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. These classes are reference counted, managed by the SCEVHandle
18 // class. We only create one SCEV of a particular shape, so pointer-comparisons
19 // for equality are legal.
21 // One important aspect of the SCEV objects is that they are never cyclic, even
22 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If
23 // the PHI node is one of the idioms that we can represent (e.g., a polynomial
24 // recurrence) then we represent it directly as a recurrence node, otherwise we
25 // represent it as a SCEVUnknown node.
27 // In addition to being able to represent expressions of various types, we also
28 // have folders that are used to build the *canonical* representation for a
29 // particular expression. These folders are capable of using a variety of
30 // rewrite rules to simplify the expressions.
32 // Once the folders are defined, we can implement the more interesting
33 // higher-level code, such as the code that recognizes PHI nodes of various
34 // types, computes the execution count of a loop, etc.
36 // TODO: We should use these routines and value representations to implement
37 // dependence analysis!
39 //===----------------------------------------------------------------------===//
41 // There are several good references for the techniques used in this analysis.
43 // Chains of recurrences -- a method to expedite the evaluation
44 // of closed-form functions
45 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima
47 // On computational properties of chains of recurrences
50 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization
51 // Robert A. van Engelen
53 // Efficient Symbolic Analysis for Optimizing Compilers
54 // Robert A. van Engelen
56 // Using the chains of recurrences algebra for data dependence testing and
57 // induction variable substitution
58 // MS Thesis, Johnie Birch
60 //===----------------------------------------------------------------------===//
62 #define DEBUG_TYPE "scalar-evolution"
63 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
64 #include "llvm/Constants.h"
65 #include "llvm/DerivedTypes.h"
66 #include "llvm/GlobalVariable.h"
67 #include "llvm/Instructions.h"
68 #include "llvm/Analysis/ConstantFolding.h"
69 #include "llvm/Analysis/Dominators.h"
70 #include "llvm/Analysis/LoopInfo.h"
71 #include "llvm/Assembly/Writer.h"
72 #include "llvm/Target/TargetData.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/Compiler.h"
75 #include "llvm/Support/ConstantRange.h"
76 #include "llvm/Support/GetElementPtrTypeIterator.h"
77 #include "llvm/Support/InstIterator.h"
78 #include "llvm/Support/ManagedStatic.h"
79 #include "llvm/Support/MathExtras.h"
80 #include "llvm/Support/raw_ostream.h"
81 #include "llvm/ADT/Statistic.h"
82 #include "llvm/ADT/STLExtras.h"
86 STATISTIC(NumArrayLenItCounts,
87 "Number of trip counts computed with array length");
88 STATISTIC(NumTripCountsComputed,
89 "Number of loops with predictable loop counts");
90 STATISTIC(NumTripCountsNotComputed,
91 "Number of loops without predictable loop counts");
92 STATISTIC(NumBruteForceTripCountsComputed,
93 "Number of loops with trip counts computed by force");
95 static cl::opt<unsigned>
96 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
97 cl::desc("Maximum number of iterations SCEV will "
98 "symbolically execute a constant derived loop"),
101 static RegisterPass<ScalarEvolution>
102 R("scalar-evolution", "Scalar Evolution Analysis", false, true);
103 char ScalarEvolution::ID = 0;
105 //===----------------------------------------------------------------------===//
106 // SCEV class definitions
107 //===----------------------------------------------------------------------===//
109 //===----------------------------------------------------------------------===//
110 // Implementation of the SCEV class.
113 void SCEV::dump() const {
118 void SCEV::print(std::ostream &o) const {
119 raw_os_ostream OS(o);
123 bool SCEV::isZero() const {
124 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
125 return SC->getValue()->isZero();
129 bool SCEV::isOne() const {
130 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
131 return SC->getValue()->isOne();
135 SCEVCouldNotCompute::SCEVCouldNotCompute() : SCEV(scCouldNotCompute) {}
136 SCEVCouldNotCompute::~SCEVCouldNotCompute() {}
138 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
139 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
143 const Type *SCEVCouldNotCompute::getType() const {
144 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
148 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
149 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
153 SCEVHandle SCEVCouldNotCompute::
154 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
155 const SCEVHandle &Conc,
156 ScalarEvolution &SE) const {
160 void SCEVCouldNotCompute::print(raw_ostream &OS) const {
161 OS << "***COULDNOTCOMPUTE***";
164 bool SCEVCouldNotCompute::classof(const SCEV *S) {
165 return S->getSCEVType() == scCouldNotCompute;
169 // SCEVConstants - Only allow the creation of one SCEVConstant for any
170 // particular value. Don't use a SCEVHandle here, or else the object will
172 static ManagedStatic<std::map<ConstantInt*, SCEVConstant*> > SCEVConstants;
175 SCEVConstant::~SCEVConstant() {
176 SCEVConstants->erase(V);
179 SCEVHandle ScalarEvolution::getConstant(ConstantInt *V) {
180 SCEVConstant *&R = (*SCEVConstants)[V];
181 if (R == 0) R = new SCEVConstant(V);
185 SCEVHandle ScalarEvolution::getConstant(const APInt& Val) {
186 return getConstant(ConstantInt::get(Val));
189 const Type *SCEVConstant::getType() const { return V->getType(); }
191 void SCEVConstant::print(raw_ostream &OS) const {
192 WriteAsOperand(OS, V, false);
195 SCEVCastExpr::SCEVCastExpr(unsigned SCEVTy,
196 const SCEVHandle &op, const Type *ty)
197 : SCEV(SCEVTy), Op(op), Ty(ty) {}
199 SCEVCastExpr::~SCEVCastExpr() {}
201 bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
202 return Op->dominates(BB, DT);
205 // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any
206 // particular input. Don't use a SCEVHandle here, or else the object will
208 static ManagedStatic<std::map<std::pair<const SCEV*, const Type*>,
209 SCEVTruncateExpr*> > SCEVTruncates;
211 SCEVTruncateExpr::SCEVTruncateExpr(const SCEVHandle &op, const Type *ty)
212 : SCEVCastExpr(scTruncate, op, ty) {
213 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
214 (Ty->isInteger() || isa<PointerType>(Ty)) &&
215 "Cannot truncate non-integer value!");
218 SCEVTruncateExpr::~SCEVTruncateExpr() {
219 SCEVTruncates->erase(std::make_pair(Op, Ty));
222 void SCEVTruncateExpr::print(raw_ostream &OS) const {
223 OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
226 // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any
227 // particular input. Don't use a SCEVHandle here, or else the object will never
229 static ManagedStatic<std::map<std::pair<const SCEV*, const Type*>,
230 SCEVZeroExtendExpr*> > SCEVZeroExtends;
232 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty)
233 : SCEVCastExpr(scZeroExtend, op, ty) {
234 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
235 (Ty->isInteger() || isa<PointerType>(Ty)) &&
236 "Cannot zero extend non-integer value!");
239 SCEVZeroExtendExpr::~SCEVZeroExtendExpr() {
240 SCEVZeroExtends->erase(std::make_pair(Op, Ty));
243 void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
244 OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
247 // SCEVSignExtends - Only allow the creation of one SCEVSignExtendExpr for any
248 // particular input. Don't use a SCEVHandle here, or else the object will never
250 static ManagedStatic<std::map<std::pair<const SCEV*, const Type*>,
251 SCEVSignExtendExpr*> > SCEVSignExtends;
253 SCEVSignExtendExpr::SCEVSignExtendExpr(const SCEVHandle &op, const Type *ty)
254 : SCEVCastExpr(scSignExtend, op, ty) {
255 assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
256 (Ty->isInteger() || isa<PointerType>(Ty)) &&
257 "Cannot sign extend non-integer value!");
260 SCEVSignExtendExpr::~SCEVSignExtendExpr() {
261 SCEVSignExtends->erase(std::make_pair(Op, Ty));
264 void SCEVSignExtendExpr::print(raw_ostream &OS) const {
265 OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
268 // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any
269 // particular input. Don't use a SCEVHandle here, or else the object will never
271 static ManagedStatic<std::map<std::pair<unsigned, std::vector<const SCEV*> >,
272 SCEVCommutativeExpr*> > SCEVCommExprs;
274 SCEVCommutativeExpr::~SCEVCommutativeExpr() {
275 std::vector<const SCEV*> SCEVOps(Operands.begin(), Operands.end());
276 SCEVCommExprs->erase(std::make_pair(getSCEVType(), SCEVOps));
279 void SCEVCommutativeExpr::print(raw_ostream &OS) const {
280 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
281 const char *OpStr = getOperationStr();
282 OS << "(" << *Operands[0];
283 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
284 OS << OpStr << *Operands[i];
288 SCEVHandle SCEVCommutativeExpr::
289 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
290 const SCEVHandle &Conc,
291 ScalarEvolution &SE) const {
292 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
294 getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
295 if (H != getOperand(i)) {
296 std::vector<SCEVHandle> NewOps;
297 NewOps.reserve(getNumOperands());
298 for (unsigned j = 0; j != i; ++j)
299 NewOps.push_back(getOperand(j));
301 for (++i; i != e; ++i)
302 NewOps.push_back(getOperand(i)->
303 replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
305 if (isa<SCEVAddExpr>(this))
306 return SE.getAddExpr(NewOps);
307 else if (isa<SCEVMulExpr>(this))
308 return SE.getMulExpr(NewOps);
309 else if (isa<SCEVSMaxExpr>(this))
310 return SE.getSMaxExpr(NewOps);
311 else if (isa<SCEVUMaxExpr>(this))
312 return SE.getUMaxExpr(NewOps);
314 assert(0 && "Unknown commutative expr!");
320 bool SCEVNAryExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
321 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
322 if (!getOperand(i)->dominates(BB, DT))
329 // SCEVUDivs - Only allow the creation of one SCEVUDivExpr for any particular
330 // input. Don't use a SCEVHandle here, or else the object will never be
332 static ManagedStatic<std::map<std::pair<const SCEV*, const SCEV*>,
333 SCEVUDivExpr*> > SCEVUDivs;
335 SCEVUDivExpr::~SCEVUDivExpr() {
336 SCEVUDivs->erase(std::make_pair(LHS, RHS));
339 bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
340 return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
343 void SCEVUDivExpr::print(raw_ostream &OS) const {
344 OS << "(" << *LHS << " /u " << *RHS << ")";
347 const Type *SCEVUDivExpr::getType() const {
348 // In most cases the types of LHS and RHS will be the same, but in some
349 // crazy cases one or the other may be a pointer. ScalarEvolution doesn't
350 // depend on the type for correctness, but handling types carefully can
351 // avoid extra casts in the SCEVExpander. The LHS is more likely to be
352 // a pointer type than the RHS, so use the RHS' type here.
353 return RHS->getType();
356 // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
357 // particular input. Don't use a SCEVHandle here, or else the object will never
359 static ManagedStatic<std::map<std::pair<const Loop *,
360 std::vector<const SCEV*> >,
361 SCEVAddRecExpr*> > SCEVAddRecExprs;
363 SCEVAddRecExpr::~SCEVAddRecExpr() {
364 std::vector<const SCEV*> SCEVOps(Operands.begin(), Operands.end());
365 SCEVAddRecExprs->erase(std::make_pair(L, SCEVOps));
368 SCEVHandle SCEVAddRecExpr::
369 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
370 const SCEVHandle &Conc,
371 ScalarEvolution &SE) const {
372 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
374 getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
375 if (H != getOperand(i)) {
376 std::vector<SCEVHandle> NewOps;
377 NewOps.reserve(getNumOperands());
378 for (unsigned j = 0; j != i; ++j)
379 NewOps.push_back(getOperand(j));
381 for (++i; i != e; ++i)
382 NewOps.push_back(getOperand(i)->
383 replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
385 return SE.getAddRecExpr(NewOps, L);
392 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
393 // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't
394 // contain L and if the start is invariant.
395 // Add recurrences are never invariant in the function-body (null loop).
397 !QueryLoop->contains(L->getHeader()) &&
398 getOperand(0)->isLoopInvariant(QueryLoop);
402 void SCEVAddRecExpr::print(raw_ostream &OS) const {
403 OS << "{" << *Operands[0];
404 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
405 OS << ",+," << *Operands[i];
406 OS << "}<" << L->getHeader()->getName() + ">";
409 // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular
410 // value. Don't use a SCEVHandle here, or else the object will never be
412 static ManagedStatic<std::map<Value*, SCEVUnknown*> > SCEVUnknowns;
414 SCEVUnknown::~SCEVUnknown() { SCEVUnknowns->erase(V); }
416 bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
417 // All non-instruction values are loop invariant. All instructions are loop
418 // invariant if they are not contained in the specified loop.
419 // Instructions are never considered invariant in the function body
420 // (null loop) because they are defined within the "loop".
421 if (Instruction *I = dyn_cast<Instruction>(V))
422 return L && !L->contains(I->getParent());
426 bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
427 if (Instruction *I = dyn_cast<Instruction>(getValue()))
428 return DT->dominates(I->getParent(), BB);
432 const Type *SCEVUnknown::getType() const {
436 void SCEVUnknown::print(raw_ostream &OS) const {
437 WriteAsOperand(OS, V, false);
440 //===----------------------------------------------------------------------===//
442 //===----------------------------------------------------------------------===//
445 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
446 /// than the complexity of the RHS. This comparator is used to canonicalize
448 class VISIBILITY_HIDDEN SCEVComplexityCompare {
451 explicit SCEVComplexityCompare(LoopInfo *li) : LI(li) {}
453 bool operator()(const SCEV *LHS, const SCEV *RHS) const {
454 // Primarily, sort the SCEVs by their getSCEVType().
455 if (LHS->getSCEVType() != RHS->getSCEVType())
456 return LHS->getSCEVType() < RHS->getSCEVType();
458 // Aside from the getSCEVType() ordering, the particular ordering
459 // isn't very important except that it's beneficial to be consistent,
460 // so that (a + b) and (b + a) don't end up as different expressions.
462 // Sort SCEVUnknown values with some loose heuristics. TODO: This is
463 // not as complete as it could be.
464 if (const SCEVUnknown *LU = dyn_cast<SCEVUnknown>(LHS)) {
465 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
467 // Order pointer values after integer values. This helps SCEVExpander
469 if (isa<PointerType>(LU->getType()) && !isa<PointerType>(RU->getType()))
471 if (isa<PointerType>(RU->getType()) && !isa<PointerType>(LU->getType()))
474 // Compare getValueID values.
475 if (LU->getValue()->getValueID() != RU->getValue()->getValueID())
476 return LU->getValue()->getValueID() < RU->getValue()->getValueID();
478 // Sort arguments by their position.
479 if (const Argument *LA = dyn_cast<Argument>(LU->getValue())) {
480 const Argument *RA = cast<Argument>(RU->getValue());
481 return LA->getArgNo() < RA->getArgNo();
484 // For instructions, compare their loop depth, and their opcode.
485 // This is pretty loose.
486 if (Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
487 Instruction *RV = cast<Instruction>(RU->getValue());
489 // Compare loop depths.
490 if (LI->getLoopDepth(LV->getParent()) !=
491 LI->getLoopDepth(RV->getParent()))
492 return LI->getLoopDepth(LV->getParent()) <
493 LI->getLoopDepth(RV->getParent());
496 if (LV->getOpcode() != RV->getOpcode())
497 return LV->getOpcode() < RV->getOpcode();
499 // Compare the number of operands.
500 if (LV->getNumOperands() != RV->getNumOperands())
501 return LV->getNumOperands() < RV->getNumOperands();
507 // Constant sorting doesn't matter since they'll be folded.
508 if (isa<SCEVConstant>(LHS))
511 // Lexicographically compare n-ary expressions.
