blob: 8e6ee5363ae4974115da9e02ae45dcb5a1a9d422 [file] [log] [blame]
Chris Lattner53e677a2004-04-02 20:23:17 +00001//===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
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.
13//
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.
20//
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.
26//
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.
31//
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.
35//
36// Orthogonal to the analysis of code above, this file also implements the
37// ScalarEvolutionRewriter class, which is used to emit code that represents the
38// various recurrences present in a loop, in canonical forms.
39//
40// TODO: We should use these routines and value representations to implement
41// dependence analysis!
42//
43//===----------------------------------------------------------------------===//
44//
45// There are several good references for the techniques used in this analysis.
46//
47// Chains of recurrences -- a method to expedite the evaluation
48// of closed-form functions
49// Olaf Bachmann, Paul S. Wang, Eugene V. Zima
50//
51// On computational properties of chains of recurrences
52// Eugene V. Zima
53//
54// Symbolic Evaluation of Chains of Recurrences for Loop Optimization
55// Robert A. van Engelen
56//
57// Efficient Symbolic Analysis for Optimizing Compilers
58// Robert A. van Engelen
59//
60// Using the chains of recurrences algebra for data dependence testing and
61// induction variable substitution
62// MS Thesis, Johnie Birch
63//
64//===----------------------------------------------------------------------===//
65
66#include "llvm/Analysis/ScalarEvolution.h"
67#include "llvm/Constants.h"
68#include "llvm/DerivedTypes.h"
69#include "llvm/Instructions.h"
70#include "llvm/Type.h"
71#include "llvm/Value.h"
72#include "llvm/Analysis/LoopInfo.h"
73#include "llvm/Assembly/Writer.h"
74#include "llvm/Transforms/Scalar.h"
75#include "llvm/Support/CFG.h"
76#include "llvm/Support/ConstantRange.h"
77#include "llvm/Support/InstIterator.h"
78#include "Support/Statistic.h"
79using namespace llvm;
80
81namespace {
82 RegisterAnalysis<ScalarEvolution>
83 R("scalar-evolution", "Scalar Evolution Analysis Printer");
84
85 Statistic<>
86 NumBruteForceEvaluations("scalar-evolution",
87 "Number of brute force evaluations needed to calculate high-order polynomial exit values");
88 Statistic<>
89 NumTripCountsComputed("scalar-evolution",
90 "Number of loops with predictable loop counts");
91 Statistic<>
92 NumTripCountsNotComputed("scalar-evolution",
93 "Number of loops without predictable loop counts");
94}
95
96//===----------------------------------------------------------------------===//
97// SCEV class definitions
98//===----------------------------------------------------------------------===//
99
100//===----------------------------------------------------------------------===//
101// Implementation of the SCEV class.
102//
103namespace {
104 enum SCEVTypes {
105 // These should be ordered in terms of increasing complexity to make the
106 // folders simpler.
107 scConstant, scTruncate, scZeroExtend, scAddExpr, scMulExpr, scUDivExpr,
108 scAddRecExpr, scUnknown, scCouldNotCompute
109 };
110
111 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
112 /// than the complexity of the RHS. If the SCEVs have identical complexity,
113 /// order them by their addresses. This comparator is used to canonicalize
114 /// expressions.
115 struct SCEVComplexityCompare {
116 bool operator()(SCEV *LHS, SCEV *RHS) {
117 if (LHS->getSCEVType() < RHS->getSCEVType())
118 return true;
119 if (LHS->getSCEVType() == RHS->getSCEVType())
120 return LHS < RHS;
121 return false;
122 }
123 };
124}
125
126SCEV::~SCEV() {}
127void SCEV::dump() const {
128 print(std::cerr);
129}
130
131/// getValueRange - Return the tightest constant bounds that this value is
132/// known to have. This method is only valid on integer SCEV objects.
133ConstantRange SCEV::getValueRange() const {
134 const Type *Ty = getType();
135 assert(Ty->isInteger() && "Can't get range for a non-integer SCEV!");
136 Ty = Ty->getUnsignedVersion();
137 // Default to a full range if no better information is available.
138 return ConstantRange(getType());
139}
140
141
142SCEVCouldNotCompute::SCEVCouldNotCompute() : SCEV(scCouldNotCompute) {}
143
144bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
145 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
146}
147
148const Type *SCEVCouldNotCompute::getType() const {
149 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
150}
151
152bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
153 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
154 return false;
155}
156
157Value *SCEVCouldNotCompute::expandCodeFor(ScalarEvolutionRewriter &SER,
158 Instruction *InsertPt) {
159 assert(0 && "Attempt to use a SCEVCouldNotCompute object!");
160 return 0;
161}
162
163
164void SCEVCouldNotCompute::print(std::ostream &OS) const {
165 OS << "***COULDNOTCOMPUTE***";
166}
167
168bool SCEVCouldNotCompute::classof(const SCEV *S) {
169 return S->getSCEVType() == scCouldNotCompute;
170}
171
172
173//===----------------------------------------------------------------------===//
174// SCEVConstant - This class represents a constant integer value.
175//
176namespace {
177 class SCEVConstant;
178 // SCEVConstants - Only allow the creation of one SCEVConstant for any
179 // particular value. Don't use a SCEVHandle here, or else the object will
180 // never be deleted!
181 std::map<ConstantInt*, SCEVConstant*> SCEVConstants;
182
183 class SCEVConstant : public SCEV {
184 ConstantInt *V;
185 SCEVConstant(ConstantInt *v) : SCEV(scConstant), V(v) {}
186
187 virtual ~SCEVConstant() {
188 SCEVConstants.erase(V);
189 }
190 public:
191 /// get method - This just gets and returns a new SCEVConstant object.
192 ///
193 static SCEVHandle get(ConstantInt *V) {
194 // Make sure that SCEVConstant instances are all unsigned.
195 if (V->getType()->isSigned()) {
196 const Type *NewTy = V->getType()->getUnsignedVersion();
197 V = cast<ConstantUInt>(ConstantExpr::getCast(V, NewTy));
198 }
199
200 SCEVConstant *&R = SCEVConstants[V];
201 if (R == 0) R = new SCEVConstant(V);
202 return R;
203 }
204
205 ConstantInt *getValue() const { return V; }
206
207 /// getValueRange - Return the tightest constant bounds that this value is
208 /// known to have. This method is only valid on integer SCEV objects.
209 virtual ConstantRange getValueRange() const {
210 return ConstantRange(V);
211 }
212
213 virtual bool isLoopInvariant(const Loop *L) const {
214 return true;
215 }
216
217 virtual bool hasComputableLoopEvolution(const Loop *L) const {
218 return false; // Not loop variant
219 }
220
221 virtual const Type *getType() const { return V->getType(); }
222
223 Value *expandCodeFor(ScalarEvolutionRewriter &SER,
224 Instruction *InsertPt) {
225 return getValue();
226 }
227
228 virtual void print(std::ostream &OS) const {
229 WriteAsOperand(OS, V, false);
230 }
231
232 /// Methods for support type inquiry through isa, cast, and dyn_cast:
233 static inline bool classof(const SCEVConstant *S) { return true; }
234 static inline bool classof(const SCEV *S) {
235 return S->getSCEVType() == scConstant;
236 }
237 };
238}
239
240
241//===----------------------------------------------------------------------===//
242// SCEVTruncateExpr - This class represents a truncation of an integer value to
243// a smaller integer value.
244//
245namespace {
246 class SCEVTruncateExpr;
247 // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any
248 // particular input. Don't use a SCEVHandle here, or else the object will
249 // never be deleted!
250 std::map<std::pair<SCEV*, const Type*>, SCEVTruncateExpr*> SCEVTruncates;
251
252 class SCEVTruncateExpr : public SCEV {
253 SCEVHandle Op;
254 const Type *Ty;
255 SCEVTruncateExpr(const SCEVHandle &op, const Type *ty)
256 : SCEV(scTruncate), Op(op), Ty(ty) {
257 assert(Op->getType()->isInteger() && Ty->isInteger() &&
258 Ty->isUnsigned() &&
259 "Cannot truncate non-integer value!");
260 assert(Op->getType()->getPrimitiveSize() > Ty->getPrimitiveSize() &&
261 "This is not a truncating conversion!");
262 }
263
264 virtual ~SCEVTruncateExpr() {
265 SCEVTruncates.erase(std::make_pair(Op, Ty));
266 }
267 public:
268 /// get method - This just gets and returns a new SCEVTruncate object
269 ///
270 static SCEVHandle get(const SCEVHandle &Op, const Type *Ty);
271
272 const SCEVHandle &getOperand() const { return Op; }
273 virtual const Type *getType() const { return Ty; }
274
275 virtual bool isLoopInvariant(const Loop *L) const {
276 return Op->isLoopInvariant(L);
277 }
278
279 virtual bool hasComputableLoopEvolution(const Loop *L) const {
280 return Op->hasComputableLoopEvolution(L);
281 }
282
283 /// getValueRange - Return the tightest constant bounds that this value is
284 /// known to have. This method is only valid on integer SCEV objects.
285 virtual ConstantRange getValueRange() const {
286 return getOperand()->getValueRange().truncate(getType());
287 }
288
289 Value *expandCodeFor(ScalarEvolutionRewriter &SER,
290 Instruction *InsertPt);
291
292 virtual void print(std::ostream &OS) const {
293 OS << "(truncate " << *Op << " to " << *Ty << ")";
294 }
295
296 /// Methods for support type inquiry through isa, cast, and dyn_cast:
297 static inline bool classof(const SCEVTruncateExpr *S) { return true; }
298 static inline bool classof(const SCEV *S) {
299 return S->getSCEVType() == scTruncate;
300 }
301 };
302}
303
304
305//===----------------------------------------------------------------------===//
306// SCEVZeroExtendExpr - This class represents a zero extension of a small
307// integer value to a larger integer value.
308//
309namespace {
310 class SCEVZeroExtendExpr;
311 // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any
312 // particular input. Don't use a SCEVHandle here, or else the object will
313 // never be deleted!
314 std::map<std::pair<SCEV*, const Type*>, SCEVZeroExtendExpr*> SCEVZeroExtends;
315
316 class SCEVZeroExtendExpr : public SCEV {
317 SCEVHandle Op;
318 const Type *Ty;
319 SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty)
320 : SCEV(scTruncate), Op(Op), Ty(ty) {
321 assert(Op->getType()->isInteger() && Ty->isInteger() &&
322 Ty->isUnsigned() &&
323 "Cannot zero extend non-integer value!");
324 assert(Op->getType()->getPrimitiveSize() < Ty->getPrimitiveSize() &&
325 "This is not an extending conversion!");
326 }
327
328 virtual ~SCEVZeroExtendExpr() {
329 SCEVZeroExtends.erase(std::make_pair(Op, Ty));
330 }
331 public:
332 /// get method - This just gets and returns a new SCEVZeroExtend object
333 ///
334 static SCEVHandle get(const SCEVHandle &Op, const Type *Ty);
335
336 const SCEVHandle &getOperand() const { return Op; }
337 virtual const Type *getType() const { return Ty; }
338
339 virtual bool isLoopInvariant(const Loop *L) const {
340 return Op->isLoopInvariant(L);
341 }
342
343 virtual bool hasComputableLoopEvolution(const Loop *L) const {
344 return Op->hasComputableLoopEvolution(L);
345 }
346
347 /// getValueRange - Return the tightest constant bounds that this value is
348 /// known to have. This method is only valid on integer SCEV objects.
349 virtual ConstantRange getValueRange() const {
350 return getOperand()->getValueRange().zeroExtend(getType());
351 }
352
353 Value *expandCodeFor(ScalarEvolutionRewriter &SER,
354 Instruction *InsertPt);
355
356 virtual void print(std::ostream &OS) const {
357 OS << "(zeroextend " << *Op << " to " << *Ty << ")";
358 }
359
360 /// Methods for support type inquiry through isa, cast, and dyn_cast:
361 static inline bool classof(const SCEVZeroExtendExpr *S) { return true; }
362 static inline bool classof(const SCEV *S) {
363 return S->getSCEVType() == scZeroExtend;
364 }
365 };
366}
367
368
369//===----------------------------------------------------------------------===//
370// SCEVCommutativeExpr - This node is the base class for n'ary commutative
371// operators.
372
373namespace {
374 class SCEVCommutativeExpr;
375 // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any
376 // particular input. Don't use a SCEVHandle here, or else the object will
377 // never be deleted!
