LLVM  6.0.0svn
LoopStrengthReduce.cpp
Go to the documentation of this file.
1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
12 // on the target.
13 //
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
19 //
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
23 //
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
25 // ...
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
28 //
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
35 //
36 // TODO: More sophistication in the way Formulae are generated and filtered.
37 //
38 // TODO: Handle multiple loops at a time.
39 //
40 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
41 // of a GlobalValue?
42 //
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
45 //
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
52 // use kind.
53 //
54 //===----------------------------------------------------------------------===//
55 
57 #include "llvm/ADT/APInt.h"
58 #include "llvm/ADT/DenseMap.h"
59 #include "llvm/ADT/DenseSet.h"
60 #include "llvm/ADT/Hashing.h"
62 #include "llvm/ADT/STLExtras.h"
63 #include "llvm/ADT/SetVector.h"
65 #include "llvm/ADT/SmallPtrSet.h"
66 #include "llvm/ADT/SmallSet.h"
67 #include "llvm/ADT/SmallVector.h"
68 #include "llvm/Analysis/IVUsers.h"
69 #include "llvm/Analysis/LoopInfo.h"
70 #include "llvm/Analysis/LoopPass.h"
76 #include "llvm/IR/BasicBlock.h"
77 #include "llvm/IR/Constant.h"
78 #include "llvm/IR/Constants.h"
79 #include "llvm/IR/DerivedTypes.h"
80 #include "llvm/IR/Dominators.h"
81 #include "llvm/IR/GlobalValue.h"
82 #include "llvm/IR/IRBuilder.h"
83 #include "llvm/IR/Instruction.h"
84 #include "llvm/IR/Instructions.h"
85 #include "llvm/IR/IntrinsicInst.h"
86 #include "llvm/IR/Module.h"
87 #include "llvm/IR/OperandTraits.h"
88 #include "llvm/IR/Operator.h"
89 #include "llvm/IR/Type.h"
90 #include "llvm/IR/Value.h"
91 #include "llvm/IR/ValueHandle.h"
92 #include "llvm/Pass.h"
93 #include "llvm/Support/Casting.h"
95 #include "llvm/Support/Compiler.h"
96 #include "llvm/Support/Debug.h"
100 #include "llvm/Transforms/Scalar.h"
104 #include <algorithm>
105 #include <cassert>
106 #include <cstddef>
107 #include <cstdint>
108 #include <cstdlib>
109 #include <iterator>
110 #include <map>
111 #include <tuple>
112 #include <utility>
113 
114 using namespace llvm;
115 
116 #define DEBUG_TYPE "loop-reduce"
117 
118 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
119 /// bail out. This threshold is far beyond the number of users that LSR can
120 /// conceivably solve, so it should not affect generated code, but catches the
121 /// worst cases before LSR burns too much compile time and stack space.
122 static const unsigned MaxIVUsers = 200;
123 
124 // Temporary flag to cleanup congruent phis after LSR phi expansion.
125 // It's currently disabled until we can determine whether it's truly useful or
126 // not. The flag should be removed after the v3.0 release.
127 // This is now needed for ivchains.
129  "enable-lsr-phielim", cl::Hidden, cl::init(true),
130  cl::desc("Enable LSR phi elimination"));
131 
132 // The flag adds instruction count to solutions cost comparision.
133 static cl::opt<bool> InsnsCost(
134  "lsr-insns-cost", cl::Hidden, cl::init(false),
135  cl::desc("Add instruction count to a LSR cost model"));
136 
137 // Flag to choose how to narrow complex lsr solution
139  "lsr-exp-narrow", cl::Hidden, cl::init(false),
140  cl::desc("Narrow LSR complex solution using"
141  " expectation of registers number"));
142 
143 // Flag to narrow search space by filtering non-optimal formulae with
144 // the same ScaledReg and Scale.
146  "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
147  cl::desc("Narrow LSR search space by filtering non-optimal formulae"
148  " with the same ScaledReg and Scale"));
149 
150 #ifndef NDEBUG
151 // Stress test IV chain generation.
153  "stress-ivchain", cl::Hidden, cl::init(false),
154  cl::desc("Stress test LSR IV chains"));
155 #else
156 static bool StressIVChain = false;
157 #endif
158 
159 namespace {
160 
161 struct MemAccessTy {
162  /// Used in situations where the accessed memory type is unknown.
163  static const unsigned UnknownAddressSpace = ~0u;
164 
165  Type *MemTy;
166  unsigned AddrSpace;
167 
168  MemAccessTy() : MemTy(nullptr), AddrSpace(UnknownAddressSpace) {}
169 
170  MemAccessTy(Type *Ty, unsigned AS) :
171  MemTy(Ty), AddrSpace(AS) {}
172 
173  bool operator==(MemAccessTy Other) const {
174  return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
175  }
176 
177  bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
178 
179  static MemAccessTy getUnknown(LLVMContext &Ctx,
180  unsigned AS = UnknownAddressSpace) {
181  return MemAccessTy(Type::getVoidTy(Ctx), AS);
182  }
183 };
184 
185 /// This class holds data which is used to order reuse candidates.