512 if (const SCEVNAryExpr *LC = dyn_cast<SCEVNAryExpr>(LHS)) {
513 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
514 for (unsigned i = 0, e = LC->getNumOperands(); i != e; ++i) {
515 if (i >= RC->getNumOperands())
517 if (operator()(LC->getOperand(i), RC->getOperand(i)))
519 if (operator()(RC->getOperand(i), LC->getOperand(i)))
522 return LC->getNumOperands() < RC->getNumOperands();
525 // Lexicographically compare udiv expressions.
526 if (const SCEVUDivExpr *LC = dyn_cast<SCEVUDivExpr>(LHS)) {
527 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
528 if (operator()(LC->getLHS(), RC->getLHS()))
530 if (operator()(RC->getLHS(), LC->getLHS()))
532 if (operator()(LC->getRHS(), RC->getRHS()))
534 if (operator()(RC->getRHS(), LC->getRHS()))
539 // Compare cast expressions by operand.
540 if (const SCEVCastExpr *LC = dyn_cast<SCEVCastExpr>(LHS)) {
541 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
542 return operator()(LC->getOperand(), RC->getOperand());
545 assert(0 && "Unknown SCEV kind!");
551 /// GroupByComplexity - Given a list of SCEV objects, order them by their
552 /// complexity, and group objects of the same complexity together by value.
553 /// When this routine is finished, we know that any duplicates in the vector are
554 /// consecutive and that complexity is monotonically increasing.
556 /// Note that we go take special precautions to ensure that we get determinstic
557 /// results from this routine. In other words, we don't want the results of
558 /// this to depend on where the addresses of various SCEV objects happened to
561 static void GroupByComplexity(std::vector<SCEVHandle> &Ops,
563 if (Ops.size() < 2) return; // Noop
564 if (Ops.size() == 2) {
565 // This is the common case, which also happens to be trivially simple.
567 if (SCEVComplexityCompare(LI)(Ops[1], Ops[0]))
568 std::swap(Ops[0], Ops[1]);
572 // Do the rough sort by complexity.
573 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
575 // Now that we are sorted by complexity, group elements of the same
576 // complexity. Note that this is, at worst, N^2, but the vector is likely to
577 // be extremely short in practice. Note that we take this approach because we
578 // do not want to depend on the addresses of the objects we are grouping.
579 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
580 const SCEV *S = Ops[i];
581 unsigned Complexity = S->getSCEVType();
583 // If there are any objects of the same complexity and same value as this
585 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
586 if (Ops[j] == S) { // Found a duplicate.
587 // Move it to immediately after i'th element.
588 std::swap(Ops[i+1], Ops[j]);
589 ++i; // no need to rescan it.
590 if (i == e-2) return; // Done!
598 //===----------------------------------------------------------------------===//
599 // Simple SCEV method implementations
600 //===----------------------------------------------------------------------===//
602 /// BinomialCoefficient - Compute BC(It, K). The result has width W.
604 static SCEVHandle BinomialCoefficient(SCEVHandle It, unsigned K,
606 const Type* ResultTy) {
607 // Handle the simplest case efficiently.
609 return SE.getTruncateOrZeroExtend(It, ResultTy);
611 // We are using the following formula for BC(It, K):
613 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
615 // Suppose, W is the bitwidth of the return value. We must be prepared for
616 // overflow. Hence, we must assure that the result of our computation is
617 // equal to the accurate one modulo 2^W. Unfortunately, division isn't
618 // safe in modular arithmetic.
620 // However, this code doesn't use exactly that formula; the formula it uses
621 // is something like the following, where T is the number of factors of 2 in
622 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
625 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
627 // This formula is trivially equivalent to the previous formula. However,
628 // this formula can be implemented much more efficiently. The trick is that
629 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
630 // arithmetic. To do exact division in modular arithmetic, all we have
631 // to do is multiply by the inverse. Therefore, this step can be done at
634 // The next issue is how to safely do the division by 2^T. The way this
635 // is done is by doing the multiplication step at a width of at least W + T
636 // bits. This way, the bottom W+T bits of the product are accurate. Then,
637 // when we perform the division by 2^T (which is equivalent to a right shift
638 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
639 // truncated out after the division by 2^T.
641 // In comparison to just directly using the first formula, this technique
642 // is much more efficient; using the first formula requires W * K bits,
643 // but this formula less than W + K bits. Also, the first formula requires
644 // a division step, whereas this formula only requires multiplies and shifts.
646 // It doesn't matter whether the subtraction step is done in the calculation
647 // width or the input iteration count's width; if the subtraction overflows,
648 // the result must be zero anyway. We prefer here to do it in the width of
649 // the induction variable because it helps a lot for certain cases; CodeGen
650 // isn't smart enough to ignore the overflow, which leads to much less
651 // efficient code if the width of the subtraction is wider than the native
654 // (It's possible to not widen at all by pulling out factors of 2 before
655 // the multiplication; for example, K=2 can be calculated as
656 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
657 // extra arithmetic, so it's not an obvious win, and it gets
658 // much more complicated for K > 3.)
660 // Protection from insane SCEVs; this bound is conservative,
661 // but it probably doesn't matter.
663 return SE.getCouldNotCompute();
665 unsigned W = SE.getTypeSizeInBits(ResultTy);
667 // Calculate K! / 2^T and T; we divide out the factors of two before
668 // multiplying for calculating K! / 2^T to avoid overflow.
669 // Other overflow doesn't matter because we only care about the bottom
670 // W bits of the result.
671 APInt OddFactorial(W, 1);
673 for (unsigned i = 3; i <= K; ++i) {
675 unsigned TwoFactors = Mult.countTrailingZeros();
677 Mult = Mult.lshr(TwoFactors);
678 OddFactorial *= Mult;
681 // We need at least W + T bits for the multiplication step
682 unsigned CalculationBits = W + T;
684 // Calcuate 2^T, at width T+W.
685 APInt DivFactor = APInt(CalculationBits, 1).shl(T);
687 // Calculate the multiplicative inverse of K! / 2^T;
688 // this multiplication factor will perform the exact division by
690 APInt Mod = APInt::getSignedMinValue(W+1);
691 APInt MultiplyFactor = OddFactorial.zext(W+1);
692 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
693 MultiplyFactor = MultiplyFactor.trunc(W);
695 // Calculate the product, at width T+W
696 const IntegerType *CalculationTy = IntegerType::get(CalculationBits);
697 SCEVHandle Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
698 for (unsigned i = 1; i != K; ++i) {
699 SCEVHandle S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
700 Dividend = SE.getMulExpr(Dividend,
701 SE.getTruncateOrZeroExtend(S, CalculationTy));
705 SCEVHandle DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
707 // Truncate the result, and divide by K! / 2^T.
709 return SE.getMulExpr(SE.getConstant(MultiplyFactor),
710 SE.getTruncateOrZeroExtend(DivResult, ResultTy));
713 /// evaluateAtIteration - Return the value of this chain of recurrences at
714 /// the specified iteration number. We can evaluate this recurrence by
715 /// multiplying each element in the chain by the binomial coefficient
716 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
718 /// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
720 /// where BC(It, k) stands for binomial coefficient.
722 SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It,
723 ScalarEvolution &SE) const {
724 SCEVHandle Result = getStart();
725 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
726 // The computation is correct in the face of overflow provided that the
727 // multiplication is performed _after_ the evaluation of the binomial
729 SCEVHandle Coeff = BinomialCoefficient(It, i, SE, getType());
730 if (isa<SCEVCouldNotCompute>(Coeff))
733 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
738 //===----------------------------------------------------------------------===//
739 // SCEV Expression folder implementations
740 //===----------------------------------------------------------------------===//
742 SCEVHandle ScalarEvolution::getTruncateExpr(const SCEVHandle &Op,
744 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
745 "This is not a truncating conversion!");
746 assert(isSCEVable(Ty) &&
747 "This is not a conversion to a SCEVable type!");
748 Ty = getEffectiveSCEVType(Ty);
750 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
752 ConstantExpr::getTrunc(SC->getValue(), Ty));
754 // trunc(trunc(x)) --> trunc(x)
755 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
756 return getTruncateExpr(ST->getOperand(), Ty);
758 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
759 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
760 return getTruncateOrSignExtend(SS->getOperand(), Ty);
762 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
763 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
764 return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
766 // If the input value is a chrec scev made out of constants, truncate
767 // all of the constants.
768 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
769 std::vector<SCEVHandle> Operands;
770 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
771 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
772 return getAddRecExpr(Operands, AddRec->getLoop());
775 SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)];
776 if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
780 SCEVHandle ScalarEvolution::getZeroExtendExpr(const SCEVHandle &Op,
782 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
783 "This is not an extending conversion!");
784 assert(isSCEVable(Ty) &&
785 "This is not a conversion to a SCEVable type!");
786 Ty = getEffectiveSCEVType(Ty);
788 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
789 const Type *IntTy = getEffectiveSCEVType(Ty);
790 Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
791 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
792 return getUnknown(C);
795 // zext(zext(x)) --> zext(x)
796 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
797 return getZeroExtendExpr(SZ->getOperand(), Ty);
799 // If the input value is a chrec scev, and we can prove that the value
800 // did not overflow the old, smaller, value, we can zero extend all of the
801 // operands (often constants). This allows analysis of something like
802 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
803 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
804 if (AR->isAffine()) {
805 // Check whether the backedge-taken count is SCEVCouldNotCompute.
806 // Note that this serves two purposes: It filters out loops that are
807 // simply not analyzable, and it covers the case where this code is
808 // being called from within backedge-taken count analysis, such that
809 // attempting to ask for the backedge-taken count would likely result
810 // in infinite recursion. In the later case, the analysis code will
811 // cope with a conservative value, and it will take care to purge
812 // that value once it has finished.
813 SCEVHandle MaxBECount = getMaxBackedgeTakenCount(AR->getLoop());
814 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
815 // Manually compute the final value for AR, checking for
817 SCEVHandle Start = AR->getStart();
818 SCEVHandle Step = AR->getStepRecurrence(*this);
820 // Check whether the backedge-taken count can be losslessly casted to
821 // the addrec's type. The count is always unsigned.
822 SCEVHandle CastedMaxBECount =
823 getTruncateOrZeroExtend(MaxBECount, Start->getType());
824 SCEVHandle RecastedMaxBECount =
825 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
826 if (MaxBECount == RecastedMaxBECount) {
828 IntegerType::get(getTypeSizeInBits(Start->getType()) * 2);
829 // Check whether Start+Step*MaxBECount has no unsigned overflow.
831 getMulExpr(CastedMaxBECount,
832 getTruncateOrZeroExtend(Step, Start->getType()));
833 SCEVHandle Add = getAddExpr(Start, ZMul);
834 SCEVHandle OperandExtendedAdd =
835 getAddExpr(getZeroExtendExpr(Start, WideTy),
836 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
837 getZeroExtendExpr(Step, WideTy)));
838 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
839 // Return the expression with the addrec on the outside.
840 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
841 getZeroExtendExpr(Step, Ty),
844 // Similar to above, only this time treat the step value as signed.
845 // This covers loops that count down.
847 getMulExpr(CastedMaxBECount,
848 getTruncateOrSignExtend(Step, Start->getType()));
849 Add = getAddExpr(Start, SMul);
851 getAddExpr(getZeroExtendExpr(Start, WideTy),
852 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
853 getSignExtendExpr(Step, WideTy)));
854 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
855 // Return the expression with the addrec on the outside.
856 return getAddRecExpr(getZeroExtendExpr(Start, Ty),
857 getSignExtendExpr(Step, Ty),
863 SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)];
864 if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
868 SCEVHandle ScalarEvolution::getSignExtendExpr(const SCEVHandle &Op,
870 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
871 "This is not an extending conversion!");
872 assert(isSCEVable(Ty) &&
873 "This is not a conversion to a SCEVable type!");
874 Ty = getEffectiveSCEVType(Ty);
876 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
877 const Type *IntTy = getEffectiveSCEVType(Ty);
878 Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
879 if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
880 return getUnknown(C);
883 // sext(sext(x)) --> sext(x)
884 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
885 return getSignExtendExpr(SS->getOperand(), Ty);
887 // If the input value is a chrec scev, and we can prove that the value
888 // did not overflow the old, smaller, value, we can sign extend all of the
889 // operands (often constants). This allows analysis of something like
890 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
891 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
892 if (AR->isAffine()) {
893 // Check whether the backedge-taken count is SCEVCouldNotCompute.
894 // Note that this serves two purposes: It filters out loops that are
895 // simply not analyzable, and it covers the case where this code is
896 // being called from within backedge-taken count analysis, such that
897 // attempting to ask for the backedge-taken count would likely result
898 // in infinite recursion. In the later case, the analysis code will
899 // cope with a conservative value, and it will take care to purge
900 // that value once it has finished.
901 SCEVHandle MaxBECount = getMaxBackedgeTakenCount(AR->getLoop());
902 if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
903 // Manually compute the final value for AR, checking for
905 SCEVHandle Start = AR->getStart();
906 SCEVHandle Step = AR->getStepRecurrence(*this);
908 // Check whether the backedge-taken count can be losslessly casted to
909 // the addrec's type. The count is always unsigned.
910 SCEVHandle CastedMaxBECount =
911 getTruncateOrZeroExtend(MaxBECount, Start->getType());
912 SCEVHandle RecastedMaxBECount =
913 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
914 if (MaxBECount == RecastedMaxBECount) {
916 IntegerType::get(getTypeSizeInBits(Start->getType()) * 2);
917 // Check whether Start+Step*MaxBECount has no signed overflow.
919 getMulExpr(CastedMaxBECount,
920 getTruncateOrSignExtend(Step, Start->getType()));
921 SCEVHandle Add = getAddExpr(Start, SMul);
922 SCEVHandle OperandExtendedAdd =
923 getAddExpr(getSignExtendExpr(Start, WideTy),
924 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
925 getSignExtendExpr(Step, WideTy)));
926 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
927 // Return the expression with the addrec on the outside.
928 return getAddRecExpr(getSignExtendExpr(Start, Ty),
929 getSignExtendExpr(Step, Ty),
935 SCEVSignExtendExpr *&Result = (*SCEVSignExtends)[std::make_pair(Op, Ty)];
936 if (Result == 0) Result = new SCEVSignExtendExpr(Op, Ty);
940 /// getAddExpr - Get a canonical add expression, or something simpler if
942 SCEVHandle ScalarEvolution::getAddExpr(std::vector<SCEVHandle> &Ops) {
943 assert(!Ops.empty() && "Cannot get empty add!");
944 if (Ops.size() == 1) return Ops[0];
946 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
947 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
948 getEffectiveSCEVType(Ops[0]->getType()) &&
949 "SCEVAddExpr operand types don't match!");
952 // Sort by complexity, this groups all similar expression types together.
953 GroupByComplexity(Ops, LI);
955 // If there are any constants, fold them together.
957 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
959 assert(Idx < Ops.size());
960 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
961 // We found two constants, fold them together!
962 ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() +
963 RHSC->getValue()->getValue());
964 Ops[0] = getConstant(Fold);
965 Ops.erase(Ops.begin()+1); // Erase the folded element
966 if (Ops.size() == 1) return Ops[0];
967 LHSC = cast<SCEVConstant>(Ops[0]);
970 // If we are left with a constant zero being added, strip it off.
971 if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
972 Ops.erase(Ops.begin());
977 if (Ops.size() == 1) return Ops[0];
979 // Okay, check to see if the same value occurs in the operand list twice. If
980 // so, merge them together into an multiply expression. Since we sorted the
981 // list, these values are required to be adjacent.
982 const Type *Ty = Ops[0]->getType();
983 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
984 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
985 // Found a match, merge the two values into a multiply, and add any
986 // remaining values to the result.