378 std::map<std::pair<unsigned, std::vector<SCEV*> >,
379 SCEVCommutativeExpr*> SCEVCommExprs;
380
381 class SCEVCommutativeExpr : public SCEV {
382 std::vector<SCEVHandle> Operands;
383
384 protected:
385 SCEVCommutativeExpr(enum SCEVTypes T, const std::vector<SCEVHandle> &ops)
386 : SCEV(T) {
387 Operands.reserve(ops.size());
388 Operands.insert(Operands.end(), ops.begin(), ops.end());
389 }
390
391 ~SCEVCommutativeExpr() {
392 SCEVCommExprs.erase(std::make_pair(getSCEVType(),
393 std::vector<SCEV*>(Operands.begin(),
394 Operands.end())));
395 }
396
397 public:
398 unsigned getNumOperands() const { return Operands.size(); }
399 const SCEVHandle &getOperand(unsigned i) const {
400 assert(i < Operands.size() && "Operand index out of range!");
401 return Operands[i];
402 }
403
404 const std::vector<SCEVHandle> &getOperands() const { return Operands; }
405 typedef std::vector<SCEVHandle>::const_iterator op_iterator;
406 op_iterator op_begin() const { return Operands.begin(); }
407 op_iterator op_end() const { return Operands.end(); }
408
409
410 virtual bool isLoopInvariant(const Loop *L) const {
411 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
412 if (!getOperand(i)->isLoopInvariant(L)) return false;
413 return true;
414 }
415
416 virtual bool hasComputableLoopEvolution(const Loop *L) const {
417 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
418 if (getOperand(i)->hasComputableLoopEvolution(L)) return true;
419 return false;
420 }
421
422 virtual const Type *getType() const { return getOperand(0)->getType(); }
423
424 virtual const char *getOperationStr() const = 0;
425
426 virtual void print(std::ostream &OS) const {
427 assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
428 const char *OpStr = getOperationStr();
429 OS << "(" << *Operands[0];
430 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
431 OS << OpStr << *Operands[i];
432 OS << ")";
433 }
434
435 /// Methods for support type inquiry through isa, cast, and dyn_cast:
436 static inline bool classof(const SCEVCommutativeExpr *S) { return true; }
437 static inline bool classof(const SCEV *S) {
438 return S->getSCEVType() == scAddExpr ||
439 S->getSCEVType() == scMulExpr;
440 }
441 };
442}
443
444//===----------------------------------------------------------------------===//
445// SCEVAddExpr - This node represents an addition of some number of SCEV's.
446//
447namespace {
448 class SCEVAddExpr : public SCEVCommutativeExpr {
449 SCEVAddExpr(const std::vector<SCEVHandle> &ops)
450 : SCEVCommutativeExpr(scAddExpr, ops) {
451 }
452
453 public:
454 static SCEVHandle get(std::vector<SCEVHandle> &Ops);
455
456 static SCEVHandle get(const SCEVHandle &LHS, const SCEVHandle &RHS) {
457 std::vector<SCEVHandle> Ops;
458 Ops.push_back(LHS);
459 Ops.push_back(RHS);
460 return get(Ops);
461 }
462
463 static SCEVHandle get(const SCEVHandle &Op0, const SCEVHandle &Op1,
464 const SCEVHandle &Op2) {
465 std::vector<SCEVHandle> Ops;
466 Ops.push_back(Op0);
467 Ops.push_back(Op1);
468 Ops.push_back(Op2);
469 return get(Ops);
470 }
471
472 virtual const char *getOperationStr() const { return " + "; }
473
474 Value *expandCodeFor(ScalarEvolutionRewriter &SER,
475 Instruction *InsertPt);
476
477 /// Methods for support type inquiry through isa, cast, and dyn_cast:
478 static inline bool classof(const SCEVAddExpr *S) { return true; }
479 static inline bool classof(const SCEV *S) {
480 return S->getSCEVType() == scAddExpr;
481 }
482 };
483}
484
485//===----------------------------------------------------------------------===//
486// SCEVMulExpr - This node represents multiplication of some number of SCEV's.
487//
488namespace {
489 class SCEVMulExpr : public SCEVCommutativeExpr {
490 SCEVMulExpr(const std::vector<SCEVHandle> &ops)
491 : SCEVCommutativeExpr(scMulExpr, ops) {
492 }
493
494 public:
495 static SCEVHandle get(std::vector<SCEVHandle> &Ops);
496
497 static SCEVHandle get(const SCEVHandle &LHS, const SCEVHandle &RHS) {
498 std::vector<SCEVHandle> Ops;
499 Ops.push_back(LHS);
500 Ops.push_back(RHS);
501 return get(Ops);
502 }
503
504 virtual const char *getOperationStr() const { return " * "; }
505
506 Value *expandCodeFor(ScalarEvolutionRewriter &SER,
507 Instruction *InsertPt);
508
509 /// Methods for support type inquiry through isa, cast, and dyn_cast:
510 static inline bool classof(const SCEVMulExpr *S) { return true; }
511 static inline bool classof(const SCEV *S) {
512 return S->getSCEVType() == scMulExpr;
513 }
514 };
515}
516
517
518//===----------------------------------------------------------------------===//
519// SCEVUDivExpr - This class represents a binary unsigned division operation.
520//
521namespace {
522 class SCEVUDivExpr;
523 // SCEVUDivs - Only allow the creation of one SCEVUDivExpr for any particular
524 // input. Don't use a SCEVHandle here, or else the object will never be
525 // deleted!
526 std::map<std::pair<SCEV*, SCEV*>, SCEVUDivExpr*> SCEVUDivs;
527
528 class SCEVUDivExpr : public SCEV {
529 SCEVHandle LHS, RHS;
530 SCEVUDivExpr(const SCEVHandle &lhs, const SCEVHandle &rhs)
531 : SCEV(scUDivExpr), LHS(lhs), RHS(rhs) {}
532
533 virtual ~SCEVUDivExpr() {
534 SCEVUDivs.erase(std::make_pair(LHS, RHS));
535 }
536 public:
537 /// get method - This just gets and returns a new SCEVUDiv object.
538 ///
539 static SCEVHandle get(const SCEVHandle &LHS, const SCEVHandle &RHS);
540
541 const SCEVHandle &getLHS() const { return LHS; }
542 const SCEVHandle &getRHS() const { return RHS; }
543
544 virtual bool isLoopInvariant(const Loop *L) const {
545 return LHS->isLoopInvariant(L) && RHS->isLoopInvariant(L);
546 }
547
548 virtual bool hasComputableLoopEvolution(const Loop *L) const {
549 return LHS->hasComputableLoopEvolution(L) &&
550 RHS->hasComputableLoopEvolution(L);
551 }
552
553 virtual const Type *getType() const {
554 const Type *Ty = LHS->getType();
555 if (Ty->isSigned()) Ty = Ty->getUnsignedVersion();
556 return Ty;
557 }
558
559 Value *expandCodeFor(ScalarEvolutionRewriter &SER,
560 Instruction *InsertPt);
561
562 virtual void print(std::ostream &OS) const {
563 OS << "(" << *LHS << " /u " << *RHS << ")";
564 }
565
566 /// Methods for support type inquiry through isa, cast, and dyn_cast:
567 static inline bool classof(const SCEVUDivExpr *S) { return true; }
568 static inline bool classof(const SCEV *S) {
569 return S->getSCEVType() == scUDivExpr;
570 }
571 };
572}
573
574
575//===----------------------------------------------------------------------===//
576
577// SCEVAddRecExpr - This node represents a polynomial recurrence on the trip
578// count of the specified loop.
579//
580// All operands of an AddRec are required to be loop invariant.
581//
582namespace {
583 class SCEVAddRecExpr;
584 // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
585 // particular input. Don't use a SCEVHandle here, or else the object will
586 // never be deleted!
587 std::map<std::pair<const Loop *, std::vector<SCEV*> >,
588 SCEVAddRecExpr*> SCEVAddRecExprs;
589
590 class SCEVAddRecExpr : public SCEV {
591 std::vector<SCEVHandle> Operands;
592 const Loop *L;
593
594 SCEVAddRecExpr(const std::vector<SCEVHandle> &ops, const Loop *l)
595 : SCEV(scAddRecExpr), Operands(ops), L(l) {
596 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
597 assert(Operands[i]->isLoopInvariant(l) &&
598 "Operands of AddRec must be loop-invariant!");
599 }
600 ~SCEVAddRecExpr() {
601 SCEVAddRecExprs.erase(std::make_pair(L,
602 std::vector<SCEV*>(Operands.begin(),
603 Operands.end())));
604 }
605 public:
606 static SCEVHandle get(const SCEVHandle &Start, const SCEVHandle &Step,
607 const Loop *);
608 static SCEVHandle get(std::vector<SCEVHandle> &Operands,
609 const Loop *);
610 static SCEVHandle get(const std::vector<SCEVHandle> &Operands,
611 const Loop *L) {
612 std::vector<SCEVHandle> NewOp(Operands);
613 return get(NewOp, L);
614 }
615
616 typedef std::vector<SCEVHandle>::const_iterator op_iterator;
617 op_iterator op_begin() const { return Operands.begin(); }
618 op_iterator op_end() const { return Operands.end(); }
619
620 unsigned getNumOperands() const { return Operands.size(); }
621 const SCEVHandle &getOperand(unsigned i) const { return Operands[i]; }
622 const SCEVHandle &getStart() const { return Operands[0]; }
623 const Loop *getLoop() const { return L; }
624
625
626 /// getStepRecurrence - This method constructs and returns the recurrence
627 /// indicating how much this expression steps by. If this is a polynomial
628 /// of degree N, it returns a chrec of degree N-1.
629 SCEVHandle getStepRecurrence() const {
630 if (getNumOperands() == 2) return getOperand(1);
631 return SCEVAddRecExpr::get(std::vector<SCEVHandle>(op_begin()+1,op_end()),
632 getLoop());
633 }
634
635 virtual bool hasComputableLoopEvolution(const Loop *QL) const {
636 if (L == QL) return true;
637 /// FIXME: What if the start or step value a recurrence for the specified
638 /// loop?
639 return false;
640 }
641
642
643 virtual bool isLoopInvariant(const Loop *QueryLoop) const {
644 // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't
645 // contain L.
646 return !QueryLoop->contains(L->getHeader());
647 }
648
649 virtual const Type *getType() const { return Operands[0]->getType(); }
650
651 Value *expandCodeFor(ScalarEvolutionRewriter &SER,
652 Instruction *InsertPt);
653
654
655 /// isAffine - Return true if this is an affine AddRec (i.e., it represents
656 /// an expressions A+B*x where A and B are loop invariant values.
657 bool isAffine() const {
658 // We know that the start value is invariant. This expression is thus
659 // affine iff the step is also invariant.
660 return getNumOperands() == 2;
661 }
662
663 /// isQuadratic - Return true if this is an quadratic AddRec (i.e., it
664 /// represents an expressions A+B*x+C*x^2 where A, B and C are loop
665 /// invariant values. This corresponds to an addrec of the form {L,+,M,+,N}
666 bool isQuadratic() const {
667 return getNumOperands() == 3;
668 }
669
670 /// evaluateAtIteration - Return the value of this chain of recurrences at
671 /// the specified iteration number.
672 SCEVHandle evaluateAtIteration(SCEVHandle It) const;
673
674 /// getNumIterationsInRange - Return the number of iterations of this loop
675 /// that produce values in the specified constant range. Another way of
676 /// looking at this is that it returns the first iteration number where the
677 /// value is not in the condition, thus computing the exit count. If the
678 /// iteration count can't be computed, an instance of SCEVCouldNotCompute is
679 /// returned.
680 SCEVHandle getNumIterationsInRange(ConstantRange Range) const;
681
682
683 virtual void print(std::ostream &OS) const {
684 OS << "{" << *Operands[0];
685 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
686 OS << ",+," << *Operands[i];
687 OS << "}<" << L->getHeader()->getName() + ">";
688 }
689
690 /// Methods for support type inquiry through isa, cast, and dyn_cast:
691 static inline bool classof(const SCEVAddRecExpr *S) { return true; }
692 static inline bool classof(const SCEV *S) {
693 return S->getSCEVType() == scAddRecExpr;
694 }
695 };
696}
697
698
699//===----------------------------------------------------------------------===//
700// SCEVUnknown - This means that we are dealing with an entirely unknown SCEV
701// value, and only represent it as it's LLVM Value. This is the "bottom" value
702// for the analysis.
703//
704namespace {
705 class SCEVUnknown;
706 // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any
707 // particular value. Don't use a SCEVHandle here, or else the object will
708 // never be deleted!
709 std::map<Value*, SCEVUnknown*> SCEVUnknowns;
710
711 class SCEVUnknown : public SCEV {
712 Value *V;
713 SCEVUnknown(Value *v) : SCEV(scUnknown), V(v) {}
714
715 protected:
716 ~SCEVUnknown() { SCEVUnknowns.erase(V); }
717 public:
718 /// get method - For SCEVUnknown, this just gets and returns a new
719 /// SCEVUnknown.