186 class RegSortData {
187 public:
188  /// This represents the set of LSRUse indices which reference
189  /// a particular register.
190  SmallBitVector UsedByIndices;
191 
192  void print(raw_ostream &OS) const;
193  void dump() const;
194 };
195 
196 } // end anonymous namespace
197 
198 void RegSortData::print(raw_ostream &OS) const {
199  OS << "[NumUses=" << UsedByIndices.count() << ']';
200 }
201 
202 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
203 LLVM_DUMP_METHOD void RegSortData::dump() const {
204  print(errs()); errs() << '\n';
205 }
206 #endif
207 
208 namespace {
209 
210 /// Map register candidates to information about how they are used.
211 class RegUseTracker {
212  typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
213 
214  RegUsesTy RegUsesMap;
216 
217 public:
218  void countRegister(const SCEV *Reg, size_t LUIdx);
219  void dropRegister(const SCEV *Reg, size_t LUIdx);
220  void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
221 
222  bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
223 
224  const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
225 
226  void clear();
227 
229  typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
230  iterator begin() { return RegSequence.begin(); }
231  iterator end() { return RegSequence.end(); }
232  const_iterator begin() const { return RegSequence.begin(); }
233  const_iterator end() const { return RegSequence.end(); }
234 };
235 
236 } // end anonymous namespace
237 
238 void
239 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
240  std::pair<RegUsesTy::iterator, bool> Pair =
241  RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
242  RegSortData &RSD = Pair.first->second;
243  if (Pair.second)
244  RegSequence.push_back(Reg);
245  RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
246  RSD.UsedByIndices.set(LUIdx);
247 }
248 
249 void
250 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
251  RegUsesTy::iterator It = RegUsesMap.find(Reg);
252  assert(It != RegUsesMap.end());
253  RegSortData &RSD = It->second;
254  assert(RSD.UsedByIndices.size() > LUIdx);
255  RSD.UsedByIndices.reset(LUIdx);
256 }
257 
258 void
259 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
260  assert(LUIdx <= LastLUIdx);
261 
262  // Update RegUses. The data structure is not optimized for this purpose;
263  // we must iterate through it and update each of the bit vectors.
264  for (auto &Pair : RegUsesMap) {
265  SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
266  if (LUIdx < UsedByIndices.size())
267  UsedByIndices[LUIdx] =
268  LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
269  UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
270  }
271 }
272 
273 bool
274 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
275  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
276  if (I == RegUsesMap.end())
277  return false;
278  const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
279  int i = UsedByIndices.find_first();
280  if (i == -1) return false;
281  if ((size_t)i != LUIdx) return true;
282  return UsedByIndices.find_next(i) != -1;
283 }
284 
285 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
286  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
287  assert(I != RegUsesMap.end() && "Unknown register!");
288  return I->second.UsedByIndices;
289 }
290 
291 void RegUseTracker::clear() {
292  RegUsesMap.clear();
293  RegSequence.clear();
294 }
295 
296 namespace {
297 
298 /// This class holds information that describes a formula for computing
299 /// satisfying a use. It may include broken-out immediates and scaled registers.
300 struct Formula {
301  /// Global base address used for complex addressing.
302  GlobalValue *BaseGV;
303 
304  /// Base offset for complex addressing.
305  int64_t BaseOffset;
306 
307  /// Whether any complex addressing has a base register.
308  bool HasBaseReg;
309 
310  /// The scale of any complex addressing.
311  int64_t Scale;
312 
313  /// The list of "base" registers for this use. When this is non-empty. The
314  /// canonical representation of a formula is
315  /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
316  /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
317  /// 3. The reg containing recurrent expr related with currect loop in the
318  /// formula should be put in the ScaledReg.
319  /// #1 enforces that the scaled register is always used when at least two
320  /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
321  /// #2 enforces that 1 * reg is reg.
322  /// #3 ensures invariant regs with respect to current loop can be combined
323  /// together in LSR codegen.
324  /// This invariant can be temporarly broken while building a formula.
325  /// However, every formula inserted into the LSRInstance must be in canonical
326  /// form.
328 
329  /// The 'scaled' register for this use. This should be non-null when Scale is
330  /// not zero.
331  const SCEV *ScaledReg;
332 
333  /// An additional constant offset which added near the use. This requires a
334  /// temporary register, but the offset itself can live in an add immediate
335  /// field rather than a register.