987 SCEVHandle Two = getIntegerSCEV(2, Ty);
988 SCEVHandle Mul = getMulExpr(Ops[i], Two);
991 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
993 return getAddExpr(Ops);
996 // Check for truncates. If all the operands are truncated from the same
997 // type, see if factoring out the truncate would permit the result to be
998 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
999 // if the contents of the resulting outer trunc fold to something simple.
1000 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
1001 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
1002 const Type *DstType = Trunc->getType();
1003 const Type *SrcType = Trunc->getOperand()->getType();
1004 std::vector<SCEVHandle> LargeOps;
1006 // Check all the operands to see if they can be represented in the
1007 // source type of the truncate.
1008 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1009 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
1010 if (T->getOperand()->getType() != SrcType) {
1014 LargeOps.push_back(T->getOperand());
1015 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
1016 // This could be either sign or zero extension, but sign extension
1017 // is much more likely to be foldable here.
1018 LargeOps.push_back(getSignExtendExpr(C, SrcType));
1019 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
1020 std::vector<SCEVHandle> LargeMulOps;
1021 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
1022 if (const SCEVTruncateExpr *T =
1023 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
1024 if (T->getOperand()->getType() != SrcType) {
1028 LargeMulOps.push_back(T->getOperand());
1029 } else if (const SCEVConstant *C =
1030 dyn_cast<SCEVConstant>(M->getOperand(j))) {
1031 // This could be either sign or zero extension, but sign extension
1032 // is much more likely to be foldable here.
1033 LargeMulOps.push_back(getSignExtendExpr(C, SrcType));
1040 LargeOps.push_back(getMulExpr(LargeMulOps));
1047 // Evaluate the expression in the larger type.
1048 SCEVHandle Fold = getAddExpr(LargeOps);
1049 // If it folds to something simple, use it. Otherwise, don't.
1050 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
1051 return getTruncateExpr(Fold, DstType);
1055 // Skip past any other cast SCEVs.
1056 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
1059 // If there are add operands they would be next.
1060 if (Idx < Ops.size()) {
1061 bool DeletedAdd = false;
1062 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
1063 // If we have an add, expand the add operands onto the end of the operands
1065 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
1066 Ops.erase(Ops.begin()+Idx);
1070 // If we deleted at least one add, we added operands to the end of the list,
1071 // and they are not necessarily sorted. Recurse to resort and resimplify
1072 // any operands we just aquired.
1074 return getAddExpr(Ops);
1077 // Skip over the add expression until we get to a multiply.
1078 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1081 // If we are adding something to a multiply expression, make sure the
1082 // something is not already an operand of the multiply. If so, merge it into
1084 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1085 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1086 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1087 const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1088 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1089 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(MulOpSCEV)) {
1090 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1091 SCEVHandle InnerMul = Mul->getOperand(MulOp == 0);
1092 if (Mul->getNumOperands() != 2) {
1093 // If the multiply has more than two operands, we must get the
1095 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
1096 MulOps.erase(MulOps.begin()+MulOp);
1097 InnerMul = getMulExpr(MulOps);
1099 SCEVHandle One = getIntegerSCEV(1, Ty);
1100 SCEVHandle AddOne = getAddExpr(InnerMul, One);
1101 SCEVHandle OuterMul = getMulExpr(AddOne, Ops[AddOp]);
1102 if (Ops.size() == 2) return OuterMul;
1104 Ops.erase(Ops.begin()+AddOp);
1105 Ops.erase(Ops.begin()+Idx-1);
1107 Ops.erase(Ops.begin()+Idx);
1108 Ops.erase(Ops.begin()+AddOp-1);
1110 Ops.push_back(OuterMul);
1111 return getAddExpr(Ops);
1114 // Check this multiply against other multiplies being added together.
1115 for (unsigned OtherMulIdx = Idx+1;
1116 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1118 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1119 // If MulOp occurs in OtherMul, we can fold the two multiplies
1121 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1122 OMulOp != e; ++OMulOp)
1123 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1124 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1125 SCEVHandle InnerMul1 = Mul->getOperand(MulOp == 0);
1126 if (Mul->getNumOperands() != 2) {
1127 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
1128 MulOps.erase(MulOps.begin()+MulOp);
1129 InnerMul1 = getMulExpr(MulOps);
1131 SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1132 if (OtherMul->getNumOperands() != 2) {
1133 std::vector<SCEVHandle> MulOps(OtherMul->op_begin(),
1134 OtherMul->op_end());
1135 MulOps.erase(MulOps.begin()+OMulOp);
1136 InnerMul2 = getMulExpr(MulOps);
1138 SCEVHandle InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
1139 SCEVHandle OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
1140 if (Ops.size() == 2) return OuterMul;
1141 Ops.erase(Ops.begin()+Idx);
1142 Ops.erase(Ops.begin()+OtherMulIdx-1);
1143 Ops.push_back(OuterMul);
1144 return getAddExpr(Ops);
1150 // If there are any add recurrences in the operands list, see if any other
1151 // added values are loop invariant. If so, we can fold them into the
1153 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1156 // Scan over all recurrences, trying to fold loop invariants into them.
1157 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1158 // Scan all of the other operands to this add and add them to the vector if
1159 // they are loop invariant w.r.t. the recurrence.
1160 std::vector<SCEVHandle> LIOps;
1161 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1162 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1163 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1164 LIOps.push_back(Ops[i]);
1165 Ops.erase(Ops.begin()+i);
1169 // If we found some loop invariants, fold them into the recurrence.
1170 if (!LIOps.empty()) {
1171 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
1172 LIOps.push_back(AddRec->getStart());
1174 std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
1175 AddRecOps[0] = getAddExpr(LIOps);
1177 SCEVHandle NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
1178 // If all of the other operands were loop invariant, we are done.
1179 if (Ops.size() == 1) return NewRec;
1181 // Otherwise, add the folded AddRec by the non-liv parts.
1182 for (unsigned i = 0;; ++i)
1183 if (Ops[i] == AddRec) {
1187 return getAddExpr(Ops);
1190 // Okay, if there weren't any loop invariants to be folded, check to see if
1191 // there are multiple AddRec's with the same loop induction variable being
1192 // added together. If so, we can fold them.
1193 for (unsigned OtherIdx = Idx+1;
1194 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1195 if (OtherIdx != Idx) {
1196 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1197 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1198 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
1199 std::vector<SCEVHandle> NewOps(AddRec->op_begin(), AddRec->op_end());
1200 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
1201 if (i >= NewOps.size()) {
1202 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
1203 OtherAddRec->op_end());
1206 NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
1208 SCEVHandle NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
1210 if (Ops.size() == 2) return NewAddRec;
1212 Ops.erase(Ops.begin()+Idx);
1213 Ops.erase(Ops.begin()+OtherIdx-1);
1214 Ops.push_back(NewAddRec);
1215 return getAddExpr(Ops);
1219 // Otherwise couldn't fold anything into this recurrence. Move onto the
1223 // Okay, it looks like we really DO need an add expr. Check to see if we
1224 // already have one, otherwise create a new one.
1225 std::vector<const SCEV*> SCEVOps(Ops.begin(), Ops.end());
1226 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scAddExpr,
1228 if (Result == 0) Result = new SCEVAddExpr(Ops);
1233 /// getMulExpr - Get a canonical multiply expression, or something simpler if
1235 SCEVHandle ScalarEvolution::getMulExpr(std::vector<SCEVHandle> &Ops) {
1236 assert(!Ops.empty() && "Cannot get empty mul!");
1238 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1239 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1240 getEffectiveSCEVType(Ops[0]->getType()) &&
1241 "SCEVMulExpr operand types don't match!");
1244 // Sort by complexity, this groups all similar expression types together.
1245 GroupByComplexity(Ops, LI);
1247 // If there are any constants, fold them together.
1249 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1251 // C1*(C2+V) -> C1*C2 + C1*V
1252 if (Ops.size() == 2)
1253 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1254 if (Add->getNumOperands() == 2 &&
1255 isa<SCEVConstant>(Add->getOperand(0)))
1256 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
1257 getMulExpr(LHSC, Add->getOperand(1)));
1261 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1262 // We found two constants, fold them together!
1263 ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() *
1264 RHSC->getValue()->getValue());
1265 Ops[0] = getConstant(Fold);
1266 Ops.erase(Ops.begin()+1); // Erase the folded element
1267 if (Ops.size() == 1) return Ops[0];
1268 LHSC = cast<SCEVConstant>(Ops[0]);
1271 // If we are left with a constant one being multiplied, strip it off.
1272 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1273 Ops.erase(Ops.begin());
1275 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
1276 // If we have a multiply of zero, it will always be zero.
1281 // Skip over the add expression until we get to a multiply.
1282 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1285 if (Ops.size() == 1)
1288 // If there are mul operands inline them all into this expression.
1289 if (Idx < Ops.size()) {
1290 bool DeletedMul = false;
1291 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1292 // If we have an mul, expand the mul operands onto the end of the operands
1294 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1295 Ops.erase(Ops.begin()+Idx);
1299 // If we deleted at least one mul, we added operands to the end of the list,
1300 // and they are not necessarily sorted. Recurse to resort and resimplify
1301 // any operands we just aquired.
1303 return getMulExpr(Ops);
1306 // If there are any add recurrences in the operands list, see if any other
1307 // added values are loop invariant. If so, we can fold them into the
1309 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1312 // Scan over all recurrences, trying to fold loop invariants into them.
1313 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1314 // Scan all of the other operands to this mul and add them to the vector if
1315 // they are loop invariant w.r.t. the recurrence.
1316 std::vector<SCEVHandle> LIOps;
1317 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1318 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1319 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1320 LIOps.push_back(Ops[i]);
1321 Ops.erase(Ops.begin()+i);
1325 // If we found some loop invariants, fold them into the recurrence.
1326 if (!LIOps.empty()) {
1327 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
1328 std::vector<SCEVHandle> NewOps;
1329 NewOps.reserve(AddRec->getNumOperands());
1330 if (LIOps.size() == 1) {
1331 const SCEV *Scale = LIOps[0];
1332 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1333 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
1335 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1336 std::vector<SCEVHandle> MulOps(LIOps);
1337 MulOps.push_back(AddRec->getOperand(i));
1338 NewOps.push_back(getMulExpr(MulOps));
1342 SCEVHandle NewRec = getAddRecExpr(NewOps, AddRec->getLoop());
1344 // If all of the other operands were loop invariant, we are done.
1345 if (Ops.size() == 1) return NewRec;
1347 // Otherwise, multiply the folded AddRec by the non-liv parts.
1348 for (unsigned i = 0;; ++i)
1349 if (Ops[i] == AddRec) {
1353 return getMulExpr(Ops);
1356 // Okay, if there weren't any loop invariants to be folded, check to see if
1357 // there are multiple AddRec's with the same loop induction variable being
1358 // multiplied together. If so, we can fold them.
1359 for (unsigned OtherIdx = Idx+1;
1360 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1361 if (OtherIdx != Idx) {
1362 const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1363 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1364 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1365 const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1366 SCEVHandle NewStart = getMulExpr(F->getStart(),
1368 SCEVHandle B = F->getStepRecurrence(*this);
1369 SCEVHandle D = G->getStepRecurrence(*this);
1370 SCEVHandle NewStep = getAddExpr(getMulExpr(F, D),
1373 SCEVHandle NewAddRec = getAddRecExpr(NewStart, NewStep,
1375 if (Ops.size() == 2) return NewAddRec;
1377 Ops.erase(Ops.begin()+Idx);
1378 Ops.erase(Ops.begin()+OtherIdx-1);
1379 Ops.push_back(NewAddRec);
1380 return getMulExpr(Ops);
1384 // Otherwise couldn't fold anything into this recurrence. Move onto the
1388 // Okay, it looks like we really DO need an mul expr. Check to see if we
1389 // already have one, otherwise create a new one.
1390 std::vector<const SCEV*> SCEVOps(Ops.begin(), Ops.end());
1391 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scMulExpr,
1394 Result = new SCEVMulExpr(Ops);
1398 /// getUDivExpr - Get a canonical multiply expression, or something simpler if
1400 SCEVHandle ScalarEvolution::getUDivExpr(const SCEVHandle &LHS,
1401 const SCEVHandle &RHS) {
1402 assert(getEffectiveSCEVType(LHS->getType()) ==
1403 getEffectiveSCEVType(RHS->getType()) &&
1404 "SCEVUDivExpr operand types don't match!");
1406 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1407 if (RHSC->getValue()->equalsInt(1))
1408 return LHS; // X udiv 1 --> x
1410 return getIntegerSCEV(0, LHS->getType()); // value is undefined
1412 // Determine if the division can be folded into the operands of
1414 // TODO: Generalize this to non-constants by using known-bits information.
1415 const Type *Ty = LHS->getType();
1416 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
1417 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ;
1418 // For non-power-of-two values, effectively round the value up to the
1419 // nearest power of two.
1420 if (!RHSC->getValue()->getValue().isPowerOf2())
1422 const IntegerType *ExtTy =
1423 IntegerType::get(getTypeSizeInBits(Ty) + MaxShiftAmt);
1424 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
1425 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
1426 if (const SCEVConstant *Step =
1427 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
1428 if (!Step->getValue()->getValue()
1429 .urem(RHSC->getValue()->getValue()) &&
1430 getZeroExtendExpr(AR, ExtTy) ==
1431 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
1432 getZeroExtendExpr(Step, ExtTy),
1434 std::vector<SCEVHandle> Operands;
1435 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
1436 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
1437 return getAddRecExpr(Operands, AR->getLoop());
1439 // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
1440 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
1441 std::vector<SCEVHandle> Operands;
1442 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
1443 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
1444 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
1445 // Find an operand that's safely divisible.
1446 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
1447 SCEVHandle Op = M->getOperand(i);
1448 SCEVHandle Div = getUDivExpr(Op, RHSC);
1449 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
1450 Operands = M->getOperands();
1452 return getMulExpr(Operands);
1456 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
1457 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
1458 std::vector<SCEVHandle> Operands;
1459 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
1460 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
1461 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
1463 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
1464 SCEVHandle Op = getUDivExpr(A->getOperand(i), RHS);
1465 if (isa<SCEVUDivExpr>(Op) || getMulExpr(Op, RHS) != A->getOperand(i))
1467 Operands.push_back(Op);
1469 if (Operands.size() == A->getNumOperands())
1470 return getAddExpr(Operands);
1474 // Fold if both operands are constant.
1475 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1476 Constant *LHSCV = LHSC->getValue();
1477 Constant *RHSCV = RHSC->getValue();
1478 return getUnknown(ConstantExpr::getUDiv(LHSCV, RHSCV));
1482 SCEVUDivExpr *&Result = (*SCEVUDivs)[std::make_pair(LHS, RHS)];
1483 if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
1488 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1489 /// Simplify the expression as much as possible.
1490 SCEVHandle ScalarEvolution::getAddRecExpr(const SCEVHandle &Start,
1491 const SCEVHandle &Step, const Loop *L) {
1492 std::vector<SCEVHandle> Operands;
1493 Operands.push_back(Start);
1494 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1495 if (StepChrec->getLoop() == L) {
1496 Operands.insert(Operands.end(), StepChrec->op_begin(),
1497 StepChrec->op_end());
1498 return getAddRecExpr(Operands, L);
1501 Operands.push_back(Step);
1502 return getAddRecExpr(Operands, L);
1505 /// getAddRecExpr - Get an add recurrence expression for the specified loop.
1506 /// Simplify the expression as much as possible.