720 static SCEVHandle get(Value *V) {
721 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
722 return SCEVConstant::get(CI);
723 SCEVUnknown *&Result = SCEVUnknowns[V];
724 if (Result == 0) Result = new SCEVUnknown(V);
725 return Result;
726 }
727
728 Value *getValue() const { return V; }
729
730 Value *expandCodeFor(ScalarEvolutionRewriter &SER,
731 Instruction *InsertPt) {
732 return V;
733 }
734
735 virtual bool isLoopInvariant(const Loop *L) const {
736 // All non-instruction values are loop invariant. All instructions are
737 // loop invariant if they are not contained in the specified loop.
738 if (Instruction *I = dyn_cast<Instruction>(V))
739 return !L->contains(I->getParent());
740 return true;
741 }
742
743 virtual bool hasComputableLoopEvolution(const Loop *QL) const {
744 return false; // not computable
745 }
746
747 virtual const Type *getType() const { return V->getType(); }
748
749 virtual void print(std::ostream &OS) const {
750 WriteAsOperand(OS, V, false);
751 }
752
753 /// Methods for support type inquiry through isa, cast, and dyn_cast:
754 static inline bool classof(const SCEVUnknown *S) { return true; }
755 static inline bool classof(const SCEV *S) {
756 return S->getSCEVType() == scUnknown;
757 }
758 };
759}
760
761//===----------------------------------------------------------------------===//
762// Simple SCEV method implementations
763//===----------------------------------------------------------------------===//
764
765/// getIntegerSCEV - Given an integer or FP type, create a constant for the
766/// specified signed integer value and return a SCEV for the constant.
767static SCEVHandle getIntegerSCEV(int Val, const Type *Ty) {
768 Constant *C;
769 if (Val == 0)
770 C = Constant::getNullValue(Ty);
771 else if (Ty->isFloatingPoint())
772 C = ConstantFP::get(Ty, Val);
773 else if (Ty->isSigned())
774 C = ConstantSInt::get(Ty, Val);
775 else {
776 C = ConstantSInt::get(Ty->getSignedVersion(), Val);
777 C = ConstantExpr::getCast(C, Ty);
778 }
779 return SCEVUnknown::get(C);
780}
781
782/// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
783/// input value to the specified type. If the type must be extended, it is zero
784/// extended.
785static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty) {
786 const Type *SrcTy = V->getType();
787 assert(SrcTy->isInteger() && Ty->isInteger() &&
788 "Cannot truncate or zero extend with non-integer arguments!");
789 if (SrcTy->getPrimitiveSize() == Ty->getPrimitiveSize())
790 return V; // No conversion
791 if (SrcTy->getPrimitiveSize() > Ty->getPrimitiveSize())
792 return SCEVTruncateExpr::get(V, Ty);
793 return SCEVZeroExtendExpr::get(V, Ty);
794}
795
796/// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
797///
798static SCEVHandle getNegativeSCEV(const SCEVHandle &V) {
799 if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
800 return SCEVUnknown::get(ConstantExpr::getNeg(VC->getValue()));
801
802 return SCEVMulExpr::get(V, getIntegerSCEV(-1, V->getType()));
803}
804
805/// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
806///
807static SCEVHandle getMinusSCEV(const SCEVHandle &LHS, const SCEVHandle &RHS) {
808 // X - Y --> X + -Y
809 return SCEVAddExpr::get(LHS, getNegativeSCEV(RHS));
810}
811
812
813/// Binomial - Evaluate N!/((N-M)!*M!) . Note that N is often large and M is
814/// often very small, so we try to reduce the number of N! terms we need to
815/// evaluate by evaluating this as (N!/(N-M)!)/M!
816static ConstantInt *Binomial(ConstantInt *N, unsigned M) {
817 uint64_t NVal = N->getRawValue();
818 uint64_t FirstTerm = 1;
819 for (unsigned i = 0; i != M; ++i)
820 FirstTerm *= NVal-i;
821
822 unsigned MFactorial = 1;
823 for (; M; --M)
824 MFactorial *= M;
825
826 Constant *Result = ConstantUInt::get(Type::ULongTy, FirstTerm/MFactorial);
827 Result = ConstantExpr::getCast(Result, N->getType());
828 assert(isa<ConstantInt>(Result) && "Cast of integer not folded??");
829 return cast<ConstantInt>(Result);
830}
831
832/// PartialFact - Compute V!/(V-NumSteps)!
833static SCEVHandle PartialFact(SCEVHandle V, unsigned NumSteps) {
834 // Handle this case efficiently, it is common to have constant iteration
835 // counts while computing loop exit values.
836 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
837 uint64_t Val = SC->getValue()->getRawValue();
838 uint64_t Result = 1;
839 for (; NumSteps; --NumSteps)
840 Result *= Val-(NumSteps-1);
841 Constant *Res = ConstantUInt::get(Type::ULongTy, Result);
842 return SCEVUnknown::get(ConstantExpr::getCast(Res, V->getType()));
843 }
844
845 const Type *Ty = V->getType();
846 if (NumSteps == 0)
847 return getIntegerSCEV(1, Ty);
848
849 SCEVHandle Result = V;
850 for (unsigned i = 1; i != NumSteps; ++i)
851 Result = SCEVMulExpr::get(Result, getMinusSCEV(V, getIntegerSCEV(i, Ty)));
852 return Result;
853}
854
855
856/// evaluateAtIteration - Return the value of this chain of recurrences at
857/// the specified iteration number. We can evaluate this recurrence by
858/// multiplying each element in the chain by the binomial coefficient
859/// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
860///
861/// A*choose(It, 0) + B*choose(It, 1) + C*choose(It, 2) + D*choose(It, 3)
862///
863/// FIXME/VERIFY: I don't trust that this is correct in the face of overflow.
864/// Is the binomial equation safe using modular arithmetic??
865///
866SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It) const {
867 SCEVHandle Result = getStart();
868 int Divisor = 1;
869 const Type *Ty = It->getType();
870 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
871 SCEVHandle BC = PartialFact(It, i);
872 Divisor *= i;
873 SCEVHandle Val = SCEVUDivExpr::get(SCEVMulExpr::get(BC, getOperand(i)),
874 getIntegerSCEV(Divisor, Ty));
875 Result = SCEVAddExpr::get(Result, Val);
876 }
877 return Result;
878}
879
880
881//===----------------------------------------------------------------------===//
882// SCEV Expression folder implementations
883//===----------------------------------------------------------------------===//
884
885SCEVHandle SCEVTruncateExpr::get(const SCEVHandle &Op, const Type *Ty) {
886 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
887 return SCEVUnknown::get(ConstantExpr::getCast(SC->getValue(), Ty));
888
889 // If the input value is a chrec scev made out of constants, truncate
890 // all of the constants.
891 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
892 std::vector<SCEVHandle> Operands;
893 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
894 // FIXME: This should allow truncation of other expression types!
895 if (isa<SCEVConstant>(AddRec->getOperand(i)))
896 Operands.push_back(get(AddRec->getOperand(i), Ty));
897 else
898 break;
899 if (Operands.size() == AddRec->getNumOperands())
900 return SCEVAddRecExpr::get(Operands, AddRec->getLoop());
901 }
902
903 SCEVTruncateExpr *&Result = SCEVTruncates[std::make_pair(Op, Ty)];
904 if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
905 return Result;
906}
907
908SCEVHandle SCEVZeroExtendExpr::get(const SCEVHandle &Op, const Type *Ty) {
909 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
910 return SCEVUnknown::get(ConstantExpr::getCast(SC->getValue(), Ty));
911
912 // FIXME: If the input value is a chrec scev, and we can prove that the value
913 // did not overflow the old, smaller, value, we can zero extend all of the
914 // operands (often constants). This would allow analysis of something like
915 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
916
917 SCEVZeroExtendExpr *&Result = SCEVZeroExtends[std::make_pair(Op, Ty)];
918 if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
919 return Result;
920}
921
922// get - Get a canonical add expression, or something simpler if possible.
923SCEVHandle SCEVAddExpr::get(std::vector<SCEVHandle> &Ops) {
924 assert(!Ops.empty() && "Cannot get empty add!");
925
926 // Sort by complexity, this groups all similar expression types together.
927 std::sort(Ops.begin(), Ops.end(), SCEVComplexityCompare());
928
929 // If there are any constants, fold them together.
930 unsigned Idx = 0;
931 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
932 ++Idx;
933 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
934 // We found two constants, fold them together!
935 Constant *Fold = ConstantExpr::getAdd(LHSC->getValue(), RHSC->getValue());
936 if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) {
937 Ops[0] = SCEVConstant::get(CI);
938 Ops.erase(Ops.begin()+1); // Erase the folded element
939 if (Ops.size() == 1) return Ops[0];
940 } else {
941 // If we couldn't fold the expression, move to the next constant. Note
942 // that this is impossible to happen in practice because we always
943 // constant fold constant ints to constant ints.
944 ++Idx;
945 }
946 }
947
948 // If we are left with a constant zero being added, strip it off.
949 if (cast<SCEVConstant>(Ops[0])->getValue()->isNullValue()) {
950 Ops.erase(Ops.begin());
951 --Idx;
952 }
953 }
954
955 if (Ops.size() == 1)
956 return Ops[0];
957
958 // Okay, check to see if the same value occurs in the operand list twice. If
959 // so, merge them together into an multiply expression. Since we sorted the
960 // list, these values are required to be adjacent.
961 const Type *Ty = Ops[0]->getType();
962 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
963 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
964 // Found a match, merge the two values into a multiply, and add any
965 // remaining values to the result.
966 SCEVHandle Two = getIntegerSCEV(2, Ty);
967 SCEVHandle Mul = SCEVMulExpr::get(Ops[i], Two);
968 if (Ops.size() == 2)
969 return Mul;
970 Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
971 Ops.push_back(Mul);
972 return SCEVAddExpr::get(Ops);
973 }
974
975 // Okay, now we know the first non-constant operand. If there are add
976 // operands they would be next.
977 if (Idx < Ops.size()) {
978 bool DeletedAdd = false;
979 while (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
980 // If we have an add, expand the add operands onto the end of the operands
981 // list.
982 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
983 Ops.erase(Ops.begin()+Idx);
984 DeletedAdd = true;
985 }
986
987 // If we deleted at least one add, we added operands to the end of the list,
988 // and they are not necessarily sorted. Recurse to resort and resimplify
989 // any operands we just aquired.
990 if (DeletedAdd)
991 return get(Ops);
992 }
993
994 // Skip over the add expression until we get to a multiply.
995 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
996 ++Idx;
997
998 // If we are adding something to a multiply expression, make sure the
999 // something is not already an operand of the multiply. If so, merge it into
1000 // the multiply.
1001 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
1002 SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
1003 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
1004 SCEV *MulOpSCEV = Mul->getOperand(MulOp);
1005 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
1006 if (MulOpSCEV == Ops[AddOp] &&
1007 (Mul->getNumOperands() != 2 || !isa<SCEVConstant>(MulOpSCEV))) {
1008 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
1009 SCEVHandle InnerMul = Mul->getOperand(MulOp == 0);
1010 if (Mul->getNumOperands() != 2) {
1011 // If the multiply has more than two operands, we must get the
1012 // Y*Z term.
1013 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
1014 MulOps.erase(MulOps.begin()+MulOp);
1015 InnerMul = SCEVMulExpr::get(MulOps);
1016 }
1017 SCEVHandle One = getIntegerSCEV(1, Ty);
1018 SCEVHandle AddOne = SCEVAddExpr::get(InnerMul, One);
1019 SCEVHandle OuterMul = SCEVMulExpr::get(AddOne, Ops[AddOp]);
1020 if (Ops.size() == 2) return OuterMul;
1021 if (AddOp < Idx) {
1022 Ops.erase(Ops.begin()+AddOp);
1023 Ops.erase(Ops.begin()+Idx-1);
1024 } else {
1025 Ops.erase(Ops.begin()+Idx);
1026 Ops.erase(Ops.begin()+AddOp-1);
1027 }
1028 Ops.push_back(OuterMul);
1029 return SCEVAddExpr::get(Ops);
1030 }
1031
1032 // Check this multiply against other multiplies being added together.
1033 for (unsigned OtherMulIdx = Idx+1;
1034 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
1035 ++OtherMulIdx) {
1036 SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
1037 // If MulOp occurs in OtherMul, we can fold the two multiplies
1038 // together.
