2333 // Sort by complexity, this groups all similar expression types together.
2334 GroupByComplexity(Ops, LI);
2336 // If there are any constants, fold them together.
2338 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2340 assert(Idx < Ops.size());
2341 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2342 // We found two constants, fold them together!
2343 ConstantInt *Fold = ConstantInt::get(getContext(),
2344 APIntOps::smax(LHSC->getValue()->getValue(),
2345 RHSC->getValue()->getValue()));
2346 Ops[0] = getConstant(Fold);
2347 Ops.erase(Ops.begin()+1); // Erase the folded element
2348 if (Ops.size() == 1) return Ops[0];
2349 LHSC = cast<SCEVConstant>(Ops[0]);
2352 // If we are left with a constant minimum-int, strip it off.
2353 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
2354 Ops.erase(Ops.begin());
2356 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
2357 // If we have an smax with a constant maximum-int, it will always be
2362 if (Ops.size() == 1) return Ops[0];
2365 // Find the first SMax
2366 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
2369 // Check to see if one of the operands is an SMax. If so, expand its operands
2370 // onto our operand list, and recurse to simplify.
2371 if (Idx < Ops.size()) {
2372 bool DeletedSMax = false;
2373 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
2374 Ops.erase(Ops.begin()+Idx);
2375 Ops.append(SMax->op_begin(), SMax->op_end());
2380 return getSMaxExpr(Ops);
2383 // Okay, check to see if the same value occurs in the operand list twice. If
2384 // so, delete one. Since we sorted the list, these values are required to
2386 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2387 // X smax Y smax Y --> X smax Y
2388 // X smax Y --> X, if X is always greater than Y
2389 if (Ops[i] == Ops[i+1] ||
2390 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) {
2391 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2393 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) {
2394 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2398 if (Ops.size() == 1) return Ops[0];
2400 assert(!Ops.empty() && "Reduced smax down to nothing!");
2402 // Okay, it looks like we really DO need an smax expr. Check to see if we
2403 // already have one, otherwise create a new one.
2404 FoldingSetNodeID ID;
2405 ID.AddInteger(scSMaxExpr);
2406 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2407 ID.AddPointer(Ops[i]);
2409 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2410 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2411 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2412 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator),
2414 UniqueSCEVs.InsertNode(S, IP);
2418 const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
2420 SmallVector<const SCEV *, 2> Ops;
2423 return getUMaxExpr(Ops);
2427 ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
2428 assert(!Ops.empty() && "Cannot get empty umax!");
2429 if (Ops.size() == 1) return Ops[0];
2431 Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
2432 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
2433 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
2434 "SCEVUMaxExpr operand types don't match!");
2437 // Sort by complexity, this groups all similar expression types together.
2438 GroupByComplexity(Ops, LI);
2440 // If there are any constants, fold them together.
2442 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
2444 assert(Idx < Ops.size());
2445 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
2446 // We found two constants, fold them together!
2447 ConstantInt *Fold = ConstantInt::get(getContext(),
2448 APIntOps::umax(LHSC->getValue()->getValue(),
2449 RHSC->getValue()->getValue()));
2450 Ops[0] = getConstant(Fold);
2451 Ops.erase(Ops.begin()+1); // Erase the folded element
2452 if (Ops.size() == 1) return Ops[0];
2453 LHSC = cast<SCEVConstant>(Ops[0]);
2456 // If we are left with a constant minimum-int, strip it off.
2457 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
2458 Ops.erase(Ops.begin());
2460 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
2461 // If we have an umax with a constant maximum-int, it will always be
2466 if (Ops.size() == 1) return Ops[0];
2469 // Find the first UMax
2470 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
2473 // Check to see if one of the operands is a UMax. If so, expand its operands
2474 // onto our operand list, and recurse to simplify.
2475 if (Idx < Ops.size()) {
2476 bool DeletedUMax = false;
2477 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
2478 Ops.erase(Ops.begin()+Idx);
2479 Ops.append(UMax->op_begin(), UMax->op_end());
2484 return getUMaxExpr(Ops);
2487 // Okay, check to see if the same value occurs in the operand list twice. If
2488 // so, delete one. Since we sorted the list, these values are required to
2490 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
2491 // X umax Y umax Y --> X umax Y
2492 // X umax Y --> X, if X is always greater than Y
2493 if (Ops[i] == Ops[i+1] ||
2494 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) {
2495 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2);
2497 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) {
2498 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
2502 if (Ops.size() == 1) return Ops[0];
2504 assert(!Ops.empty() && "Reduced umax down to nothing!");
2506 // Okay, it looks like we really DO need a umax expr. Check to see if we
2507 // already have one, otherwise create a new one.
2508 FoldingSetNodeID ID;
2509 ID.AddInteger(scUMaxExpr);
2510 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2511 ID.AddPointer(Ops[i]);
2513 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
2514 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size());
2515 std::uninitialized_copy(Ops.begin(), Ops.end(), O);
2516 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator),
2518 UniqueSCEVs.InsertNode(S, IP);
2522 const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
2524 // ~smax(~x, ~y) == smin(x, y).
2525 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2528 const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
2530 // ~umax(~x, ~y) == umin(x, y)
2531 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
2534 const SCEV *ScalarEvolution::getSizeOfExpr(Type *AllocTy) {
2535 // If we have TargetData, we can bypass creating a target-independent
2536 // constant expression and then folding it back into a ConstantInt.
2537 // This is just a compile-time optimization.
2539 return getConstant(TD->getIntPtrType(getContext()),
2540 TD->getTypeAllocSize(AllocTy));
2542 Constant *C = ConstantExpr::getSizeOf(AllocTy);
2543 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2544 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2546 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2547 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2550 const SCEV *ScalarEvolution::getAlignOfExpr(Type *AllocTy) {
2551 Constant *C = ConstantExpr::getAlignOf(AllocTy);
2552 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2553 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2555 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
2556 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2559 const SCEV *ScalarEvolution::getOffsetOfExpr(StructType *STy,
2561 // If we have TargetData, we can bypass creating a target-independent
2562 // constant expression and then folding it back into a ConstantInt.
2563 // This is just a compile-time optimization.
2565 return getConstant(TD->getIntPtrType(getContext()),
2566 TD->getStructLayout(STy)->getElementOffset(FieldNo));
2568 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
2569 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2570 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2572 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
2573 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2576 const SCEV *ScalarEvolution::getOffsetOfExpr(Type *CTy,
2577 Constant *FieldNo) {
2578 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
2579 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2580 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
2582 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
2583 return getTruncateOrZeroExtend(getSCEV(C), Ty);
2586 const SCEV *ScalarEvolution::getUnknown(Value *V) {
2587 // Don't attempt to do anything other than create a SCEVUnknown object
2588 // here. createSCEV only calls getUnknown after checking for all other
2589 // interesting possibilities, and any other code that calls getUnknown
2590 // is doing so in order to hide a value from SCEV canonicalization.
2592 FoldingSetNodeID ID;
2593 ID.AddInteger(scUnknown);
2596 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
2597 assert(cast<SCEVUnknown>(S)->getValue() == V &&
2598 "Stale SCEVUnknown in uniquing map!");
2601 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
2603 FirstUnknown = cast<SCEVUnknown>(S);
2604 UniqueSCEVs.InsertNode(S, IP);
2608 //===----------------------------------------------------------------------===//
2609 // Basic SCEV Analysis and PHI Idiom Recognition Code
2612 /// isSCEVable - Test if values of the given type are analyzable within
2613 /// the SCEV framework. This primarily includes integer types, and it
2614 /// can optionally include pointer types if the ScalarEvolution class
2615 /// has access to target-specific information.
2616 bool ScalarEvolution::isSCEVable(Type *Ty) const {
2617 // Integers and pointers are always SCEVable.
2618 return Ty->isIntegerTy() || Ty->isPointerTy();
2621 /// getTypeSizeInBits - Return the size in bits of the specified type,
2622 /// for which isSCEVable must return true.
2623 uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const {
2624 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2626 // If we have a TargetData, use it!
2628 return TD->getTypeSizeInBits(Ty);
2630 // Integer types have fixed sizes.
2631 if (Ty->isIntegerTy())
2632 return Ty->getPrimitiveSizeInBits();
2634 // The only other support type is pointer. Without TargetData, conservatively
2635 // assume pointers are 64-bit.
2636 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
2640 /// getEffectiveSCEVType - Return a type with the same bitwidth as
2641 /// the given type and which represents how SCEV will treat the given
2642 /// type, for which isSCEVable must return true. For pointer types,
2643 /// this is the pointer-sized integer type.
2644 Type *ScalarEvolution::getEffectiveSCEVType(Type *Ty) const {
2645 assert(isSCEVable(Ty) && "Type is not SCEVable!");
2647 if (Ty->isIntegerTy())
2650 // The only other support type is pointer.
2651 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
2652 if (TD) return TD->getIntPtrType(getContext());
2654 // Without TargetData, conservatively assume pointers are 64-bit.
2655 return Type::getInt64Ty(getContext());
2658 const SCEV *ScalarEvolution::getCouldNotCompute() {
2659 return &CouldNotCompute;
2662 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
2663 /// expression and create a new one.
2664 const SCEV *ScalarEvolution::getSCEV(Value *V) {
2665 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
2667 ValueExprMapType::const_iterator I = ValueExprMap.find(V);
2668 if (I != ValueExprMap.end()) return I->second;
2669 const SCEV *S = createSCEV(V);
2671 // The process of creating a SCEV for V may have caused other SCEVs
2672 // to have been created, so it's necessary to insert the new entry
2673 // from scratch, rather than trying to remember the insert position
2675 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S));
2679 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
2681 const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
2682 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2684 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
2686 Type *Ty = V->getType();
2687 Ty = getEffectiveSCEVType(Ty);
2688 return getMulExpr(V,
2689 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
2692 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
2693 const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
2694 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
2696 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
2698 Type *Ty = V->getType();
2699 Ty = getEffectiveSCEVType(Ty);
2700 const SCEV *AllOnes =
2701 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
2702 return getMinusSCEV(AllOnes, V);
2705 /// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1.
2706 const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
2707 SCEV::NoWrapFlags Flags) {
2708 assert(!maskFlags(Flags, SCEV::FlagNUW) && "subtraction does not have NUW");
2710 // Fast path: X - X --> 0.
2712 return getConstant(LHS->getType(), 0);
2715 return getAddExpr(LHS, getNegativeSCEV(RHS), Flags);
2718 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
2719 /// input value to the specified type. If the type must be extended, it is zero
2722 ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, Type *Ty) {
2723 Type *SrcTy = V->getType();
2724 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2725 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2726 "Cannot truncate or zero extend with non-integer arguments!");
2727 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2728 return V; // No conversion
2729 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2730 return getTruncateExpr(V, Ty);
2731 return getZeroExtendExpr(V, Ty);
2734 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
2735 /// input value to the specified type. If the type must be extended, it is sign
2738 ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
2740 Type *SrcTy = V->getType();
2741 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2742 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2743 "Cannot truncate or zero extend with non-integer arguments!");
2744 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2745 return V; // No conversion
2746 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
2747 return getTruncateExpr(V, Ty);
2748 return getSignExtendExpr(V, Ty);
2751 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
2752 /// input value to the specified type. If the type must be extended, it is zero
2753 /// extended. The conversion must not be narrowing.
2755 ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, Type *Ty) {
2756 Type *SrcTy = V->getType();
2757 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2758 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2759 "Cannot noop or zero extend with non-integer arguments!");
2760 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2761 "getNoopOrZeroExtend cannot truncate!");
2762 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2763 return V; // No conversion
2764 return getZeroExtendExpr(V, Ty);
2767 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
2768 /// input value to the specified type. If the type must be extended, it is sign
2769 /// extended. The conversion must not be narrowing.
2771 ScalarEvolution::getNoopOrSignExtend(const SCEV *V, Type *Ty) {
2772 Type *SrcTy = V->getType();
2773 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2774 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2775 "Cannot noop or sign extend with non-integer arguments!");
2776 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2777 "getNoopOrSignExtend cannot truncate!");
2778 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2779 return V; // No conversion
2780 return getSignExtendExpr(V, Ty);
2783 /// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
2784 /// the input value to the specified type. If the type must be extended,
2785 /// it is extended with unspecified bits. The conversion must not be
2788 ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, Type *Ty) {
2789 Type *SrcTy = V->getType();
2790 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2791 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2792 "Cannot noop or any extend with non-integer arguments!");
2793 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
2794 "getNoopOrAnyExtend cannot truncate!");
2795 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2796 return V; // No conversion
2797 return getAnyExtendExpr(V, Ty);
2800 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
2801 /// input value to the specified type. The conversion must not be widening.
2803 ScalarEvolution::getTruncateOrNoop(const SCEV *V, Type *Ty) {
2804 Type *SrcTy = V->getType();
2805 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
2806 (Ty->isIntegerTy() || Ty->isPointerTy()) &&
2807 "Cannot truncate or noop with non-integer arguments!");
2808 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
2809 "getTruncateOrNoop cannot extend!");
2810 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
2811 return V; // No conversion
2812 return getTruncateExpr(V, Ty);
2815 /// getUMaxFromMismatchedTypes - Promote the operands to the wider of
2816 /// the types using zero-extension, and then perform a umax operation
2818 const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
2820 const SCEV *PromotedLHS = LHS;
2821 const SCEV *PromotedRHS = RHS;
2823 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2824 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2826 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2828 return getUMaxExpr(PromotedLHS, PromotedRHS);
2831 /// getUMinFromMismatchedTypes - Promote the operands to the wider of
2832 /// the types using zero-extension, and then perform a umin operation
2834 const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
2836 const SCEV *PromotedLHS = LHS;
2837 const SCEV *PromotedRHS = RHS;
2839 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
2840 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
2842 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
2844 return getUMinExpr(PromotedLHS, PromotedRHS);
2847 /// getPointerBase - Transitively follow the chain of pointer-type operands
2848 /// until reaching a SCEV that does not have a single pointer operand. This
2849 /// returns a SCEVUnknown pointer for well-formed pointer-type expressions,
2850 /// but corner cases do exist.
2851 const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) {
2852 // A pointer operand may evaluate to a nonpointer expression, such as null.
2853 if (!V->getType()->isPointerTy())
2856 if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(V)) {
2857 return getPointerBase(Cast->getOperand());
2859 else if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(V)) {
2860 const SCEV *PtrOp = 0;
2861 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
2863 if ((*I)->getType()->isPointerTy()) {
2864 // Cannot find the base of an expression with multiple pointer operands.
2872 return getPointerBase(PtrOp);
2877 /// PushDefUseChildren - Push users of the given Instruction
2878 /// onto the given Worklist.
2880 PushDefUseChildren(Instruction *I,
2881 SmallVectorImpl<Instruction *> &Worklist) {
2882 // Push the def-use children onto the Worklist stack.
2883 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2885 Worklist.push_back(cast<Instruction>(*UI));
2888 /// ForgetSymbolicValue - This looks up computed SCEV values for all
2889 /// instructions that depend on the given instruction and removes them from
2890 /// the ValueExprMapType map if they reference SymName. This is used during PHI
2893 ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) {
2894 SmallVector<Instruction *, 16> Worklist;
2895 PushDefUseChildren(PN, Worklist);
2897 SmallPtrSet<Instruction *, 8> Visited;
2899 while (!Worklist.empty()) {
2900 Instruction *I = Worklist.pop_back_val();
2901 if (!Visited.insert(I)) continue;
2903 ValueExprMapType::iterator It =
2904 ValueExprMap.find(static_cast<Value *>(I));
2905 if (It != ValueExprMap.end()) {
2906 const SCEV *Old = It->second;
2908 // Short-circuit the def-use traversal if the symbolic name
2909 // ceases to appear in expressions.
2910 if (Old != SymName && !hasOperand(Old, SymName))
2913 // SCEVUnknown for a PHI either means that it has an unrecognized
2914 // structure, it's a PHI that's in the progress of being computed
2915 // by createNodeForPHI, or it's a single-value PHI. In the first case,
2916 // additional loop trip count information isn't going to change anything.
2917 // In the second case, createNodeForPHI will perform the necessary
2918 // updates on its own when it gets to that point. In the third, we do
2919 // want to forget the SCEVUnknown.