1507 SCEVHandle ScalarEvolution::getAddRecExpr(std::vector<SCEVHandle> &Operands,
1509 if (Operands.size() == 1) return Operands[0];
1511 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
1512 assert(getEffectiveSCEVType(Operands[i]->getType()) ==
1513 getEffectiveSCEVType(Operands[0]->getType()) &&
1514 "SCEVAddRecExpr operand types don't match!");
1517 if (Operands.back()->isZero()) {
1518 Operands.pop_back();
1519 return getAddRecExpr(Operands, L); // {X,+,0} --> X
1522 // Canonicalize nested AddRecs in by nesting them in order of loop depth.
1523 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
1524 const Loop* NestedLoop = NestedAR->getLoop();
1525 if (L->getLoopDepth() < NestedLoop->getLoopDepth()) {
1526 std::vector<SCEVHandle> NestedOperands(NestedAR->op_begin(),
1527 NestedAR->op_end());
1528 SCEVHandle NestedARHandle(NestedAR);
1529 Operands[0] = NestedAR->getStart();
1530 NestedOperands[0] = getAddRecExpr(Operands, L);
1531 return getAddRecExpr(NestedOperands, NestedLoop);
1535 std::vector<const SCEV*> SCEVOps(Operands.begin(), Operands.end());
1536 SCEVAddRecExpr *&Result = (*SCEVAddRecExprs)[std::make_pair(L, SCEVOps)];
1537 if (Result == 0) Result = new SCEVAddRecExpr(Operands, L);
1541 SCEVHandle ScalarEvolution::getSMaxExpr(const SCEVHandle &LHS,
1542 const SCEVHandle &RHS) {
1543 std::vector<SCEVHandle> Ops;
1546 return getSMaxExpr(Ops);
1549 SCEVHandle ScalarEvolution::getSMaxExpr(std::vector<SCEVHandle> Ops) {
1550 assert(!Ops.empty() && "Cannot get empty smax!");
1551 if (Ops.size() == 1) return Ops[0];
1553 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1554 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1555 getEffectiveSCEVType(Ops[0]->getType()) &&
1556 "SCEVSMaxExpr operand types don't match!");
1559 // Sort by complexity, this groups all similar expression types together.
1560 GroupByComplexity(Ops, LI);
1562 // If there are any constants, fold them together.
1564 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1566 assert(Idx < Ops.size());
1567 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1568 // We found two constants, fold them together!
1569 ConstantInt *Fold = ConstantInt::get(
1570 APIntOps::smax(LHSC->getValue()->getValue(),
1571 RHSC->getValue()->getValue()));
1572 Ops[0] = getConstant(Fold);
1573 Ops.erase(Ops.begin()+1); // Erase the folded element
1574 if (Ops.size() == 1) return Ops[0];
1575 LHSC = cast<SCEVConstant>(Ops[0]);
1578 // If we are left with a constant -inf, strip it off.
1579 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
1580 Ops.erase(Ops.begin());
1585 if (Ops.size() == 1) return Ops[0];
1587 // Find the first SMax
1588 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
1591 // Check to see if one of the operands is an SMax. If so, expand its operands
1592 // onto our operand list, and recurse to simplify.
1593 if (Idx < Ops.size()) {
1594 bool DeletedSMax = false;
1595 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
1596 Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
1597 Ops.erase(Ops.begin()+Idx);
1602 return getSMaxExpr(Ops);
1605 // Okay, check to see if the same value occurs in the operand list twice. If
1606 // so, delete one. Since we sorted the list, these values are required to
1608 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1609 if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
1610 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1614 if (Ops.size() == 1) return Ops[0];
1616 assert(!Ops.empty() && "Reduced smax down to nothing!");
1618 // Okay, it looks like we really DO need an smax expr. Check to see if we
1619 // already have one, otherwise create a new one.
1620 std::vector<const SCEV*> SCEVOps(Ops.begin(), Ops.end());
1621 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scSMaxExpr,
1623 if (Result == 0) Result = new SCEVSMaxExpr(Ops);
1627 SCEVHandle ScalarEvolution::getUMaxExpr(const SCEVHandle &LHS,
1628 const SCEVHandle &RHS) {
1629 std::vector<SCEVHandle> Ops;
1632 return getUMaxExpr(Ops);
1635 SCEVHandle ScalarEvolution::getUMaxExpr(std::vector<SCEVHandle> Ops) {
1636 assert(!Ops.empty() && "Cannot get empty umax!");
1637 if (Ops.size() == 1) return Ops[0];
1639 for (unsigned i = 1, e = Ops.size(); i != e; ++i)
1640 assert(getEffectiveSCEVType(Ops[i]->getType()) ==
1641 getEffectiveSCEVType(Ops[0]->getType()) &&
1642 "SCEVUMaxExpr operand types don't match!");
1645 // Sort by complexity, this groups all similar expression types together.
1646 GroupByComplexity(Ops, LI);
1648 // If there are any constants, fold them together.
1650 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1652 assert(Idx < Ops.size());
1653 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1654 // We found two constants, fold them together!
1655 ConstantInt *Fold = ConstantInt::get(
1656 APIntOps::umax(LHSC->getValue()->getValue(),
1657 RHSC->getValue()->getValue()));
1658 Ops[0] = getConstant(Fold);
1659 Ops.erase(Ops.begin()+1); // Erase the folded element
1660 if (Ops.size() == 1) return Ops[0];
1661 LHSC = cast<SCEVConstant>(Ops[0]);
1664 // If we are left with a constant zero, strip it off.
1665 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
1666 Ops.erase(Ops.begin());
1671 if (Ops.size() == 1) return Ops[0];
1673 // Find the first UMax
1674 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
1677 // Check to see if one of the operands is a UMax. If so, expand its operands
1678 // onto our operand list, and recurse to simplify.
1679 if (Idx < Ops.size()) {
1680 bool DeletedUMax = false;
1681 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
1682 Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
1683 Ops.erase(Ops.begin()+Idx);
1688 return getUMaxExpr(Ops);
1691 // Okay, check to see if the same value occurs in the operand list twice. If
1692 // so, delete one. Since we sorted the list, these values are required to
1694 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
1695 if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
1696 Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
1700 if (Ops.size() == 1) return Ops[0];
1702 assert(!Ops.empty() && "Reduced umax down to nothing!");
1704 // Okay, it looks like we really DO need a umax expr. Check to see if we
1705 // already have one, otherwise create a new one.
1706 std::vector<const SCEV*> SCEVOps(Ops.begin(), Ops.end());
1707 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scUMaxExpr,
1709 if (Result == 0) Result = new SCEVUMaxExpr(Ops);
1713 SCEVHandle ScalarEvolution::getUnknown(Value *V) {
1714 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
1715 return getConstant(CI);
1716 if (isa<ConstantPointerNull>(V))
1717 return getIntegerSCEV(0, V->getType());
1718 SCEVUnknown *&Result = (*SCEVUnknowns)[V];
1719 if (Result == 0) Result = new SCEVUnknown(V);
1723 //===----------------------------------------------------------------------===//
1724 // Basic SCEV Analysis and PHI Idiom Recognition Code
1727 /// isSCEVable - Test if values of the given type are analyzable within
1728 /// the SCEV framework. This primarily includes integer types, and it
1729 /// can optionally include pointer types if the ScalarEvolution class
1730 /// has access to target-specific information.
1731 bool ScalarEvolution::isSCEVable(const Type *Ty) const {
1732 // Integers are always SCEVable.
1733 if (Ty->isInteger())
1736 // Pointers are SCEVable if TargetData information is available
1737 // to provide pointer size information.
1738 if (isa<PointerType>(Ty))
1741 // Otherwise it's not SCEVable.
1745 /// getTypeSizeInBits - Return the size in bits of the specified type,
1746 /// for which isSCEVable must return true.
1747 uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
1748 assert(isSCEVable(Ty) && "Type is not SCEVable!");
1750 // If we have a TargetData, use it!
1752 return TD->getTypeSizeInBits(Ty);
1754 // Otherwise, we support only integer types.
1755 assert(Ty->isInteger() && "isSCEVable permitted a non-SCEVable type!");
1756 return Ty->getPrimitiveSizeInBits();
1759 /// getEffectiveSCEVType - Return a type with the same bitwidth as
1760 /// the given type and which represents how SCEV will treat the given
1761 /// type, for which isSCEVable must return true. For pointer types,
1762 /// this is the pointer-sized integer type.
1763 const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
1764 assert(isSCEVable(Ty) && "Type is not SCEVable!");
1766 if (Ty->isInteger())
1769 assert(isa<PointerType>(Ty) && "Unexpected non-pointer non-integer type!");
1770 return TD->getIntPtrType();
1773 SCEVHandle ScalarEvolution::getCouldNotCompute() {
1774 return UnknownValue;
1777 /// hasSCEV - Return true if the SCEV for this value has already been
1779 bool ScalarEvolution::hasSCEV(Value *V) const {
1780 return Scalars.count(V);
1783 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1784 /// expression and create a new one.
1785 SCEVHandle ScalarEvolution::getSCEV(Value *V) {
1786 assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
1788 std::map<SCEVCallbackVH, SCEVHandle>::iterator I = Scalars.find(V);
1789 if (I != Scalars.end()) return I->second;
1790 SCEVHandle S = createSCEV(V);
1791 Scalars.insert(std::make_pair(SCEVCallbackVH(V, this), S));
1795 /// getIntegerSCEV - Given an integer or FP type, create a constant for the
1796 /// specified signed integer value and return a SCEV for the constant.
1797 SCEVHandle ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
1798 Ty = getEffectiveSCEVType(Ty);
1801 C = Constant::getNullValue(Ty);
1802 else if (Ty->isFloatingPoint())
1803 C = ConstantFP::get(APFloat(Ty==Type::FloatTy ? APFloat::IEEEsingle :
1804 APFloat::IEEEdouble, Val));
1806 C = ConstantInt::get(Ty, Val);
1807 return getUnknown(C);
1810 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
1812 SCEVHandle ScalarEvolution::getNegativeSCEV(const SCEVHandle &V) {
1813 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
1814 return getUnknown(ConstantExpr::getNeg(VC->getValue()));
1816 const Type *Ty = V->getType();
1817 Ty = getEffectiveSCEVType(Ty);
1818 return getMulExpr(V, getConstant(ConstantInt::getAllOnesValue(Ty)));
1821 /// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
1822 SCEVHandle ScalarEvolution::getNotSCEV(const SCEVHandle &V) {
1823 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
1824 return getUnknown(ConstantExpr::getNot(VC->getValue()));
1826 const Type *Ty = V->getType();
1827 Ty = getEffectiveSCEVType(Ty);
1828 SCEVHandle AllOnes = getConstant(ConstantInt::getAllOnesValue(Ty));
1829 return getMinusSCEV(AllOnes, V);
1832 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
1834 SCEVHandle ScalarEvolution::getMinusSCEV(const SCEVHandle &LHS,
1835 const SCEVHandle &RHS) {
1837 return getAddExpr(LHS, getNegativeSCEV(RHS));
1840 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
1841 /// input value to the specified type. If the type must be extended, it is zero
1844 ScalarEvolution::getTruncateOrZeroExtend(const SCEVHandle &V,
1846 const Type *SrcTy = V->getType();
1847 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
1848 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
1849 "Cannot truncate or zero extend with non-integer arguments!");
1850 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
1851 return V; // No conversion
1852 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
1853 return getTruncateExpr(V, Ty);
1854 return getZeroExtendExpr(V, Ty);
1857 /// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
1858 /// input value to the specified type. If the type must be extended, it is sign
1861 ScalarEvolution::getTruncateOrSignExtend(const SCEVHandle &V,
1863 const Type *SrcTy = V->getType();
1864 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
1865 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
1866 "Cannot truncate or zero extend with non-integer arguments!");
1867 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
1868 return V; // No conversion
1869 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
1870 return getTruncateExpr(V, Ty);
1871 return getSignExtendExpr(V, Ty);
1874 /// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
1875 /// input value to the specified type. If the type must be extended, it is zero
1876 /// extended. The conversion must not be narrowing.
1878 ScalarEvolution::getNoopOrZeroExtend(const SCEVHandle &V, const Type *Ty) {
1879 const Type *SrcTy = V->getType();
1880 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
1881 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
1882 "Cannot noop or zero extend with non-integer arguments!");
1883 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
1884 "getNoopOrZeroExtend cannot truncate!");
1885 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
1886 return V; // No conversion
1887 return getZeroExtendExpr(V, Ty);
1890 /// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
1891 /// input value to the specified type. If the type must be extended, it is sign
1892 /// extended. The conversion must not be narrowing.
1894 ScalarEvolution::getNoopOrSignExtend(const SCEVHandle &V, const Type *Ty) {
1895 const Type *SrcTy = V->getType();
1896 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
1897 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
1898 "Cannot noop or sign extend with non-integer arguments!");
1899 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
1900 "getNoopOrSignExtend cannot truncate!");
1901 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
1902 return V; // No conversion
1903 return getSignExtendExpr(V, Ty);
1906 /// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
1907 /// input value to the specified type. The conversion must not be widening.
1909 ScalarEvolution::getTruncateOrNoop(const SCEVHandle &V, const Type *Ty) {
1910 const Type *SrcTy = V->getType();
1911 assert((SrcTy->isInteger() || (TD && isa<PointerType>(SrcTy))) &&
1912 (Ty->isInteger() || (TD && isa<PointerType>(Ty))) &&
1913 "Cannot truncate or noop with non-integer arguments!");
1914 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
1915 "getTruncateOrNoop cannot extend!");
1916 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
1917 return V; // No conversion
1918 return getTruncateExpr(V, Ty);
1921 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value for
1922 /// the specified instruction and replaces any references to the symbolic value
1923 /// SymName with the specified value. This is used during PHI resolution.
1924 void ScalarEvolution::
1925 ReplaceSymbolicValueWithConcrete(Instruction *I, const SCEVHandle &SymName,
1926 const SCEVHandle &NewVal) {
1927 std::map<SCEVCallbackVH, SCEVHandle>::iterator SI =
1928 Scalars.find(SCEVCallbackVH(I, this));
1929 if (SI == Scalars.end()) return;
1932 SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal, *this);
1933 if (NV == SI->second) return; // No change.
1935 SI->second = NV; // Update the scalars map!
1937 // Any instruction values that use this instruction might also need to be
1939 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1941 ReplaceSymbolicValueWithConcrete(cast<Instruction>(*UI), SymName, NewVal);
1944 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
1945 /// a loop header, making it a potential recurrence, or it doesn't.
1947 SCEVHandle ScalarEvolution::createNodeForPHI(PHINode *PN) {
1948 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
1949 if (const Loop *L = LI->getLoopFor(PN->getParent()))
1950 if (L->getHeader() == PN->getParent()) {
1951 // If it lives in the loop header, it has two incoming values, one
1952 // from outside the loop, and one from inside.
1953 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1954 unsigned BackEdge = IncomingEdge^1;
1956 // While we are analyzing this PHI node, handle its value symbolically.
1957 SCEVHandle SymbolicName = getUnknown(PN);
1958 assert(Scalars.find(PN) == Scalars.end() &&
1959 "PHI node already processed?");
1960 Scalars.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
1962 // Using this symbolic name for the PHI, analyze the value coming around
1964 SCEVHandle BEValue = getSCEV(PN->getIncomingValue(BackEdge));
1966 // NOTE: If BEValue is loop invariant, we know that the PHI node just
1967 // has a special value for the first iteration of the loop.
1969 // If the value coming around the backedge is an add with the symbolic
1970 // value we just inserted, then we found a simple induction variable!
1971 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
1972 // If there is a single occurrence of the symbolic value, replace it
1973 // with a recurrence.