1039 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
1040 OMulOp != e; ++OMulOp)
1041 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
1042 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
1043 SCEVHandle InnerMul1 = Mul->getOperand(MulOp == 0);
1044 if (Mul->getNumOperands() != 2) {
1045 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
1046 MulOps.erase(MulOps.begin()+MulOp);
1047 InnerMul1 = SCEVMulExpr::get(MulOps);
1048 }
1049 SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0);
1050 if (OtherMul->getNumOperands() != 2) {
1051 std::vector<SCEVHandle> MulOps(OtherMul->op_begin(),
1052 OtherMul->op_end());
1053 MulOps.erase(MulOps.begin()+OMulOp);
1054 InnerMul2 = SCEVMulExpr::get(MulOps);
1055 }
1056 SCEVHandle InnerMulSum = SCEVAddExpr::get(InnerMul1,InnerMul2);
1057 SCEVHandle OuterMul = SCEVMulExpr::get(MulOpSCEV, InnerMulSum);
1058 if (Ops.size() == 2) return OuterMul;
1059 Ops.erase(Ops.begin()+Idx);
1060 Ops.erase(Ops.begin()+OtherMulIdx-1);
1061 Ops.push_back(OuterMul);
1062 return SCEVAddExpr::get(Ops);
1063 }
1064 }
1065 }
1066 }
1067
1068 // If there are any add recurrences in the operands list, see if any other
1069 // added values are loop invariant. If so, we can fold them into the
1070 // recurrence.
1071 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1072 ++Idx;
1073
1074 // Scan over all recurrences, trying to fold loop invariants into them.
1075 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1076 // Scan all of the other operands to this add and add them to the vector if
1077 // they are loop invariant w.r.t. the recurrence.
1078 std::vector<SCEVHandle> LIOps;
1079 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1080 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1081 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1082 LIOps.push_back(Ops[i]);
1083 Ops.erase(Ops.begin()+i);
1084 --i; --e;
1085 }
1086
1087 // If we found some loop invariants, fold them into the recurrence.
1088 if (!LIOps.empty()) {
1089 // NLI + LI + { Start,+,Step} --> NLI + { LI+Start,+,Step }
1090 LIOps.push_back(AddRec->getStart());
1091
1092 std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
1093 AddRecOps[0] = SCEVAddExpr::get(LIOps);
1094
1095 SCEVHandle NewRec = SCEVAddRecExpr::get(AddRecOps, AddRec->getLoop());
1096 // If all of the other operands were loop invariant, we are done.
1097 if (Ops.size() == 1) return NewRec;
1098
1099 // Otherwise, add the folded AddRec by the non-liv parts.
1100 for (unsigned i = 0;; ++i)
1101 if (Ops[i] == AddRec) {
1102 Ops[i] = NewRec;
1103 break;
1104 }
1105 return SCEVAddExpr::get(Ops);
1106 }
1107
1108 // Okay, if there weren't any loop invariants to be folded, check to see if
1109 // there are multiple AddRec's with the same loop induction variable being
1110 // added together. If so, we can fold them.
1111 for (unsigned OtherIdx = Idx+1;
1112 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1113 if (OtherIdx != Idx) {
1114 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1115 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1116 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
1117 std::vector<SCEVHandle> NewOps(AddRec->op_begin(), AddRec->op_end());
1118 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
1119 if (i >= NewOps.size()) {
1120 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
1121 OtherAddRec->op_end());
1122 break;
1123 }
1124 NewOps[i] = SCEVAddExpr::get(NewOps[i], OtherAddRec->getOperand(i));
1125 }
1126 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop());
1127
1128 if (Ops.size() == 2) return NewAddRec;
1129
1130 Ops.erase(Ops.begin()+Idx);
1131 Ops.erase(Ops.begin()+OtherIdx-1);
1132 Ops.push_back(NewAddRec);
1133 return SCEVAddExpr::get(Ops);
1134 }
1135 }
1136
1137 // Otherwise couldn't fold anything into this recurrence. Move onto the
1138 // next one.
1139 }
1140
1141 // Okay, it looks like we really DO need an add expr. Check to see if we
1142 // already have one, otherwise create a new one.
1143 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1144 SCEVCommutativeExpr *&Result = SCEVCommExprs[std::make_pair(scAddExpr,
1145 SCEVOps)];
1146 if (Result == 0) Result = new SCEVAddExpr(Ops);
1147 return Result;
1148}
1149
1150
1151SCEVHandle SCEVMulExpr::get(std::vector<SCEVHandle> &Ops) {
1152 assert(!Ops.empty() && "Cannot get empty mul!");
1153
1154 // Sort by complexity, this groups all similar expression types together.
1155 std::sort(Ops.begin(), Ops.end(), SCEVComplexityCompare());
1156
1157 // If there are any constants, fold them together.
1158 unsigned Idx = 0;
1159 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
1160
1161 // C1*(C2+V) -> C1*C2 + C1*V
1162 if (Ops.size() == 2)
1163 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
1164 if (Add->getNumOperands() == 2 &&
1165 isa<SCEVConstant>(Add->getOperand(0)))
1166 return SCEVAddExpr::get(SCEVMulExpr::get(LHSC, Add->getOperand(0)),
1167 SCEVMulExpr::get(LHSC, Add->getOperand(1)));
1168
1169
1170 ++Idx;
1171 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
1172 // We found two constants, fold them together!
1173 Constant *Fold = ConstantExpr::getMul(LHSC->getValue(), RHSC->getValue());
1174 if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) {
1175 Ops[0] = SCEVConstant::get(CI);
1176 Ops.erase(Ops.begin()+1); // Erase the folded element
1177 if (Ops.size() == 1) return Ops[0];
1178 } else {
1179 // If we couldn't fold the expression, move to the next constant. Note
1180 // that this is impossible to happen in practice because we always
1181 // constant fold constant ints to constant ints.
1182 ++Idx;
1183 }
1184 }
1185
1186 // If we are left with a constant one being multiplied, strip it off.
1187 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
1188 Ops.erase(Ops.begin());
1189 --Idx;
1190 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isNullValue()) {
1191 // If we have a multiply of zero, it will always be zero.
1192 return Ops[0];
1193 }
1194 }
1195
1196 // Skip over the add expression until we get to a multiply.
1197 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
1198 ++Idx;
1199
1200 if (Ops.size() == 1)
1201 return Ops[0];
1202
1203 // If there are mul operands inline them all into this expression.
1204 if (Idx < Ops.size()) {
1205 bool DeletedMul = false;
1206 while (SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
1207 // If we have an mul, expand the mul operands onto the end of the operands
1208 // list.
1209 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
1210 Ops.erase(Ops.begin()+Idx);
1211 DeletedMul = true;
1212 }
1213
1214 // If we deleted at least one mul, we added operands to the end of the list,
1215 // and they are not necessarily sorted. Recurse to resort and resimplify
1216 // any operands we just aquired.
1217 if (DeletedMul)
1218 return get(Ops);
1219 }
1220
1221 // If there are any add recurrences in the operands list, see if any other
1222 // added values are loop invariant. If so, we can fold them into the
1223 // recurrence.
1224 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
1225 ++Idx;
1226
1227 // Scan over all recurrences, trying to fold loop invariants into them.
1228 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
1229 // Scan all of the other operands to this mul and add them to the vector if
1230 // they are loop invariant w.r.t. the recurrence.
1231 std::vector<SCEVHandle> LIOps;
1232 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
1233 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1234 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
1235 LIOps.push_back(Ops[i]);
1236 Ops.erase(Ops.begin()+i);
1237 --i; --e;
1238 }
1239
1240 // If we found some loop invariants, fold them into the recurrence.
1241 if (!LIOps.empty()) {
1242 // NLI * LI * { Start,+,Step} --> NLI * { LI*Start,+,LI*Step }
1243 std::vector<SCEVHandle> NewOps;
1244 NewOps.reserve(AddRec->getNumOperands());
1245 if (LIOps.size() == 1) {
1246 SCEV *Scale = LIOps[0];
1247 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
1248 NewOps.push_back(SCEVMulExpr::get(Scale, AddRec->getOperand(i)));
1249 } else {
1250 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
1251 std::vector<SCEVHandle> MulOps(LIOps);
1252 MulOps.push_back(AddRec->getOperand(i));
1253 NewOps.push_back(SCEVMulExpr::get(MulOps));
1254 }
1255 }
1256
1257 SCEVHandle NewRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop());
1258
1259 // If all of the other operands were loop invariant, we are done.
1260 if (Ops.size() == 1) return NewRec;
1261
1262 // Otherwise, multiply the folded AddRec by the non-liv parts.
1263 for (unsigned i = 0;; ++i)
1264 if (Ops[i] == AddRec) {
1265 Ops[i] = NewRec;
1266 break;
1267 }
1268 return SCEVMulExpr::get(Ops);
1269 }
1270
1271 // Okay, if there weren't any loop invariants to be folded, check to see if
1272 // there are multiple AddRec's with the same loop induction variable being
1273 // multiplied together. If so, we can fold them.
1274 for (unsigned OtherIdx = Idx+1;
1275 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
1276 if (OtherIdx != Idx) {
1277 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
1278 if (AddRec->getLoop() == OtherAddRec->getLoop()) {
1279 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
1280 SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
1281 SCEVHandle NewStart = SCEVMulExpr::get(F->getStart(),
1282 G->getStart());
1283 SCEVHandle B = F->getStepRecurrence();
1284 SCEVHandle D = G->getStepRecurrence();
1285 SCEVHandle NewStep = SCEVAddExpr::get(SCEVMulExpr::get(F, D),
1286 SCEVMulExpr::get(G, B),
1287 SCEVMulExpr::get(B, D));
1288 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewStart, NewStep,
1289 F->getLoop());
1290 if (Ops.size() == 2) return NewAddRec;
1291
1292 Ops.erase(Ops.begin()+Idx);
1293 Ops.erase(Ops.begin()+OtherIdx-1);
1294 Ops.push_back(NewAddRec);
1295 return SCEVMulExpr::get(Ops);
1296 }
1297 }
1298
1299 // Otherwise couldn't fold anything into this recurrence. Move onto the
1300 // next one.
1301 }
1302
1303 // Okay, it looks like we really DO need an mul expr. Check to see if we
1304 // already have one, otherwise create a new one.
1305 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
1306 SCEVCommutativeExpr *&Result = SCEVCommExprs[std::make_pair(scMulExpr,
1307 SCEVOps)];
1308 if (Result == 0) Result = new SCEVMulExpr(Ops);
1309 return Result;
1310}
1311
1312SCEVHandle SCEVUDivExpr::get(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1313 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
1314 if (RHSC->getValue()->equalsInt(1))
1315 return LHS; // X /u 1 --> x
1316 if (RHSC->getValue()->isAllOnesValue())
1317 return getNegativeSCEV(LHS); // X /u -1 --> -x
1318
1319 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
1320 Constant *LHSCV = LHSC->getValue();
1321 Constant *RHSCV = RHSC->getValue();
1322 if (LHSCV->getType()->isSigned())
1323 LHSCV = ConstantExpr::getCast(LHSCV,
1324 LHSCV->getType()->getUnsignedVersion());
1325 if (RHSCV->getType()->isSigned())
1326 RHSCV = ConstantExpr::getCast(RHSCV, LHSCV->getType());
1327 return SCEVUnknown::get(ConstantExpr::getDiv(LHSCV, RHSCV));
1328 }
1329 }
1330
1331 // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow.
1332
1333 SCEVUDivExpr *&Result = SCEVUDivs[std::make_pair(LHS, RHS)];
1334 if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
1335 return Result;
1336}
1337
1338
1339/// SCEVAddRecExpr::get - Get a add recurrence expression for the
1340/// specified loop. Simplify the expression as much as possible.
1341SCEVHandle SCEVAddRecExpr::get(const SCEVHandle &Start,
1342 const SCEVHandle &Step, const Loop *L) {
1343 std::vector<SCEVHandle> Operands;
1344 Operands.push_back(Start);
1345 if (SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
1346 if (StepChrec->getLoop() == L) {
1347 Operands.insert(Operands.end(), StepChrec->op_begin(),
1348 StepChrec->op_end());
1349 return get(Operands, L);
1350 }
1351
1352 Operands.push_back(Step);
1353 return get(Operands, L);
1354}
1355
1356/// SCEVAddRecExpr::get - Get a add recurrence expression for the
1357/// specified loop. Simplify the expression as much as possible.