2920 if (!isa<PHINode>(I) ||
2921 !isa<SCEVUnknown>(Old) ||
2922 (I != PN && Old == SymName)) {
2923 forgetMemoizedResults(Old);
2924 ValueExprMap.erase(It);
2928 PushDefUseChildren(I, Worklist);
2932 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
2933 /// a loop header, making it a potential recurrence, or it doesn't.
2935 const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
2936 if (const Loop *L = LI->getLoopFor(PN->getParent()))
2937 if (L->getHeader() == PN->getParent()) {
2938 // The loop may have multiple entrances or multiple exits; we can analyze
2939 // this phi as an addrec if it has a unique entry value and a unique
2941 Value *BEValueV = 0, *StartValueV = 0;
2942 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2943 Value *V = PN->getIncomingValue(i);
2944 if (L->contains(PN->getIncomingBlock(i))) {
2947 } else if (BEValueV != V) {
2951 } else if (!StartValueV) {
2953 } else if (StartValueV != V) {
2958 if (BEValueV && StartValueV) {
2959 // While we are analyzing this PHI node, handle its value symbolically.
2960 const SCEV *SymbolicName = getUnknown(PN);
2961 assert(ValueExprMap.find(PN) == ValueExprMap.end() &&
2962 "PHI node already processed?");
2963 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
2965 // Using this symbolic name for the PHI, analyze the value coming around
2967 const SCEV *BEValue = getSCEV(BEValueV);
2969 // NOTE: If BEValue is loop invariant, we know that the PHI node just
2970 // has a special value for the first iteration of the loop.
2972 // If the value coming around the backedge is an add with the symbolic
2973 // value we just inserted, then we found a simple induction variable!
2974 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
2975 // If there is a single occurrence of the symbolic value, replace it
2976 // with a recurrence.
2977 unsigned FoundIndex = Add->getNumOperands();
2978 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2979 if (Add->getOperand(i) == SymbolicName)
2980 if (FoundIndex == e) {
2985 if (FoundIndex != Add->getNumOperands()) {
2986 // Create an add with everything but the specified operand.
2987 SmallVector<const SCEV *, 8> Ops;
2988 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
2989 if (i != FoundIndex)
2990 Ops.push_back(Add->getOperand(i));
2991 const SCEV *Accum = getAddExpr(Ops);
2993 // This is not a valid addrec if the step amount is varying each
2994 // loop iteration, but is not itself an addrec in this loop.
2995 if (isLoopInvariant(Accum, L) ||
2996 (isa<SCEVAddRecExpr>(Accum) &&
2997 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
2998 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap;
3000 // If the increment doesn't overflow, then neither the addrec nor
3001 // the post-increment will overflow.
3002 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
3003 if (OBO->hasNoUnsignedWrap())
3004 Flags = setFlags(Flags, SCEV::FlagNUW);
3005 if (OBO->hasNoSignedWrap())
3006 Flags = setFlags(Flags, SCEV::FlagNSW);
3007 } else if (const GEPOperator *GEP =
3008 dyn_cast<GEPOperator>(BEValueV)) {
3009 // If the increment is an inbounds GEP, then we know the address
3010 // space cannot be wrapped around. We cannot make any guarantee
3011 // about signed or unsigned overflow because pointers are
3012 // unsigned but we may have a negative index from the base
3014 if (GEP->isInBounds())
3015 Flags = setFlags(Flags, SCEV::FlagNW);
3018 const SCEV *StartVal = getSCEV(StartValueV);
3019 const SCEV *PHISCEV = getAddRecExpr(StartVal, Accum, L, Flags);
3021 // Since the no-wrap flags are on the increment, they apply to the
3022 // post-incremented value as well.
3023 if (isLoopInvariant(Accum, L))
3024 (void)getAddRecExpr(getAddExpr(StartVal, Accum),
3027 // Okay, for the entire analysis of this edge we assumed the PHI
3028 // to be symbolic. We now need to go back and purge all of the
3029 // entries for the scalars that use the symbolic expression.
3030 ForgetSymbolicName(PN, SymbolicName);
3031 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3035 } else if (const SCEVAddRecExpr *AddRec =
3036 dyn_cast<SCEVAddRecExpr>(BEValue)) {
3037 // Otherwise, this could be a loop like this:
3038 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
3039 // In this case, j = {1,+,1} and BEValue is j.
3040 // Because the other in-value of i (0) fits the evolution of BEValue
3041 // i really is an addrec evolution.
3042 if (AddRec->getLoop() == L && AddRec->isAffine()) {
3043 const SCEV *StartVal = getSCEV(StartValueV);
3045 // If StartVal = j.start - j.stride, we can use StartVal as the
3046 // initial step of the addrec evolution.
3047 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
3048 AddRec->getOperand(1))) {
3049 // FIXME: For constant StartVal, we should be able to infer
3051 const SCEV *PHISCEV =
3052 getAddRecExpr(StartVal, AddRec->getOperand(1), L,
3055 // Okay, for the entire analysis of this edge we assumed the PHI
3056 // to be symbolic. We now need to go back and purge all of the
3057 // entries for the scalars that use the symbolic expression.
3058 ForgetSymbolicName(PN, SymbolicName);
3059 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV;
3067 // If the PHI has a single incoming value, follow that value, unless the
3068 // PHI's incoming blocks are in a different loop, in which case doing so
3069 // risks breaking LCSSA form. Instcombine would normally zap these, but
3070 // it doesn't have DominatorTree information, so it may miss cases.
3071 if (Value *V = SimplifyInstruction(PN, TD, DT))
3072 if (LI->replacementPreservesLCSSAForm(PN, V))
3075 // If it's not a loop phi, we can't handle it yet.
3076 return getUnknown(PN);
3079 /// createNodeForGEP - Expand GEP instructions into add and multiply
3080 /// operations. This allows them to be analyzed by regular SCEV code.
3082 const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
3084 // Don't blindly transfer the inbounds flag from the GEP instruction to the
3085 // Add expression, because the Instruction may be guarded by control flow
3086 // and the no-overflow bits may not be valid for the expression in any
3088 bool isInBounds = GEP->isInBounds();
3090 Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
3091 Value *Base = GEP->getOperand(0);
3092 // Don't attempt to analyze GEPs over unsized objects.
3093 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
3094 return getUnknown(GEP);
3095 const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
3096 gep_type_iterator GTI = gep_type_begin(GEP);
3097 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
3101 // Compute the (potentially symbolic) offset in bytes for this index.
3102 if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
3103 // For a struct, add the member offset.
3104 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
3105 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
3107 // Add the field offset to the running total offset.
3108 TotalOffset = getAddExpr(TotalOffset, FieldOffset);
3110 // For an array, add the element offset, explicitly scaled.
3111 const SCEV *ElementSize = getSizeOfExpr(*GTI);
3112 const SCEV *IndexS = getSCEV(Index);
3113 // Getelementptr indices are signed.
3114 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
3116 // Multiply the index by the element size to compute the element offset.
3117 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize,
3118 isInBounds ? SCEV::FlagNSW :
3121 // Add the element offset to the running total offset.
3122 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
3126 // Get the SCEV for the GEP base.
3127 const SCEV *BaseS = getSCEV(Base);
3129 // Add the total offset from all the GEP indices to the base.
3130 return getAddExpr(BaseS, TotalOffset,
3131 isInBounds ? SCEV::FlagNSW : SCEV::FlagAnyWrap);
3134 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
3135 /// guaranteed to end in (at every loop iteration). It is, at the same time,
3136 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
3137 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
3139 ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
3140 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3141 return C->getValue()->getValue().countTrailingZeros();
3143 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
3144 return std::min(GetMinTrailingZeros(T->getOperand()),
3145 (uint32_t)getTypeSizeInBits(T->getType()));
3147 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
3148 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3149 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3150 getTypeSizeInBits(E->getType()) : OpRes;
3153 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
3154 uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
3155 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
3156 getTypeSizeInBits(E->getType()) : OpRes;
3159 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
3160 // The result is the min of all operands results.
3161 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3162 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3163 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3167 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
3168 // The result is the sum of all operands results.
3169 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
3170 uint32_t BitWidth = getTypeSizeInBits(M->getType());
3171 for (unsigned i = 1, e = M->getNumOperands();
3172 SumOpRes != BitWidth && i != e; ++i)
3173 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
3178 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
3179 // The result is the min of all operands results.
3180 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
3181 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
3182 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
3186 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
3187 // The result is the min of all operands results.
3188 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3189 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3190 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3194 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
3195 // The result is the min of all operands results.
3196 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
3197 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
3198 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
3202 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3203 // For a SCEVUnknown, ask ValueTracking.
3204 unsigned BitWidth = getTypeSizeInBits(U->getType());
3205 APInt Mask = APInt::getAllOnesValue(BitWidth);
3206 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3207 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
3208 return Zeros.countTrailingOnes();
3215 /// getUnsignedRange - Determine the unsigned range for a particular SCEV.
3218 ScalarEvolution::getUnsignedRange(const SCEV *S) {
3219 // See if we've computed this range already.
3220 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S);
3221 if (I != UnsignedRanges.end())
3224 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3225 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue()));
3227 unsigned BitWidth = getTypeSizeInBits(S->getType());
3228 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3230 // If the value has known zeros, the maximum unsigned value will have those
3231 // known zeros as well.
3232 uint32_t TZ = GetMinTrailingZeros(S);
3234 ConservativeResult =
3235 ConstantRange(APInt::getMinValue(BitWidth),
3236 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
3238 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3239 ConstantRange X = getUnsignedRange(Add->getOperand(0));
3240 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3241 X = X.add(getUnsignedRange(Add->getOperand(i)));
3242 return setUnsignedRange(Add, ConservativeResult.intersectWith(X));
3245 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3246 ConstantRange X = getUnsignedRange(Mul->getOperand(0));
3247 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3248 X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
3249 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X));
3252 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3253 ConstantRange X = getUnsignedRange(SMax->getOperand(0));
3254 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3255 X = X.smax(getUnsignedRange(SMax->getOperand(i)));
3256 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X));
3259 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3260 ConstantRange X = getUnsignedRange(UMax->getOperand(0));
3261 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3262 X = X.umax(getUnsignedRange(UMax->getOperand(i)));
3263 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X));
3266 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3267 ConstantRange X = getUnsignedRange(UDiv->getLHS());
3268 ConstantRange Y = getUnsignedRange(UDiv->getRHS());
3269 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3272 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3273 ConstantRange X = getUnsignedRange(ZExt->getOperand());
3274 return setUnsignedRange(ZExt,
3275 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3278 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3279 ConstantRange X = getUnsignedRange(SExt->getOperand());
3280 return setUnsignedRange(SExt,
3281 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3284 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3285 ConstantRange X = getUnsignedRange(Trunc->getOperand());
3286 return setUnsignedRange(Trunc,
3287 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3290 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3291 // If there's no unsigned wrap, the value will never be less than its
3293 if (AddRec->getNoWrapFlags(SCEV::FlagNUW))
3294 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
3295 if (!C->getValue()->isZero())
3296 ConservativeResult =
3297 ConservativeResult.intersectWith(
3298 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
3300 // TODO: non-affine addrec
3301 if (AddRec->isAffine()) {
3302 Type *Ty = AddRec->getType();
3303 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3304 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3305 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3306 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3308 const SCEV *Start = AddRec->getStart();
3309 const SCEV *Step = AddRec->getStepRecurrence(*this);
3311 ConstantRange StartRange = getUnsignedRange(Start);
3312 ConstantRange StepRange = getSignedRange(Step);
3313 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3314 ConstantRange EndRange =
3315 StartRange.add(MaxBECountRange.multiply(StepRange));
3317 // Check for overflow. This must be done with ConstantRange arithmetic
3318 // because we could be called from within the ScalarEvolution overflow
3320 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1);
3321 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3322 ConstantRange ExtMaxBECountRange =
3323 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3324 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1);
3325 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3327 return setUnsignedRange(AddRec, ConservativeResult);
3329 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
3330 EndRange.getUnsignedMin());
3331 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
3332 EndRange.getUnsignedMax());
3333 if (Min.isMinValue() && Max.isMaxValue())
3334 return setUnsignedRange(AddRec, ConservativeResult);
3335 return setUnsignedRange(AddRec,
3336 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3340 return setUnsignedRange(AddRec, ConservativeResult);
3343 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3344 // For a SCEVUnknown, ask ValueTracking.
3345 APInt Mask = APInt::getAllOnesValue(BitWidth);
3346 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
3347 ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
3348 if (Ones == ~Zeros + 1)
3349 return setUnsignedRange(U, ConservativeResult);
3350 return setUnsignedRange(U,
3351 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1)));
3354 return setUnsignedRange(S, ConservativeResult);
3357 /// getSignedRange - Determine the signed range for a particular SCEV.
3360 ScalarEvolution::getSignedRange(const SCEV *S) {
3361 // See if we've computed this range already.
3362 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S);
3363 if (I != SignedRanges.end())
3366 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
3367 return setSignedRange(C, ConstantRange(C->getValue()->getValue()));
3369 unsigned BitWidth = getTypeSizeInBits(S->getType());
3370 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
3372 // If the value has known zeros, the maximum signed value will have those
3373 // known zeros as well.
3374 uint32_t TZ = GetMinTrailingZeros(S);
3376 ConservativeResult =
3377 ConstantRange(APInt::getSignedMinValue(BitWidth),
3378 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
3380 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3381 ConstantRange X = getSignedRange(Add->getOperand(0));
3382 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
3383 X = X.add(getSignedRange(Add->getOperand(i)));
3384 return setSignedRange(Add, ConservativeResult.intersectWith(X));
3387 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3388 ConstantRange X = getSignedRange(Mul->getOperand(0));
3389 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
3390 X = X.multiply(getSignedRange(Mul->getOperand(i)));
3391 return setSignedRange(Mul, ConservativeResult.intersectWith(X));
3394 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
3395 ConstantRange X = getSignedRange(SMax->getOperand(0));
3396 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
3397 X = X.smax(getSignedRange(SMax->getOperand(i)));
3398 return setSignedRange(SMax, ConservativeResult.intersectWith(X));
3401 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
3402 ConstantRange X = getSignedRange(UMax->getOperand(0));
3403 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
3404 X = X.umax(getSignedRange(UMax->getOperand(i)));
3405 return setSignedRange(UMax, ConservativeResult.intersectWith(X));
3408 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
3409 ConstantRange X = getSignedRange(UDiv->getLHS());
3410 ConstantRange Y = getSignedRange(UDiv->getRHS());
3411 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y)));
3414 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
3415 ConstantRange X = getSignedRange(ZExt->getOperand());
3416 return setSignedRange(ZExt,
3417 ConservativeResult.intersectWith(X.zeroExtend(BitWidth)));
3420 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
3421 ConstantRange X = getSignedRange(SExt->getOperand());
3422 return setSignedRange(SExt,
3423 ConservativeResult.intersectWith(X.signExtend(BitWidth)));
3426 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
3427 ConstantRange X = getSignedRange(Trunc->getOperand());
3428 return setSignedRange(Trunc,
3429 ConservativeResult.intersectWith(X.truncate(BitWidth)));
3432 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
3433 // If there's no signed wrap, and all the operands have the same sign or
3434 // zero, the value won't ever change sign.
3435 if (AddRec->getNoWrapFlags(SCEV::FlagNSW)) {
3436 bool AllNonNeg = true;
3437 bool AllNonPos = true;
3438 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
3439 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
3440 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
3443 ConservativeResult = ConservativeResult.intersectWith(
3444 ConstantRange(APInt(BitWidth, 0),
3445 APInt::getSignedMinValue(BitWidth)));
3447 ConservativeResult = ConservativeResult.intersectWith(
3448 ConstantRange(APInt::getSignedMinValue(BitWidth),
3449 APInt(BitWidth, 1)));
3452 // TODO: non-affine addrec
3453 if (AddRec->isAffine()) {
3454 Type *Ty = AddRec->getType();
3455 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
3456 if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
3457 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
3458 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
3460 const SCEV *Start = AddRec->getStart();
3461 const SCEV *Step = AddRec->getStepRecurrence(*this);
3463 ConstantRange StartRange = getSignedRange(Start);
3464 ConstantRange StepRange = getSignedRange(Step);
3465 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount);
3466 ConstantRange EndRange =
3467 StartRange.add(MaxBECountRange.multiply(StepRange));
3469 // Check for overflow. This must be done with ConstantRange arithmetic
3470 // because we could be called from within the ScalarEvolution overflow
3472 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1);
3473 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1);
3474 ConstantRange ExtMaxBECountRange =
3475 MaxBECountRange.zextOrTrunc(BitWidth*2+1);
3476 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1);
3477 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) !=
3479 return setSignedRange(AddRec, ConservativeResult);
3481 APInt Min = APIntOps::smin(StartRange.getSignedMin(),
3482 EndRange.getSignedMin());
3483 APInt Max = APIntOps::smax(StartRange.getSignedMax(),
3484 EndRange.getSignedMax());
3485 if (Min.isMinSignedValue() && Max.isMaxSignedValue())
3486 return setSignedRange(AddRec, ConservativeResult);
3487 return setSignedRange(AddRec,
3488 ConservativeResult.intersectWith(ConstantRange(Min, Max+1)));
3492 return setSignedRange(AddRec, ConservativeResult);
3495 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3496 // For a SCEVUnknown, ask ValueTracking.