336  int64_t UnfoldedOffset;
337 
338  Formula()
339  : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
340  ScaledReg(nullptr), UnfoldedOffset(0) {}
341 
342  void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
343 
344  bool isCanonical(const Loop &L) const;
345 
346  void canonicalize(const Loop &L);
347 
348  bool unscale();
349 
350  bool hasZeroEnd() const;
351 
352  size_t getNumRegs() const;
353  Type *getType() const;
354 
355  void deleteBaseReg(const SCEV *&S);
356 
357  bool referencesReg(const SCEV *S) const;
358  bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
359  const RegUseTracker &RegUses) const;
360 
361  void print(raw_ostream &OS) const;
362  void dump() const;
363 };
364 
365 } // end anonymous namespace
366 
367 /// Recursion helper for initialMatch.
368 static void DoInitialMatch(const SCEV *S, Loop *L,
371  ScalarEvolution &SE) {
372  // Collect expressions which properly dominate the loop header.
373  if (SE.properlyDominates(S, L->getHeader())) {
374  Good.push_back(S);
375  return;
376  }
377 
378  // Look at add operands.
379  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
380  for (const SCEV *S : Add->operands())
381  DoInitialMatch(S, L, Good, Bad, SE);
382  return;
383  }
384 
385  // Look at addrec operands.
386  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
387  if (!AR->getStart()->isZero() && AR->isAffine()) {
388  DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
389  DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
390  AR->getStepRecurrence(SE),
391  // FIXME: AR->getNoWrapFlags()
392  AR->getLoop(), SCEV::FlagAnyWrap),
393  L, Good, Bad, SE);
394  return;
395  }
396 
397  // Handle a multiplication by -1 (negation) if it didn't fold.
398  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
399  if (Mul->getOperand(0)->isAllOnesValue()) {
400  SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
401  const SCEV *NewMul = SE.getMulExpr(Ops);
402 
405  DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
406  const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
407  SE.getEffectiveSCEVType(NewMul->getType())));
408  for (const SCEV *S : MyGood)
409  Good.push_back(SE.getMulExpr(NegOne, S));
410  for (const SCEV *S : MyBad)
411  Bad.push_back(SE.getMulExpr(NegOne, S));
412  return;
413  }
414 
415  // Ok, we can't do anything interesting. Just stuff the whole thing into a
416  // register and hope for the best.
417  Bad.push_back(S);
418 }
419 
420 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
421 /// all loop-invariant and loop-computable values in a single base register.
422 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
425  DoInitialMatch(S, L, Good, Bad, SE);
426  if (!Good.empty()) {
427  const SCEV *Sum = SE.getAddExpr(Good);
428  if (!Sum->isZero())
429  BaseRegs.push_back(Sum);
430  HasBaseReg = true;
431  }
432  if (!Bad.empty()) {
433  const SCEV *Sum = SE.getAddExpr(Bad);
434  if (!Sum->isZero())
435  BaseRegs.push_back(Sum);
436  HasBaseReg = true;
437  }
438  canonicalize(*L);
439 }
440 
441 /// \brief Check whether or not this formula statisfies the canonical
442 /// representation.
443 /// \see Formula::BaseRegs.
444 bool Formula::isCanonical(const Loop &L) const {
445  if (!ScaledReg)
446  return BaseRegs.size() <= 1;
447 
448  if (Scale != 1)
449  return true;
450 
451  if (Scale == 1 && BaseRegs.empty())
452  return false;
453 
454  const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
455  if (SAR && SAR->getLoop() == &L)
456  return true;
457 
458  // If ScaledReg is not a recurrent expr, or it is but its loop is not current
459  // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
460  // loop, we want to swap the reg in BaseRegs with ScaledReg.
461  auto I =
462  find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
463  return isa<const SCEVAddRecExpr>(S) &&
464  (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
465  });
466  return I == BaseRegs.end();
467 }
468 
469 /// \brief Helper method to morph a formula into its canonical representation.
470 /// \see Formula::BaseRegs.
471 /// Every formula having more than one base register, must use the ScaledReg
472 /// field. Otherwise, we would have to do special cases everywhere in LSR
473 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
474 /// On the other hand, 1*reg should be canonicalized into reg.
475 void Formula::canonicalize(const Loop &L) {
476  if (isCanonical(L))
477  return;
478  // So far we did not need this case. This is easy to implement but it is
479  // useless to maintain dead code. Beside it could hurt compile time.
480  assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
481 
482  // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
483  if (!ScaledReg) {
484  ScaledReg = BaseRegs.back();
485  BaseRegs.pop_back();
486  Scale = 1;
487  }
488 
489  // If ScaledReg is an invariant with respect to L, find the reg from
490  // BaseRegs containing the recurrent expr related with Loop L. Swap the
491  // reg with ScaledReg.