1358SCEVHandle SCEVAddRecExpr::get(std::vector<SCEVHandle> &Operands,
1359 const Loop *L) {
1360 if (Operands.size() == 1) return Operands[0];
1361
1362 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Operands.back()))
1363 if (StepC->getValue()->isNullValue()) {
1364 Operands.pop_back();
1365 return get(Operands, L); // { X,+,0 } --> X
1366 }
1367
1368 SCEVAddRecExpr *&Result =
1369 SCEVAddRecExprs[std::make_pair(L, std::vector<SCEV*>(Operands.begin(),
1370 Operands.end()))];
1371 if (Result == 0) Result = new SCEVAddRecExpr(Operands, L);
1372 return Result;
1373}
1374
1375
1376//===----------------------------------------------------------------------===//
1377// Non-trivial closed-form SCEV Expanders
1378//===----------------------------------------------------------------------===//
1379
1380Value *SCEVTruncateExpr::expandCodeFor(ScalarEvolutionRewriter &SER,
1381 Instruction *InsertPt) {
1382 Value *V = SER.ExpandCodeFor(getOperand(), InsertPt);
1383 return new CastInst(V, getType(), "tmp.", InsertPt);
1384}
1385
1386Value *SCEVZeroExtendExpr::expandCodeFor(ScalarEvolutionRewriter &SER,
1387 Instruction *InsertPt) {
1388 Value *V = SER.ExpandCodeFor(getOperand(), InsertPt,
1389 getOperand()->getType()->getUnsignedVersion());
1390 return new CastInst(V, getType(), "tmp.", InsertPt);
1391}
1392
1393Value *SCEVAddExpr::expandCodeFor(ScalarEvolutionRewriter &SER,
1394 Instruction *InsertPt) {
1395 const Type *Ty = getType();
1396 Value *V = SER.ExpandCodeFor(getOperand(getNumOperands()-1), InsertPt, Ty);
1397
1398 // Emit a bunch of add instructions
1399 for (int i = getNumOperands()-2; i >= 0; --i)
1400 V = BinaryOperator::create(Instruction::Add, V,
1401 SER.ExpandCodeFor(getOperand(i), InsertPt, Ty),
1402 "tmp.", InsertPt);
1403 return V;
1404}
1405
1406Value *SCEVMulExpr::expandCodeFor(ScalarEvolutionRewriter &SER,
1407 Instruction *InsertPt) {
1408 const Type *Ty = getType();
1409 int FirstOp = 0; // Set if we should emit a subtract.
1410 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getOperand(0)))
1411 if (SC->getValue()->isAllOnesValue())
1412 FirstOp = 1;
1413
1414 int i = getNumOperands()-2;
1415 Value *V = SER.ExpandCodeFor(getOperand(i+1), InsertPt, Ty);
1416
1417 // Emit a bunch of multiply instructions
1418 for (; i >= FirstOp; --i)
1419 V = BinaryOperator::create(Instruction::Mul, V,
1420 SER.ExpandCodeFor(getOperand(i), InsertPt, Ty),
1421 "tmp.", InsertPt);
1422 // -1 * ... ---> 0 - ...
1423 if (FirstOp == 1)
1424 V = BinaryOperator::create(Instruction::Sub, Constant::getNullValue(Ty), V,
1425 "tmp.", InsertPt);
1426 return V;
1427}
1428
1429Value *SCEVUDivExpr::expandCodeFor(ScalarEvolutionRewriter &SER,
1430 Instruction *InsertPt) {
1431 const Type *Ty = getType();
1432 Value *LHS = SER.ExpandCodeFor(getLHS(), InsertPt, Ty);
1433 Value *RHS = SER.ExpandCodeFor(getRHS(), InsertPt, Ty);
1434 return BinaryOperator::create(Instruction::Div, LHS, RHS, "tmp.", InsertPt);
1435}
1436
1437Value *SCEVAddRecExpr::expandCodeFor(ScalarEvolutionRewriter &SER,
1438 Instruction *InsertPt) {
1439 const Type *Ty = getType();
1440 // We cannot yet do fp recurrences, e.g. the xform of {X,+,F} --> X+{0,+,F}
1441 assert(Ty->isIntegral() && "Cannot expand fp recurrences yet!");
1442
1443 // {X,+,F} --> X + {0,+,F}
1444 if (!isa<SCEVConstant>(getStart()) ||
1445 !cast<SCEVConstant>(getStart())->getValue()->isNullValue()) {
1446 Value *Start = SER.ExpandCodeFor(getStart(), InsertPt, Ty);
1447 std::vector<SCEVHandle> NewOps(op_begin(), op_end());
1448 NewOps[0] = getIntegerSCEV(0, getType());
1449 Value *Rest = SER.ExpandCodeFor(SCEVAddRecExpr::get(NewOps, getLoop()),
1450 InsertPt, getType());
1451
1452 // FIXME: look for an existing add to use.
1453 return BinaryOperator::create(Instruction::Add, Rest, Start, "tmp.",
1454 InsertPt);
1455 }
1456
1457 // {0,+,1} --> Insert a canonical induction variable into the loop!
1458 if (getNumOperands() == 2 && getOperand(1) == getIntegerSCEV(1, getType())) {
1459 // Create and insert the PHI node for the induction variable in the
1460 // specified loop.
1461 BasicBlock *Header = getLoop()->getHeader();
1462 PHINode *PN = new PHINode(Ty, "indvar", Header->begin());
1463 PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader());
1464
1465 // Insert a unit add instruction after the PHI nodes in the header block.
1466 BasicBlock::iterator I = PN;
1467 while (isa<PHINode>(I)) ++I;
1468
1469 Constant *One = Ty->isFloatingPoint() ?(Constant*)ConstantFP::get(Ty, 1.0)
1470 :(Constant*)ConstantInt::get(Ty, 1);
1471 Instruction *Add = BinaryOperator::create(Instruction::Add, PN, One,
1472 "indvar.next", I);
1473
1474 pred_iterator PI = pred_begin(Header);
1475 if (*PI == L->getLoopPreheader())
1476 ++PI;
1477 PN->addIncoming(Add, *PI);
1478 return PN;
1479 }
1480
1481 // Get the canonical induction variable I for this loop.
1482 Value *I = SER.GetOrInsertCanonicalInductionVariable(getLoop(), Ty);
1483
1484 if (getNumOperands() == 2) { // {0,+,F} --> i*F
1485 Value *F = SER.ExpandCodeFor(getOperand(1), InsertPt, Ty);
1486 return BinaryOperator::create(Instruction::Mul, I, F, "tmp.", InsertPt);
1487 }
1488
1489 // If this is a chain of recurrences, turn it into a closed form, using the
1490 // folders, then expandCodeFor the closed form. This allows the folders to
1491 // simplify the expression without having to build a bunch of special code
1492 // into this folder.
1493 SCEVHandle IH = SCEVUnknown::get(I); // Get I as a "symbolic" SCEV.
1494
1495 SCEVHandle V = evaluateAtIteration(IH);
Chris Lattnerd18d9dc2004-04-02 20:26:46 +00001496 //std::cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
Chris Lattner53e677a2004-04-02 20:23:17 +00001497
1498 return SER.ExpandCodeFor(V, InsertPt, Ty);
1499}
1500
1501
1502//===----------------------------------------------------------------------===//
1503// ScalarEvolutionsImpl Definition and Implementation
1504//===----------------------------------------------------------------------===//
1505//
1506/// ScalarEvolutionsImpl - This class implements the main driver for the scalar
1507/// evolution code.
1508///
1509namespace {
1510 struct ScalarEvolutionsImpl {
1511 /// F - The function we are analyzing.
1512 ///
1513 Function &F;
1514
1515 /// LI - The loop information for the function we are currently analyzing.
1516 ///
1517 LoopInfo &LI;
1518
1519 /// UnknownValue - This SCEV is used to represent unknown trip counts and
1520 /// things.
1521 SCEVHandle UnknownValue;
1522
1523 /// Scalars - This is a cache of the scalars we have analyzed so far.
1524 ///
1525 std::map<Value*, SCEVHandle> Scalars;
1526
1527 /// IterationCounts - Cache the iteration count of the loops for this
1528 /// function as they are computed.
1529 std::map<const Loop*, SCEVHandle> IterationCounts;
1530
1531 public:
1532 ScalarEvolutionsImpl(Function &f, LoopInfo &li)
1533 : F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {}
1534
1535 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1536 /// expression and create a new one.
1537 SCEVHandle getSCEV(Value *V);
1538
1539 /// getSCEVAtScope - Compute the value of the specified expression within
1540 /// the indicated loop (which may be null to indicate in no loop). If the
1541 /// expression cannot be evaluated, return UnknownValue itself.
1542 SCEVHandle getSCEVAtScope(SCEV *V, const Loop *L);
1543
1544
1545 /// hasLoopInvariantIterationCount - Return true if the specified loop has
1546 /// an analyzable loop-invariant iteration count.
1547 bool hasLoopInvariantIterationCount(const Loop *L);
1548
1549 /// getIterationCount - If the specified loop has a predictable iteration
1550 /// count, return it. Note that it is not valid to call this method on a
1551 /// loop without a loop-invariant iteration count.
1552 SCEVHandle getIterationCount(const Loop *L);
1553
1554 /// deleteInstructionFromRecords - This method should be called by the
1555 /// client before it removes an instruction from the program, to make sure
1556 /// that no dangling references are left around.
1557 void deleteInstructionFromRecords(Instruction *I);
1558
1559 private:
1560 /// createSCEV - We know that there is no SCEV for the specified value.
1561 /// Analyze the expression.
1562 SCEVHandle createSCEV(Value *V);
1563 SCEVHandle createNodeForCast(CastInst *CI);
1564
1565 /// createNodeForPHI - Provide the special handling we need to analyze PHI
1566 /// SCEVs.
1567 SCEVHandle createNodeForPHI(PHINode *PN);
1568 void UpdatePHIUserScalarEntries(Instruction *I, PHINode *PN,
1569 std::set<Instruction*> &UpdatedInsts);
1570
1571 /// ComputeIterationCount - Compute the number of times the specified loop
1572 /// will iterate.
1573 SCEVHandle ComputeIterationCount(const Loop *L);
1574
1575 /// HowFarToZero - Return the number of times a backedge comparing the
1576 /// specified value to zero will execute. If not computable, return
1577 /// UnknownValue
1578 SCEVHandle HowFarToZero(SCEV *V, const Loop *L);
1579
1580 /// HowFarToNonZero - Return the number of times a backedge checking the
1581 /// specified value for nonzero will execute. If not computable, return
1582 /// UnknownValue
1583 SCEVHandle HowFarToNonZero(SCEV *V, const Loop *L);
1584 };
1585}
1586
1587//===----------------------------------------------------------------------===//
1588// Basic SCEV Analysis and PHI Idiom Recognition Code
1589//
1590
1591/// deleteInstructionFromRecords - This method should be called by the
1592/// client before it removes an instruction from the program, to make sure
1593/// that no dangling references are left around.
1594void ScalarEvolutionsImpl::deleteInstructionFromRecords(Instruction *I) {
1595 Scalars.erase(I);
1596}
1597
1598
1599/// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
1600/// expression and create a new one.
1601SCEVHandle ScalarEvolutionsImpl::getSCEV(Value *V) {
1602 assert(V->getType() != Type::VoidTy && "Can't analyze void expressions!");
1603
1604 std::map<Value*, SCEVHandle>::iterator I = Scalars.find(V);
1605 if (I != Scalars.end()) return I->second;
1606 SCEVHandle S = createSCEV(V);
1607 Scalars.insert(std::make_pair(V, S));
1608 return S;
1609}
1610
1611
1612/// UpdatePHIUserScalarEntries - After PHI node analysis, we have a bunch of
1613/// entries in the scalar map that refer to the "symbolic" PHI value instead of
1614/// the recurrence value. After we resolve the PHI we must loop over all of the
1615/// using instructions that have scalar map entries and update them.
1616void ScalarEvolutionsImpl::UpdatePHIUserScalarEntries(Instruction *I,
1617 PHINode *PN,
1618 std::set<Instruction*> &UpdatedInsts) {
1619 std::map<Value*, SCEVHandle>::iterator SI = Scalars.find(I);
1620 if (SI == Scalars.end()) return; // This scalar wasn't previous processed.
1621 if (UpdatedInsts.insert(I).second) {
1622 Scalars.erase(SI); // Remove the old entry
1623 getSCEV(I); // Calculate the new entry
1624
1625 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1626 UI != E; ++UI)
1627 UpdatePHIUserScalarEntries(cast<Instruction>(*UI), PN, UpdatedInsts);
1628 }
1629}
1630
1631
1632/// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
1633/// a loop header, making it a potential recurrence, or it doesn't.
1634///
1635SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) {
1636 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
1637 if (const Loop *L = LI.getLoopFor(PN->getParent()))
1638 if (L->getHeader() == PN->getParent()) {
1639 // If it lives in the loop header, it has two incoming values, one
1640 // from outside the loop, and one from inside.