3497 if (!U->getValue()->getType()->isIntegerTy() && !TD)
3498 return setSignedRange(U, ConservativeResult);
3499 unsigned NS = ComputeNumSignBits(U->getValue(), TD);
3501 return setSignedRange(U, ConservativeResult);
3502 return setSignedRange(U, ConservativeResult.intersectWith(
3503 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
3504 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1)));
3507 return setSignedRange(S, ConservativeResult);
3510 /// createSCEV - We know that there is no SCEV for the specified value.
3511 /// Analyze the expression.
3513 const SCEV *ScalarEvolution::createSCEV(Value *V) {
3514 if (!isSCEVable(V->getType()))
3515 return getUnknown(V);
3517 unsigned Opcode = Instruction::UserOp1;
3518 if (Instruction *I = dyn_cast<Instruction>(V)) {
3519 Opcode = I->getOpcode();
3521 // Don't attempt to analyze instructions in blocks that aren't
3522 // reachable. Such instructions don't matter, and they aren't required
3523 // to obey basic rules for definitions dominating uses which this
3524 // analysis depends on.
3525 if (!DT->isReachableFromEntry(I->getParent()))
3526 return getUnknown(V);
3527 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
3528 Opcode = CE->getOpcode();
3529 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
3530 return getConstant(CI);
3531 else if (isa<ConstantPointerNull>(V))
3532 return getConstant(V->getType(), 0);
3533 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
3534 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
3536 return getUnknown(V);
3538 Operator *U = cast<Operator>(V);
3540 case Instruction::Add: {
3541 // The simple thing to do would be to just call getSCEV on both operands
3542 // and call getAddExpr with the result. However if we're looking at a
3543 // bunch of things all added together, this can be quite inefficient,
3544 // because it leads to N-1 getAddExpr calls for N ultimate operands.
3545 // Instead, gather up all the operands and make a single getAddExpr call.
3546 // LLVM IR canonical form means we need only traverse the left operands.
3547 SmallVector<const SCEV *, 4> AddOps;
3548 AddOps.push_back(getSCEV(U->getOperand(1)));
3549 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
3550 unsigned Opcode = Op->getValueID() - Value::InstructionVal;
3551 if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
3553 U = cast<Operator>(Op);
3554 const SCEV *Op1 = getSCEV(U->getOperand(1));
3555 if (Opcode == Instruction::Sub)
3556 AddOps.push_back(getNegativeSCEV(Op1));
3558 AddOps.push_back(Op1);
3560 AddOps.push_back(getSCEV(U->getOperand(0)));
3561 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap;
3562 OverflowingBinaryOperator *OBO = cast<OverflowingBinaryOperator>(V);
3563 if (OBO->hasNoSignedWrap())
3564 setFlags(Flags, SCEV::FlagNSW);
3565 if (OBO->hasNoUnsignedWrap())
3566 setFlags(Flags, SCEV::FlagNUW);
3567 return getAddExpr(AddOps, Flags);
3569 case Instruction::Mul: {
3570 // See the Add code above.
3571 SmallVector<const SCEV *, 4> MulOps;
3572 MulOps.push_back(getSCEV(U->getOperand(1)));
3573 for (Value *Op = U->getOperand(0);
3574 Op->getValueID() == Instruction::Mul + Value::InstructionVal;
3575 Op = U->getOperand(0)) {
3576 U = cast<Operator>(Op);
3577 MulOps.push_back(getSCEV(U->getOperand(1)));
3579 MulOps.push_back(getSCEV(U->getOperand(0)));
3580 return getMulExpr(MulOps);
3582 case Instruction::UDiv:
3583 return getUDivExpr(getSCEV(U->getOperand(0)),
3584 getSCEV(U->getOperand(1)));
3585 case Instruction::Sub:
3586 return getMinusSCEV(getSCEV(U->getOperand(0)),
3587 getSCEV(U->getOperand(1)));
3588 case Instruction::And:
3589 // For an expression like x&255 that merely masks off the high bits,
3590 // use zext(trunc(x)) as the SCEV expression.
3591 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3592 if (CI->isNullValue())
3593 return getSCEV(U->getOperand(1));
3594 if (CI->isAllOnesValue())
3595 return getSCEV(U->getOperand(0));
3596 const APInt &A = CI->getValue();
3598 // Instcombine's ShrinkDemandedConstant may strip bits out of
3599 // constants, obscuring what would otherwise be a low-bits mask.
3600 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
3601 // knew about to reconstruct a low-bits mask value.
3602 unsigned LZ = A.countLeadingZeros();
3603 unsigned BitWidth = A.getBitWidth();
3604 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
3605 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3606 ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
3608 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
3610 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
3612 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
3613 IntegerType::get(getContext(), BitWidth - LZ)),
3618 case Instruction::Or:
3619 // If the RHS of the Or is a constant, we may have something like:
3620 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
3621 // optimizations will transparently handle this case.
3623 // In order for this transformation to be safe, the LHS must be of the
3624 // form X*(2^n) and the Or constant must be less than 2^n.
3625 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3626 const SCEV *LHS = getSCEV(U->getOperand(0));
3627 const APInt &CIVal = CI->getValue();
3628 if (GetMinTrailingZeros(LHS) >=
3629 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
3630 // Build a plain add SCEV.
3631 const SCEV *S = getAddExpr(LHS, getSCEV(CI));
3632 // If the LHS of the add was an addrec and it has no-wrap flags,
3633 // transfer the no-wrap flags, since an or won't introduce a wrap.
3634 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
3635 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
3636 const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags(
3637 OldAR->getNoWrapFlags());
3643 case Instruction::Xor:
3644 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
3645 // If the RHS of the xor is a signbit, then this is just an add.
3646 // Instcombine turns add of signbit into xor as a strength reduction step.
3647 if (CI->getValue().isSignBit())
3648 return getAddExpr(getSCEV(U->getOperand(0)),
3649 getSCEV(U->getOperand(1)));
3651 // If the RHS of xor is -1, then this is a not operation.
3652 if (CI->isAllOnesValue())
3653 return getNotSCEV(getSCEV(U->getOperand(0)));
3655 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
3656 // This is a variant of the check for xor with -1, and it handles
3657 // the case where instcombine has trimmed non-demanded bits out
3658 // of an xor with -1.
3659 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
3660 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
3661 if (BO->getOpcode() == Instruction::And &&
3662 LCI->getValue() == CI->getValue())
3663 if (const SCEVZeroExtendExpr *Z =
3664 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
3665 Type *UTy = U->getType();
3666 const SCEV *Z0 = Z->getOperand();
3667 Type *Z0Ty = Z0->getType();
3668 unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
3670 // If C is a low-bits mask, the zero extend is serving to
3671 // mask off the high bits. Complement the operand and
3672 // re-apply the zext.
3673 if (APIntOps::isMask(Z0TySize, CI->getValue()))
3674 return getZeroExtendExpr(getNotSCEV(Z0), UTy);
3676 // If C is a single bit, it may be in the sign-bit position
3677 // before the zero-extend. In this case, represent the xor
3678 // using an add, which is equivalent, and re-apply the zext.
3679 APInt Trunc = CI->getValue().trunc(Z0TySize);
3680 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
3682 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
3688 case Instruction::Shl:
3689 // Turn shift left of a constant amount into a multiply.
3690 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3691 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3693 // If the shift count is not less than the bitwidth, the result of
3694 // the shift is undefined. Don't try to analyze it, because the
3695 // resolution chosen here may differ from the resolution chosen in
3696 // other parts of the compiler.
3697 if (SA->getValue().uge(BitWidth))
3700 Constant *X = ConstantInt::get(getContext(),
3701 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3702 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3706 case Instruction::LShr:
3707 // Turn logical shift right of a constant into a unsigned divide.
3708 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
3709 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
3711 // If the shift count is not less than the bitwidth, the result of
3712 // the shift is undefined. Don't try to analyze it, because the
3713 // resolution chosen here may differ from the resolution chosen in
3714 // other parts of the compiler.
3715 if (SA->getValue().uge(BitWidth))
3718 Constant *X = ConstantInt::get(getContext(),
3719 APInt(BitWidth, 1).shl(SA->getZExtValue()));
3720 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
3724 case Instruction::AShr:
3725 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
3726 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
3727 if (Operator *L = dyn_cast<Operator>(U->getOperand(0)))
3728 if (L->getOpcode() == Instruction::Shl &&
3729 L->getOperand(1) == U->getOperand(1)) {
3730 uint64_t BitWidth = getTypeSizeInBits(U->getType());
3732 // If the shift count is not less than the bitwidth, the result of
3733 // the shift is undefined. Don't try to analyze it, because the
3734 // resolution chosen here may differ from the resolution chosen in
3735 // other parts of the compiler.
3736 if (CI->getValue().uge(BitWidth))
3739 uint64_t Amt = BitWidth - CI->getZExtValue();
3740 if (Amt == BitWidth)
3741 return getSCEV(L->getOperand(0)); // shift by zero --> noop
3743 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
3744 IntegerType::get(getContext(),
3750 case Instruction::Trunc:
3751 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
3753 case Instruction::ZExt:
3754 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3756 case Instruction::SExt:
3757 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
3759 case Instruction::BitCast:
3760 // BitCasts are no-op casts so we just eliminate the cast.
3761 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
3762 return getSCEV(U->getOperand(0));
3765 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
3766 // lead to pointer expressions which cannot safely be expanded to GEPs,
3767 // because ScalarEvolution doesn't respect the GEP aliasing rules when
3768 // simplifying integer expressions.
3770 case Instruction::GetElementPtr:
3771 return createNodeForGEP(cast<GEPOperator>(U));
3773 case Instruction::PHI:
3774 return createNodeForPHI(cast<PHINode>(U));
3776 case Instruction::Select:
3777 // This could be a smax or umax that was lowered earlier.
3778 // Try to recover it.
3779 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
3780 Value *LHS = ICI->getOperand(0);
3781 Value *RHS = ICI->getOperand(1);
3782 switch (ICI->getPredicate()) {
3783 case ICmpInst::ICMP_SLT:
3784 case ICmpInst::ICMP_SLE:
3785 std::swap(LHS, RHS);
3787 case ICmpInst::ICMP_SGT:
3788 case ICmpInst::ICMP_SGE:
3789 // a >s b ? a+x : b+x -> smax(a, b)+x
3790 // a >s b ? b+x : a+x -> smin(a, b)+x
3791 if (LHS->getType() == U->getType()) {
3792 const SCEV *LS = getSCEV(LHS);
3793 const SCEV *RS = getSCEV(RHS);
3794 const SCEV *LA = getSCEV(U->getOperand(1));
3795 const SCEV *RA = getSCEV(U->getOperand(2));
3796 const SCEV *LDiff = getMinusSCEV(LA, LS);
3797 const SCEV *RDiff = getMinusSCEV(RA, RS);
3799 return getAddExpr(getSMaxExpr(LS, RS), LDiff);
3800 LDiff = getMinusSCEV(LA, RS);
3801 RDiff = getMinusSCEV(RA, LS);
3803 return getAddExpr(getSMinExpr(LS, RS), LDiff);
3806 case ICmpInst::ICMP_ULT:
3807 case ICmpInst::ICMP_ULE:
3808 std::swap(LHS, RHS);
3810 case ICmpInst::ICMP_UGT:
3811 case ICmpInst::ICMP_UGE:
3812 // a >u b ? a+x : b+x -> umax(a, b)+x
3813 // a >u b ? b+x : a+x -> umin(a, b)+x
3814 if (LHS->getType() == U->getType()) {
3815 const SCEV *LS = getSCEV(LHS);
3816 const SCEV *RS = getSCEV(RHS);
3817 const SCEV *LA = getSCEV(U->getOperand(1));
3818 const SCEV *RA = getSCEV(U->getOperand(2));
3819 const SCEV *LDiff = getMinusSCEV(LA, LS);
3820 const SCEV *RDiff = getMinusSCEV(RA, RS);
3822 return getAddExpr(getUMaxExpr(LS, RS), LDiff);
3823 LDiff = getMinusSCEV(LA, RS);
3824 RDiff = getMinusSCEV(RA, LS);
3826 return getAddExpr(getUMinExpr(LS, RS), LDiff);
3829 case ICmpInst::ICMP_NE:
3830 // n != 0 ? n+x : 1+x -> umax(n, 1)+x
3831 if (LHS->getType() == U->getType() &&
3832 isa<ConstantInt>(RHS) &&
3833 cast<ConstantInt>(RHS)->isZero()) {
3834 const SCEV *One = getConstant(LHS->getType(), 1);
3835 const SCEV *LS = getSCEV(LHS);
3836 const SCEV *LA = getSCEV(U->getOperand(1));
3837 const SCEV *RA = getSCEV(U->getOperand(2));
3838 const SCEV *LDiff = getMinusSCEV(LA, LS);
3839 const SCEV *RDiff = getMinusSCEV(RA, One);
3841 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3844 case ICmpInst::ICMP_EQ:
3845 // n == 0 ? 1+x : n+x -> umax(n, 1)+x
3846 if (LHS->getType() == U->getType() &&
3847 isa<ConstantInt>(RHS) &&
3848 cast<ConstantInt>(RHS)->isZero()) {
3849 const SCEV *One = getConstant(LHS->getType(), 1);
3850 const SCEV *LS = getSCEV(LHS);
3851 const SCEV *LA = getSCEV(U->getOperand(1));
3852 const SCEV *RA = getSCEV(U->getOperand(2));
3853 const SCEV *LDiff = getMinusSCEV(LA, One);
3854 const SCEV *RDiff = getMinusSCEV(RA, LS);
3856 return getAddExpr(getUMaxExpr(One, LS), LDiff);
3864 default: // We cannot analyze this expression.
3868 return getUnknown(V);
3873 //===----------------------------------------------------------------------===//
3874 // Iteration Count Computation Code
3877 /// getSmallConstantTripCount - Returns the maximum trip count of this loop as a
3878 /// normal unsigned value, if possible. Returns 0 if the trip count is unknown
3879 /// or not constant. Will also return 0 if the maximum trip count is very large
3881 unsigned ScalarEvolution::getSmallConstantTripCount(Loop *L,
3882 BasicBlock *ExitBlock) {
3883 const SCEVConstant *ExitCount =
3884 dyn_cast<SCEVConstant>(getExitCount(L, ExitBlock));
3888 ConstantInt *ExitConst = ExitCount->getValue();
3890 // Guard against huge trip counts.
3891 if (ExitConst->getValue().getActiveBits() > 32)
3894 // In case of integer overflow, this returns 0, which is correct.
3895 return ((unsigned)ExitConst->getZExtValue()) + 1;
3898 /// getSmallConstantTripMultiple - Returns the largest constant divisor of the
3899 /// trip count of this loop as a normal unsigned value, if possible. This
3900 /// means that the actual trip count is always a multiple of the returned
3901 /// value (don't forget the trip count could very well be zero as well!).
3903 /// Returns 1 if the trip count is unknown or not guaranteed to be the
3904 /// multiple of a constant (which is also the case if the trip count is simply
3905 /// constant, use getSmallConstantTripCount for that case), Will also return 1
3906 /// if the trip count is very large (>= 2^32).
3907 unsigned ScalarEvolution::getSmallConstantTripMultiple(Loop *L,
3908 BasicBlock *ExitBlock) {
3909 const SCEV *ExitCount = getExitCount(L, ExitBlock);
3910 if (ExitCount == getCouldNotCompute())
3913 // Get the trip count from the BE count by adding 1.