492  const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
493  if (!SAR || SAR->getLoop() != &L) {
494  auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
495  [&](const SCEV *S) {
496  return isa<const SCEVAddRecExpr>(S) &&
497  (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
498  });
499  if (I != BaseRegs.end())
500  std::swap(ScaledReg, *I);
501  }
502 }
503 
504 /// \brief Get rid of the scale in the formula.
505 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
506 /// \return true if it was possible to get rid of the scale, false otherwise.
507 /// \note After this operation the formula may not be in the canonical form.
508 bool Formula::unscale() {
509  if (Scale != 1)
510  return false;
511  Scale = 0;
512  BaseRegs.push_back(ScaledReg);
513  ScaledReg = nullptr;
514  return true;
515 }
516 
517 bool Formula::hasZeroEnd() const {
518  if (UnfoldedOffset || BaseOffset)
519  return false;
520  if (BaseRegs.size() != 1 || ScaledReg)
521  return false;
522  return true;
523 }
524 
525 /// Return the total number of register operands used by this formula. This does
526 /// not include register uses implied by non-constant addrec strides.
527 size_t Formula::getNumRegs() const {
528  return !!ScaledReg + BaseRegs.size();
529 }
530 
531 /// Return the type of this formula, if it has one, or null otherwise. This type
532 /// is meaningless except for the bit size.
533 Type *Formula::getType() const {
534  return !BaseRegs.empty() ? BaseRegs.front()->getType() :
535  ScaledReg ? ScaledReg->getType() :
536  BaseGV ? BaseGV->getType() :
537  nullptr;
538 }
539 
540 /// Delete the given base reg from the BaseRegs list.
541 void Formula::deleteBaseReg(const SCEV *&S) {
542  if (&S != &BaseRegs.back())
543  std::swap(S, BaseRegs.back());
544  BaseRegs.pop_back();
545 }
546 
547 /// Test if this formula references the given register.
548 bool Formula::referencesReg(const SCEV *S) const {
549  return S == ScaledReg || is_contained(BaseRegs, S);
550 }
551 
552 /// Test whether this formula uses registers which are used by uses other than
553 /// the use with the given index.
554 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
555  const RegUseTracker &RegUses) const {
556  if (ScaledReg)
557  if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
558  return true;
559  for (const SCEV *BaseReg : BaseRegs)
560  if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
561  return true;
562  return false;
563 }
564 
565 void Formula::print(raw_ostream &OS) const {
566  bool First = true;
567  if (BaseGV) {
568  if (!First) OS << " + "; else First = false;
569  BaseGV->printAsOperand(OS, /*PrintType=*/false);
570  }
571  if (BaseOffset != 0) {
572  if (!First) OS << " + "; else First = false;
573  OS << BaseOffset;
574  }
575  for (const SCEV *BaseReg : BaseRegs) {
576  if (!First) OS << " + "; else First = false;
577  OS << "reg(" << *BaseReg << ')';
578  }
579  if (HasBaseReg && BaseRegs.empty()) {
580  if (!First) OS << " + "; else First = false;
581  OS << "**error: HasBaseReg**";
582  } else if (!HasBaseReg && !BaseRegs.empty()) {
583  if (!First) OS << " + "; else First = false;
584  OS << "**error: !HasBaseReg**";
585  }
586  if (Scale != 0) {
587  if (!First) OS << " + "; else First = false;
588  OS << Scale << "*reg(";
589  if (ScaledReg)
590  OS << *ScaledReg;
591  else
592  OS << "<unknown>";
593  OS << ')';
594  }
595  if (UnfoldedOffset != 0) {
596  if (!First) OS << " + ";
597  OS << "imm(" << UnfoldedOffset << ')';
598  }
599 }
600 
601 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
602 LLVM_DUMP_METHOD void Formula::dump() const {
603  print(errs()); errs() << '\n';
604 }
605 #endif
606 
607 /// Return true if the given addrec can be sign-extended without changing its
608 /// value.
609 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
610  Type *WideTy =
612  return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
613 }
614 
615 /// Return true if the given add can be sign-extended without changing its
616 /// value.
617 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
618  Type *WideTy =
620  return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
621 }
622 
623 /// Return true if the given mul can be sign-extended without changing its
624 /// value.
625 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
626  Type *WideTy =
628  SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
629  return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
630 }
631 
632 /// Return an expression for LHS /s RHS, if it can be determined and if the
633 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
634 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
635 /// the multiplication may overflow, which is useful when the result will be
636 /// used in a context where the most significant bits are ignored.
637 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
638  ScalarEvolution &SE,
639  bool IgnoreSignificantBits = false) {
640  // Handle the trivial case, which works for any SCEV type.
641  if (LHS == RHS)
642  return SE.getConstant(LHS->getType(), 1);
643 
644  // Handle a few RHS special cases.