1974 unsigned FoundIndex = Add->getNumOperands();
1975 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1976 if (Add->getOperand(i) == SymbolicName)
1977 if (FoundIndex == e) {
1982 if (FoundIndex != Add->getNumOperands()) {
1983 // Create an add with everything but the specified operand.
1984 std::vector<SCEVHandle> Ops;
1985 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1986 if (i != FoundIndex)
1987 Ops.push_back(Add->getOperand(i));
1988 SCEVHandle Accum = getAddExpr(Ops);
1990 // This is not a valid addrec if the step amount is varying each
1991 // loop iteration, but is not itself an addrec in this loop.
1992 if (Accum->isLoopInvariant(L) ||
1993 (isa<SCEVAddRecExpr>(Accum) &&
1994 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
1995 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1996 SCEVHandle PHISCEV = getAddRecExpr(StartVal, Accum, L);
1998 // Okay, for the entire analysis of this edge we assumed the PHI
1999 // to be symbolic. We now need to go back and update all of the
2000 // entries for the scalars that use the PHI (except for the PHI
2001 // itself) to use the new analyzed value instead of the "symbolic"
2003 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
2007 } else if (const SCEVAddRecExpr *AddRec =
2008 dyn_cast<SCEVAddRecExpr>(BEValue)) {
2009 // Otherwise, this could be a loop like this:
2010 // i = 0; for (j = 1; ..; ++j) { .... i = j; }
2011 // In this case, j = {1,+,1} and BEValue is j.
2012 // Because the other in-value of i (0) fits the evolution of BEValue
2013 // i really is an addrec evolution.
2014 if (AddRec->getLoop() == L && AddRec->isAffine()) {
2015 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
2017 // If StartVal = j.start - j.stride, we can use StartVal as the
2018 // initial step of the addrec evolution.
2019 if (StartVal == getMinusSCEV(AddRec->getOperand(0),
2020 AddRec->getOperand(1))) {
2021 SCEVHandle PHISCEV =
2022 getAddRecExpr(StartVal, AddRec->getOperand(1), L);
2024 // Okay, for the entire analysis of this edge we assumed the PHI
2025 // to be symbolic. We now need to go back and update all of the
2026 // entries for the scalars that use the PHI (except for the PHI
2027 // itself) to use the new analyzed value instead of the "symbolic"
2029 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV);
2035 return SymbolicName;
2038 // If it's not a loop phi, we can't handle it yet.
2039 return getUnknown(PN);
2042 /// createNodeForGEP - Expand GEP instructions into add and multiply
2043 /// operations. This allows them to be analyzed by regular SCEV code.
2045 SCEVHandle ScalarEvolution::createNodeForGEP(User *GEP) {
2047 const Type *IntPtrTy = TD->getIntPtrType();
2048 Value *Base = GEP->getOperand(0);
2049 // Don't attempt to analyze GEPs over unsized objects.
2050 if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
2051 return getUnknown(GEP);
2052 SCEVHandle TotalOffset = getIntegerSCEV(0, IntPtrTy);
2053 gep_type_iterator GTI = gep_type_begin(GEP);
2054 for (GetElementPtrInst::op_iterator I = next(GEP->op_begin()),
2058 // Compute the (potentially symbolic) offset in bytes for this index.
2059 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
2060 // For a struct, add the member offset.
2061 const StructLayout &SL = *TD->getStructLayout(STy);
2062 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
2063 uint64_t Offset = SL.getElementOffset(FieldNo);
2064 TotalOffset = getAddExpr(TotalOffset,
2065 getIntegerSCEV(Offset, IntPtrTy));
2067 // For an array, add the element offset, explicitly scaled.
2068 SCEVHandle LocalOffset = getSCEV(Index);
2069 if (!isa<PointerType>(LocalOffset->getType()))
2070 // Getelementptr indicies are signed.
2071 LocalOffset = getTruncateOrSignExtend(LocalOffset,
2074 getMulExpr(LocalOffset,
2075 getIntegerSCEV(TD->getTypeAllocSize(*GTI),
2077 TotalOffset = getAddExpr(TotalOffset, LocalOffset);
2080 return getAddExpr(getSCEV(Base), TotalOffset);
2083 /// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
2084 /// guaranteed to end in (at every loop iteration). It is, at the same time,
2085 /// the minimum number of times S is divisible by 2. For example, given {4,+,8}
2086 /// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
2087 static uint32_t GetMinTrailingZeros(SCEVHandle S, const ScalarEvolution &SE) {
2088 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
2089 return C->getValue()->getValue().countTrailingZeros();
2091 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
2092 return std::min(GetMinTrailingZeros(T->getOperand(), SE),
2093 (uint32_t)SE.getTypeSizeInBits(T->getType()));
2095 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
2096 uint32_t OpRes = GetMinTrailingZeros(E->getOperand(), SE);
2097 return OpRes == SE.getTypeSizeInBits(E->getOperand()->getType()) ?
2098 SE.getTypeSizeInBits(E->getType()) : OpRes;
2101 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
2102 uint32_t OpRes = GetMinTrailingZeros(E->getOperand(), SE);
2103 return OpRes == SE.getTypeSizeInBits(E->getOperand()->getType()) ?
2104 SE.getTypeSizeInBits(E->getType()) : OpRes;
2107 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
2108 // The result is the min of all operands results.
2109 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0), SE);
2110 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2111 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i), SE));
2115 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
2116 // The result is the sum of all operands results.
2117 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0), SE);
2118 uint32_t BitWidth = SE.getTypeSizeInBits(M->getType());
2119 for (unsigned i = 1, e = M->getNumOperands();
2120 SumOpRes != BitWidth && i != e; ++i)
2121 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i), SE),
2126 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
2127 // The result is the min of all operands results.
2128 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0), SE);
2129 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
2130 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i), SE));
2134 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
2135 // The result is the min of all operands results.
2136 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0), SE);
2137 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2138 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i), SE));
2142 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
2143 // The result is the min of all operands results.
2144 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0), SE);
2145 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
2146 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i), SE));
2150 // SCEVUDivExpr, SCEVUnknown
2154 /// createSCEV - We know that there is no SCEV for the specified value.
2155 /// Analyze the expression.
2157 SCEVHandle ScalarEvolution::createSCEV(Value *V) {
2158 if (!isSCEVable(V->getType()))
2159 return getUnknown(V);
2161 unsigned Opcode = Instruction::UserOp1;
2162 if (Instruction *I = dyn_cast<Instruction>(V))
2163 Opcode = I->getOpcode();
2164 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2165 Opcode = CE->getOpcode();
2167 return getUnknown(V);
2169 User *U = cast<User>(V);
2171 case Instruction::Add:
2172 return getAddExpr(getSCEV(U->getOperand(0)),
2173 getSCEV(U->getOperand(1)));
2174 case Instruction::Mul:
2175 return getMulExpr(getSCEV(U->getOperand(0)),
2176 getSCEV(U->getOperand(1)));
2177 case Instruction::UDiv:
2178 return getUDivExpr(getSCEV(U->getOperand(0)),
2179 getSCEV(U->getOperand(1)));
2180 case Instruction::Sub:
2181 return getMinusSCEV(getSCEV(U->getOperand(0)),
2182 getSCEV(U->getOperand(1)));
2183 case Instruction::And:
2184 // For an expression like x&255 that merely masks off the high bits,
2185 // use zext(trunc(x)) as the SCEV expression.
2186 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2187 if (CI->isNullValue())
2188 return getSCEV(U->getOperand(1));
2189 if (CI->isAllOnesValue())
2190 return getSCEV(U->getOperand(0));
2191 const APInt &A = CI->getValue();
2192 unsigned Ones = A.countTrailingOnes();
2193 if (APIntOps::isMask(Ones, A))
2195 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
2196 IntegerType::get(Ones)),
2200 case Instruction::Or:
2201 // If the RHS of the Or is a constant, we may have something like:
2202 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
2203 // optimizations will transparently handle this case.
2205 // In order for this transformation to be safe, the LHS must be of the
2206 // form X*(2^n) and the Or constant must be less than 2^n.
2207 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2208 SCEVHandle LHS = getSCEV(U->getOperand(0));
2209 const APInt &CIVal = CI->getValue();
2210 if (GetMinTrailingZeros(LHS, *this) >=
2211 (CIVal.getBitWidth() - CIVal.countLeadingZeros()))
2212 return getAddExpr(LHS, getSCEV(U->getOperand(1)));
2215 case Instruction::Xor:
2216 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
2217 // If the RHS of the xor is a signbit, then this is just an add.
2218 // Instcombine turns add of signbit into xor as a strength reduction step.
2219 if (CI->getValue().isSignBit())
2220 return getAddExpr(getSCEV(U->getOperand(0)),
2221 getSCEV(U->getOperand(1)));
2223 // If the RHS of xor is -1, then this is a not operation.
2224 if (CI->isAllOnesValue())
2225 return getNotSCEV(getSCEV(U->getOperand(0)));
2227 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
2228 // This is a variant of the check for xor with -1, and it handles
2229 // the case where instcombine has trimmed non-demanded bits out
2230 // of an xor with -1.
2231 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
2232 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
2233 if (BO->getOpcode() == Instruction::And &&
2234 LCI->getValue() == CI->getValue())
2235 if (const SCEVZeroExtendExpr *Z =
2236 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0))))
2237 return getZeroExtendExpr(getNotSCEV(Z->getOperand()),
2242 case Instruction::Shl:
2243 // Turn shift left of a constant amount into a multiply.
2244 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
2245 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
2246 Constant *X = ConstantInt::get(
2247 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
2248 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
2252 case Instruction::LShr:
2253 // Turn logical shift right of a constant into a unsigned divide.
2254 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
2255 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
2256 Constant *X = ConstantInt::get(
2257 APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
2258 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
2262 case Instruction::AShr:
2263 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
2264 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
2265 if (Instruction *L = dyn_cast<Instruction>(U->getOperand(0)))
2266 if (L->getOpcode() == Instruction::Shl &&
2267 L->getOperand(1) == U->getOperand(1)) {
2268 unsigned BitWidth = getTypeSizeInBits(U->getType());
2269 uint64_t Amt = BitWidth - CI->getZExtValue();
2270 if (Amt == BitWidth)
2271 return getSCEV(L->getOperand(0)); // shift by zero --> noop
2273 return getIntegerSCEV(0, U->getType()); // value is undefined
2275 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
2276 IntegerType::get(Amt)),
2281 case Instruction::Trunc:
2282 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
2284 case Instruction::ZExt:
2285 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
2287 case Instruction::SExt:
2288 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
2290 case Instruction::BitCast:
2291 // BitCasts are no-op casts so we just eliminate the cast.
2292 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
2293 return getSCEV(U->getOperand(0));
2296 case Instruction::IntToPtr:
2297 if (!TD) break; // Without TD we can't analyze pointers.
2298 return getTruncateOrZeroExtend(getSCEV(U->getOperand(0)),
2299 TD->getIntPtrType());
2301 case Instruction::PtrToInt:
2302 if (!TD) break; // Without TD we can't analyze pointers.
2303 return getTruncateOrZeroExtend(getSCEV(U->getOperand(0)),
2306 case Instruction::GetElementPtr:
2307 if (!TD) break; // Without TD we can't analyze pointers.
2308 return createNodeForGEP(U);
2310 case Instruction::PHI:
2311 return createNodeForPHI(cast<PHINode>(U));
2313 case Instruction::Select:
2314 // This could be a smax or umax that was lowered earlier.
2315 // Try to recover it.
2316 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
2317 Value *LHS = ICI->getOperand(0);
2318 Value *RHS = ICI->getOperand(1);
2319 switch (ICI->getPredicate()) {
2320 case ICmpInst::ICMP_SLT:
2321 case ICmpInst::ICMP_SLE:
2322 std::swap(LHS, RHS);
2324 case ICmpInst::ICMP_SGT:
2325 case ICmpInst::ICMP_SGE:
2326 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
2327 return getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
2328 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
2329 // ~smax(~x, ~y) == smin(x, y).
2330 return getNotSCEV(getSMaxExpr(
2331 getNotSCEV(getSCEV(LHS)),
2332 getNotSCEV(getSCEV(RHS))));
2334 case ICmpInst::ICMP_ULT:
2335 case ICmpInst::ICMP_ULE:
2336 std::swap(LHS, RHS);
2338 case ICmpInst::ICMP_UGT:
2339 case ICmpInst::ICMP_UGE:
2340 if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
2341 return getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
2342 else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
2343 // ~umax(~x, ~y) == umin(x, y)
2344 return getNotSCEV(getUMaxExpr(getNotSCEV(getSCEV(LHS)),
2345 getNotSCEV(getSCEV(RHS))));
2352 default: // We cannot analyze this expression.
2356 return getUnknown(V);
2361 //===----------------------------------------------------------------------===//
2362 // Iteration Count Computation Code
2365 /// getBackedgeTakenCount - If the specified loop has a predictable
2366 /// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
2367 /// object. The backedge-taken count is the number of times the loop header
2368 /// will be branched to from within the loop. This is one less than the
2369 /// trip count of the loop, since it doesn't count the first iteration,
2370 /// when the header is branched to from outside the loop.
2372 /// Note that it is not valid to call this method on a loop without a
2373 /// loop-invariant backedge-taken count (see
2374 /// hasLoopInvariantBackedgeTakenCount).
2376 SCEVHandle ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
2377 return getBackedgeTakenInfo(L).Exact;
2380 /// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
2381 /// return the least SCEV value that is known never to be less than the
2382 /// actual backedge taken count.
2383 SCEVHandle ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
2384 return getBackedgeTakenInfo(L).Max;
2387 const ScalarEvolution::BackedgeTakenInfo &
2388 ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
2389 // Initially insert a CouldNotCompute for this loop. If the insertion
2390 // succeeds, procede to actually compute a backedge-taken count and
2391 // update the value. The temporary CouldNotCompute value tells SCEV
2392 // code elsewhere that it shouldn't attempt to request a new
2393 // backedge-taken count, which could result in infinite recursion.
2394 std::pair<std::map<const Loop*, BackedgeTakenInfo>::iterator, bool> Pair =
2395 BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
2397 BackedgeTakenInfo ItCount = ComputeBackedgeTakenCount(L);
2398 if (ItCount.Exact != UnknownValue) {
2399 assert(ItCount.Exact->isLoopInvariant(L) &&
2400 ItCount.Max->isLoopInvariant(L) &&
2401 "Computed trip count isn't loop invariant for loop!");
2402 ++NumTripCountsComputed;
2404 // Update the value in the map.
2405 Pair.first->second = ItCount;
2406 } else if (isa<PHINode>(L->getHeader()->begin())) {
2407 // Only count loops that have phi nodes as not being computable.
2408 ++NumTripCountsNotComputed;
2411 // Now that we know more about the trip count for this loop, forget any
2412 // existing SCEV values for PHI nodes in this loop since they are only
2413 // conservative estimates made without the benefit
2414 // of trip count information.
2415 if (ItCount.hasAnyInfo())
2418 return Pair.first->second;
2421 /// forgetLoopBackedgeTakenCount - This method should be called by the
2422 /// client when it has changed a loop in a way that may effect
2423 /// ScalarEvolution's ability to compute a trip count, or if the loop
2425 void ScalarEvolution::forgetLoopBackedgeTakenCount(const Loop *L) {
2426 BackedgeTakenCounts.erase(L);
2430 /// forgetLoopPHIs - Delete the memoized SCEVs associated with the
2431 /// PHI nodes in the given loop. This is used when the trip count of
2432 /// the loop may have changed.
2433 void ScalarEvolution::forgetLoopPHIs(const Loop *L) {
2434 BasicBlock *Header = L->getHeader();
2436 // Push all Loop-header PHIs onto the Worklist stack, except those
2437 // that are presently represented via a SCEVUnknown. SCEVUnknown for
2438 // a PHI either means that it has an unrecognized structure, or it's
2439 // a PHI that's in the progress of being computed by createNodeForPHI.