3914 const SCEV *TCMul = getAddExpr(ExitCount,
3915 getConstant(ExitCount->getType(), 1));
3916 // FIXME: SCEV distributes multiplication as V1*C1 + V2*C1. We could attempt
3917 // to factor simple cases.
3918 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(TCMul))
3919 TCMul = Mul->getOperand(0);
3921 const SCEVConstant *MulC = dyn_cast<SCEVConstant>(TCMul);
3925 ConstantInt *Result = MulC->getValue();
3927 // Guard against huge trip counts.
3928 if (!Result || Result->getValue().getActiveBits() > 32)
3931 return (unsigned)Result->getZExtValue();
3934 // getExitCount - Get the expression for the number of loop iterations for which
3935 // this loop is guaranteed not to exit via ExitintBlock. Otherwise return
3936 // SCEVCouldNotCompute.
3937 const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) {
3938 return getBackedgeTakenInfo(L).getExact(ExitingBlock, this);
3941 /// getBackedgeTakenCount - If the specified loop has a predictable
3942 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
3943 /// object. The backedge-taken count is the number of times the loop header
3944 /// will be branched to from within the loop. This is one less than the
3945 /// trip count of the loop, since it doesn't count the first iteration,
3946 /// when the header is branched to from outside the loop.
3948 /// Note that it is not valid to call this method on a loop without a
3949 /// loop-invariant backedge-taken count (see
3950 /// hasLoopInvariantBackedgeTakenCount).
3952 const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
3953 return getBackedgeTakenInfo(L).getExact(this);
3956 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
3957 /// return the least SCEV value that is known never to be less than the
3958 /// actual backedge taken count.
3959 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
3960 return getBackedgeTakenInfo(L).getMax(this);
3963 /// PushLoopPHIs - Push PHI nodes in the header of the given loop
3964 /// onto the given Worklist.
3966 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
3967 BasicBlock *Header = L->getHeader();
3969 // Push all Loop-header PHIs onto the Worklist stack.
3970 for (BasicBlock::iterator I = Header->begin();
3971 PHINode *PN = dyn_cast<PHINode>(I); ++I)
3972 Worklist.push_back(PN);
3975 const ScalarEvolution::BackedgeTakenInfo &
3976 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
3977 // Initially insert an invalid entry for this loop. If the insertion
3978 // succeeds, proceed to actually compute a backedge-taken count and
3979 // update the value. The temporary CouldNotCompute value tells SCEV
3980 // code elsewhere that it shouldn't attempt to request a new
3981 // backedge-taken count, which could result in infinite recursion.
3982 std::pair<DenseMap<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
3983 BackedgeTakenCounts.insert(std::make_pair(L, BackedgeTakenInfo()));
3985 return Pair.first->second;
3987 // ComputeBackedgeTakenCount may allocate memory for its result. Inserting it
3988 // into the BackedgeTakenCounts map transfers ownership. Otherwise, the result
3989 // must be cleared in this scope.
3990 BackedgeTakenInfo Result = ComputeBackedgeTakenCount(L);
3992 if (Result.getExact(this) != getCouldNotCompute()) {
3993 assert(isLoopInvariant(Result.getExact(this), L) &&
3994 isLoopInvariant(Result.getMax(this), L) &&
3995 "Computed backedge-taken count isn't loop invariant for loop!");
3996 ++NumTripCountsComputed;
3998 else if (Result.getMax(this) == getCouldNotCompute() &&
3999 isa<PHINode>(L->getHeader()->begin())) {
4000 // Only count loops that have phi nodes as not being computable.
4001 ++NumTripCountsNotComputed;
4004 // Now that we know more about the trip count for this loop, forget any
4005 // existing SCEV values for PHI nodes in this loop since they are only
4006 // conservative estimates made without the benefit of trip count
4007 // information. This is similar to the code in forgetLoop, except that
4008 // it handles SCEVUnknown PHI nodes specially.
4009 if (Result.hasAnyInfo()) {
4010 SmallVector<Instruction *, 16> Worklist;
4011 PushLoopPHIs(L, Worklist);
4013 SmallPtrSet<Instruction *, 8> Visited;
4014 while (!Worklist.empty()) {
4015 Instruction *I = Worklist.pop_back_val();
4016 if (!Visited.insert(I)) continue;
4018 ValueExprMapType::iterator It =
4019 ValueExprMap.find(static_cast<Value *>(I));
4020 if (It != ValueExprMap.end()) {
4021 const SCEV *Old = It->second;
4023 // SCEVUnknown for a PHI either means that it has an unrecognized
4024 // structure, or it's a PHI that's in the progress of being computed
4025 // by createNodeForPHI. In the former case, additional loop trip
4026 // count information isn't going to change anything. In the later
4027 // case, createNodeForPHI will perform the necessary updates on its
4028 // own when it gets to that point.
4029 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
4030 forgetMemoizedResults(Old);
4031 ValueExprMap.erase(It);
4033 if (PHINode *PN = dyn_cast<PHINode>(I))
4034 ConstantEvolutionLoopExitValue.erase(PN);
4037 PushDefUseChildren(I, Worklist);
4041 // Re-lookup the insert position, since the call to
4042 // ComputeBackedgeTakenCount above could result in a
4043 // recusive call to getBackedgeTakenInfo (on a different
4044 // loop), which would invalidate the iterator computed
4046 return BackedgeTakenCounts.find(L)->second = Result;
4049 /// forgetLoop - This method should be called by the client when it has
4050 /// changed a loop in a way that may effect ScalarEvolution's ability to
4051 /// compute a trip count, or if the loop is deleted.
4052 void ScalarEvolution::forgetLoop(const Loop *L) {
4053 // Drop any stored trip count value.
4054 DenseMap<const Loop*, BackedgeTakenInfo>::iterator BTCPos =
4055 BackedgeTakenCounts.find(L);
4056 if (BTCPos != BackedgeTakenCounts.end()) {
4057 BTCPos->second.clear();
4058 BackedgeTakenCounts.erase(BTCPos);
4061 // Drop information about expressions based on loop-header PHIs.
4062 SmallVector<Instruction *, 16> Worklist;
4063 PushLoopPHIs(L, Worklist);
4065 SmallPtrSet<Instruction *, 8> Visited;
4066 while (!Worklist.empty()) {
4067 Instruction *I = Worklist.pop_back_val();
4068 if (!Visited.insert(I)) continue;
4070 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
4071 if (It != ValueExprMap.end()) {
4072 forgetMemoizedResults(It->second);
4073 ValueExprMap.erase(It);
4074 if (PHINode *PN = dyn_cast<PHINode>(I))
4075 ConstantEvolutionLoopExitValue.erase(PN);
4078 PushDefUseChildren(I, Worklist);
4081 // Forget all contained loops too, to avoid dangling entries in the
4082 // ValuesAtScopes map.
4083 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
4087 /// forgetValue - This method should be called by the client when it has
4088 /// changed a value in a way that may effect its value, or which may
4089 /// disconnect it from a def-use chain linking it to a loop.
4090 void ScalarEvolution::forgetValue(Value *V) {
4091 Instruction *I = dyn_cast<Instruction>(V);
4094 // Drop information about expressions based on loop-header PHIs.
4095 SmallVector<Instruction *, 16> Worklist;
4096 Worklist.push_back(I);
4098 SmallPtrSet<Instruction *, 8> Visited;
4099 while (!Worklist.empty()) {
4100 I = Worklist.pop_back_val();
4101 if (!Visited.insert(I)) continue;
4103 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
4104 if (It != ValueExprMap.end()) {
4105 forgetMemoizedResults(It->second);
4106 ValueExprMap.erase(It);
4107 if (PHINode *PN = dyn_cast<PHINode>(I))
4108 ConstantEvolutionLoopExitValue.erase(PN);
4111 PushDefUseChildren(I, Worklist);
4115 /// getExact - Get the exact loop backedge taken count considering all loop
4116 /// exits. If all exits are computable, this is the minimum computed count.
4118 ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
4119 // If any exits were not computable, the loop is not computable.
4120 if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute();
4122 // We need at least one computable exit.
4123 if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
4124 assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
4126 const SCEV *BECount = 0;
4127 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4128 ENT != 0; ENT = ENT->getNextExit()) {
4130 assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV");
4133 BECount = ENT->ExactNotTaken;
4135 BECount = SE->getUMinFromMismatchedTypes(BECount, ENT->ExactNotTaken);
4137 assert(BECount && "Invalid not taken count for loop exit");
4141 /// getExact - Get the exact not taken count for this loop exit.
4143 ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock,
4144 ScalarEvolution *SE) const {
4145 for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
4146 ENT != 0; ENT = ENT->getNextExit()) {
4148 if (ENT->ExitingBlock == ExitingBlock)
4149 return ENT->ExactNotTaken;
4151 return SE->getCouldNotCompute();
4154 /// getMax - Get the max backedge taken count for the loop.
4156 ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const {
4157 return Max ? Max : SE->getCouldNotCompute();
4160 /// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each
4161 /// computable exit into a persistent ExitNotTakenInfo array.
4162 ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo(
4163 SmallVectorImpl< std::pair<BasicBlock *, const SCEV *> > &ExitCounts,
4164 bool Complete, const SCEV *MaxCount) : Max(MaxCount) {
4167 ExitNotTaken.setIncomplete();
4169 unsigned NumExits = ExitCounts.size();
4170 if (NumExits == 0) return;
4172 ExitNotTaken.ExitingBlock = ExitCounts[0].first;
4173 ExitNotTaken.ExactNotTaken = ExitCounts[0].second;
4174 if (NumExits == 1) return;
4176 // Handle the rare case of multiple computable exits.
4177 ExitNotTakenInfo *ENT = new ExitNotTakenInfo[NumExits-1];
4179 ExitNotTakenInfo *PrevENT = &ExitNotTaken;
4180 for (unsigned i = 1; i < NumExits; ++i, PrevENT = ENT, ++ENT) {
4181 PrevENT->setNextExit(ENT);
4182 ENT->ExitingBlock = ExitCounts[i].first;
4183 ENT->ExactNotTaken = ExitCounts[i].second;
4187 /// clear - Invalidate this result and free the ExitNotTakenInfo array.
4188 void ScalarEvolution::BackedgeTakenInfo::clear() {
4189 ExitNotTaken.ExitingBlock = 0;
4190 ExitNotTaken.ExactNotTaken = 0;
4191 delete[] ExitNotTaken.getNextExit();
4194 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
4195 /// of the specified loop will execute.
4196 ScalarEvolution::BackedgeTakenInfo
4197 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
4198 SmallVector<BasicBlock *, 8> ExitingBlocks;
4199 L->getExitingBlocks(ExitingBlocks);
4201 // Examine all exits and pick the most conservative values.
4202 const SCEV *MaxBECount = getCouldNotCompute();
4203 bool CouldComputeBECount = true;
4204 SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts;
4205 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
4206 ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]);
4207 if (EL.Exact == getCouldNotCompute())
4208 // We couldn't compute an exact value for this exit, so
4209 // we won't be able to compute an exact value for the loop.
4210 CouldComputeBECount = false;
4212 ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact));
4214 if (MaxBECount == getCouldNotCompute())
4215 MaxBECount = EL.Max;
4216 else if (EL.Max != getCouldNotCompute())
4217 MaxBECount = getUMinFromMismatchedTypes(MaxBECount, EL.Max);
4220 return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
4223 /// ComputeExitLimit - Compute the number of times the backedge of the specified
4224 /// loop will execute if it exits via the specified block.
4225 ScalarEvolution::ExitLimit
4226 ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) {
4228 // Okay, we've chosen an exiting block. See what condition causes us to
4229 // exit at this block.
4231 // FIXME: we should be able to handle switch instructions (with a single exit)
4232 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
4233 if (ExitBr == 0) return getCouldNotCompute();
4234 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
4236 // At this point, we know we have a conditional branch that determines whether
4237 // the loop is exited. However, we don't know if the branch is executed each
4238 // time through the loop. If not, then the execution count of the branch will
4239 // not be equal to the trip count of the loop.
4241 // Currently we check for this by checking to see if the Exit branch goes to
4242 // the loop header. If so, we know it will always execute the same number of
4243 // times as the loop. We also handle the case where the exit block *is* the
4244 // loop header. This is common for un-rotated loops.
4246 // If both of those tests fail, walk up the unique predecessor chain to the
4247 // header, stopping if there is an edge that doesn't exit the loop. If the
4248 // header is reached, the execution count of the branch will be equal to the
4249 // trip count of the loop.
4251 // More extensive analysis could be done to handle more cases here.
4253 if (ExitBr->getSuccessor(0) != L->getHeader() &&
4254 ExitBr->getSuccessor(1) != L->getHeader() &&
4255 ExitBr->getParent() != L->getHeader()) {
4256 // The simple checks failed, try climbing the unique predecessor chain
4257 // up to the header.
4259 for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
4260 BasicBlock *Pred = BB->getUniquePredecessor();
4262 return getCouldNotCompute();
4263 TerminatorInst *PredTerm = Pred->getTerminator();
4264 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
4265 BasicBlock *PredSucc = PredTerm->getSuccessor(i);
4268 // If the predecessor has a successor that isn't BB and isn't
4269 // outside the loop, assume the worst.
4270 if (L->contains(PredSucc))
4271 return getCouldNotCompute();
4273 if (Pred == L->getHeader()) {
4280 return getCouldNotCompute();
4283 // Proceed to the next level to examine the exit condition expression.
4284 return ComputeExitLimitFromCond(L, ExitBr->getCondition(),
4285 ExitBr->getSuccessor(0),
4286 ExitBr->getSuccessor(1));
4289 /// ComputeExitLimitFromCond - Compute the number of times the
4290 /// backedge of the specified loop will execute if its exit condition
4291 /// were a conditional branch of ExitCond, TBB, and FBB.
4292 ScalarEvolution::ExitLimit
4293 ScalarEvolution::ComputeExitLimitFromCond(const Loop *L,
4297 // Check if the controlling expression for this loop is an And or Or.
4298 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
4299 if (BO->getOpcode() == Instruction::And) {
4300 // Recurse on the operands of the and.
4301 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB);
4302 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB);
4303 const SCEV *BECount = getCouldNotCompute();
4304 const SCEV *MaxBECount = getCouldNotCompute();
4305 if (L->contains(TBB)) {
4306 // Both conditions must be true for the loop to continue executing.
4307 // Choose the less conservative count.
4308 if (EL0.Exact == getCouldNotCompute() ||
4309 EL1.Exact == getCouldNotCompute())
4310 BECount = getCouldNotCompute();
4312 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4313 if (EL0.Max == getCouldNotCompute())
4314 MaxBECount = EL1.Max;
4315 else if (EL1.Max == getCouldNotCompute())
4316 MaxBECount = EL0.Max;
4318 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4320 // Both conditions must be true at the same time for the loop to exit.
4321 // For now, be conservative.
4322 assert(L->contains(FBB) && "Loop block has no successor in loop!");
4323 if (EL0.Max == EL1.Max)
4324 MaxBECount = EL0.Max;
4325 if (EL0.Exact == EL1.Exact)
4326 BECount = EL0.Exact;
4329 return ExitLimit(BECount, MaxBECount);
4331 if (BO->getOpcode() == Instruction::Or) {
4332 // Recurse on the operands of the or.
4333 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB);
4334 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB);
4335 const SCEV *BECount = getCouldNotCompute();
4336 const SCEV *MaxBECount = getCouldNotCompute();
4337 if (L->contains(FBB)) {
4338 // Both conditions must be false for the loop to continue executing.
4339 // Choose the less conservative count.
4340 if (EL0.Exact == getCouldNotCompute() ||
4341 EL1.Exact == getCouldNotCompute())
4342 BECount = getCouldNotCompute();
4344 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
4345 if (EL0.Max == getCouldNotCompute())
4346 MaxBECount = EL1.Max;
4347 else if (EL1.Max == getCouldNotCompute())
4348 MaxBECount = EL0.Max;
4350 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
4352 // Both conditions must be false at the same time for the loop to exit.
4353 // For now, be conservative.
4354 assert(L->contains(TBB) && "Loop block has no successor in loop!");
4355 if (EL0.Max == EL1.Max)
4356 MaxBECount = EL0.Max;
4357 if (EL0.Exact == EL1.Exact)
4358 BECount = EL0.Exact;
4361 return ExitLimit(BECount, MaxBECount);
4365 // With an icmp, it may be feasible to compute an exact backedge-taken count.