645  const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
646  if (RC) {
647  const APInt &RA = RC->getAPInt();
648  // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
649  // some folding.
650  if (RA.isAllOnesValue())
651  return SE.getMulExpr(LHS, RC);
652  // Handle x /s 1 as x.
653  if (RA == 1)
654  return LHS;
655  }
656 
657  // Check for a division of a constant by a constant.
658  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
659  if (!RC)
660  return nullptr;
661  const APInt &LA = C->getAPInt();
662  const APInt &RA = RC->getAPInt();
663  if (LA.srem(RA) != 0)
664  return nullptr;
665  return SE.getConstant(LA.sdiv(RA));
666  }
667 
668  // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
669  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
670  if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
671  const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
672  IgnoreSignificantBits);
673  if (!Step) return nullptr;
674  const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
675  IgnoreSignificantBits);
676  if (!Start) return nullptr;
677  // FlagNW is independent of the start value, step direction, and is
678  // preserved with smaller magnitude steps.
679  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
680  return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
681  }
682  return nullptr;
683  }
684 
685  // Distribute the sdiv over add operands, if the add doesn't overflow.
686  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
687  if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
689  for (const SCEV *S : Add->operands()) {
690  const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
691  if (!Op) return nullptr;
692  Ops.push_back(Op);
693  }
694  return SE.getAddExpr(Ops);
695  }
696  return nullptr;
697  }
698 
699  // Check for a multiply operand that we can pull RHS out of.
700  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
701  if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
703  bool Found = false;
704  for (const SCEV *S : Mul->operands()) {
705  if (!Found)
706  if (const SCEV *Q = getExactSDiv(S, RHS, SE,
707  IgnoreSignificantBits)) {
708  S = Q;
709  Found = true;
710  }
711  Ops.push_back(S);
712  }
713  return Found ? SE.getMulExpr(Ops) : nullptr;
714  }
715  return nullptr;
716  }
717 
718  // Otherwise we don't know.
719  return nullptr;
720 }
721 
722 /// If S involves the addition of a constant integer value, return that integer
723 /// value, and mutate S to point to a new SCEV with that value excluded.
724 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
725  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
726  if (C->getAPInt().getMinSignedBits() <= 64) {
727  S = SE.getConstant(C->getType(), 0);
728  return C->getValue()->getSExtValue();
729  }
730  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
731  SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
732  int64_t Result = ExtractImmediate(NewOps.front(), SE);
733  if (Result != 0)
734  S = SE.getAddExpr(NewOps);
735  return Result;
736  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
737  SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
738  int64_t Result = ExtractImmediate(NewOps.front(), SE);
739  if (Result != 0)
740  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
741  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
743  return Result;
744  }
745  return 0;
746 }
747 
748 /// If S involves the addition of a GlobalValue address, return that symbol, and
749 /// mutate S to point to a new SCEV with that value excluded.
751  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
752  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
753  S = SE.getConstant(GV->getType(), 0);
754  return GV;
755  }
756  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
757  SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
758  GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
759  if (Result)
760  S = SE.getAddExpr(NewOps);
761  return Result;
762  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
763  SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
764  GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
765  if (Result)
766  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
767  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
769  return Result;
770  }
771  return nullptr;
772 }
773 
774 /// Returns true if the specified instruction is using the specified value as an
775 /// address.
776 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
777  bool isAddress = isa<LoadInst>(Inst);
778  if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
779  if (SI->getPointerOperand() == OperandVal)
780  isAddress = true;
781  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
782  // Addressing modes can also be folded into prefetches and a variety
783  // of intrinsics.
784  switch (II->getIntrinsicID()) {
785  default: break;
786  case Intrinsic::memset:
787  case Intrinsic::prefetch:
788  if (II->getArgOperand(0) == OperandVal)
789  isAddress = true;
790  break;
791  case Intrinsic::memmove:
792  case Intrinsic::memcpy:
793  if (II->getArgOperand(0) == OperandVal ||
794  II->getArgOperand(1) == OperandVal)
795  isAddress = true;
796  break;
797  }
798  } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
799  if (RMW->getPointerOperand() == OperandVal)
800  isAddress = true;
801  } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
802  if (CmpX->getPointerOperand() == OperandVal)
803  isAddress = true;
804  }
805  return isAddress;
806 }
807 
808 /// Return the type of the memory being accessed.
809 static MemAccessTy getAccessType(const Instruction *Inst) {
810  MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
811  if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
812  AccessTy.MemTy = SI->getOperand(0)->getType();
813  AccessTy.AddrSpace = SI->getPointerAddressSpace();
814  } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
815  AccessTy.AddrSpace = LI->getPointerAddressSpace();
816  } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
817  AccessTy.AddrSpace = RMW->getPointerAddressSpace();
818  } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
819  AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
820  }
821 
822  // All pointers have the same requirements, so canonicalize them to an
823  // arbitrary pointer type to minimize variation.