2440 // In the former case, additional loop trip count information isn't
2441 // going to change anything. In the later case, createNodeForPHI will
2442 // perform the necessary updates on its own when it gets to that point.
2443 SmallVector<Instruction *, 16> Worklist;
2444 for (BasicBlock::iterator I = Header->begin();
2445 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2446 std::map<SCEVCallbackVH, SCEVHandle>::iterator It = Scalars.find((Value*)I);
2447 if (It != Scalars.end() && !isa<SCEVUnknown>(It->second))
2448 Worklist.push_back(PN);
2451 while (!Worklist.empty()) {
2452 Instruction *I = Worklist.pop_back_val();
2453 if (Scalars.erase(I))
2454 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
2456 Worklist.push_back(cast<Instruction>(UI));
2460 /// ComputeBackedgeTakenCount - Compute the number of times the backedge
2461 /// of the specified loop will execute.
2462 ScalarEvolution::BackedgeTakenInfo
2463 ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
2464 // If the loop has a non-one exit block count, we can't analyze it.
2465 SmallVector<BasicBlock*, 8> ExitBlocks;
2466 L->getExitBlocks(ExitBlocks);
2467 if (ExitBlocks.size() != 1) return UnknownValue;
2469 // Okay, there is one exit block. Try to find the condition that causes the
2470 // loop to be exited.
2471 BasicBlock *ExitBlock = ExitBlocks[0];
2473 BasicBlock *ExitingBlock = 0;
2474 for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock);
2476 if (L->contains(*PI)) {
2477 if (ExitingBlock == 0)
2480 return UnknownValue; // More than one block exiting!
2482 assert(ExitingBlock && "No exits from loop, something is broken!");
2484 // Okay, we've computed the exiting block. See what condition causes us to
2487 // FIXME: we should be able to handle switch instructions (with a single exit)
2488 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2489 if (ExitBr == 0) return UnknownValue;
2490 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
2492 // At this point, we know we have a conditional branch that determines whether
2493 // the loop is exited. However, we don't know if the branch is executed each
2494 // time through the loop. If not, then the execution count of the branch will
2495 // not be equal to the trip count of the loop.
2497 // Currently we check for this by checking to see if the Exit branch goes to
2498 // the loop header. If so, we know it will always execute the same number of
2499 // times as the loop. We also handle the case where the exit block *is* the
2500 // loop header. This is common for un-rotated loops. More extensive analysis
2501 // could be done to handle more cases here.
2502 if (ExitBr->getSuccessor(0) != L->getHeader() &&
2503 ExitBr->getSuccessor(1) != L->getHeader() &&
2504 ExitBr->getParent() != L->getHeader())
2505 return UnknownValue;
2507 ICmpInst *ExitCond = dyn_cast<ICmpInst>(ExitBr->getCondition());
2509 // If it's not an integer or pointer comparison then compute it the hard way.
2511 return ComputeBackedgeTakenCountExhaustively(L, ExitBr->getCondition(),
2512 ExitBr->getSuccessor(0) == ExitBlock);
2514 // If the condition was exit on true, convert the condition to exit on false
2515 ICmpInst::Predicate Cond;
2516 if (ExitBr->getSuccessor(1) == ExitBlock)
2517 Cond = ExitCond->getPredicate();
2519 Cond = ExitCond->getInversePredicate();
2521 // Handle common loops like: for (X = "string"; *X; ++X)
2522 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
2523 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
2525 ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
2526 if (!isa<SCEVCouldNotCompute>(ItCnt)) return ItCnt;
2529 SCEVHandle LHS = getSCEV(ExitCond->getOperand(0));
2530 SCEVHandle RHS = getSCEV(ExitCond->getOperand(1));
2532 // Try to evaluate any dependencies out of the loop.
2533 LHS = getSCEVAtScope(LHS, L);
2534 RHS = getSCEVAtScope(RHS, L);
2536 // At this point, we would like to compute how many iterations of the
2537 // loop the predicate will return true for these inputs.
2538 if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
2539 // If there is a loop-invariant, force it into the RHS.
2540 std::swap(LHS, RHS);
2541 Cond = ICmpInst::getSwappedPredicate(Cond);
2544 // If we have a comparison of a chrec against a constant, try to use value
2545 // ranges to answer this query.
2546 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
2547 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
2548 if (AddRec->getLoop() == L) {
2549 // Form the constant range.
2550 ConstantRange CompRange(
2551 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
2553 SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange, *this);
2554 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
2558 case ICmpInst::ICMP_NE: { // while (X != Y)
2559 // Convert to: while (X-Y != 0)
2560 SCEVHandle TC = HowFarToZero(getMinusSCEV(LHS, RHS), L);
2561 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2564 case ICmpInst::ICMP_EQ: {
2565 // Convert to: while (X-Y == 0) // while (X == Y)
2566 SCEVHandle TC = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
2567 if (!isa<SCEVCouldNotCompute>(TC)) return TC;
2570 case ICmpInst::ICMP_SLT: {
2571 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
2572 if (BTI.hasAnyInfo()) return BTI;
2575 case ICmpInst::ICMP_SGT: {
2576 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
2577 getNotSCEV(RHS), L, true);
2578 if (BTI.hasAnyInfo()) return BTI;
2581 case ICmpInst::ICMP_ULT: {
2582 BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
2583 if (BTI.hasAnyInfo()) return BTI;
2586 case ICmpInst::ICMP_UGT: {
2587 BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
2588 getNotSCEV(RHS), L, false);
2589 if (BTI.hasAnyInfo()) return BTI;
2594 errs() << "ComputeBackedgeTakenCount ";
2595 if (ExitCond->getOperand(0)->getType()->isUnsigned())
2596 errs() << "[unsigned] ";
2597 errs() << *LHS << " "
2598 << Instruction::getOpcodeName(Instruction::ICmp)
2599 << " " << *RHS << "\n";
2604 ComputeBackedgeTakenCountExhaustively(L, ExitCond,
2605 ExitBr->getSuccessor(0) == ExitBlock);
2608 static ConstantInt *
2609 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
2610 ScalarEvolution &SE) {
2611 SCEVHandle InVal = SE.getConstant(C);
2612 SCEVHandle Val = AddRec->evaluateAtIteration(InVal, SE);
2613 assert(isa<SCEVConstant>(Val) &&
2614 "Evaluation of SCEV at constant didn't fold correctly?");
2615 return cast<SCEVConstant>(Val)->getValue();
2618 /// GetAddressedElementFromGlobal - Given a global variable with an initializer
2619 /// and a GEP expression (missing the pointer index) indexing into it, return
2620 /// the addressed element of the initializer or null if the index expression is
2623 GetAddressedElementFromGlobal(GlobalVariable *GV,
2624 const std::vector<ConstantInt*> &Indices) {
2625 Constant *Init = GV->getInitializer();
2626 for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
2627 uint64_t Idx = Indices[i]->getZExtValue();
2628 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2629 assert(Idx < CS->getNumOperands() && "Bad struct index!");
2630 Init = cast<Constant>(CS->getOperand(Idx));
2631 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2632 if (Idx >= CA->getNumOperands()) return 0; // Bogus program
2633 Init = cast<Constant>(CA->getOperand(Idx));
2634 } else if (isa<ConstantAggregateZero>(Init)) {
2635 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2636 assert(Idx < STy->getNumElements() && "Bad struct index!");
2637 Init = Constant::getNullValue(STy->getElementType(Idx));
2638 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
2639 if (Idx >= ATy->getNumElements()) return 0; // Bogus program
2640 Init = Constant::getNullValue(ATy->getElementType());
2642 assert(0 && "Unknown constant aggregate type!");
2646 return 0; // Unknown initializer type
2652 /// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
2653 /// 'icmp op load X, cst', try to see if we can compute the backedge
2654 /// execution count.
2655 SCEVHandle ScalarEvolution::
2656 ComputeLoadConstantCompareBackedgeTakenCount(LoadInst *LI, Constant *RHS,
2658 ICmpInst::Predicate predicate) {
2659 if (LI->isVolatile()) return UnknownValue;
2661 // Check to see if the loaded pointer is a getelementptr of a global.
2662 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
2663 if (!GEP) return UnknownValue;
2665 // Make sure that it is really a constant global we are gepping, with an
2666 // initializer, and make sure the first IDX is really 0.
2667 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
2668 if (!GV || !GV->isConstant() || !GV->hasInitializer() ||
2669 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
2670 !cast<Constant>(GEP->getOperand(1))->isNullValue())
2671 return UnknownValue;
2673 // Okay, we allow one non-constant index into the GEP instruction.
2675 std::vector<ConstantInt*> Indexes;
2676 unsigned VarIdxNum = 0;
2677 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
2678 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
2679 Indexes.push_back(CI);
2680 } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
2681 if (VarIdx) return UnknownValue; // Multiple non-constant idx's.
2682 VarIdx = GEP->getOperand(i);
2684 Indexes.push_back(0);
2687 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
2688 // Check to see if X is a loop variant variable value now.
2689 SCEVHandle Idx = getSCEV(VarIdx);
2690 Idx = getSCEVAtScope(Idx, L);
2692 // We can only recognize very limited forms of loop index expressions, in
2693 // particular, only affine AddRec's like {C1,+,C2}.
2694 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
2695 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
2696 !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
2697 !isa<SCEVConstant>(IdxExpr->getOperand(1)))
2698 return UnknownValue;
2700 unsigned MaxSteps = MaxBruteForceIterations;
2701 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
2702 ConstantInt *ItCst =
2703 ConstantInt::get(IdxExpr->getType(), IterationNum);
2704 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
2706 // Form the GEP offset.
2707 Indexes[VarIdxNum] = Val;
2709 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
2710 if (Result == 0) break; // Cannot compute!
2712 // Evaluate the condition for this iteration.
2713 Result = ConstantExpr::getICmp(predicate, Result, RHS);
2714 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
2715 if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
2717 errs() << "\n***\n*** Computed loop count " << *ItCst
2718 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
2721 ++NumArrayLenItCounts;
2722 return getConstant(ItCst); // Found terminating iteration!
2725 return UnknownValue;
2729 /// CanConstantFold - Return true if we can constant fold an instruction of the
2730 /// specified type, assuming that all operands were constants.
2731 static bool CanConstantFold(const Instruction *I) {
2732 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
2733 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
2736 if (const CallInst *CI = dyn_cast<CallInst>(I))
2737 if (const Function *F = CI->getCalledFunction())
2738 return canConstantFoldCallTo(F);
2742 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
2743 /// in the loop that V is derived from. We allow arbitrary operations along the
2744 /// way, but the operands of an operation must either be constants or a value
2745 /// derived from a constant PHI. If this expression does not fit with these
2746 /// constraints, return null.
2747 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
2748 // If this is not an instruction, or if this is an instruction outside of the
2749 // loop, it can't be derived from a loop PHI.
2750 Instruction *I = dyn_cast<Instruction>(V);
2751 if (I == 0 || !L->contains(I->getParent())) return 0;
2753 if (PHINode *PN = dyn_cast<PHINode>(I)) {
2754 if (L->getHeader() == I->getParent())
2757 // We don't currently keep track of the control flow needed to evaluate
2758 // PHIs, so we cannot handle PHIs inside of loops.
2762 // If we won't be able to constant fold this expression even if the operands
2763 // are constants, return early.
2764 if (!CanConstantFold(I)) return 0;
2766 // Otherwise, we can evaluate this instruction if all of its operands are
2767 // constant or derived from a PHI node themselves.
2769 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
2770 if (!(isa<Constant>(I->getOperand(Op)) ||
2771 isa<GlobalValue>(I->getOperand(Op)))) {
2772 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
2773 if (P == 0) return 0; // Not evolving from PHI
2777 return 0; // Evolving from multiple different PHIs.
2780 // This is a expression evolving from a constant PHI!
2784 /// EvaluateExpression - Given an expression that passes the
2785 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
2786 /// in the loop has the value PHIVal. If we can't fold this expression for some
2787 /// reason, return null.
2788 static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
2789 if (isa<PHINode>(V)) return PHIVal;
2790 if (Constant *C = dyn_cast<Constant>(V)) return C;
2791 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
2792 Instruction *I = cast<Instruction>(V);
2794 std::vector<Constant*> Operands;
2795 Operands.resize(I->getNumOperands());
2797 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2798 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal);
2799 if (Operands[i] == 0) return 0;
2802 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
2803 return ConstantFoldCompareInstOperands(CI->getPredicate(),
2804 &Operands[0], Operands.size());
2806 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
2807 &Operands[0], Operands.size());
2810 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
2811 /// in the header of its containing loop, we know the loop executes a
2812 /// constant number of times, and the PHI node is just a recurrence
2813 /// involving constants, fold it.
2814 Constant *ScalarEvolution::
2815 getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& BEs, const Loop *L){
2816 std::map<PHINode*, Constant*>::iterator I =
2817 ConstantEvolutionLoopExitValue.find(PN);
2818 if (I != ConstantEvolutionLoopExitValue.end())
2821 if (BEs.ugt(APInt(BEs.getBitWidth(),MaxBruteForceIterations)))
2822 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
2824 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
2826 // Since the loop is canonicalized, the PHI node must have two entries. One
2827 // entry must be a constant (coming in from outside of the loop), and the
2828 // second must be derived from the same PHI.
2829 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
2830 Constant *StartCST =
2831 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
2833 return RetVal = 0; // Must be a constant.
2835 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
2836 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
2838 return RetVal = 0; // Not derived from same PHI.
2840 // Execute the loop symbolically to determine the exit value.
2841 if (BEs.getActiveBits() >= 32)
2842 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
2844 unsigned NumIterations = BEs.getZExtValue(); // must be in range
2845 unsigned IterationNum = 0;
2846 for (Constant *PHIVal = StartCST; ; ++IterationNum) {
2847 if (IterationNum == NumIterations)
2848 return RetVal = PHIVal; // Got exit value!
2850 // Compute the value of the PHI node for the next iteration.
2851 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
2852 if (NextPHI == PHIVal)
2853 return RetVal = NextPHI; // Stopped evolving!
2855 return 0; // Couldn't evaluate!
2860 /// ComputeBackedgeTakenCountExhaustively - If the trip is known to execute a
2861 /// constant number of times (the condition evolves only from constants),
2862 /// try to evaluate a few iterations of the loop until we get the exit
2863 /// condition gets a value of ExitWhen (true or false). If we cannot
2864 /// evaluate the trip count of the loop, return UnknownValue.
2865 SCEVHandle ScalarEvolution::
2866 ComputeBackedgeTakenCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) {
2867 PHINode *PN = getConstantEvolvingPHI(Cond, L);
2868 if (PN == 0) return UnknownValue;
2870 // Since the loop is canonicalized, the PHI node must have two entries. One
2871 // entry must be a constant (coming in from outside of the loop), and the
2872 // second must be derived from the same PHI.
2873 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
2874 Constant *StartCST =
2875 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
2876 if (StartCST == 0) return UnknownValue; // Must be a constant.
2878 Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
2879 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
2880 if (PN2 != PN) return UnknownValue; // Not derived from same PHI.
2882 // Okay, we find a PHI node that defines the trip count of this loop. Execute
2883 // the loop symbolically to determine when the condition gets a value of
2885 unsigned IterationNum = 0;
2886 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
2887 for (Constant *PHIVal = StartCST;
2888 IterationNum != MaxIterations; ++IterationNum) {
2889 ConstantInt *CondVal =
2890 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal));
2892 // Couldn't symbolically evaluate.