1641 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1642 unsigned BackEdge = IncomingEdge^1;
1643
1644 // While we are analyzing this PHI node, handle its value symbolically.
1645 SCEVHandle SymbolicName = SCEVUnknown::get(PN);
1646 assert(Scalars.find(PN) == Scalars.end() &&
1647 "PHI node already processed?");
1648 Scalars.insert(std::make_pair(PN, SymbolicName));
1649
1650 // Using this symbolic name for the PHI, analyze the value coming around
1651 // the back-edge.
1652 SCEVHandle BEValue = getSCEV(PN->getIncomingValue(BackEdge));
1653
1654 // NOTE: If BEValue is loop invariant, we know that the PHI node just
1655 // has a special value for the first iteration of the loop.
1656
1657 // If the value coming around the backedge is an add with the symbolic
1658 // value we just inserted, then we found a simple induction variable!
1659 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
1660 // If there is a single occurrence of the symbolic value, replace it
1661 // with a recurrence.
1662 unsigned FoundIndex = Add->getNumOperands();
1663 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1664 if (Add->getOperand(i) == SymbolicName)
1665 if (FoundIndex == e) {
1666 FoundIndex = i;
1667 break;
1668 }
1669
1670 if (FoundIndex != Add->getNumOperands()) {
1671 // Create an add with everything but the specified operand.
1672 std::vector<SCEVHandle> Ops;
1673 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
1674 if (i != FoundIndex)
1675 Ops.push_back(Add->getOperand(i));
1676 SCEVHandle Accum = SCEVAddExpr::get(Ops);
1677
1678 // This is not a valid addrec if the step amount is varying each
1679 // loop iteration, but is not itself an addrec in this loop.
1680 if (Accum->isLoopInvariant(L) ||
1681 (isa<SCEVAddRecExpr>(Accum) &&
1682 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
1683 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
1684 SCEVHandle PHISCEV = SCEVAddRecExpr::get(StartVal, Accum, L);
1685
1686 // Okay, for the entire analysis of this edge we assumed the PHI
1687 // to be symbolic. We now need to go back and update all of the
1688 // entries for the scalars that use the PHI (except for the PHI
1689 // itself) to use the new analyzed value instead of the "symbolic"
1690 // value.
1691 Scalars.find(PN)->second = PHISCEV; // Update the PHI value
1692 std::set<Instruction*> UpdatedInsts;
1693 UpdatedInsts.insert(PN);
1694 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
1695 UI != E; ++UI)
1696 UpdatePHIUserScalarEntries(cast<Instruction>(*UI), PN,
1697 UpdatedInsts);
1698 return PHISCEV;
1699 }
1700 }
1701 }
1702
1703 return SymbolicName;
1704 }
1705
1706 // If it's not a loop phi, we can't handle it yet.
1707 return SCEVUnknown::get(PN);
1708}
1709
1710/// createNodeForCast - Handle the various forms of casts that we support.
1711///
1712SCEVHandle ScalarEvolutionsImpl::createNodeForCast(CastInst *CI) {
1713 const Type *SrcTy = CI->getOperand(0)->getType();
1714 const Type *DestTy = CI->getType();
1715
1716 // If this is a noop cast (ie, conversion from int to uint), ignore it.
1717 if (SrcTy->isLosslesslyConvertibleTo(DestTy))
1718 return getSCEV(CI->getOperand(0));
1719
1720 if (SrcTy->isInteger() && DestTy->isInteger()) {
1721 // Otherwise, if this is a truncating integer cast, we can represent this
1722 // cast.
1723 if (SrcTy->getPrimitiveSize() > DestTy->getPrimitiveSize())
1724 return SCEVTruncateExpr::get(getSCEV(CI->getOperand(0)),
1725 CI->getType()->getUnsignedVersion());
1726 if (SrcTy->isUnsigned() &&
1727 SrcTy->getPrimitiveSize() > DestTy->getPrimitiveSize())
1728 return SCEVZeroExtendExpr::get(getSCEV(CI->getOperand(0)),
1729 CI->getType()->getUnsignedVersion());
1730 }
1731
1732 // If this is an sign or zero extending cast and we can prove that the value
1733 // will never overflow, we could do similar transformations.
1734
1735 // Otherwise, we can't handle this cast!
1736 return SCEVUnknown::get(CI);
1737}
1738
1739
1740/// createSCEV - We know that there is no SCEV for the specified value.
1741/// Analyze the expression.
1742///
1743SCEVHandle ScalarEvolutionsImpl::createSCEV(Value *V) {
1744 if (Instruction *I = dyn_cast<Instruction>(V)) {
1745 switch (I->getOpcode()) {
1746 case Instruction::Add:
1747 return SCEVAddExpr::get(getSCEV(I->getOperand(0)),
1748 getSCEV(I->getOperand(1)));
1749 case Instruction::Mul:
1750 return SCEVMulExpr::get(getSCEV(I->getOperand(0)),
1751 getSCEV(I->getOperand(1)));
1752 case Instruction::Div:
1753 if (V->getType()->isInteger() && V->getType()->isUnsigned())
1754 return SCEVUDivExpr::get(getSCEV(I->getOperand(0)),
1755 getSCEV(I->getOperand(1)));
1756 break;
1757
1758 case Instruction::Sub:
1759 return getMinusSCEV(getSCEV(I->getOperand(0)), getSCEV(I->getOperand(1)));
1760
1761 case Instruction::Shl:
1762 // Turn shift left of a constant amount into a multiply.
1763 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1764 Constant *X = ConstantInt::get(V->getType(), 1);
1765 X = ConstantExpr::getShl(X, SA);
1766 return SCEVMulExpr::get(getSCEV(I->getOperand(0)), getSCEV(X));
1767 }
1768 break;
1769
1770 case Instruction::Shr:
1771 if (ConstantUInt *SA = dyn_cast<ConstantUInt>(I->getOperand(1)))
1772 if (V->getType()->isUnsigned()) {
1773 Constant *X = ConstantInt::get(V->getType(), 1);
1774 X = ConstantExpr::getShl(X, SA);
1775 return SCEVUDivExpr::get(getSCEV(I->getOperand(0)), getSCEV(X));
1776 }
1777 break;
1778
1779 case Instruction::Cast:
1780 return createNodeForCast(cast<CastInst>(I));
1781
1782 case Instruction::PHI:
1783 return createNodeForPHI(cast<PHINode>(I));
1784
1785 default: // We cannot analyze this expression.
1786 break;
1787 }
1788 }
1789
1790 return SCEVUnknown::get(V);
1791}
1792
1793
1794
1795//===----------------------------------------------------------------------===//
1796// Iteration Count Computation Code
1797//
1798
1799/// getIterationCount - If the specified loop has a predictable iteration
1800/// count, return it. Note that it is not valid to call this method on a
1801/// loop without a loop-invariant iteration count.
1802SCEVHandle ScalarEvolutionsImpl::getIterationCount(const Loop *L) {
1803 std::map<const Loop*, SCEVHandle>::iterator I = IterationCounts.find(L);
1804 if (I == IterationCounts.end()) {
1805 SCEVHandle ItCount = ComputeIterationCount(L);
1806 I = IterationCounts.insert(std::make_pair(L, ItCount)).first;
1807 if (ItCount != UnknownValue) {
1808 assert(ItCount->isLoopInvariant(L) &&
1809 "Computed trip count isn't loop invariant for loop!");
1810 ++NumTripCountsComputed;
1811 } else if (isa<PHINode>(L->getHeader()->begin())) {
1812 // Only count loops that have phi nodes as not being computable.
1813 ++NumTripCountsNotComputed;
1814 }
1815 }
1816 return I->second;
1817}
1818
1819/// ComputeIterationCount - Compute the number of times the specified loop
1820/// will iterate.
1821SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) {
1822 // If the loop has a non-one exit block count, we can't analyze it.
1823 if (L->getExitBlocks().size() != 1) return UnknownValue;
1824
1825 // Okay, there is one exit block. Try to find the condition that causes the
1826 // loop to be exited.
1827 BasicBlock *ExitBlock = L->getExitBlocks()[0];
1828
1829 BasicBlock *ExitingBlock = 0;
1830 for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock);
1831 PI != E; ++PI)
1832 if (L->contains(*PI)) {
1833 if (ExitingBlock == 0)
1834 ExitingBlock = *PI;
1835 else
1836 return UnknownValue; // More than one block exiting!
1837 }
1838 assert(ExitingBlock && "No exits from loop, something is broken!");
1839
1840 // Okay, we've computed the exiting block. See what condition causes us to
1841 // exit.
1842 //
1843 // FIXME: we should be able to handle switch instructions (with a single exit)
1844 // FIXME: We should handle cast of int to bool as well
1845 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1846 if (ExitBr == 0) return UnknownValue;
1847 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
1848 SetCondInst *ExitCond = dyn_cast<SetCondInst>(ExitBr->getCondition());
1849 if (ExitCond == 0) return UnknownValue;
1850
1851 SCEVHandle LHS = getSCEV(ExitCond->getOperand(0));
1852 SCEVHandle RHS = getSCEV(ExitCond->getOperand(1));
1853
1854 // Try to evaluate any dependencies out of the loop.
1855 SCEVHandle Tmp = getSCEVAtScope(LHS, L);
1856 if (!isa<SCEVCouldNotCompute>(Tmp)) LHS = Tmp;
1857 Tmp = getSCEVAtScope(RHS, L);
1858 if (!isa<SCEVCouldNotCompute>(Tmp)) RHS = Tmp;
1859
1860 // If the condition was exit on true, convert the condition to exit on false.
1861 Instruction::BinaryOps Cond;
1862 if (ExitBr->getSuccessor(1) == ExitBlock)
1863 Cond = ExitCond->getOpcode();
1864 else
1865 Cond = ExitCond->getInverseCondition();
1866
1867 // At this point, we would like to compute how many iterations of the loop the
1868 // predicate will return true for these inputs.
1869 if (isa<SCEVConstant>(LHS) && !isa<SCEVConstant>(RHS)) {
1870 // If there is a constant, force it into the RHS.
1871 std::swap(LHS, RHS);
1872 Cond = SetCondInst::getSwappedCondition(Cond);
1873 }
1874
1875 // FIXME: think about handling pointer comparisons! i.e.:
1876 // while (P != P+100) ++P;
1877
1878 // If we have a comparison of a chrec against a constant, try to use value
1879 // ranges to answer this query.
1880 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
1881 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
1882 if (AddRec->getLoop() == L) {
1883 // Form the comparison range using the constant of the correct type so
1884 // that the ConstantRange class knows to do a signed or unsigned
1885 // comparison.
1886 ConstantInt *CompVal = RHSC->getValue();
1887 const Type *RealTy = ExitCond->getOperand(0)->getType();
1888 CompVal = dyn_cast<ConstantInt>(ConstantExpr::getCast(CompVal, RealTy));
1889 if (CompVal) {
1890 // Form the constant range.
1891 ConstantRange CompRange(Cond, CompVal);
1892
1893 // Now that we have it, if it's signed, convert it to an unsigned
1894 // range.
1895 if (CompRange.getLower()->getType()->isSigned()) {
1896 const Type *NewTy = RHSC->getValue()->getType();
1897 Constant *NewL = ConstantExpr::getCast(CompRange.getLower(), NewTy);
1898 Constant *NewU = ConstantExpr::getCast(CompRange.getUpper(), NewTy);
1899 CompRange = ConstantRange(NewL, NewU);
1900 }
1901
1902 SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange);
1903 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
1904 }
1905 }
1906
1907 switch (Cond) {
1908 case Instruction::SetNE: // while (X != Y)
1909 // Convert to: while (X-Y != 0)
1910 if (LHS->getType()->isInteger())
1911 return HowFarToZero(getMinusSCEV(LHS, RHS), L);
1912 break;
1913 case Instruction::SetEQ:
1914 // Convert to: while (X-Y == 0) // while (X == Y)
1915 if (LHS->getType()->isInteger())
1916 return HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
1917 break;
1918 default:
Chris Lattnerd18d9dc2004-04-02 20:26:46 +00001919#if 0
Chris Lattner53e677a2004-04-02 20:23:17 +00001920 std::cerr << "ComputeIterationCount ";
1921 if (ExitCond->getOperand(0)->getType()->isUnsigned())
1922 std::cerr << "[unsigned] ";
1923 std::cerr << *LHS << " "
1924 << Instruction::getOpcodeName(Cond) << " " << *RHS << "\n";
Chris Lattnerd18d9dc2004-04-02 20:26:46 +00001925#endif
Chris Lattnere34c0b42004-04-03 00:43:03 +00001926 break;
Chris Lattner53e677a2004-04-02 20:23:17 +00001927 }
1928 return UnknownValue;
1929}
1930
1931/// getSCEVAtScope - Compute the value of the specified expression within the
1932/// indicated loop (which may be null to indicate in no loop). If the
1933/// expression cannot be evaluated, return UnknownValue.