824  if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
825  AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
826  PTy->getAddressSpace());
827 
828  return AccessTy;
829 }
830 
831 /// Return true if this AddRec is already a phi in its loop.
832 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
833  for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
834  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
835  if (SE.isSCEVable(PN->getType()) &&
836  (SE.getEffectiveSCEVType(PN->getType()) ==
837  SE.getEffectiveSCEVType(AR->getType())) &&
838  SE.getSCEV(PN) == AR)
839  return true;
840  }
841  return false;
842 }
843 
844 /// Check if expanding this expression is likely to incur significant cost. This
845 /// is tricky because SCEV doesn't track which expressions are actually computed
846 /// by the current IR.
847 ///
848 /// We currently allow expansion of IV increments that involve adds,
849 /// multiplication by constants, and AddRecs from existing phis.
850 ///
851 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
852 /// obvious multiple of the UDivExpr.
853 static bool isHighCostExpansion(const SCEV *S,
854  SmallPtrSetImpl<const SCEV*> &Processed,
855  ScalarEvolution &SE) {
856  // Zero/One operand expressions
857  switch (S->getSCEVType()) {
858  case scUnknown:
859  case scConstant:
860  return false;
861  case scTruncate:
862  return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
863  Processed, SE);
864  case scZeroExtend:
865  return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
866  Processed, SE);
867  case scSignExtend:
868  return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
869  Processed, SE);
870  }
871 
872  if (!Processed.insert(S).second)
873  return false;
874 
875  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
876  for (const SCEV *S : Add->operands()) {
877  if (isHighCostExpansion(S, Processed, SE))
878  return true;
879  }
880  return false;
881  }
882 
883  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
884  if (Mul->getNumOperands() == 2) {
885  // Multiplication by a constant is ok
886  if (isa<SCEVConstant>(Mul->getOperand(0)))
887  return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
888 
889  // If we have the value of one operand, check if an existing
890  // multiplication already generates this expression.
891  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
892  Value *UVal = U->getValue();
893  for (User *UR : UVal->users()) {
894  // If U is a constant, it may be used by a ConstantExpr.
895  Instruction *UI = dyn_cast<Instruction>(UR);
896  if (UI && UI->getOpcode() == Instruction::Mul &&
897  SE.isSCEVable(UI->getType())) {
898  return SE.getSCEV(UI) == Mul;
899  }
900  }
901  }
902  }
903  }
904 
905  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
906  if (isExistingPhi(AR, SE))
907  return false;
908  }
909 
910  // Fow now, consider any other type of expression (div/mul/min/max) high cost.
911  return true;
912 }
913 
914 /// If any of the instructions is the specified set are trivially dead, delete
915 /// them and see if this makes any of their operands subsequently dead.
916 static bool
918  bool Changed = false;
919 
920  while (!DeadInsts.empty()) {
921  Value *V = DeadInsts.pop_back_val();
922  Instruction *I = dyn_cast_or_null<Instruction>(V);
923 
924  if (!I || !isInstructionTriviallyDead(I))
925  continue;
926 
927  for (Use &O : I->operands())
928  if (Instruction *U = dyn_cast<Instruction>(O)) {
929  O = nullptr;
930  if (U->use_empty())
931  DeadInsts.emplace_back(U);
932  }
933 
934  I->eraseFromParent();
935  Changed = true;
936  }
937 
938  return Changed;
939 }
940 
941 namespace {
942 
943 class LSRUse;
944 
945 } // end anonymous namespace
946 
947 /// \brief Check if the addressing mode defined by \p F is completely
948 /// folded in \p LU at isel time.
949 /// This includes address-mode folding and special icmp tricks.
950 /// This function returns true if \p LU can accommodate what \p F
951 /// defines and up to 1 base + 1 scaled + offset.
952 /// In other words, if \p F has several base registers, this function may
953 /// still return true. Therefore, users still need to account for
954 /// additional base registers and/or unfolded offsets to derive an
955 /// accurate cost model.
956 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
957  const LSRUse &LU, const Formula &F);
958 // Get the cost of the scaling factor used in F for LU.
959 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
960  const LSRUse &LU, const Formula &F,
961  const Loop &L);
962 
963 namespace {
964 
965 /// This class is used to measure and compare candidate formulae.
966 class Cost {
968 
969 public:
970  Cost() {
971  C.Insns = 0;
972  C.NumRegs = 0;
973  C.AddRecCost = 0;
974  C.NumIVMuls = 0;
975  C.NumBaseAdds = 0;
976  C.ImmCost = 0;
977  C.SetupCost = 0;
978  C.ScaleCost = 0;
979  }
980 
981  bool isLess(Cost &Other, const TargetTransformInfo &TTI);
982 
983  void Lose();
984 
985 #ifndef NDEBUG
986  // Once any of the metrics loses, they must all remain losers.