2893 if (!CondVal) return UnknownValue;
2895 if (CondVal->getValue() == uint64_t(ExitWhen)) {
2896 ConstantEvolutionLoopExitValue[PN] = PHIVal;
2897 ++NumBruteForceTripCountsComputed;
2898 return getConstant(ConstantInt::get(Type::Int32Ty, IterationNum));
2901 // Compute the value of the PHI node for the next iteration.
2902 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal);
2903 if (NextPHI == 0 || NextPHI == PHIVal)
2904 return UnknownValue; // Couldn't evaluate or not making progress...
2908 // Too many iterations were needed to evaluate.
2909 return UnknownValue;
2912 /// getSCEVAtScope - Return a SCEV expression handle for the specified value
2913 /// at the specified scope in the program. The L value specifies a loop
2914 /// nest to evaluate the expression at, where null is the top-level or a
2915 /// specified loop is immediately inside of the loop.
2917 /// This method can be used to compute the exit value for a variable defined
2918 /// in a loop by querying what the value will hold in the parent loop.
2920 /// In the case that a relevant loop exit value cannot be computed, the
2921 /// original value V is returned.
2922 SCEVHandle ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
2923 // FIXME: this should be turned into a virtual method on SCEV!
2925 if (isa<SCEVConstant>(V)) return V;
2927 // If this instruction is evolved from a constant-evolving PHI, compute the
2928 // exit value from the loop without using SCEVs.
2929 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
2930 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
2931 const Loop *LI = (*this->LI)[I->getParent()];
2932 if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
2933 if (PHINode *PN = dyn_cast<PHINode>(I))
2934 if (PN->getParent() == LI->getHeader()) {
2935 // Okay, there is no closed form solution for the PHI node. Check
2936 // to see if the loop that contains it has a known backedge-taken
2937 // count. If so, we may be able to force computation of the exit
2939 SCEVHandle BackedgeTakenCount = getBackedgeTakenCount(LI);
2940 if (const SCEVConstant *BTCC =
2941 dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
2942 // Okay, we know how many times the containing loop executes. If
2943 // this is a constant evolving PHI node, get the final value at
2944 // the specified iteration number.
2945 Constant *RV = getConstantEvolutionLoopExitValue(PN,
2946 BTCC->getValue()->getValue(),
2948 if (RV) return getUnknown(RV);
2952 // Okay, this is an expression that we cannot symbolically evaluate
2953 // into a SCEV. Check to see if it's possible to symbolically evaluate
2954 // the arguments into constants, and if so, try to constant propagate the
2955 // result. This is particularly useful for computing loop exit values.
2956 if (CanConstantFold(I)) {
2957 // Check to see if we've folded this instruction at this loop before.
2958 std::map<const Loop *, Constant *> &Values = ValuesAtScopes[I];
2959 std::pair<std::map<const Loop *, Constant *>::iterator, bool> Pair =
2960 Values.insert(std::make_pair(L, static_cast<Constant *>(0)));
2962 return Pair.first->second ? &*getUnknown(Pair.first->second) : V;
2964 std::vector<Constant*> Operands;
2965 Operands.reserve(I->getNumOperands());
2966 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2967 Value *Op = I->getOperand(i);
2968 if (Constant *C = dyn_cast<Constant>(Op)) {
2969 Operands.push_back(C);
2971 // If any of the operands is non-constant and if they are
2972 // non-integer and non-pointer, don't even try to analyze them
2973 // with scev techniques.
2974 if (!isSCEVable(Op->getType()))
2977 SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L);
2978 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
2979 Constant *C = SC->getValue();
2980 if (C->getType() != Op->getType())
2981 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
2985 Operands.push_back(C);
2986 } else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
2987 if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
2988 if (C->getType() != Op->getType())
2990 ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
2994 Operands.push_back(C);
3004 if (const CmpInst *CI = dyn_cast<CmpInst>(I))
3005 C = ConstantFoldCompareInstOperands(CI->getPredicate(),
3006 &Operands[0], Operands.size());
3008 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
3009 &Operands[0], Operands.size());
3010 Pair.first->second = C;
3011 return getUnknown(C);
3015 // This is some other type of SCEVUnknown, just return it.
3019 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
3020 // Avoid performing the look-up in the common case where the specified
3021 // expression has no loop-variant portions.
3022 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
3023 SCEVHandle OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
3024 if (OpAtScope != Comm->getOperand(i)) {
3025 // Okay, at least one of these operands is loop variant but might be
3026 // foldable. Build a new instance of the folded commutative expression.
3027 std::vector<SCEVHandle> NewOps(Comm->op_begin(), Comm->op_begin()+i);
3028 NewOps.push_back(OpAtScope);
3030 for (++i; i != e; ++i) {
3031 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
3032 NewOps.push_back(OpAtScope);
3034 if (isa<SCEVAddExpr>(Comm))
3035 return getAddExpr(NewOps);
3036 if (isa<SCEVMulExpr>(Comm))
3037 return getMulExpr(NewOps);
3038 if (isa<SCEVSMaxExpr>(Comm))
3039 return getSMaxExpr(NewOps);
3040 if (isa<SCEVUMaxExpr>(Comm))
3041 return getUMaxExpr(NewOps);
3042 assert(0 && "Unknown commutative SCEV type!");
3045 // If we got here, all operands are loop invariant.
3049 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
3050 SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L);
3051 SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L);
3052 if (LHS == Div->getLHS() && RHS == Div->getRHS())
3053 return Div; // must be loop invariant
3054 return getUDivExpr(LHS, RHS);
3057 // If this is a loop recurrence for a loop that does not contain L, then we
3058 // are dealing with the final value computed by the loop.
3059 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
3060 if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
3061 // To evaluate this recurrence, we need to know how many times the AddRec
3062 // loop iterates. Compute this now.
3063 SCEVHandle BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
3064 if (BackedgeTakenCount == UnknownValue) return AddRec;
3066 // Then, evaluate the AddRec.
3067 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
3072 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
3073 SCEVHandle Op = getSCEVAtScope(Cast->getOperand(), L);
3074 if (Op == Cast->getOperand())
3075 return Cast; // must be loop invariant
3076 return getZeroExtendExpr(Op, Cast->getType());
3079 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
3080 SCEVHandle Op = getSCEVAtScope(Cast->getOperand(), L);
3081 if (Op == Cast->getOperand())
3082 return Cast; // must be loop invariant
3083 return getSignExtendExpr(Op, Cast->getType());
3086 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
3087 SCEVHandle Op = getSCEVAtScope(Cast->getOperand(), L);
3088 if (Op == Cast->getOperand())
3089 return Cast; // must be loop invariant
3090 return getTruncateExpr(Op, Cast->getType());
3093 assert(0 && "Unknown SCEV type!");
3097 /// getSCEVAtScope - This is a convenience function which does
3098 /// getSCEVAtScope(getSCEV(V), L).
3099 SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
3100 return getSCEVAtScope(getSCEV(V), L);
3103 /// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
3104 /// following equation:
3106 /// A * X = B (mod N)
3108 /// where N = 2^BW and BW is the common bit width of A and B. The signedness of
3109 /// A and B isn't important.
3111 /// If the equation does not have a solution, SCEVCouldNotCompute is returned.
3112 static SCEVHandle SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
3113 ScalarEvolution &SE) {
3114 uint32_t BW = A.getBitWidth();
3115 assert(BW == B.getBitWidth() && "Bit widths must be the same.");
3116 assert(A != 0 && "A must be non-zero.");
3120 // The gcd of A and N may have only one prime factor: 2. The number of
3121 // trailing zeros in A is its multiplicity
3122 uint32_t Mult2 = A.countTrailingZeros();
3125 // 2. Check if B is divisible by D.
3127 // B is divisible by D if and only if the multiplicity of prime factor 2 for B
3128 // is not less than multiplicity of this prime factor for D.
3129 if (B.countTrailingZeros() < Mult2)
3130 return SE.getCouldNotCompute();
3132 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
3135 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
3136 // bit width during computations.
3137 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
3138 APInt Mod(BW + 1, 0);
3139 Mod.set(BW - Mult2); // Mod = N / D
3140 APInt I = AD.multiplicativeInverse(Mod);
3142 // 4. Compute the minimum unsigned root of the equation:
3143 // I * (B / D) mod (N / D)
3144 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
3146 // The result is guaranteed to be less than 2^BW so we may truncate it to BW
3148 return SE.getConstant(Result.trunc(BW));
3151 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the
3152 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
3153 /// might be the same) or two SCEVCouldNotCompute objects.
3155 static std::pair<SCEVHandle,SCEVHandle>
3156 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
3157 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
3158 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
3159 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
3160 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
3162 // We currently can only solve this if the coefficients are constants.
3163 if (!LC || !MC || !NC) {
3164 const SCEV *CNC = SE.getCouldNotCompute();
3165 return std::make_pair(CNC, CNC);
3168 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
3169 const APInt &L = LC->getValue()->getValue();
3170 const APInt &M = MC->getValue()->getValue();
3171 const APInt &N = NC->getValue()->getValue();
3172 APInt Two(BitWidth, 2);
3173 APInt Four(BitWidth, 4);
3176 using namespace APIntOps;
3178 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
3179 // The B coefficient is M-N/2
3183 // The A coefficient is N/2
3184 APInt A(N.sdiv(Two));
3186 // Compute the B^2-4ac term.
3189 SqrtTerm -= Four * (A * C);
3191 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
3192 // integer value or else APInt::sqrt() will assert.
3193 APInt SqrtVal(SqrtTerm.sqrt());
3195 // Compute the two solutions for the quadratic formula.
3196 // The divisions must be performed as signed divisions.
3198 APInt TwoA( A << 1 );
3199 if (TwoA.isMinValue()) {
3200 const SCEV *CNC = SE.getCouldNotCompute();
3201 return std::make_pair(CNC, CNC);
3204 ConstantInt *Solution1 = ConstantInt::get((NegB + SqrtVal).sdiv(TwoA));
3205 ConstantInt *Solution2 = ConstantInt::get((NegB - SqrtVal).sdiv(TwoA));
3207 return std::make_pair(SE.getConstant(Solution1),
3208 SE.getConstant(Solution2));
3209 } // end APIntOps namespace
3212 /// HowFarToZero - Return the number of times a backedge comparing the specified
3213 /// value to zero will execute. If not computable, return UnknownValue.
3214 SCEVHandle ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
3215 // If the value is a constant
3216 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
3217 // If the value is already zero, the branch will execute zero times.
3218 if (C->getValue()->isZero()) return C;
3219 return UnknownValue; // Otherwise it will loop infinitely.
3222 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
3223 if (!AddRec || AddRec->getLoop() != L)
3224 return UnknownValue;
3226 if (AddRec->isAffine()) {
3227 // If this is an affine expression, the execution count of this branch is
3228 // the minimum unsigned root of the following equation:
3230 // Start + Step*N = 0 (mod 2^BW)
3234 // Step*N = -Start (mod 2^BW)
3236 // where BW is the common bit width of Start and Step.
3238 // Get the initial value for the loop.
3239 SCEVHandle Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
3240 SCEVHandle Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
3242 if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
3243 // For now we handle only constant steps.
3245 // First, handle unitary steps.
3246 if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
3247 return getNegativeSCEV(Start); // N = -Start (as unsigned)
3248 if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
3249 return Start; // N = Start (as unsigned)
3251 // Then, try to solve the above equation provided that Start is constant.
3252 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
3253 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
3254 -StartC->getValue()->getValue(),
3257 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
3258 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
3259 // the quadratic equation to solve it.
3260 std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec,
3262 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
3263 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
3266 errs() << "HFTZ: " << *V << " - sol#1: " << *R1
3267 << " sol#2: " << *R2 << "\n";
3269 // Pick the smallest positive root value.
3270 if (ConstantInt *CB =
3271 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
3272 R1->getValue(), R2->getValue()))) {
3273 if (CB->getZExtValue() == false)
3274 std::swap(R1, R2); // R1 is the minimum root now.
3276 // We can only use this value if the chrec ends up with an exact zero
3277 // value at this index. When solving for "X*X != 5", for example, we
3278 // should not accept a root of 2.
3279 SCEVHandle Val = AddRec->evaluateAtIteration(R1, *this);
3281 return R1; // We found a quadratic root!
3286 return UnknownValue;
3289 /// HowFarToNonZero - Return the number of times a backedge checking the
3290 /// specified value for nonzero will execute. If not computable, return
3292 SCEVHandle ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
3293 // Loops that look like: while (X == 0) are very strange indeed. We don't
3294 // handle them yet except for the trivial case. This could be expanded in the
3295 // future as needed.
3297 // If the value is a constant, check to see if it is known to be non-zero
3298 // already. If so, the backedge will execute zero times.
3299 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
3300 if (!C->getValue()->isNullValue())
3301 return getIntegerSCEV(0, C->getType());
3302 return UnknownValue; // Otherwise it will loop infinitely.
3305 // We could implement others, but I really doubt anyone writes loops like
3306 // this, and if they did, they would already be constant folded.
3307 return UnknownValue;
3310 /// getLoopPredecessor - If the given loop's header has exactly one unique
3311 /// predecessor outside the loop, return it. Otherwise return null.
3313 BasicBlock *ScalarEvolution::getLoopPredecessor(const Loop *L) {
3314 BasicBlock *Header = L->getHeader();
3315 BasicBlock *Pred = 0;
3316 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header);
3318 if (!L->contains(*PI)) {
3319 if (Pred && Pred != *PI) return 0; // Multiple predecessors.
3325 /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
3326 /// (which may not be an immediate predecessor) which has exactly one
3327 /// successor from which BB is reachable, or null if no such block is
3331 ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
3332 // If the block has a unique predecessor, then there is no path from the
3333 // predecessor to the block that does not go through the direct edge
3334 // from the predecessor to the block.
3335 if (BasicBlock *Pred = BB->getSinglePredecessor())
3338 // A loop's header is defined to be a block that dominates the loop.
3339 // If the header has a unique predecessor outside the loop, it must be
3340 // a block that has exactly one successor that can reach the loop.
3341 if (Loop *L = LI->getLoopFor(BB))
3342 return getLoopPredecessor(L);
3347 /// isLoopGuardedByCond - Test whether entry to the loop is protected by
3348 /// a conditional between LHS and RHS. This is used to help avoid max
3349 /// expressions in loop trip counts.
3350 bool ScalarEvolution::isLoopGuardedByCond(const Loop *L,
3351 ICmpInst::Predicate Pred,
3352 const SCEV *LHS, const SCEV *RHS) {
3353 // Interpret a null as meaning no loop, where there is obviously no guard
3354 // (interprocedural conditions notwithstanding).
3355 if (!L) return false;
3357 BasicBlock *Predecessor = getLoopPredecessor(L);
3358 BasicBlock *PredecessorDest = L->getHeader();
3360 // Starting at the loop predecessor, climb up the predecessor chain, as long
3361 // as there are predecessors that can be found that have unique successors
3362 // leading to the original header.
3364 PredecessorDest = Predecessor,
3365 Predecessor = getPredecessorWithUniqueSuccessorForBB(Predecessor)) {
3367 BranchInst *LoopEntryPredicate =
3368 dyn_cast<BranchInst>(Predecessor->getTerminator());
3369 if (!LoopEntryPredicate ||
3370 LoopEntryPredicate->isUnconditional())
3373 ICmpInst *ICI = dyn_cast<ICmpInst>(LoopEntryPredicate->getCondition());
3376 // Now that we found a conditional branch that dominates the loop, check to
3377 // see if it is the comparison we are looking for.