1934SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) {
1935 // FIXME: this should be turned into a virtual method on SCEV!
1936
1937 if (isa<SCEVConstant>(V) || isa<SCEVUnknown>(V)) return V;
1938 if (SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
1939 // Avoid performing the look-up in the common case where the specified
1940 // expression has no loop-variant portions.
1941 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
1942 SCEVHandle OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
1943 if (OpAtScope != Comm->getOperand(i)) {
1944 if (OpAtScope == UnknownValue) return UnknownValue;
1945 // Okay, at least one of these operands is loop variant but might be
1946 // foldable. Build a new instance of the folded commutative expression.
1947 std::vector<SCEVHandle> NewOps(Comm->op_begin(), Comm->op_begin()+i-1);
1948 NewOps.push_back(OpAtScope);
1949
1950 for (++i; i != e; ++i) {
1951 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
1952 if (OpAtScope == UnknownValue) return UnknownValue;
1953 NewOps.push_back(OpAtScope);
1954 }
1955 if (isa<SCEVAddExpr>(Comm))
1956 return SCEVAddExpr::get(NewOps);
1957 assert(isa<SCEVMulExpr>(Comm) && "Only know about add and mul!");
1958 return SCEVMulExpr::get(NewOps);
1959 }
1960 }
1961 // If we got here, all operands are loop invariant.
1962 return Comm;
1963 }
1964
1965 if (SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(V)) {
1966 SCEVHandle LHS = getSCEVAtScope(UDiv->getLHS(), L);
1967 if (LHS == UnknownValue) return LHS;
1968 SCEVHandle RHS = getSCEVAtScope(UDiv->getRHS(), L);
1969 if (RHS == UnknownValue) return RHS;
1970 if (LHS == UDiv->getLHS() && RHS == UDiv->getRHS())
1971 return UDiv; // must be loop invariant
1972 return SCEVUDivExpr::get(LHS, RHS);
1973 }
1974
1975 // If this is a loop recurrence for a loop that does not contain L, then we
1976 // are dealing with the final value computed by the loop.
1977 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
1978 if (!L || !AddRec->getLoop()->contains(L->getHeader())) {
1979 // To evaluate this recurrence, we need to know how many times the AddRec
1980 // loop iterates. Compute this now.
1981 SCEVHandle IterationCount = getIterationCount(AddRec->getLoop());
1982 if (IterationCount == UnknownValue) return UnknownValue;
1983 IterationCount = getTruncateOrZeroExtend(IterationCount,
1984 AddRec->getType());
1985
1986 // If the value is affine, simplify the expression evaluation to just
1987 // Start + Step*IterationCount.
1988 if (AddRec->isAffine())
1989 return SCEVAddExpr::get(AddRec->getStart(),
1990 SCEVMulExpr::get(IterationCount,
1991 AddRec->getOperand(1)));
1992
1993 // Otherwise, evaluate it the hard way.
1994 return AddRec->evaluateAtIteration(IterationCount);
1995 }
1996 return UnknownValue;
1997 }
1998
1999 //assert(0 && "Unknown SCEV type!");
2000 return UnknownValue;
2001}
2002
2003
2004/// SolveQuadraticEquation - Find the roots of the quadratic equation for the
2005/// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
2006/// might be the same) or two SCEVCouldNotCompute objects.
2007///
2008static std::pair<SCEVHandle,SCEVHandle>
2009SolveQuadraticEquation(const SCEVAddRecExpr *AddRec) {
2010 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
2011 SCEVConstant *L = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
2012 SCEVConstant *M = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
2013 SCEVConstant *N = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
2014
2015 // We currently can only solve this if the coefficients are constants.
2016 if (!L || !M || !N) {
2017 SCEV *CNC = new SCEVCouldNotCompute();
2018 return std::make_pair(CNC, CNC);
2019 }
2020
2021 Constant *Two = ConstantInt::get(L->getValue()->getType(), 2);
2022
2023 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
2024 Constant *C = L->getValue();
2025 // The B coefficient is M-N/2
2026 Constant *B = ConstantExpr::getSub(M->getValue(),
2027 ConstantExpr::getDiv(N->getValue(),
2028 Two));
2029 // The A coefficient is N/2
2030 Constant *A = ConstantExpr::getDiv(N->getValue(), Two);
2031
2032 // Compute the B^2-4ac term.
2033 Constant *SqrtTerm =
2034 ConstantExpr::getMul(ConstantInt::get(C->getType(), 4),
2035 ConstantExpr::getMul(A, C));
2036 SqrtTerm = ConstantExpr::getSub(ConstantExpr::getMul(B, B), SqrtTerm);
2037
2038 // Compute floor(sqrt(B^2-4ac))
2039 ConstantUInt *SqrtVal =
2040 cast<ConstantUInt>(ConstantExpr::getCast(SqrtTerm,
2041 SqrtTerm->getType()->getUnsignedVersion()));
2042 uint64_t SqrtValV = SqrtVal->getValue();
2043 uint64_t SqrtValV2 = (uint64_t)sqrtl(SqrtValV);
2044 // The square root might not be precise for arbitrary 64-bit integer
2045 // values. Do some sanity checks to ensure it's correct.
2046 if (SqrtValV2*SqrtValV2 > SqrtValV ||
2047 (SqrtValV2+1)*(SqrtValV2+1) <= SqrtValV) {
2048 SCEV *CNC = new SCEVCouldNotCompute();
2049 return std::make_pair(CNC, CNC);
2050 }
2051
2052 SqrtVal = ConstantUInt::get(Type::ULongTy, SqrtValV2);
2053 SqrtTerm = ConstantExpr::getCast(SqrtVal, SqrtTerm->getType());
2054
2055 Constant *NegB = ConstantExpr::getNeg(B);
2056 Constant *TwoA = ConstantExpr::getMul(A, Two);
2057
2058 // The divisions must be performed as signed divisions.
2059 const Type *SignedTy = NegB->getType()->getSignedVersion();
2060 NegB = ConstantExpr::getCast(NegB, SignedTy);
2061 TwoA = ConstantExpr::getCast(TwoA, SignedTy);
2062 SqrtTerm = ConstantExpr::getCast(SqrtTerm, SignedTy);
2063
2064 Constant *Solution1 =
2065 ConstantExpr::getDiv(ConstantExpr::getAdd(NegB, SqrtTerm), TwoA);
2066 Constant *Solution2 =
2067 ConstantExpr::getDiv(ConstantExpr::getSub(NegB, SqrtTerm), TwoA);
2068 return std::make_pair(SCEVUnknown::get(Solution1),
2069 SCEVUnknown::get(Solution2));
2070}
2071
2072/// HowFarToZero - Return the number of times a backedge comparing the specified
2073/// value to zero will execute. If not computable, return UnknownValue
2074SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) {
2075 // If the value is a constant
2076 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2077 // If the value is already zero, the branch will execute zero times.
2078 if (C->getValue()->isNullValue()) return C;
2079 return UnknownValue; // Otherwise it will loop infinitely.
2080 }
2081
2082 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
2083 if (!AddRec || AddRec->getLoop() != L)
2084 return UnknownValue;
2085
2086 if (AddRec->isAffine()) {
2087 // If this is an affine expression the execution count of this branch is
2088 // equal to:
2089 //
2090 // (0 - Start/Step) iff Start % Step == 0
2091 //
2092 // Get the initial value for the loop.
2093 SCEVHandle Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
2094 SCEVHandle Step = AddRec->getOperand(1);
2095
2096 Step = getSCEVAtScope(Step, L->getParentLoop());
2097
2098 // Figure out if Start % Step == 0.
2099 // FIXME: We should add DivExpr and RemExpr operations to our AST.
2100 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
2101 if (StepC->getValue()->equalsInt(1)) // N % 1 == 0
2102 return getNegativeSCEV(Start); // 0 - Start/1 == -Start
2103 if (StepC->getValue()->isAllOnesValue()) // N % -1 == 0
2104 return Start; // 0 - Start/-1 == Start
2105
2106 // Check to see if Start is divisible by SC with no remainder.
2107 if (SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start)) {
2108 ConstantInt *StartCC = StartC->getValue();
2109 Constant *StartNegC = ConstantExpr::getNeg(StartCC);
2110 Constant *Rem = ConstantExpr::getRem(StartNegC, StepC->getValue());
2111 if (Rem->isNullValue()) {
2112 Constant *Result =ConstantExpr::getDiv(StartNegC,StepC->getValue());
2113 return SCEVUnknown::get(Result);
2114 }
2115 }
2116 }
2117 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
2118 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
2119 // the quadratic equation to solve it.
2120 std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec);
2121 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2122 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2123 if (R1) {
Chris Lattnerd18d9dc2004-04-02 20:26:46 +00002124#if 0
Chris Lattner53e677a2004-04-02 20:23:17 +00002125 std::cerr << "HFTZ: " << *V << " - sol#1: " << *R1
2126 << " sol#2: " << *R2 << "\n";
Chris Lattnerd18d9dc2004-04-02 20:26:46 +00002127#endif
Chris Lattner53e677a2004-04-02 20:23:17 +00002128 // Pick the smallest positive root value.
2129 assert(R1->getType()->isUnsigned()&&"Didn't canonicalize to unsigned?");
2130 if (ConstantBool *CB =
2131 dyn_cast<ConstantBool>(ConstantExpr::getSetLT(R1->getValue(),
2132 R2->getValue()))) {
2133 if (CB != ConstantBool::True)
2134 std::swap(R1, R2); // R1 is the minimum root now.
2135
2136 // We can only use this value if the chrec ends up with an exact zero
2137 // value at this index. When solving for "X*X != 5", for example, we
2138 // should not accept a root of 2.
2139 SCEVHandle Val = AddRec->evaluateAtIteration(R1);
2140 if (SCEVConstant *EvalVal = dyn_cast<SCEVConstant>(Val))
2141 if (EvalVal->getValue()->isNullValue())
2142 return R1; // We found a quadratic root!
2143 }
2144 }
2145 }
2146
2147 return UnknownValue;
2148}
2149
2150/// HowFarToNonZero - Return the number of times a backedge checking the
2151/// specified value for nonzero will execute. If not computable, return
2152/// UnknownValue
2153SCEVHandle ScalarEvolutionsImpl::HowFarToNonZero(SCEV *V, const Loop *L) {
2154 // Loops that look like: while (X == 0) are very strange indeed. We don't
2155 // handle them yet except for the trivial case. This could be expanded in the
2156 // future as needed.
2157
2158 // If the value is a constant, check to see if it is known to be non-zero
2159 // already. If so, the backedge will execute zero times.
2160 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
2161 Constant *Zero = Constant::getNullValue(C->getValue()->getType());
2162 Constant *NonZero = ConstantExpr::getSetNE(C->getValue(), Zero);
2163 if (NonZero == ConstantBool::True)
2164 return getSCEV(Zero);
2165 return UnknownValue; // Otherwise it will loop infinitely.
2166 }
2167
2168 // We could implement others, but I really doubt anyone writes loops like
2169 // this, and if they did, they would already be constant folded.
2170 return UnknownValue;
2171}
2172
2173static ConstantInt *
2174EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, Constant *C) {
2175 SCEVHandle InVal = SCEVConstant::get(cast<ConstantInt>(C));
2176 SCEVHandle Val = AddRec->evaluateAtIteration(InVal);
2177 assert(isa<SCEVConstant>(Val) &&
2178 "Evaluation of SCEV at constant didn't fold correctly?");
2179 return cast<SCEVConstant>(Val)->getValue();
2180}
2181
2182
2183/// getNumIterationsInRange - Return the number of iterations of this loop that
2184/// produce values in the specified constant range. Another way of looking at
2185/// this is that it returns the first iteration number where the value is not in
2186/// the condition, thus computing the exit count. If the iteration count can't
2187/// be computed, an instance of SCEVCouldNotCompute is returned.
2188SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range) const {
2189 if (Range.isFullSet()) // Infinite loop.
2190 return new SCEVCouldNotCompute();
2191
2192 // If the start is a non-zero constant, shift the range to simplify things.
2193 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
2194 if (!SC->getValue()->isNullValue()) {
2195 std::vector<SCEVHandle> Operands(op_begin(), op_end());
2196 Operands[0] = getIntegerSCEV(0, SC->getType());
2197 SCEVHandle Shifted = SCEVAddRecExpr::get(Operands, getLoop());
2198 if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast<SCEVAddRecExpr>(Shifted))
2199 return ShiftedAddRec->getNumIterationsInRange(
2200 Range.subtract(SC->getValue()));
2201 // This is strange and shouldn't happen.