987  bool isValid() {
988  return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
989  | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
990  || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
991  & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
992  }
993 #endif
994 
995  bool isLoser() {
996  assert(isValid() && "invalid cost");
997  return C.NumRegs == ~0u;
998  }
999 
1000  void RateFormula(const TargetTransformInfo &TTI,
1001  const Formula &F,
1003  const DenseSet<const SCEV *> &VisitedRegs,
1004  const Loop *L,
1005  ScalarEvolution &SE, DominatorTree &DT,
1006  const LSRUse &LU,
1007  SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1008 
1009  void print(raw_ostream &OS) const;
1010  void dump() const;
1011 
1012 private:
1013  void RateRegister(const SCEV *Reg,
1015  const Loop *L,
1016  ScalarEvolution &SE, DominatorTree &DT);
1017  void RatePrimaryRegister(const SCEV *Reg,
1019  const Loop *L,
1020  ScalarEvolution &SE, DominatorTree &DT,
1021  SmallPtrSetImpl<const SCEV *> *LoserRegs);
1022 };
1023 
1024 /// An operand value in an instruction which is to be replaced with some
1025 /// equivalent, possibly strength-reduced, replacement.
1026 struct LSRFixup {
1027  /// The instruction which will be updated.
1028  Instruction *UserInst;
1029 
1030  /// The operand of the instruction which will be replaced. The operand may be
1031  /// used more than once; every instance will be replaced.
1032  Value *OperandValToReplace;
1033 
1034  /// If this user is to use the post-incremented value of an induction
1035  /// variable, this variable is non-null and holds the loop associated with the
1036  /// induction variable.
1037  PostIncLoopSet PostIncLoops;
1038 
1039  /// A constant offset to be added to the LSRUse expression. This allows
1040  /// multiple fixups to share the same LSRUse with different offsets, for
1041  /// example in an unrolled loop.
1042  int64_t Offset;
1043 
1044  bool isUseFullyOutsideLoop(const Loop *L) const;
1045 
1046  LSRFixup();
1047 
1048  void print(raw_ostream &OS) const;
1049  void dump() const;
1050 };
1051 
1052 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1053 /// SmallVectors of const SCEV*.
1054 struct UniquifierDenseMapInfo {
1055  static SmallVector<const SCEV *, 4> getEmptyKey() {
1057  V.push_back(reinterpret_cast<const SCEV *>(-1));
1058  return V;
1059  }
1060 
1061  static SmallVector<const SCEV *, 4> getTombstoneKey() {
1063  V.push_back(reinterpret_cast<const SCEV *>(-2));
1064  return V;
1065  }
1066 
1067  static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1068  return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1069  }
1070 
1071  static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1072  const SmallVector<const SCEV *, 4> &RHS) {
1073  return LHS == RHS;
1074  }
1075 };
1076 
1077 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1078 /// as uses invented by LSR itself. It includes information about what kinds of
1079 /// things can be folded into the user, information about the user itself, and
1080 /// information about how the use may be satisfied. TODO: Represent multiple
1081 /// users of the same expression in common?
1082 class LSRUse {
1083  DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1084 
1085 public:
1086  /// An enum for a kind of use, indicating what types of scaled and immediate
1087  /// operands it might support.
1088  enum KindType {
1089  Basic, ///< A normal use, with no folding.
1090  Special, ///< A special case of basic, allowing -1 scales.
1091  Address, ///< An address use; folding according to TargetLowering
1092  ICmpZero ///< An equality icmp with both operands folded into one.
1093  // TODO: Add a generic icmp too?
1094  };
1095 
1096  typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1097 
1098  KindType Kind;
1099  MemAccessTy AccessTy;
1100 
1101  /// The list of operands which are to be replaced.
1103 
1104  /// Keep track of the min and max offsets of the fixups.
1105  int64_t MinOffset;
1106  int64_t MaxOffset;
1107 
1108  /// This records whether all of the fixups using this LSRUse are outside of
1109  /// the loop, in which case some special-case heuristics may be used.
1110  bool AllFixupsOutsideLoop;
1111 
1112  /// RigidFormula is set to true to guarantee that this use will be associated
1113  /// with a single formula--the one that initially matched. Some SCEV
1114  /// expressions cannot be expanded. This allows LSR to consider the registers
1115  /// used by those expressions without the need to expand them later after
1116  /// changing the formula.
1117  bool RigidFormula;
1118 
1119  /// This records the widest use type for any fixup using this
1120  /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1121  /// fixup widths to be equivalent, because the narrower one may be relying on
1122  /// the implicit truncation to truncate away bogus bits.