3378 Value *PreCondLHS = ICI->getOperand(0);
3379 Value *PreCondRHS = ICI->getOperand(1);
3380 ICmpInst::Predicate Cond;
3381 if (LoopEntryPredicate->getSuccessor(0) == PredecessorDest)
3382 Cond = ICI->getPredicate();
3384 Cond = ICI->getInversePredicate();
3387 ; // An exact match.
3388 else if (!ICmpInst::isTrueWhenEqual(Cond) && Pred == ICmpInst::ICMP_NE)
3389 ; // The actual condition is beyond sufficient.
3391 // Check a few special cases.
3393 case ICmpInst::ICMP_UGT:
3394 if (Pred == ICmpInst::ICMP_ULT) {
3395 std::swap(PreCondLHS, PreCondRHS);
3396 Cond = ICmpInst::ICMP_ULT;
3400 case ICmpInst::ICMP_SGT:
3401 if (Pred == ICmpInst::ICMP_SLT) {
3402 std::swap(PreCondLHS, PreCondRHS);
3403 Cond = ICmpInst::ICMP_SLT;
3407 case ICmpInst::ICMP_NE:
3408 // Expressions like (x >u 0) are often canonicalized to (x != 0),
3409 // so check for this case by checking if the NE is comparing against
3410 // a minimum or maximum constant.
3411 if (!ICmpInst::isTrueWhenEqual(Pred))
3412 if (ConstantInt *CI = dyn_cast<ConstantInt>(PreCondRHS)) {
3413 const APInt &A = CI->getValue();
3415 case ICmpInst::ICMP_SLT:
3416 if (A.isMaxSignedValue()) break;
3418 case ICmpInst::ICMP_SGT:
3419 if (A.isMinSignedValue()) break;
3421 case ICmpInst::ICMP_ULT:
3422 if (A.isMaxValue()) break;
3424 case ICmpInst::ICMP_UGT:
3425 if (A.isMinValue()) break;
3430 Cond = ICmpInst::ICMP_NE;
3431 // NE is symmetric but the original comparison may not be. Swap
3432 // the operands if necessary so that they match below.
3433 if (isa<SCEVConstant>(LHS))
3434 std::swap(PreCondLHS, PreCondRHS);
3439 // We weren't able to reconcile the condition.
3443 if (!PreCondLHS->getType()->isInteger()) continue;
3445 SCEVHandle PreCondLHSSCEV = getSCEV(PreCondLHS);
3446 SCEVHandle PreCondRHSSCEV = getSCEV(PreCondRHS);
3447 if ((LHS == PreCondLHSSCEV && RHS == PreCondRHSSCEV) ||
3448 (LHS == getNotSCEV(PreCondRHSSCEV) &&
3449 RHS == getNotSCEV(PreCondLHSSCEV)))
3456 /// HowManyLessThans - Return the number of times a backedge containing the
3457 /// specified less-than comparison will execute. If not computable, return
3459 ScalarEvolution::BackedgeTakenInfo ScalarEvolution::
3460 HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
3461 const Loop *L, bool isSigned) {
3462 // Only handle: "ADDREC < LoopInvariant".
3463 if (!RHS->isLoopInvariant(L)) return UnknownValue;
3465 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
3466 if (!AddRec || AddRec->getLoop() != L)
3467 return UnknownValue;
3469 if (AddRec->isAffine()) {
3470 // FORNOW: We only support unit strides.
3471 unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
3472 SCEVHandle Step = AddRec->getStepRecurrence(*this);
3473 SCEVHandle NegOne = getIntegerSCEV(-1, AddRec->getType());
3475 // TODO: handle non-constant strides.
3476 const SCEVConstant *CStep = dyn_cast<SCEVConstant>(Step);
3477 if (!CStep || CStep->isZero())
3478 return UnknownValue;
3479 if (CStep->isOne()) {
3480 // With unit stride, the iteration never steps past the limit value.
3481 } else if (CStep->getValue()->getValue().isStrictlyPositive()) {
3482 if (const SCEVConstant *CLimit = dyn_cast<SCEVConstant>(RHS)) {
3483 // Test whether a positive iteration iteration can step past the limit
3484 // value and past the maximum value for its type in a single step.
3486 APInt Max = APInt::getSignedMaxValue(BitWidth);
3487 if ((Max - CStep->getValue()->getValue())
3488 .slt(CLimit->getValue()->getValue()))
3489 return UnknownValue;
3491 APInt Max = APInt::getMaxValue(BitWidth);
3492 if ((Max - CStep->getValue()->getValue())
3493 .ult(CLimit->getValue()->getValue()))
3494 return UnknownValue;
3497 // TODO: handle non-constant limit values below.
3498 return UnknownValue;
3500 // TODO: handle negative strides below.
3501 return UnknownValue;
3503 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
3504 // m. So, we count the number of iterations in which {n,+,s} < m is true.
3505 // Note that we cannot simply return max(m-n,0)/s because it's not safe to
3506 // treat m-n as signed nor unsigned due to overflow possibility.
3508 // First, we get the value of the LHS in the first iteration: n
3509 SCEVHandle Start = AddRec->getOperand(0);
3511 // Determine the minimum constant start value.
3512 SCEVHandle MinStart = isa<SCEVConstant>(Start) ? Start :
3513 getConstant(isSigned ? APInt::getSignedMinValue(BitWidth) :
3514 APInt::getMinValue(BitWidth));
3516 // If we know that the condition is true in order to enter the loop,
3517 // then we know that it will run exactly (m-n)/s times. Otherwise, we
3518 // only know that it will execute (max(m,n)-n)/s times. In both cases,
3519 // the division must round up.
3520 SCEVHandle End = RHS;
3521 if (!isLoopGuardedByCond(L,
3522 isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
3523 getMinusSCEV(Start, Step), RHS))
3524 End = isSigned ? getSMaxExpr(RHS, Start)
3525 : getUMaxExpr(RHS, Start);
3527 // Determine the maximum constant end value.
3528 SCEVHandle MaxEnd = isa<SCEVConstant>(End) ? End :
3529 getConstant(isSigned ? APInt::getSignedMaxValue(BitWidth) :
3530 APInt::getMaxValue(BitWidth));
3532 // Finally, we subtract these two values and divide, rounding up, to get
3533 // the number of times the backedge is executed.
3534 SCEVHandle BECount = getUDivExpr(getAddExpr(getMinusSCEV(End, Start),
3535 getAddExpr(Step, NegOne)),
3538 // The maximum backedge count is similar, except using the minimum start
3539 // value and the maximum end value.
3540 SCEVHandle MaxBECount = getUDivExpr(getAddExpr(getMinusSCEV(MaxEnd,
3542 getAddExpr(Step, NegOne)),
3545 return BackedgeTakenInfo(BECount, MaxBECount);
3548 return UnknownValue;
3551 /// getNumIterationsInRange - Return the number of iterations of this loop that
3552 /// produce values in the specified constant range. Another way of looking at
3553 /// this is that it returns the first iteration number where the value is not in
3554 /// the condition, thus computing the exit count. If the iteration count can't
3555 /// be computed, an instance of SCEVCouldNotCompute is returned.
3556 SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
3557 ScalarEvolution &SE) const {
3558 if (Range.isFullSet()) // Infinite loop.
3559 return SE.getCouldNotCompute();
3561 // If the start is a non-zero constant, shift the range to simplify things.
3562 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
3563 if (!SC->getValue()->isZero()) {
3564 std::vector<SCEVHandle> Operands(op_begin(), op_end());
3565 Operands[0] = SE.getIntegerSCEV(0, SC->getType());
3566 SCEVHandle Shifted = SE.getAddRecExpr(Operands, getLoop());
3567 if (const SCEVAddRecExpr *ShiftedAddRec =
3568 dyn_cast<SCEVAddRecExpr>(Shifted))
3569 return ShiftedAddRec->getNumIterationsInRange(
3570 Range.subtract(SC->getValue()->getValue()), SE);
3571 // This is strange and shouldn't happen.
3572 return SE.getCouldNotCompute();
3575 // The only time we can solve this is when we have all constant indices.
3576 // Otherwise, we cannot determine the overflow conditions.
3577 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
3578 if (!isa<SCEVConstant>(getOperand(i)))
3579 return SE.getCouldNotCompute();
3582 // Okay at this point we know that all elements of the chrec are constants and
3583 // that the start element is zero.
3585 // First check to see if the range contains zero. If not, the first
3587 unsigned BitWidth = SE.getTypeSizeInBits(getType());
3588 if (!Range.contains(APInt(BitWidth, 0)))
3589 return SE.getConstant(ConstantInt::get(getType(),0));
3592 // If this is an affine expression then we have this situation:
3593 // Solve {0,+,A} in Range === Ax in Range
3595 // We know that zero is in the range. If A is positive then we know that
3596 // the upper value of the range must be the first possible exit value.
3597 // If A is negative then the lower of the range is the last possible loop
3598 // value. Also note that we already checked for a full range.
3599 APInt One(BitWidth,1);
3600 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
3601 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
3603 // The exit value should be (End+A)/A.
3604 APInt ExitVal = (End + A).udiv(A);
3605 ConstantInt *ExitValue = ConstantInt::get(ExitVal);
3607 // Evaluate at the exit value. If we really did fall out of the valid
3608 // range, then we computed our trip count, otherwise wrap around or other
3609 // things must have happened.
3610 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
3611 if (Range.contains(Val->getValue()))
3612 return SE.getCouldNotCompute(); // Something strange happened
3614 // Ensure that the previous value is in the range. This is a sanity check.
3615 assert(Range.contains(
3616 EvaluateConstantChrecAtConstant(this,
3617 ConstantInt::get(ExitVal - One), SE)->getValue()) &&
3618 "Linear scev computation is off in a bad way!");
3619 return SE.getConstant(ExitValue);
3620 } else if (isQuadratic()) {
3621 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
3622 // quadratic equation to solve it. To do this, we must frame our problem in
3623 // terms of figuring out when zero is crossed, instead of when
3624 // Range.getUpper() is crossed.
3625 std::vector<SCEVHandle> NewOps(op_begin(), op_end());
3626 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
3627 SCEVHandle NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
3629 // Next, solve the constructed addrec
3630 std::pair<SCEVHandle,SCEVHandle> Roots =
3631 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
3632 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
3633 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
3635 // Pick the smallest positive root value.
3636 if (ConstantInt *CB =
3637 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
3638 R1->getValue(), R2->getValue()))) {
3639 if (CB->getZExtValue() == false)
3640 std::swap(R1, R2); // R1 is the minimum root now.
3642 // Make sure the root is not off by one. The returned iteration should
3643 // not be in the range, but the previous one should be. When solving
3644 // for "X*X < 5", for example, we should not return a root of 2.
3645 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
3648 if (Range.contains(R1Val->getValue())) {
3649 // The next iteration must be out of the range...
3650 ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()+1);
3652 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
3653 if (!Range.contains(R1Val->getValue()))
3654 return SE.getConstant(NextVal);
3655 return SE.getCouldNotCompute(); // Something strange happened
3658 // If R1 was not in the range, then it is a good return value. Make
3659 // sure that R1-1 WAS in the range though, just in case.
3660 ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()-1);
3661 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
3662 if (Range.contains(R1Val->getValue()))
3664 return SE.getCouldNotCompute(); // Something strange happened
3669 return SE.getCouldNotCompute();
3674 //===----------------------------------------------------------------------===//
3675 // SCEVCallbackVH Class Implementation
3676 //===----------------------------------------------------------------------===//
3678 void ScalarEvolution::SCEVCallbackVH::deleted() {
3679 assert(SE && "SCEVCallbackVH called with a non-null ScalarEvolution!");
3680 if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
3681 SE->ConstantEvolutionLoopExitValue.erase(PN);
3682 if (Instruction *I = dyn_cast<Instruction>(getValPtr()))
3683 SE->ValuesAtScopes.erase(I);
3684 SE->Scalars.erase(getValPtr());
3685 // this now dangles!
3688 void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *) {
3689 assert(SE && "SCEVCallbackVH called with a non-null ScalarEvolution!");
3691 // Forget all the expressions associated with users of the old value,
3692 // so that future queries will recompute the expressions using the new
3694 SmallVector<User *, 16> Worklist;
3695 Value *Old = getValPtr();
3696 bool DeleteOld = false;
3697 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
3699 Worklist.push_back(*UI);
3700 while (!Worklist.empty()) {
3701 User *U = Worklist.pop_back_val();
3702 // Deleting the Old value will cause this to dangle. Postpone
3703 // that until everything else is done.
3708 if (PHINode *PN = dyn_cast<PHINode>(U))
3709 SE->ConstantEvolutionLoopExitValue.erase(PN);
3710 if (Instruction *I = dyn_cast<Instruction>(U))
3711 SE->ValuesAtScopes.erase(I);
3712 if (SE->Scalars.erase(U))
3713 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
3715 Worklist.push_back(*UI);
3718 if (PHINode *PN = dyn_cast<PHINode>(Old))
3719 SE->ConstantEvolutionLoopExitValue.erase(PN);
3720 if (Instruction *I = dyn_cast<Instruction>(Old))
3721 SE->ValuesAtScopes.erase(I);
3722 SE->Scalars.erase(Old);
3723 // this now dangles!
3728 ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
3729 : CallbackVH(V), SE(se) {}
3731 //===----------------------------------------------------------------------===//
3732 // ScalarEvolution Class Implementation
3733 //===----------------------------------------------------------------------===//
3735 ScalarEvolution::ScalarEvolution()
3736 : FunctionPass(&ID), UnknownValue(new SCEVCouldNotCompute()) {
3739 bool ScalarEvolution::runOnFunction(Function &F) {
3741 LI = &getAnalysis<LoopInfo>();
3742 TD = getAnalysisIfAvailable<TargetData>();
3746 void ScalarEvolution::releaseMemory() {
3748 BackedgeTakenCounts.clear();
3749 ConstantEvolutionLoopExitValue.clear();
3750 ValuesAtScopes.clear();
3753 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
3754 AU.setPreservesAll();
3755 AU.addRequiredTransitive<LoopInfo>();
3758 bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
3759 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
3762 static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
3764 // Print all inner loops first
3765 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
3766 PrintLoopInfo(OS, SE, *I);
3768 OS << "Loop " << L->getHeader()->getName() << ": ";
3770 SmallVector<BasicBlock*, 8> ExitBlocks;
3771 L->getExitBlocks(ExitBlocks);
3772 if (ExitBlocks.size() != 1)
3773 OS << "<multiple exits> ";
3775 if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
3776 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
3778 OS << "Unpredictable backedge-taken count. ";
3784 void ScalarEvolution::print(raw_ostream &OS, const Module* ) const {
3785 // ScalarEvolution's implementaiton of the print method is to print
3786 // out SCEV values of all instructions that are interesting. Doing
3787 // this potentially causes it to create new SCEV objects though,
3788 // which technically conflicts with the const qualifier. This isn't
3789 // observable from outside the class though (the hasSCEV function
3790 // notwithstanding), so casting away the const isn't dangerous.
3791 ScalarEvolution &SE = *const_cast<ScalarEvolution*>(this);
3793 OS << "Classifying expressions for: " << F->getName() << "\n";
3794 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
3795 if (isSCEVable(I->getType())) {
3798 SCEVHandle SV = SE.getSCEV(&*I);
3802 if (const Loop *L = LI->getLoopFor((*I).getParent())) {
3804 SCEVHandle ExitValue = SE.getSCEVAtScope(&*I, L->getParentLoop());
3805 if (!ExitValue->isLoopInvariant(L)) {
3806 OS << "<<Unknown>>";
3815 OS << "Determining loop execution counts for: " << F->getName() << "\n";
3816 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
3817 PrintLoopInfo(OS, &SE, *I);
3820 void ScalarEvolution::print(std::ostream &o, const Module *M) const {
3821 raw_os_ostream OS(o);