4366 // Proceed to the next level to examine the icmp.
4367 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
4368 return ComputeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB);
4370 // Check for a constant condition. These are normally stripped out by
4371 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
4372 // preserve the CFG and is temporarily leaving constant conditions
4374 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) {
4375 if (L->contains(FBB) == !CI->getZExtValue())
4376 // The backedge is always taken.
4377 return getCouldNotCompute();
4379 // The backedge is never taken.
4380 return getConstant(CI->getType(), 0);
4383 // If it's not an integer or pointer comparison then compute it the hard way.
4384 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4387 /// ComputeExitLimitFromICmp - Compute the number of times the
4388 /// backedge of the specified loop will execute if its exit condition
4389 /// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
4390 ScalarEvolution::ExitLimit
4391 ScalarEvolution::ComputeExitLimitFromICmp(const Loop *L,
4396 // If the condition was exit on true, convert the condition to exit on false
4397 ICmpInst::Predicate Cond;
4398 if (!L->contains(FBB))
4399 Cond = ExitCond->getPredicate();
4401 Cond = ExitCond->getInversePredicate();
4403 // Handle common loops like: for (X = "string"; *X; ++X)
4404 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
4405 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
4407 ComputeLoadConstantCompareExitLimit(LI, RHS, L, Cond);
4408 if (ItCnt.hasAnyInfo())
4412 const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
4413 const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
4415 // Try to evaluate any dependencies out of the loop.
4416 LHS = getSCEVAtScope(LHS, L);
4417 RHS = getSCEVAtScope(RHS, L);
4419 // At this point, we would like to compute how many iterations of the
4420 // loop the predicate will return true for these inputs.
4421 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) {
4422 // If there is a loop-invariant, force it into the RHS.
4423 std::swap(LHS, RHS);
4424 Cond = ICmpInst::getSwappedPredicate(Cond);
4427 // Simplify the operands before analyzing them.
4428 (void)SimplifyICmpOperands(Cond, LHS, RHS);
4430 // If we have a comparison of a chrec against a constant, try to use value
4431 // ranges to answer this query.
4432 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
4433 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
4434 if (AddRec->getLoop() == L) {
4435 // Form the constant range.
4436 ConstantRange CompRange(
4437 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
4439 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
4440 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
4444 case ICmpInst::ICMP_NE: { // while (X != Y)
4445 // Convert to: while (X-Y != 0)
4446 ExitLimit EL = HowFarToZero(getMinusSCEV(LHS, RHS), L);
4447 if (EL.hasAnyInfo()) return EL;
4450 case ICmpInst::ICMP_EQ: { // while (X == Y)
4451 // Convert to: while (X-Y == 0)
4452 ExitLimit EL = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
4453 if (EL.hasAnyInfo()) return EL;
4456 case ICmpInst::ICMP_SLT: {
4457 ExitLimit EL = HowManyLessThans(LHS, RHS, L, true);
4458 if (EL.hasAnyInfo()) return EL;
4461 case ICmpInst::ICMP_SGT: {
4462 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4463 getNotSCEV(RHS), L, true);
4464 if (EL.hasAnyInfo()) return EL;
4467 case ICmpInst::ICMP_ULT: {
4468 ExitLimit EL = HowManyLessThans(LHS, RHS, L, false);
4469 if (EL.hasAnyInfo()) return EL;
4472 case ICmpInst::ICMP_UGT: {
4473 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
4474 getNotSCEV(RHS), L, false);
4475 if (EL.hasAnyInfo()) return EL;
4480 dbgs() << "ComputeBackedgeTakenCount ";
4481 if (ExitCond->getOperand(0)->getType()->isUnsigned())
4482 dbgs() << "[unsigned] ";
4483 dbgs() << *LHS << " "
4484 << Instruction::getOpcodeName(Instruction::ICmp)
4485 << " " << *RHS << "\n";
4489 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
4492 static ConstantInt *
4493 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
4494 ScalarEvolution &SE) {
4495 const SCEV *InVal = SE.getConstant(C);
4496 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
4497 assert(isa<SCEVConstant>(Val) &&
4498 "Evaluation of SCEV at constant didn't fold correctly?");
4499 return cast<SCEVConstant>(Val)->getValue();
4502 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
4503 /// and a GEP expression (missing the pointer index) indexing into it, return
4504 /// the addressed element of the initializer or null if the index expression is
4507 GetAddressedElementFromGlobal(GlobalVariable *GV,
4508 const std::vector<ConstantInt*> &Indices) {
4509 Constant *Init = GV->getInitializer();
4510 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
4511 uint64_t Idx = Indices[i]->getZExtValue();
4512 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
4513 assert(Idx < CS->getNumOperands() && "Bad struct index!");
4514 Init = cast<Constant>(CS->getOperand(Idx));
4515 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
4516 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
4517 Init = cast<Constant>(CA->getOperand(Idx));
4518 } else if (isa<ConstantAggregateZero>(Init)) {
4519 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
4520 assert(Idx < STy->getNumElements() && "Bad struct index!");
4521 Init = Constant::getNullValue(STy->getElementType(Idx));
4522 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
4523 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
4524 Init = Constant::getNullValue(ATy->getElementType());
4526 llvm_unreachable("Unknown constant aggregate type!");
4530 return 0; // Unknown initializer type
4536 /// ComputeLoadConstantCompareExitLimit - Given an exit condition of
4537 /// 'icmp op load X, cst', try to see if we can compute the backedge
4538 /// execution count.
4539 ScalarEvolution::ExitLimit
4540 ScalarEvolution::ComputeLoadConstantCompareExitLimit(
4544 ICmpInst::Predicate predicate) {
4546 if (LI->isVolatile()) return getCouldNotCompute();
4548 // Check to see if the loaded pointer is a getelementptr of a global.
4549 // TODO: Use SCEV instead of manually grubbing with GEPs.
4550 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
4551 if (!GEP) return getCouldNotCompute();
4553 // Make sure that it is really a constant global we are gepping, with an
4554 // initializer, and make sure the first IDX is really 0.
4555 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
4556 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
4557 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
4558 !cast<Constant>(GEP->getOperand(1))->isNullValue())
4559 return getCouldNotCompute();
4561 // Okay, we allow one non-constant index into the GEP instruction.
4563 std::vector<ConstantInt*> Indexes;
4564 unsigned VarIdxNum = 0;
4565 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
4566 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4567 Indexes.push_back(CI);
4568 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
4569 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
4570 VarIdx = GEP->getOperand(i);
4572 Indexes.push_back(0);
4575 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
4576 // Check to see if X is a loop variant variable value now.
4577 const SCEV *Idx = getSCEV(VarIdx);
4578 Idx = getSCEVAtScope(Idx, L);
4580 // We can only recognize very limited forms of loop index expressions, in
4581 // particular, only affine AddRec's like {C1,+,C2}.
4582 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
4583 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) ||
4584 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
4585 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
4586 return getCouldNotCompute();
4588 unsigned MaxSteps = MaxBruteForceIterations;
4589 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
4590 ConstantInt *ItCst = ConstantInt::get(
4591 cast<IntegerType>(IdxExpr->getType()), IterationNum);
4592 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
4594 // Form the GEP offset.
4595 Indexes[VarIdxNum] = Val;
4597 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
4598 if (Result == 0) break; // Cannot compute!
4600 // Evaluate the condition for this iteration.
4601 Result = ConstantExpr::getICmp(predicate, Result, RHS);
4602 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
4603 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
4605 dbgs() << "\n***\n*** Computed loop count " << *ItCst
4606 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
4609 ++NumArrayLenItCounts;
4610 return getConstant(ItCst); // Found terminating iteration!
4613 return getCouldNotCompute();
4617 /// CanConstantFold - Return true if we can constant fold an instruction of the
4618 /// specified type, assuming that all operands were constants.
4619 static bool CanConstantFold(const Instruction *I) {
4620 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
4621 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
4624 if (const CallInst *CI = dyn_cast<CallInst>(I))
4625 if (const Function *F = CI->getCalledFunction())
4626 return canConstantFoldCallTo(F);
4630 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
4631 /// in the loop that V is derived from. We allow arbitrary operations along the
4632 /// way, but the operands of an operation must either be constants or a value
4633 /// derived from a constant PHI. If this expression does not fit with these
4634 /// constraints, return null.
4635 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
4636 // If this is not an instruction, or if this is an instruction outside of the
4637 // loop, it can't be derived from a loop PHI.
4638 Instruction *I = dyn_cast<Instruction>(V);
4639 if (I == 0 || !L->contains(I)) return 0;
4641 if (PHINode *PN = dyn_cast<PHINode>(I)) {
4642 if (L->getHeader() == I->getParent())
4645 // We don't currently keep track of the control flow needed to evaluate
4646 // PHIs, so we cannot handle PHIs inside of loops.
4650 // If we won't be able to constant fold this expression even if the operands
4651 // are constants, return early.
4652 if (!CanConstantFold(I)) return 0;
4654 // Otherwise, we can evaluate this instruction if all of its operands are
4655 // constant or derived from a PHI node themselves.
4657 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
4658 if (!isa<Constant>(I->getOperand(Op))) {
4659 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
4660 if (P == 0) return 0; // Not evolving from PHI
4664 return 0; // Evolving from multiple different PHIs.
4667 // This is a expression evolving from a constant PHI!
4671 /// EvaluateExpression - Given an expression that passes the
4672 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
4673 /// in the loop has the value PHIVal. If we can't fold this expression for some
4674 /// reason, return null.
4675 static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
4676 const TargetData *TD) {
4677 if (isa<PHINode>(V)) return PHIVal;
4678 if (Constant *C = dyn_cast<Constant>(V)) return C;
4679 Instruction *I = cast<Instruction>(V);
4681 std::vector<Constant*> Operands(I->getNumOperands());
4683 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4684 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
4685 if (Operands[i] == 0) return 0;
4688 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4689 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
4691 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Operands, TD);
4694 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
4695 /// in the header of its containing loop, we know the loop executes a
4696 /// constant number of times, and the PHI node is just a recurrence
4697 /// involving constants, fold it.
4699 ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
4702 DenseMap<PHINode*, Constant*>::const_iterator I =
4703 ConstantEvolutionLoopExitValue.find(PN);
4704 if (I != ConstantEvolutionLoopExitValue.end())
4707 if (BEs.ugt(MaxBruteForceIterations))
4708 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
4710 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
4712 // Since the loop is canonicalized, the PHI node must have two entries. One
4713 // entry must be a constant (coming in from outside of the loop), and the
4714 // second must be derived from the same PHI.
4715 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4716 Constant *StartCST =
4717 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4719 return RetVal = 0; // Must be a constant.
4721 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4722 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4723 !isa<Constant>(BEValue))
4724 return RetVal = 0; // Not derived from same PHI.
4726 // Execute the loop symbolically to determine the exit value.
4727 if (BEs.getActiveBits() >= 32)
4728 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
4730 unsigned NumIterations = BEs.getZExtValue(); // must be in range
4731 unsigned IterationNum = 0;
4732 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
4733 if (IterationNum == NumIterations)
4734 return RetVal = PHIVal; // Got exit value!
4736 // Compute the value of the PHI node for the next iteration.
4737 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4738 if (NextPHI == PHIVal)
4739 return RetVal = NextPHI; // Stopped evolving!
4741 return 0; // Couldn't evaluate!
4746 /// ComputeExitCountExhaustively - If the loop is known to execute a
4747 /// constant number of times (the condition evolves only from constants),
4748 /// try to evaluate a few iterations of the loop until we get the exit
4749 /// condition gets a value of ExitWhen (true or false). If we cannot
4750 /// evaluate the trip count of the loop, return getCouldNotCompute().
4751 const SCEV * ScalarEvolution::ComputeExitCountExhaustively(const Loop *L,
4754 PHINode *PN = getConstantEvolvingPHI(Cond, L);
4755 if (PN == 0) return getCouldNotCompute();
4757 // If the loop is canonicalized, the PHI will have exactly two entries.
4758 // That's the only form we support here.
4759 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
4761 // One entry must be a constant (coming in from outside of the loop), and the
4762 // second must be derived from the same PHI.
4763 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
4764 Constant *StartCST =
4765 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
4766 if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
4768 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
4769 if (getConstantEvolvingPHI(BEValue, L) != PN &&
4770 !isa<Constant>(BEValue))
4771 return getCouldNotCompute(); // Not derived from same PHI.
4773 // Okay, we find a PHI node that defines the trip count of this loop. Execute
4774 // the loop symbolically to determine when the condition gets a value of
4776 unsigned IterationNum = 0;
4777 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
4778 for (Constant *PHIVal = StartCST;
4779 IterationNum != MaxIterations; ++IterationNum) {
4780 ConstantInt *CondVal =
4781 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
4783 // Couldn't symbolically evaluate.
4784 if (!CondVal) return getCouldNotCompute();
4786 if (CondVal->getValue() == uint64_t(ExitWhen)) {
4787 ++NumBruteForceTripCountsComputed;
4788 return getConstant(Type::getInt32Ty(getContext()), IterationNum);
4791 // Compute the value of the PHI node for the next iteration.
4792 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
4793 if (NextPHI == 0 || NextPHI == PHIVal)
4794 return getCouldNotCompute();// Couldn't evaluate or not making progress...
4798 // Too many iterations were needed to evaluate.
4799 return getCouldNotCompute();
4802 /// getSCEVAtScope - Return a SCEV expression for the specified value
4803 /// at the specified scope in the program. The L value specifies a loop
4804 /// nest to evaluate the expression at, where null is the top-level or a
4805 /// specified loop is immediately inside of the loop.
4807 /// This method can be used to compute the exit value for a variable defined
4808 /// in a loop by querying what the value will hold in the parent loop.
4810 /// In the case that a relevant loop exit value cannot be computed, the
4811 /// original value V is returned.
4812 const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
4813 // Check to see if we've folded this expression at this loop before.
4814 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
4815 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
4816 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
4818 return Pair.first->second ? Pair.first->second : V;
4820 // Otherwise compute it.
4821 const SCEV *C = computeSCEVAtScope(V, L);
4822 ValuesAtScopes[V][L] = C;
4826 const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
4827 if (isa<SCEVConstant>(V)) return V;
4829 // If this instruction is evolved from a constant-evolving PHI, compute the
4830 // exit value from the loop without using SCEVs.
4831 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
4832 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
4833 const Loop *LI = (*this->LI)[I->getParent()];
4834 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
4835 if (PHINode *PN = dyn_cast<PHINode>(I))
4836 if (PN->getParent() == LI->getHeader()) {
4837 // Okay, there is no closed form solution for the PHI node. Check
4838 // to see if the loop that contains it has a known backedge-taken
4839 // count. If so, we may be able to force computation of the exit
4841 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
4842 if (const SCEVConstant *BTCC =
4843 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
4844 // Okay, we know how many times the containing loop executes. If
4845 // this is a constant evolving PHI node, get the final value at
4846 // the specified iteration number.
4847 Constant *RV = getConstantEvolutionLoopExitValue(PN,
4848 BTCC->getValue()->getValue(),
4850 if (RV) return getSCEV(RV);
4854 // Okay, this is an expression that we cannot symbolically evaluate
4855 // into a SCEV. Check to see if it's possible to symbolically evaluate
4856 // the arguments into constants, and if so, try to constant propagate the
4857 // result. This is particularly useful for computing loop exit values.
4858 if (CanConstantFold(I)) {
4859 SmallVector<Constant *, 4> Operands;
4860 bool MadeImprovement = false;
4861 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
4862 Value *Op = I->getOperand(i);
4863 if (Constant *C = dyn_cast<Constant>(Op)) {
4864 Operands.push_back(C);
4868 // If any of the operands is non-constant and if they are
4869 // non-integer and non-pointer, don't even try to analyze them
4870 // with scev techniques.
4871 if (!isSCEVable(Op->getType()))
4874 const SCEV *OrigV = getSCEV(Op);
4875 const SCEV *OpV = getSCEVAtScope(OrigV, L);
4876 MadeImprovement |= OrigV != OpV;
4879 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
4881 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV))
4882 C = dyn_cast<Constant>(SU->getValue());
4884 if (C->getType() != Op->getType())
4885 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4889 Operands.push_back(C);
4892 // Check to see if getSCEVAtScope actually made an improvement.
4893 if (MadeImprovement) {
4895 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
4896 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
4897 Operands[0], Operands[1], TD);
4899 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
4907 // This is some other type of SCEVUnknown, just return it.