2202 return new SCEVCouldNotCompute();
2203 }
2204
2205 // The only time we can solve this is when we have all constant indices.
2206 // Otherwise, we cannot determine the overflow conditions.
2207 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
2208 if (!isa<SCEVConstant>(getOperand(i)))
2209 return new SCEVCouldNotCompute();
2210
2211
2212 // Okay at this point we know that all elements of the chrec are constants and
2213 // that the start element is zero.
2214
2215 // First check to see if the range contains zero. If not, the first
2216 // iteration exits.
2217 ConstantInt *Zero = ConstantInt::get(getType(), 0);
2218 if (!Range.contains(Zero)) return SCEVConstant::get(Zero);
2219
2220 if (isAffine()) {
2221 // If this is an affine expression then we have this situation:
2222 // Solve {0,+,A} in Range === Ax in Range
2223
2224 // Since we know that zero is in the range, we know that the upper value of
2225 // the range must be the first possible exit value. Also note that we
2226 // already checked for a full range.
2227 ConstantInt *Upper = cast<ConstantInt>(Range.getUpper());
2228 ConstantInt *A = cast<SCEVConstant>(getOperand(1))->getValue();
2229 ConstantInt *One = ConstantInt::get(getType(), 1);
2230
2231 // The exit value should be (Upper+A-1)/A.
2232 Constant *ExitValue = Upper;
2233 if (A != One) {
2234 ExitValue = ConstantExpr::getSub(ConstantExpr::getAdd(Upper, A), One);
2235 ExitValue = ConstantExpr::getDiv(ExitValue, A);
2236 }
2237 assert(isa<ConstantInt>(ExitValue) &&
2238 "Constant folding of integers not implemented?");
2239
2240 // Evaluate at the exit value. If we really did fall out of the valid
2241 // range, then we computed our trip count, otherwise wrap around or other
2242 // things must have happened.
2243 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue);
2244 if (Range.contains(Val))
2245 return new SCEVCouldNotCompute(); // Something strange happened
2246
2247 // Ensure that the previous value is in the range. This is a sanity check.
2248 assert(Range.contains(EvaluateConstantChrecAtConstant(this,
2249 ConstantExpr::getSub(ExitValue, One))) &&
2250 "Linear scev computation is off in a bad way!");
2251 return SCEVConstant::get(cast<ConstantInt>(ExitValue));
2252 } else if (isQuadratic()) {
2253 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
2254 // quadratic equation to solve it. To do this, we must frame our problem in
2255 // terms of figuring out when zero is crossed, instead of when
2256 // Range.getUpper() is crossed.
2257 std::vector<SCEVHandle> NewOps(op_begin(), op_end());
2258 NewOps[0] = getNegativeSCEV(SCEVUnknown::get(Range.getUpper()));
2259 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, getLoop());
2260
2261 // Next, solve the constructed addrec
2262 std::pair<SCEVHandle,SCEVHandle> Roots =
2263 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec));
2264 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
2265 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
2266 if (R1) {
2267 // Pick the smallest positive root value.
2268 assert(R1->getType()->isUnsigned() && "Didn't canonicalize to unsigned?");
2269 if (ConstantBool *CB =
2270 dyn_cast<ConstantBool>(ConstantExpr::getSetLT(R1->getValue(),
2271 R2->getValue()))) {
2272 if (CB != ConstantBool::True)
2273 std::swap(R1, R2); // R1 is the minimum root now.
2274
2275 // Make sure the root is not off by one. The returned iteration should
2276 // not be in the range, but the previous one should be. When solving
2277 // for "X*X < 5", for example, we should not return a root of 2.
2278 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
2279 R1->getValue());
2280 if (Range.contains(R1Val)) {
2281 // The next iteration must be out of the range...
2282 Constant *NextVal =
2283 ConstantExpr::getAdd(R1->getValue(),
2284 ConstantInt::get(R1->getType(), 1));
2285
2286 R1Val = EvaluateConstantChrecAtConstant(this, NextVal);
2287 if (!Range.contains(R1Val))
2288 return SCEVUnknown::get(NextVal);
2289 return new SCEVCouldNotCompute(); // Something strange happened
2290 }
2291
2292 // If R1 was not in the range, then it is a good return value. Make
2293 // sure that R1-1 WAS in the range though, just in case.
2294 Constant *NextVal =
2295 ConstantExpr::getSub(R1->getValue(),
2296 ConstantInt::get(R1->getType(), 1));
2297 R1Val = EvaluateConstantChrecAtConstant(this, NextVal);
2298 if (Range.contains(R1Val))
2299 return R1;
2300 return new SCEVCouldNotCompute(); // Something strange happened
2301 }
2302 }
2303 }
2304
2305 // Fallback, if this is a general polynomial, figure out the progression
2306 // through brute force: evaluate until we find an iteration that fails the
2307 // test. This is likely to be slow, but getting an accurate trip count is
2308 // incredibly important, we will be able to simplify the exit test a lot, and
2309 // we are almost guaranteed to get a trip count in this case.
2310 ConstantInt *TestVal = ConstantInt::get(getType(), 0);
2311 ConstantInt *One = ConstantInt::get(getType(), 1);
2312 ConstantInt *EndVal = TestVal; // Stop when we wrap around.
2313 do {
2314 ++NumBruteForceEvaluations;
2315 SCEVHandle Val = evaluateAtIteration(SCEVConstant::get(TestVal));
2316 if (!isa<SCEVConstant>(Val)) // This shouldn't happen.
2317 return new SCEVCouldNotCompute();
2318
2319 // Check to see if we found the value!
2320 if (!Range.contains(cast<SCEVConstant>(Val)->getValue()))
2321 return SCEVConstant::get(TestVal);
2322
2323 // Increment to test the next index.
2324 TestVal = cast<ConstantInt>(ConstantExpr::getAdd(TestVal, One));
2325 } while (TestVal != EndVal);
2326
2327 return new SCEVCouldNotCompute();
2328}
2329
2330
2331
2332//===----------------------------------------------------------------------===//
2333// ScalarEvolution Class Implementation
2334//===----------------------------------------------------------------------===//
2335
2336bool ScalarEvolution::runOnFunction(Function &F) {
2337 Impl = new ScalarEvolutionsImpl(F, getAnalysis<LoopInfo>());
2338 return false;
2339}
2340
2341void ScalarEvolution::releaseMemory() {
2342 delete (ScalarEvolutionsImpl*)Impl;
2343 Impl = 0;
2344}
2345
2346void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
2347 AU.setPreservesAll();
2348 AU.addRequiredID(LoopSimplifyID);
2349 AU.addRequiredTransitive<LoopInfo>();
2350}
2351
2352SCEVHandle ScalarEvolution::getSCEV(Value *V) const {
2353 return ((ScalarEvolutionsImpl*)Impl)->getSCEV(V);
2354}
2355
2356SCEVHandle ScalarEvolution::getIterationCount(const Loop *L) const {
2357 return ((ScalarEvolutionsImpl*)Impl)->getIterationCount(L);
2358}
2359
2360bool ScalarEvolution::hasLoopInvariantIterationCount(const Loop *L) const {
2361 return !isa<SCEVCouldNotCompute>(getIterationCount(L));
2362}
2363
2364SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) const {
2365 return ((ScalarEvolutionsImpl*)Impl)->getSCEVAtScope(getSCEV(V), L);
2366}
2367
2368void ScalarEvolution::deleteInstructionFromRecords(Instruction *I) const {
2369 return ((ScalarEvolutionsImpl*)Impl)->deleteInstructionFromRecords(I);
2370}
2371
2372
2373/// shouldSubstituteIndVar - Return true if we should perform induction variable
2374/// substitution for this variable. This is a hack because we don't have a
2375/// strength reduction pass yet. When we do we will promote all vars, because
2376/// we can strength reduce them later as desired.
2377bool ScalarEvolution::shouldSubstituteIndVar(const SCEV *S) const {
2378 // Don't substitute high degree polynomials.
2379 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S))
2380 if (AddRec->getNumOperands() > 3) return false;
2381 return true;
2382}
2383
2384
2385static void PrintLoopInfo(std::ostream &OS, const ScalarEvolution *SE,
2386 const Loop *L) {
2387 // Print all inner loops first
2388 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
2389 PrintLoopInfo(OS, SE, *I);
2390
2391 std::cerr << "Loop " << L->getHeader()->getName() << ": ";
2392 if (L->getExitBlocks().size() != 1)
2393 std::cerr << "<multiple exits> ";
2394
2395 if (SE->hasLoopInvariantIterationCount(L)) {
2396 std::cerr << *SE->getIterationCount(L) << " iterations! ";
2397 } else {
2398 std::cerr << "Unpredictable iteration count. ";
2399 }
2400
2401 std::cerr << "\n";
2402}
2403
2404void ScalarEvolution::print(std::ostream &OS) const {
2405 Function &F = ((ScalarEvolutionsImpl*)Impl)->F;
2406 LoopInfo &LI = ((ScalarEvolutionsImpl*)Impl)->LI;
2407
2408 OS << "Classifying expressions for: " << F.getName() << "\n";
2409 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
2410 if ((*I)->getType()->isInteger()) {
2411 OS << **I;
2412 OS << " --> ";
2413 SCEVHandle SV = getSCEV(*I);
2414 SV->print(OS);
2415 OS << "\t\t";
2416
2417 if ((*I)->getType()->isIntegral()) {
2418 ConstantRange Bounds = SV->getValueRange();
2419 if (!Bounds.isFullSet())
2420 OS << "Bounds: " << Bounds << " ";
2421 }
2422
2423 if (const Loop *L = LI.getLoopFor((*I)->getParent())) {
2424 OS << "Exits: ";
2425 SCEVHandle ExitValue = getSCEVAtScope(*I, L->getParentLoop());
2426 if (isa<SCEVCouldNotCompute>(ExitValue)) {
2427 OS << "<<Unknown>>";
2428 } else {
2429 OS << *ExitValue;
2430 }
2431 }
2432
2433
2434 OS << "\n";
2435 }
2436
2437 OS << "Determining loop execution counts for: " << F.getName() << "\n";
2438 for (LoopInfo::iterator I = LI.begin(), E = LI.end(); I != E; ++I)
2439 PrintLoopInfo(OS, this, *I);
2440}
2441
2442//===----------------------------------------------------------------------===//
2443// ScalarEvolutionRewriter Class Implementation
2444//===----------------------------------------------------------------------===//
2445
2446Value *ScalarEvolutionRewriter::
2447GetOrInsertCanonicalInductionVariable(const Loop *L, const Type *Ty) {
2448 assert((Ty->isInteger() || Ty->isFloatingPoint()) &&
2449 "Can only insert integer or floating point induction variables!");
2450
2451 // Check to see if we already inserted one.
2452 SCEVHandle H = SCEVAddRecExpr::get(getIntegerSCEV(0, Ty),
2453 getIntegerSCEV(1, Ty), L);
2454 return ExpandCodeFor(H, 0, Ty);
2455}
2456
2457/// ExpandCodeFor - Insert code to directly compute the specified SCEV
2458/// expression into the program. The inserted code is inserted into the
2459/// specified block.
2460Value *ScalarEvolutionRewriter::ExpandCodeFor(SCEVHandle SH,
2461 Instruction *InsertPt,
2462 const Type *Ty) {
2463 std::map<SCEVHandle, Value*>::iterator ExistVal =InsertedExpressions.find(SH);
2464 Value *V;
2465 if (ExistVal != InsertedExpressions.end()) {
2466 V = ExistVal->second;
2467 } else {
2468 // Ask the recurrence object to expand the code for itself.
2469 V = SH->expandCodeFor(*this, InsertPt);
2470 // Cache the generated result.
2471 InsertedExpressions.insert(std::make_pair(SH, V));
2472 }
2473
2474 if (Ty == 0 || V->getType() == Ty)
2475 return V;
2476 if (Constant *C = dyn_cast<Constant>(V))
2477 return ConstantExpr::getCast(C, Ty);
2478 else if (Instruction *I = dyn_cast<Instruction>(V)) {
2479 // FIXME: check to see if there is already a cast!
2480 BasicBlock::iterator IP = I; ++IP;
Chris Lattnerddd947f2004-04-05 18:46:55 +00002481 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
2482 IP = II->getNormalDest()->begin();
Chris Lattner53e677a2004-04-02 20:23:17 +00002483 while (isa<PHINode>(IP)) ++IP;
2484 return new CastInst(V, Ty, V->getName(), IP);
2485 } else {
2486 // FIXME: check to see if there is already a cast!
2487 return new CastInst(V, Ty, V->getName(), InsertPt);
2488 }
2489}