4911 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
4912 // Avoid performing the look-up in the common case where the specified
4913 // expression has no loop-variant portions.
4914 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
4915 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4916 if (OpAtScope != Comm->getOperand(i)) {
4917 // Okay, at least one of these operands is loop variant but might be
4918 // foldable. Build a new instance of the folded commutative expression.
4919 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
4920 Comm->op_begin()+i);
4921 NewOps.push_back(OpAtScope);
4923 for (++i; i != e; ++i) {
4924 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
4925 NewOps.push_back(OpAtScope);
4927 if (isa<SCEVAddExpr>(Comm))
4928 return getAddExpr(NewOps);
4929 if (isa<SCEVMulExpr>(Comm))
4930 return getMulExpr(NewOps);
4931 if (isa<SCEVSMaxExpr>(Comm))
4932 return getSMaxExpr(NewOps);
4933 if (isa<SCEVUMaxExpr>(Comm))
4934 return getUMaxExpr(NewOps);
4935 llvm_unreachable("Unknown commutative SCEV type!");
4938 // If we got here, all operands are loop invariant.
4942 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
4943 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
4944 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
4945 if (LHS == Div->getLHS() && RHS == Div->getRHS())
4946 return Div; // must be loop invariant
4947 return getUDivExpr(LHS, RHS);
4950 // If this is a loop recurrence for a loop that does not contain L, then we
4951 // are dealing with the final value computed by the loop.
4952 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
4953 // First, attempt to evaluate each operand.
4954 // Avoid performing the look-up in the common case where the specified
4955 // expression has no loop-variant portions.
4956 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
4957 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
4958 if (OpAtScope == AddRec->getOperand(i))
4961 // Okay, at least one of these operands is loop variant but might be
4962 // foldable. Build a new instance of the folded commutative expression.
4963 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
4964 AddRec->op_begin()+i);
4965 NewOps.push_back(OpAtScope);
4966 for (++i; i != e; ++i)
4967 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
4969 const SCEV *FoldedRec =
4970 getAddRecExpr(NewOps, AddRec->getLoop(),
4971 AddRec->getNoWrapFlags(SCEV::FlagNW));
4972 AddRec = dyn_cast<SCEVAddRecExpr>(FoldedRec);
4973 // The addrec may be folded to a nonrecurrence, for example, if the
4974 // induction variable is multiplied by zero after constant folding. Go
4975 // ahead and return the folded value.
4981 // If the scope is outside the addrec's loop, evaluate it by using the
4982 // loop exit value of the addrec.
4983 if (!AddRec->getLoop()->contains(L)) {
4984 // To evaluate this recurrence, we need to know how many times the AddRec
4985 // loop iterates. Compute this now.
4986 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
4987 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
4989 // Then, evaluate the AddRec.
4990 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
4996 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
4997 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
4998 if (Op == Cast->getOperand())
4999 return Cast; // must be loop invariant
5000 return getZeroExtendExpr(Op, Cast->getType());
5003 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
5004 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5005 if (Op == Cast->getOperand())
5006 return Cast; // must be loop invariant
5007 return getSignExtendExpr(Op, Cast->getType());
5010 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
5011 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
5012 if (Op == Cast->getOperand())
5013 return Cast; // must be loop invariant
5014 return getTruncateExpr(Op, Cast->getType());
5017 llvm_unreachable("Unknown SCEV type!");
5021 /// getSCEVAtScope - This is a convenience function which does
5022 /// getSCEVAtScope(getSCEV(V), L).
5023 const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
5024 return getSCEVAtScope(getSCEV(V), L);
5027 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
5028 /// following equation:
5030 /// A * X = B (mod N)
5032 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
5033 /// A and B isn't important.
5035 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
5036 static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
5037 ScalarEvolution &SE) {
5038 uint32_t BW = A.getBitWidth();
5039 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
5040 assert(A != 0 && "A must be non-zero.");
5044 // The gcd of A and N may have only one prime factor: 2. The number of
5045 // trailing zeros in A is its multiplicity
5046 uint32_t Mult2 = A.countTrailingZeros();
5049 // 2. Check if B is divisible by D.
5051 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
5052 // is not less than multiplicity of this prime factor for D.
5053 if (B.countTrailingZeros() < Mult2)
5054 return SE.getCouldNotCompute();
5056 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
5059 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
5060 // bit width during computations.
5061 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
5062 APInt Mod(BW + 1, 0);
5063 Mod.setBit(BW - Mult2); // Mod = N / D
5064 APInt I = AD.multiplicativeInverse(Mod);
5066 // 4. Compute the minimum unsigned root of the equation:
5067 // I * (B / D) mod (N / D)
5068 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
5070 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
5072 return SE.getConstant(Result.trunc(BW));
5075 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
5076 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
5077 /// might be the same) or two SCEVCouldNotCompute objects.
5079 static std::pair<const SCEV *,const SCEV *>
5080 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
5081 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
5082 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
5083 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
5084 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
5086 // We currently can only solve this if the coefficients are constants.
5087 if (!LC || !MC || !NC) {
5088 const SCEV *CNC = SE.getCouldNotCompute();
5089 return std::make_pair(CNC, CNC);
5092 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
5093 const APInt &L = LC->getValue()->getValue();
5094 const APInt &M = MC->getValue()->getValue();
5095 const APInt &N = NC->getValue()->getValue();
5096 APInt Two(BitWidth, 2);
5097 APInt Four(BitWidth, 4);
5100 using namespace APIntOps;
5102 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
5103 // The B coefficient is M-N/2
5107 // The A coefficient is N/2
5108 APInt A(N.sdiv(Two));
5110 // Compute the B^2-4ac term.
5113 SqrtTerm -= Four * (A * C);
5115 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
5116 // integer value or else APInt::sqrt() will assert.
5117 APInt SqrtVal(SqrtTerm.sqrt());
5119 // Compute the two solutions for the quadratic formula.
5120 // The divisions must be performed as signed divisions.
5123 if (TwoA.isMinValue()) {
5124 const SCEV *CNC = SE.getCouldNotCompute();
5125 return std::make_pair(CNC, CNC);
5128 LLVMContext &Context = SE.getContext();
5130 ConstantInt *Solution1 =
5131 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
5132 ConstantInt *Solution2 =
5133 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
5135 return std::make_pair(SE.getConstant(Solution1),
5136 SE.getConstant(Solution2));
5137 } // end APIntOps namespace
5140 /// HowFarToZero - Return the number of times a backedge comparing the specified
5141 /// value to zero will execute. If not computable, return CouldNotCompute.
5143 /// This is only used for loops with a "x != y" exit test. The exit condition is
5144 /// now expressed as a single expression, V = x-y. So the exit test is
5145 /// effectively V != 0. We know and take advantage of the fact that this
5146 /// expression only being used in a comparison by zero context.
5147 ScalarEvolution::ExitLimit
5148 ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
5149 // If the value is a constant
5150 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5151 // If the value is already zero, the branch will execute zero times.
5152 if (C->getValue()->isZero()) return C;
5153 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5156 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
5157 if (!AddRec || AddRec->getLoop() != L)
5158 return getCouldNotCompute();
5160 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
5161 // the quadratic equation to solve it.
5162 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) {
5163 std::pair<const SCEV *,const SCEV *> Roots =
5164 SolveQuadraticEquation(AddRec, *this);
5165 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
5166 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
5169 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
5170 << " sol#2: " << *R2 << "\n";
5172 // Pick the smallest positive root value.
5173 if (ConstantInt *CB =
5174 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT,
5177 if (CB->getZExtValue() == false)
5178 std::swap(R1, R2); // R1 is the minimum root now.
5180 // We can only use this value if the chrec ends up with an exact zero
5181 // value at this index. When solving for "X*X != 5", for example, we
5182 // should not accept a root of 2.
5183 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
5185 return R1; // We found a quadratic root!
5188 return getCouldNotCompute();
5191 // Otherwise we can only handle this if it is affine.
5192 if (!AddRec->isAffine())
5193 return getCouldNotCompute();
5195 // If this is an affine expression, the execution count of this branch is
5196 // the minimum unsigned root of the following equation:
5198 // Start + Step*N = 0 (mod 2^BW)
5202 // Step*N = -Start (mod 2^BW)
5204 // where BW is the common bit width of Start and Step.
5206 // Get the initial value for the loop.
5207 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
5208 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
5210 // For now we handle only constant steps.
5212 // TODO: Handle a nonconstant Step given AddRec<NUW>. If the
5213 // AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap
5214 // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step.
5215 // We have not yet seen any such cases.
5216 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
5218 return getCouldNotCompute();
5220 // For positive steps (counting up until unsigned overflow):
5221 // N = -Start/Step (as unsigned)
5222 // For negative steps (counting down to zero):
5224 // First compute the unsigned distance from zero in the direction of Step.
5225 bool CountDown = StepC->getValue()->getValue().isNegative();
5226 const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
5228 // Handle unitary steps, which cannot wraparound.
5229 // 1*N = -Start; -1*N = Start (mod 2^BW), so:
5230 // N = Distance (as unsigned)
5231 if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) {
5232 ConstantRange CR = getUnsignedRange(Start);
5233 const SCEV *MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
5234 : ~CR.getUnsignedMin());
5235 return ExitLimit(Distance, MaxBECount);
5238 // If the recurrence is known not to wraparound, unsigned divide computes the
5239 // back edge count. We know that the value will either become zero (and thus
5240 // the loop terminates), that the loop will terminate through some other exit
5241 // condition first, or that the loop has undefined behavior. This means
5242 // we can't "miss" the exit value, even with nonunit stride.
5244 // FIXME: Prove that loops always exhibits *acceptable* undefined
5245 // behavior. Loops must exhibit defined behavior until a wrapped value is
5246 // actually used. So the trip count computed by udiv could be smaller than the
5247 // number of well-defined iterations.
5248 if (AddRec->getNoWrapFlags(SCEV::FlagNW))
5249 // FIXME: We really want an "isexact" bit for udiv.
5250 return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
5252 // Then, try to solve the above equation provided that Start is constant.
5253 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
5254 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
5255 -StartC->getValue()->getValue(),
5257 return getCouldNotCompute();
5260 /// HowFarToNonZero - Return the number of times a backedge checking the
5261 /// specified value for nonzero will execute. If not computable, return
5263 ScalarEvolution::ExitLimit
5264 ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
5265 // Loops that look like: while (X == 0) are very strange indeed. We don't
5266 // handle them yet except for the trivial case. This could be expanded in the
5267 // future as needed.
5269 // If the value is a constant, check to see if it is known to be non-zero
5270 // already. If so, the backedge will execute zero times.
5271 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
5272 if (!C->getValue()->isNullValue())
5273 return getConstant(C->getType(), 0);
5274 return getCouldNotCompute(); // Otherwise it will loop infinitely.
5277 // We could implement others, but I really doubt anyone writes loops like
5278 // this, and if they did, they would already be constant folded.
5279 return getCouldNotCompute();
5282 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
5283 /// (which may not be an immediate predecessor) which has exactly one
5284 /// successor from which BB is reachable, or null if no such block is
5287 std::pair<BasicBlock *, BasicBlock *>
5288 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
5289 // If the block has a unique predecessor, then there is no path from the
5290 // predecessor to the block that does not go through the direct edge
5291 // from the predecessor to the block.
5292 if (BasicBlock *Pred = BB->getSinglePredecessor())
5293 return std::make_pair(Pred, BB);
5295 // A loop's header is defined to be a block that dominates the loop.
5296 // If the header has a unique predecessor outside the loop, it must be
5297 // a block that has exactly one successor that can reach the loop.
5298 if (Loop *L = LI->getLoopFor(BB))
5299 return std::make_pair(L->getLoopPredecessor(), L->getHeader());
5301 return std::pair<BasicBlock *, BasicBlock *>();
5304 /// HasSameValue - SCEV structural equivalence is usually sufficient for
5305 /// testing whether two expressions are equal, however for the purposes of
5306 /// looking for a condition guarding a loop, it can be useful to be a little
5307 /// more general, since a front-end may have replicated the controlling
5310 static bool HasSameValue(const SCEV *A, const SCEV *B) {
5311 // Quick check to see if they are the same SCEV.
5312 if (A == B) return true;
5314 // Otherwise, if they're both SCEVUnknown, it's possible that they hold
5315 // two different instructions with the same value. Check for this case.
5316 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
5317 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
5318 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
5319 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
5320 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
5323 // Otherwise assume they may have a different value.
5327 /// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with
5328 /// predicate Pred. Return true iff any changes were made.
5330 bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
5331 const SCEV *&LHS, const SCEV *&RHS) {
5332 bool Changed = false;
5334 // Canonicalize a constant to the right side.
5335 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
5336 // Check for both operands constant.
5337 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
5338 if (ConstantExpr::getICmp(Pred,
5340 RHSC->getValue())->isNullValue())
5341 goto trivially_false;
5343 goto trivially_true;
5345 // Otherwise swap the operands to put the constant on the right.
5346 std::swap(LHS, RHS);
5347 Pred = ICmpInst::getSwappedPredicate(Pred);
5351 // If we're comparing an addrec with a value which is loop-invariant in the
5352 // addrec's loop, put the addrec on the left. Also make a dominance check,
5353 // as both operands could be addrecs loop-invariant in each other's loop.
5354 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) {
5355 const Loop *L = AR->getLoop();
5356 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) {
5357 std::swap(LHS, RHS);
5358 Pred = ICmpInst::getSwappedPredicate(Pred);
5363 // If there's a constant operand, canonicalize comparisons with boundary
5364 // cases, and canonicalize *-or-equal comparisons to regular comparisons.
5365 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
5366 const APInt &RA = RC->getValue()->getValue();
5368 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5369 case ICmpInst::ICMP_EQ:
5370 case ICmpInst::ICMP_NE:
5372 case ICmpInst::ICMP_UGE:
5373 if ((RA - 1).isMinValue()) {
5374 Pred = ICmpInst::ICMP_NE;
5375 RHS = getConstant(RA - 1);
5379 if (RA.isMaxValue()) {
5380 Pred = ICmpInst::ICMP_EQ;
5384 if (RA.isMinValue()) goto trivially_true;
5386 Pred = ICmpInst::ICMP_UGT;
5387 RHS = getConstant(RA - 1);
5390 case ICmpInst::ICMP_ULE:
5391 if ((RA + 1).isMaxValue()) {
5392 Pred = ICmpInst::ICMP_NE;
5393 RHS = getConstant(RA + 1);
5397 if (RA.isMinValue()) {
5398 Pred = ICmpInst::ICMP_EQ;
5402 if (RA.isMaxValue()) goto trivially_true;
5404 Pred = ICmpInst::ICMP_ULT;
5405 RHS = getConstant(RA + 1);
5408 case ICmpInst::ICMP_SGE:
5409 if ((RA - 1).isMinSignedValue()) {
5410 Pred = ICmpInst::ICMP_NE;
5411 RHS = getConstant(RA - 1);
5415 if (RA.isMaxSignedValue()) {
5416 Pred = ICmpInst::ICMP_EQ;
5420 if (RA.isMinSignedValue()) goto trivially_true;
5422 Pred = ICmpInst::ICMP_SGT;
5423 RHS = getConstant(RA - 1);
5426 case ICmpInst::ICMP_SLE:
5427 if ((RA + 1).isMaxSignedValue()) {
5428 Pred = ICmpInst::ICMP_NE;
5429 RHS = getConstant(RA + 1);
5433 if (RA.isMinSignedValue()) {
5434 Pred = ICmpInst::ICMP_EQ;
5438 if (RA.isMaxSignedValue()) goto trivially_true;
5440 Pred = ICmpInst::ICMP_SLT;
5441 RHS = getConstant(RA + 1);
5444 case ICmpInst::ICMP_UGT:
5445 if (RA.isMinValue()) {
5446 Pred = ICmpInst::ICMP_NE;
5450 if ((RA + 1).isMaxValue()) {
5451 Pred = ICmpInst::ICMP_EQ;
5452 RHS = getConstant(RA + 1);
5456 if (RA.isMaxValue()) goto trivially_false;
5458 case ICmpInst::ICMP_ULT:
5459 if (RA.isMaxValue()) {
5460 Pred = ICmpInst::ICMP_NE;
5464 if ((RA - 1).isMinValue()) {
5465 Pred = ICmpInst::ICMP_EQ;
5466 RHS = getConstant(RA - 1);
5470 if (RA.isMinValue()) goto trivially_false;
5472 case ICmpInst::ICMP_SGT:
5473 if (RA.isMinSignedValue()) {
5474 Pred = ICmpInst::ICMP_NE;
5478 if ((RA + 1).isMaxSignedValue()) {
5479 Pred = ICmpInst::ICMP_EQ;
5480 RHS = getConstant(RA + 1);
5484 if (RA.isMaxSignedValue()) goto trivially_false;
5486 case ICmpInst::ICMP_SLT:
5487 if (RA.isMaxSignedValue()) {
5488 Pred = ICmpInst::ICMP_NE;
5492 if ((RA - 1).isMinSignedValue()) {
5493 Pred = ICmpInst::ICMP_EQ;
5494 RHS = getConstant(RA - 1);
5498 if (RA.isMinSignedValue()) goto trivially_false;
5503 // Check for obvious equality.
5504 if (HasSameValue(LHS, RHS)) {
5505 if (ICmpInst::isTrueWhenEqual(Pred))
5506 goto trivially_true;
5507 if (ICmpInst::isFalseWhenEqual(Pred))
5508 goto trivially_false;
5511 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by
5512 // adding or subtracting 1 from one of the operands.
5514 case ICmpInst::ICMP_SLE:
5515 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
5516 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5518 Pred = ICmpInst::ICMP_SLT;
5520 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
5521 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5523 Pred = ICmpInst::ICMP_SLT;
5527 case ICmpInst::ICMP_SGE:
5528 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
5529 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5531 Pred = ICmpInst::ICMP_SGT;
5533 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
5534 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5536 Pred = ICmpInst::ICMP_SGT;
5540 case ICmpInst::ICMP_ULE:
5541 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
5542 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
5544 Pred = ICmpInst::ICMP_ULT;
5546 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
5547 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
5549 Pred = ICmpInst::ICMP_ULT;
5553 case ICmpInst::ICMP_UGE:
5554 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
5555 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
5557 Pred = ICmpInst::ICMP_UGT;
5559 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
5560 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
5562 Pred = ICmpInst::ICMP_UGT;
5570 // TODO: More simplifications are possible here.
5576 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5577 Pred = ICmpInst::ICMP_EQ;
5582 LHS = RHS = getConstant(ConstantInt::getFalse(getContext()));
5583 Pred = ICmpInst::ICMP_NE;
5587 bool ScalarEvolution::isKnownNegative(const SCEV *S) {
5588 return getSignedRange(S).getSignedMax().isNegative();
5591 bool ScalarEvolution::isKnownPositive(const SCEV *S) {
5592 return getSignedRange(S).getSignedMin().isStrictlyPositive();
5595 bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
5596 return !getSignedRange(S).getSignedMin().isNegative();
5599 bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
5600 return !getSignedRange(S).getSignedMax().isStrictlyPositive();
5603 bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
5604 return isKnownNegative(S) || isKnownPositive(S);
5607 bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
5608 const SCEV *LHS, const SCEV *RHS) {
5609 // Canonicalize the inputs first.
5610 (void)SimplifyICmpOperands(Pred, LHS, RHS);
5612 // If LHS or RHS is an addrec, check to see if the condition is true in
5613 // every iteration of the loop.
5614 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
5615 if (isLoopEntryGuardedByCond(
5616 AR->getLoop(), Pred, AR->getStart(), RHS) &&
5617 isLoopBackedgeGuardedByCond(
5618 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
5620 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
5621 if (isLoopEntryGuardedByCond(
5622 AR->getLoop(), Pred, LHS, AR->getStart()) &&
5623 isLoopBackedgeGuardedByCond(
5624 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
5627 // Otherwise see what can be done with known constant ranges.
5628 return isKnownPredicateWithRanges(Pred, LHS, RHS);
5632 ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred,
5633 const SCEV *LHS, const SCEV *RHS) {
5634 if (HasSameValue(LHS, RHS))
5635 return ICmpInst::isTrueWhenEqual(Pred);
5637 // This code is split out from isKnownPredicate because it is called from
5638 // within isLoopEntryGuardedByCond.
5641 llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5643 case ICmpInst::ICMP_SGT:
5644 Pred = ICmpInst::ICMP_SLT;
5645 std::swap(LHS, RHS);
5646 case ICmpInst::ICMP_SLT: {
5647 ConstantRange LHSRange = getSignedRange(LHS);
5648 ConstantRange RHSRange = getSignedRange(RHS);
5649 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
5651 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
5655 case ICmpInst::ICMP_SGE:
5656 Pred = ICmpInst::ICMP_SLE;
5657 std::swap(LHS, RHS);
5658 case ICmpInst::ICMP_SLE: {
5659 ConstantRange LHSRange = getSignedRange(LHS);
5660 ConstantRange RHSRange = getSignedRange(RHS);
5661 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
5663 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
5667 case ICmpInst::ICMP_UGT:
5668 Pred = ICmpInst::ICMP_ULT;
5669 std::swap(LHS, RHS);
5670 case ICmpInst::ICMP_ULT: {
5671 ConstantRange LHSRange = getUnsignedRange(LHS);
5672 ConstantRange RHSRange = getUnsignedRange(RHS);
5673 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
5675 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
5679 case ICmpInst::ICMP_UGE:
5680 Pred = ICmpInst::ICMP_ULE;
5681 std::swap(LHS, RHS);
5682 case ICmpInst::ICMP_ULE: {
5683 ConstantRange LHSRange = getUnsignedRange(LHS);
5684 ConstantRange RHSRange = getUnsignedRange(RHS);
5685 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
5687 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
5691 case ICmpInst::ICMP_NE: {
5692 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
5694 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
5697 const SCEV *Diff = getMinusSCEV(LHS, RHS);
5698 if (isKnownNonZero(Diff))
5702 case ICmpInst::ICMP_EQ:
5703 // The check at the top of the function catches the case where
5704 // the values are known to be equal.
5710 /// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
5711 /// protected by a conditional between LHS and RHS. This is used to
5712 /// to eliminate casts.
5714 ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
5715 ICmpInst::Predicate Pred,
5716 const SCEV *LHS, const SCEV *RHS) {
5717 // Interpret a null as meaning no loop, where there is obviously no guard
5718 // (interprocedural conditions notwithstanding).
5719 if (!L) return true;
5721 BasicBlock *Latch = L->getLoopLatch();
5725 BranchInst *LoopContinuePredicate =
5726 dyn_cast<BranchInst>(Latch->getTerminator());
5727 if (!LoopContinuePredicate ||
5728 LoopContinuePredicate->isUnconditional())
5731 return isImpliedCond(Pred, LHS, RHS,
5732 LoopContinuePredicate->getCondition(),
5733 LoopContinuePredicate->getSuccessor(0) != L->getHeader());
5736 /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected
5737 /// by a conditional between LHS and RHS. This is used to help avoid max
5738 /// expressions in loop trip counts, and to eliminate casts.
5740 ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L,
5741 ICmpInst::Predicate Pred,
5742 const SCEV *LHS, const SCEV *RHS) {
5743 // Interpret a null as meaning no loop, where there is obviously no guard
5744 // (interprocedural conditions notwithstanding).
5745 if (!L) return false;
5747 // Starting at the loop predecessor, climb up the predecessor chain, as long
5748 // as there are predecessors that can be found that have unique successors
5749 // leading to the original header.
5750 for (std::pair<BasicBlock *, BasicBlock *>
5751 Pair(L->getLoopPredecessor(), L->getHeader());
5753 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
5755 BranchInst *LoopEntryPredicate =
5756 dyn_cast<BranchInst>(Pair.first->getTerminator());
5757 if (!LoopEntryPredicate ||
5758 LoopEntryPredicate->isUnconditional())
5761 if (isImpliedCond(Pred, LHS, RHS,
5762 LoopEntryPredicate->getCondition(),
5763 LoopEntryPredicate->getSuccessor(0) != Pair.second))
5770 /// isImpliedCond - Test whether the condition described by Pred, LHS,
5771 /// and RHS is true whenever the given Cond value evaluates to true.
5772 bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
5773 const SCEV *LHS, const SCEV *RHS,
5774 Value *FoundCondValue,
5776 // Recursively handle And and Or conditions.
5777 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
5778 if (BO->getOpcode() == Instruction::And) {
5780 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5781 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5782 } else if (BO->getOpcode() == Instruction::Or) {
5784 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) ||
5785 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse);
5789 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue);
5790 if (!ICI) return false;
5792 // Bail if the ICmp's operands' types are wider than the needed type
5793 // before attempting to call getSCEV on them. This avoids infinite
5794 // recursion, since the analysis of widening casts can require loop
5795 // exit condition information for overflow checking, which would
5797 if (getTypeSizeInBits(LHS->getType()) <
5798 getTypeSizeInBits(ICI->getOperand(0)->getType()))
5801 // Now that we found a conditional branch that dominates the loop, check to
5802 // see if it is the comparison we are looking for.
5803 ICmpInst::Predicate FoundPred;
5805 FoundPred = ICI->getInversePredicate();
5807 FoundPred = ICI->getPredicate();
5809 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
5810 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
5812 // Balance the types. The case where FoundLHS' type is wider than
5813 // LHS' type is checked for above.
5814 if (getTypeSizeInBits(LHS->getType()) >
5815 getTypeSizeInBits(FoundLHS->getType())) {
5816 if (CmpInst::isSigned(Pred)) {
5817 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
5818 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
5820 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
5821 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
5825 // Canonicalize the query to match the way instcombine will have
5826 // canonicalized the comparison.
5827 if (SimplifyICmpOperands(Pred, LHS, RHS))
5829 return CmpInst::isTrueWhenEqual(Pred);
5830 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS))
5831 if (FoundLHS == FoundRHS)
5832 return CmpInst::isFalseWhenEqual(Pred);
5834 // Check to see if we can make the LHS or RHS match.
5835 if (LHS == FoundRHS || RHS == FoundLHS) {
5836 if (isa<SCEVConstant>(RHS)) {
5837 std::swap(FoundLHS, FoundRHS);
5838 FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
5840 std::swap(LHS, RHS);
5841 Pred = ICmpInst::getSwappedPredicate(Pred);
5845 // Check whether the found predicate is the same as the desired predicate.
5846 if (FoundPred == Pred)
5847 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
5849 // Check whether swapping the found predicate makes it the same as the
5850 // desired predicate.
5851 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
5852 if (isa<SCEVConstant>(RHS))
5853 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
5855 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
5856 RHS, LHS, FoundLHS, FoundRHS);
5859 // Check whether the actual condition is beyond sufficient.
5860 if (FoundPred == ICmpInst::ICMP_EQ)
5861 if (ICmpInst::isTrueWhenEqual(Pred))
5862 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
5864 if (Pred == ICmpInst::ICMP_NE)
5865 if (!ICmpInst::isTrueWhenEqual(FoundPred))
5866 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
5869 // Otherwise assume the worst.
5873 /// isImpliedCondOperands - Test whether the condition described by Pred,
5874 /// LHS, and RHS is true whenever the condition described by Pred, FoundLHS,
5875 /// and FoundRHS is true.
5876 bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
5877 const SCEV *LHS, const SCEV *RHS,
5878 const SCEV *FoundLHS,
5879 const SCEV *FoundRHS) {
5880 return isImpliedCondOperandsHelper(Pred, LHS, RHS,
5881 FoundLHS, FoundRHS) ||
5882 // ~x < ~y --> x > y
5883 isImpliedCondOperandsHelper(Pred, LHS, RHS,
5884 getNotSCEV(FoundRHS),
5885 getNotSCEV(FoundLHS));
5888 /// isImpliedCondOperandsHelper - Test whether the condition described by
5889 /// Pred, LHS, and RHS is true whenever the condition described by Pred,
5890 /// FoundLHS, and FoundRHS is true.
5892 ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
5893 const SCEV *LHS, const SCEV *RHS,
5894 const SCEV *FoundLHS,
5895 const SCEV *FoundRHS) {
5897 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
5898 case ICmpInst::ICMP_EQ:
5899 case ICmpInst::ICMP_NE:
5900 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
5903 case ICmpInst::ICMP_SLT:
5904 case ICmpInst::ICMP_SLE:
5905 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
5906 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS))
5909 case ICmpInst::ICMP_SGT:
5910 case ICmpInst::ICMP_SGE:
5911 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
5912 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS))
5915 case ICmpInst::ICMP_ULT:
5916 case ICmpInst::ICMP_ULE:
5917 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
5918 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS))
5921 case ICmpInst::ICMP_UGT:
5922 case ICmpInst::ICMP_UGE:
5923 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
5924 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS))
5932 /// getBECount - Subtract the end and start values and divide by the step,
5933 /// rounding up, to get the number of times the backedge is executed. Return
5934 /// CouldNotCompute if an intermediate computation overflows.
5935 const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
5939 assert(!isKnownNegative(Step) &&
5940 "This code doesn't handle negative strides yet!");
5942 Type *Ty = Start->getType();
5944 // When Start == End, we have an exact BECount == 0. Short-circuit this case
5945 // here because SCEV may not be able to determine that the unsigned division
5946 // after rounding is zero.
5948 return getConstant(Ty, 0);
5950 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
5951 const SCEV *Diff = getMinusSCEV(End, Start);
5952 const SCEV *RoundUp = getAddExpr(Step, NegOne);
5954 // Add an adjustment to the difference between End and Start so that
5955 // the division will effectively round up.
5956 const SCEV *Add = getAddExpr(Diff, RoundUp);
5959 // Check Add for unsigned overflow.
5960 // TODO: More sophisticated things could be done here.
5961 Type *WideTy = IntegerType::get(getContext(),
5962 getTypeSizeInBits(Ty) + 1);
5963 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
5964 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
5965 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
5966 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
5967 return getCouldNotCompute();
5970 return getUDivExpr(Add, Step);
5973 /// HowManyLessThans - Return the number of times a backedge containing the
5974 /// specified less-than comparison will execute. If not computable, return
5975 /// CouldNotCompute.
5976 ScalarEvolution::ExitLimit
5977 ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
5978 const Loop *L, bool isSigned) {
5979 // Only handle: "ADDREC < LoopInvariant".
5980 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute();
5982 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
5983 if (!AddRec || AddRec->getLoop() != L)
5984 return getCouldNotCompute();
5986 // Check to see if we have a flag which makes analysis easy.
5987 bool NoWrap = isSigned ? AddRec->getNoWrapFlags(SCEV::FlagNSW) :
5988 AddRec->getNoWrapFlags(SCEV::FlagNUW);
5990 if (AddRec->isAffine()) {
5991 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
5992 const SCEV *Step = AddRec->getStepRecurrence(*this);
5995 return getCouldNotCompute();
5996 if (Step->isOne()) {
5997 // With unit stride, the iteration never steps past the limit value.
5998 } else if (isKnownPositive(Step)) {
5999 // Test whether a positive iteration can step past the limit
6000 // value and past the maximum value for its type in a single step.
6001 // Note that it's not sufficient to check NoWrap here, because even
6002 // though the value after a wrap is undefined, it's not undefined
6003 // behavior, so if wrap does occur, the loop could either terminate or
6004 // loop infinitely, but in either case, the loop is guaranteed to
6005 // iterate at least until the iteration where the wrapping occurs.
6006 const SCEV *One = getConstant(Step->getType(), 1);
6008 APInt Max = APInt::getSignedMaxValue(BitWidth);
6009 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
6010 .slt(getSignedRange(RHS).getSignedMax()))
6011 return getCouldNotCompute();
6013 APInt Max = APInt::getMaxValue(BitWidth);
6014 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
6015 .ult(getUnsignedRange(RHS).getUnsignedMax()))
6016 return getCouldNotCompute();
6019 // TODO: Handle negative strides here and below.
6020 return getCouldNotCompute();
6022 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
6023 // m. So, we count the number of iterations in which {n,+,s} < m is true.
6024 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
6025 // treat m-n as signed nor unsigned due to overflow possibility.
6027 // First, we get the value of the LHS in the first iteration: n
6028 const SCEV *Start = AddRec->getOperand(0);
6030 // Determine the minimum constant start value.
6031 const SCEV *MinStart = getConstant(isSigned ?
6032 getSignedRange(Start).getSignedMin() :
6033 getUnsignedRange(Start).getUnsignedMin());
6035 // If we know that the condition is true in order to enter the loop,
6036 // then we know that it will run exactly (m-n)/s times. Otherwise, we
6037 // only know that it will execute (max(m,n)-n)/s times. In both cases,
6038 // the division must round up.
6039 const SCEV *End = RHS;
6040 if (!isLoopEntryGuardedByCond(L,
6041 isSigned ? ICmpInst::ICMP_SLT :
6043 getMinusSCEV(Start, Step), RHS))
6044 End = isSigned ? getSMaxExpr(RHS, Start)
6045 : getUMaxExpr(RHS, Start);
6047 // Determine the maximum constant end value.
6048 const SCEV *MaxEnd = getConstant(isSigned ?
6049 getSignedRange(End).getSignedMax() :
6050 getUnsignedRange(End).getUnsignedMax());
6052 // If MaxEnd is within a step of the maximum integer value in its type,
6053 // adjust it down to the minimum value which would produce the same effect.
6054 // This allows the subsequent ceiling division of (N+(step-1))/step to
6055 // compute the correct value.
6056 const SCEV *StepMinusOne = getMinusSCEV(Step,
6057 getConstant(Step->getType(), 1));
6060 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
6063 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
6066 // Finally, we subtract these two values and divide, rounding up, to get
6067 // the number of times the backedge is executed.
6068 const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
6070 // The maximum backedge count is similar, except using the minimum start
6071 // value and the maximum end value.
6072 // If we already have an exact constant BECount, use it instead.
6073 const SCEV *MaxBECount = isa<SCEVConstant>(BECount) ? BECount
6074 : getBECount(MinStart, MaxEnd, Step, NoWrap);
6076 // If the stride is nonconstant, and NoWrap == true, then
6077 // getBECount(MinStart, MaxEnd) may not compute. This would result in an
6078 // exact BECount and invalid MaxBECount, which should be avoided to catch
6079 // more optimization opportunities.
6080 if (isa<SCEVCouldNotCompute>(MaxBECount))
6081 MaxBECount = BECount;
6083 return ExitLimit(BECount, MaxBECount);
6086 return getCouldNotCompute();
6089 /// getNumIterationsInRange - Return the number of iterations of this loop that
6090 /// produce values in the specified constant range. Another way of looking at
6091 /// this is that it returns the first iteration number where the value is not in
6092 /// the condition, thus computing the exit count. If the iteration count can't
6093 /// be computed, an instance of SCEVCouldNotCompute is returned.
6094 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
6095 ScalarEvolution &SE) const {
6096 if (Range.isFullSet()) // Infinite loop.
6097 return SE.getCouldNotCompute();
6099 // If the start is a non-zero constant, shift the range to simplify things.
6100 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
6101 if (!SC->getValue()->isZero()) {
6102 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
6103 Operands[0] = SE.getConstant(SC->getType(), 0);
6104 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop(),
6105 getNoWrapFlags(FlagNW));
6106 if (const SCEVAddRecExpr *ShiftedAddRec =
6107 dyn_cast<SCEVAddRecExpr>(Shifted))
6108 return ShiftedAddRec->getNumIterationsInRange(
6109 Range.subtract(SC->getValue()->getValue()), SE);
6110 // This is strange and shouldn't happen.
6111 return SE.getCouldNotCompute();
6114 // The only time we can solve this is when we have all constant indices.
6115 // Otherwise, we cannot determine the overflow conditions.
6116 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
6117 if (!isa<SCEVConstant>(getOperand(i)))
6118 return SE.getCouldNotCompute();
6121 // Okay at this point we know that all elements of the chrec are constants and
6122 // that the start element is zero.
6124 // First check to see if the range contains zero. If not, the first
6126 unsigned BitWidth = SE.getTypeSizeInBits(getType());
6127 if (!Range.contains(APInt(BitWidth, 0)))
6128 return SE.getConstant(getType(), 0);
6131 // If this is an affine expression then we have this situation:
6132 // Solve {0,+,A} in Range === Ax in Range
6134 // We know that zero is in the range. If A is positive then we know that
6135 // the upper value of the range must be the first possible exit value.
6136 // If A is negative then the lower of the range is the last possible loop
6137 // value. Also note that we already checked for a full range.
6138 APInt One(BitWidth,1);
6139 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
6140 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
6142 // The exit value should be (End+A)/A.
6143 APInt ExitVal = (End + A).udiv(A);
6144 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
6146 // Evaluate at the exit value. If we really did fall out of the valid
6147 // range, then we computed our trip count, otherwise wrap around or other
6148 // things must have happened.
6149 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
6150 if (Range.contains(Val->getValue()))
6151 return SE.getCouldNotCompute(); // Something strange happened
6153 // Ensure that the previous value is in the range. This is a sanity check.
6154 assert(Range.contains(
6155 EvaluateConstantChrecAtConstant(this,
6156 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
6157 "Linear scev computation is off in a bad way!");
6158 return SE.getConstant(ExitValue);
6159 } else if (isQuadratic()) {
6160 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
6161 // quadratic equation to solve it. To do this, we must frame our problem in
6162 // terms of figuring out when zero is crossed, instead of when
6163 // Range.getUpper() is crossed.
6164 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
6165 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
6166 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop(),
6167 // getNoWrapFlags(FlagNW)
6170 // Next, solve the constructed addrec
6171 std::pair<const SCEV *,const SCEV *> Roots =
6172 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
6173 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
6174 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
6176 // Pick the smallest positive root value.
6177 if (ConstantInt *CB =
6178 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
6179 R1->getValue(), R2->getValue()))) {
6180 if (CB->getZExtValue() == false)
6181 std::swap(R1, R2); // R1 is the minimum root now.
6183 // Make sure the root is not off by one. The returned iteration should
6184 // not be in the range, but the previous one should be. When solving
6185 // for "X*X < 5", for example, we should not return a root of 2.
6186 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
6189 if (Range.contains(R1Val->getValue())) {
6190 // The next iteration must be out of the range...
6191 ConstantInt *NextVal =
6192 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
6194 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6195 if (!Range.contains(R1Val->getValue()))
6196 return SE.getConstant(NextVal);
6197 return SE.getCouldNotCompute(); // Something strange happened
6200 // If R1 was not in the range, then it is a good return value. Make
6201 // sure that R1-1 WAS in the range though, just in case.
6202 ConstantInt *NextVal =
6203 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
6204 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
6205 if (Range.contains(R1Val->getValue()))
6207 return SE.getCouldNotCompute(); // Something strange happened
6212 return SE.getCouldNotCompute();
6217 //===----------------------------------------------------------------------===//
6218 // SCEVCallbackVH Class Implementation
6219 //===----------------------------------------------------------------------===//
6221 void ScalarEvolution::SCEVCallbackVH::deleted() {
6222 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6223 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
6224 SE->ConstantEvolutionLoopExitValue.erase(PN);
6225 SE->ValueExprMap.erase(getValPtr());
6226 // this now dangles!
6229 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
6230 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
6232 // Forget all the expressions associated with users of the old value,
6233 // so that future queries will recompute the expressions using the new
6235 Value *Old = getValPtr();
6236 SmallVector<User *, 16> Worklist;
6237 SmallPtrSet<User *, 8> Visited;
6238 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
6240 Worklist.push_back(*UI);
6241 while (!Worklist.empty()) {
6242 User *U = Worklist.pop_back_val();
6243 // Deleting the Old value will cause this to dangle. Postpone
6244 // that until everything else is done.
6247 if (!Visited.insert(U))
6249 if (PHINode *PN = dyn_cast<PHINode>(U))
6250 SE->ConstantEvolutionLoopExitValue.erase(PN);
6251 SE->ValueExprMap.erase(U);
6252 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
6254 Worklist.push_back(*UI);
6256 // Delete the Old value.
6257 if (PHINode *PN = dyn_cast<PHINode>(Old))
6258 SE->ConstantEvolutionLoopExitValue.erase(PN);
6259 SE->ValueExprMap.erase(Old);
6260 // this now dangles!
6263 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
6264 : CallbackVH(V), SE(se) {}
6266 //===----------------------------------------------------------------------===//
6267 // ScalarEvolution Class Implementation
6268 //===----------------------------------------------------------------------===//
6270 ScalarEvolution::ScalarEvolution()
6271 : FunctionPass(ID), FirstUnknown(0) {
6272 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry());
6275 bool ScalarEvolution::runOnFunction(Function &F) {
6277 LI = &getAnalysis<LoopInfo>();
6278 TD = getAnalysisIfAvailable<TargetData>();
6279 DT = &getAnalysis<DominatorTree>();
6283 void ScalarEvolution::releaseMemory() {
6284 // Iterate through all the SCEVUnknown instances and call their
6285 // destructors, so that they release their references to their values.
6286 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
6290 ValueExprMap.clear();
6292 // Free any extra memory created for ExitNotTakenInfo in the unlikely event
6293 // that a loop had multiple computable exits.
6294 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
6295 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end();
6300 BackedgeTakenCounts.clear();
6301 ConstantEvolutionLoopExitValue.clear();
6302 ValuesAtScopes.clear();
6303 LoopDispositions.clear();
6304 BlockDispositions.clear();
6305 UnsignedRanges.clear();
6306 SignedRanges.clear();
6307 UniqueSCEVs.clear();
6308 SCEVAllocator.Reset();
6311 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
6312 AU.setPreservesAll();
6313 AU.addRequiredTransitive<LoopInfo>();
6314 AU.addRequiredTransitive<DominatorTree>();
6317 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
6318 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
6321 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
6323 // Print all inner loops first
6324 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
6325 PrintLoopInfo(OS, SE, *I);
6328 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6331 SmallVector<BasicBlock *, 8> ExitBlocks;
6332 L->getExitBlocks(ExitBlocks);
6333 if (ExitBlocks.size() != 1)
6334 OS << "<multiple exits> ";
6336 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
6337 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
6339 OS << "Unpredictable backedge-taken count. ";
6344 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
6347 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
6348 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
6350 OS << "Unpredictable max backedge-taken count. ";
6356 void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
6357 // ScalarEvolution's implementation of the print method is to print
6358 // out SCEV values of all instructions that are interesting. Doing
6359 // this potentially causes it to create new SCEV objects though,
6360 // which technically conflicts with the const qualifier. This isn't
6361 // observable from outside the class though, so casting away the
6362 // const isn't dangerous.
6363 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
6365 OS << "Classifying expressions for: ";
6366 WriteAsOperand(OS, F, /*PrintType=*/false);
6368 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
6369 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
6372 const SCEV *SV = SE.getSCEV(&*I);
6375 const Loop *L = LI->getLoopFor((*I).getParent());
6377 const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
6384 OS << "\t\t" "Exits: ";
6385 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
6386 if (!SE.isLoopInvariant(ExitValue, L)) {
6387 OS << "<<Unknown>>";
6396 OS << "Determining loop execution counts for: ";
6397 WriteAsOperand(OS, F, /*PrintType=*/false);
6399 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
6400 PrintLoopInfo(OS, &SE, *I);
6403 ScalarEvolution::LoopDisposition
6404 ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) {
6405 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S];
6406 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair =
6407 Values.insert(std::make_pair(L, LoopVariant));
6409 return Pair.first->second;
6411 LoopDisposition D = computeLoopDisposition(S, L);
6412 return LoopDispositions[S][L] = D;
6415 ScalarEvolution::LoopDisposition
6416 ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) {
6417 switch (S->getSCEVType()) {
6419 return LoopInvariant;
6423 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L);
6424 case scAddRecExpr: {
6425 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6427 // If L is the addrec's loop, it's computable.
6428 if (AR->getLoop() == L)
6429 return LoopComputable;
6431 // Add recurrences are never invariant in the function-body (null loop).
6435 // This recurrence is variant w.r.t. L if L contains AR's loop.
6436 if (L->contains(AR->getLoop()))
6439 // This recurrence is invariant w.r.t. L if AR's loop contains L.
6440 if (AR->getLoop()->contains(L))
6441 return LoopInvariant;
6443 // This recurrence is variant w.r.t. L if any of its operands
6445 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
6447 if (!isLoopInvariant(*I, L))
6450 // Otherwise it's loop-invariant.
6451 return LoopInvariant;
6457 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6458 bool HasVarying = false;
6459 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6461 LoopDisposition D = getLoopDisposition(*I, L);
6462 if (D == LoopVariant)
6464 if (D == LoopComputable)
6467 return HasVarying ? LoopComputable : LoopInvariant;
6470 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6471 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L);
6472 if (LD == LoopVariant)
6474 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L);
6475 if (RD == LoopVariant)
6477 return (LD == LoopInvariant && RD == LoopInvariant) ?
6478 LoopInvariant : LoopComputable;
6481 // All non-instruction values are loop invariant. All instructions are loop
6482 // invariant if they are not contained in the specified loop.
6483 // Instructions are never considered invariant in the function body
6484 // (null loop) because they are defined within the "loop".
6485 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue()))
6486 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant;
6487 return LoopInvariant;
6488 case scCouldNotCompute:
6489 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6493 llvm_unreachable("Unknown SCEV kind!");
6497 bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
6498 return getLoopDisposition(S, L) == LoopInvariant;
6501 bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) {
6502 return getLoopDisposition(S, L) == LoopComputable;
6505 ScalarEvolution::BlockDisposition
6506 ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6507 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S];
6508 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool>
6509 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock));
6511 return Pair.first->second;
6513 BlockDisposition D = computeBlockDisposition(S, BB);
6514 return BlockDispositions[S][BB] = D;
6517 ScalarEvolution::BlockDisposition
6518 ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) {
6519 switch (S->getSCEVType()) {
6521 return ProperlyDominatesBlock;
6525 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB);
6526 case scAddRecExpr: {
6527 // This uses a "dominates" query instead of "properly dominates" query
6528 // to test for proper dominance too, because the instruction which
6529 // produces the addrec's value is a PHI, and a PHI effectively properly
6530 // dominates its entire containing block.
6531 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S);
6532 if (!DT->dominates(AR->getLoop()->getHeader(), BB))
6533 return DoesNotDominateBlock;
6535 // FALL THROUGH into SCEVNAryExpr handling.
6540 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6542 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6544 BlockDisposition D = getBlockDisposition(*I, BB);
6545 if (D == DoesNotDominateBlock)
6546 return DoesNotDominateBlock;
6547 if (D == DominatesBlock)
6550 return Proper ? ProperlyDominatesBlock : DominatesBlock;
6553 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6554 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6555 BlockDisposition LD = getBlockDisposition(LHS, BB);
6556 if (LD == DoesNotDominateBlock)
6557 return DoesNotDominateBlock;
6558 BlockDisposition RD = getBlockDisposition(RHS, BB);
6559 if (RD == DoesNotDominateBlock)
6560 return DoesNotDominateBlock;
6561 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ?
6562 ProperlyDominatesBlock : DominatesBlock;
6565 if (Instruction *I =
6566 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) {
6567 if (I->getParent() == BB)
6568 return DominatesBlock;
6569 if (DT->properlyDominates(I->getParent(), BB))
6570 return ProperlyDominatesBlock;
6571 return DoesNotDominateBlock;
6573 return ProperlyDominatesBlock;
6574 case scCouldNotCompute:
6575 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6576 return DoesNotDominateBlock;
6579 llvm_unreachable("Unknown SCEV kind!");
6580 return DoesNotDominateBlock;
6583 bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
6584 return getBlockDisposition(S, BB) >= DominatesBlock;
6587 bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) {
6588 return getBlockDisposition(S, BB) == ProperlyDominatesBlock;
6591 bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
6592 switch (S->getSCEVType()) {
6597 case scSignExtend: {
6598 const SCEVCastExpr *Cast = cast<SCEVCastExpr>(S);
6599 const SCEV *CastOp = Cast->getOperand();
6600 return Op == CastOp || hasOperand(CastOp, Op);
6607 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
6608 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
6610 const SCEV *NAryOp = *I;
6611 if (NAryOp == Op || hasOperand(NAryOp, Op))
6617 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
6618 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
6619 return LHS == Op || hasOperand(LHS, Op) ||
6620 RHS == Op || hasOperand(RHS, Op);
6624 case scCouldNotCompute:
6625 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
6629 llvm_unreachable("Unknown SCEV kind!");
6633 void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
6634 ValuesAtScopes.erase(S);
6635 LoopDispositions.erase(S);
6636 BlockDispositions.erase(S);
6637 UnsignedRanges.erase(S);
6638 SignedRanges.erase(S);