1123  Type *WidestFixupType;
1124 
1125  /// A list of ways to build a value that can satisfy this user. After the
1126  /// list is populated, one of these is selected heuristically and used to
1127  /// formulate a replacement for OperandValToReplace in UserInst.
1128  SmallVector<Formula, 12> Formulae;
1129 
1130  /// The set of register candidates used by all formulae in this LSRUse.
1132 
1133  LSRUse(KindType K, MemAccessTy AT)
1134  : Kind(K), AccessTy(AT), MinOffset(INT64_MAX), MaxOffset(INT64_MIN),
1135  AllFixupsOutsideLoop(true), RigidFormula(false),
1136  WidestFixupType(nullptr) {}
1137 
1138  LSRFixup &getNewFixup() {
1139  Fixups.push_back(LSRFixup());
1140  return Fixups.back();
1141  }
1142 
1143  void pushFixup(LSRFixup &f) {
1144  Fixups.push_back(f);
1145  if (f.Offset > MaxOffset)
1146  MaxOffset = f.Offset;
1147  if (f.Offset < MinOffset)
1148  MinOffset = f.Offset;
1149  }
1150 
1151  bool HasFormulaWithSameRegs(const Formula &F) const;
1152  float getNotSelectedProbability(const SCEV *Reg) const;
1153  bool InsertFormula(const Formula &F, const Loop &L);
1154  void DeleteFormula(Formula &F);
1155  void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1156 
1157  void print(raw_ostream &OS) const;
1158  void dump() const;
1159 };
1160 
1161 } // end anonymous namespace
1162 
1163 /// Tally up interesting quantities from the given register.
1164 void Cost::RateRegister(const SCEV *Reg,
1166  const Loop *L,
1167  ScalarEvolution &SE, DominatorTree &DT) {
1168  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1169  // If this is an addrec for another loop, it should be an invariant
1170  // with respect to L since L is the innermost loop (at least
1171  // for now LSR only handles innermost loops).
1172  if (AR->getLoop() != L) {
1173  // If the AddRec exists, consider it's register free and leave it alone.
1174  if (isExistingPhi(AR, SE))
1175  return;
1176 
1177  // It is bad to allow LSR for current loop to add induction variables
1178  // for its sibling loops.
1179  if (!AR->getLoop()->contains(L)) {
1180  Lose();
1181  return;
1182  }
1183 
1184  // Otherwise, it will be an invariant with respect to Loop L.
1185  ++C.NumRegs;
1186  return;
1187  }
1188  C.AddRecCost += 1; /// TODO: This should be a function of the stride.
1189 
1190  // Add the step value register, if it needs one.
1191  // TODO: The non-affine case isn't precisely modeled here.
1192  if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1193  if (!Regs.count(AR->getOperand(1))) {
1194  RateRegister(AR->getOperand(1), Regs, L, SE, DT);
1195  if (isLoser())
1196  return;
1197  }
1198  }
1199  }
1200  ++C.NumRegs;
1201 
1202  // Rough heuristic; favor registers which don't require extra setup
1203  // instructions in the preheader.
1204  if (!isa<SCEVUnknown>(Reg) &&
1205  !isa<SCEVConstant>(Reg) &&
1206  !(isa<SCEVAddRecExpr>(Reg) &&
1207  (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
1208  isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
1209  ++C.SetupCost;
1210 
1211  C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1212  SE.hasComputableLoopEvolution(Reg, L);
1213 }
1214 
1215 /// Record this register in the set. If we haven't seen it before, rate
1216 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1217 /// one of those regs an instant loser.
1218 void Cost::RatePrimaryRegister(const SCEV *Reg,
1220  const Loop *L,
1221  ScalarEvolution &SE, DominatorTree &DT,
1222  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1223  if (LoserRegs && LoserRegs->count(Reg)) {
1224  Lose();
1225  return;
1226  }
1227  if (Regs.insert(Reg).second) {
1228  RateRegister(Reg, Regs, L, SE, DT);
1229  if (LoserRegs && isLoser())
1230  LoserRegs->insert(Reg);
1231  }
1232 }
1233 
1234 void Cost::RateFormula(const TargetTransformInfo &TTI,
1235  const Formula &F,
1237  const DenseSet<const SCEV *> &VisitedRegs,
1238  const Loop *L,
1239  ScalarEvolution &SE, DominatorTree &DT,
1240  const LSRUse &LU,
1241  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1242  assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1243  // Tally up the registers.
1244  unsigned PrevAddRecCost = C.AddRecCost;
1245  unsigned PrevNumRegs = C.NumRegs;
1246  unsigned PrevNumBaseAdds = C.NumBaseAdds;
1247  if (const SCEV *ScaledReg = F.ScaledReg) {
1248  if (VisitedRegs.count(ScaledReg)) {
1249  Lose();
1250  return;
1251  }
1252  RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1253  if (isLoser())
1254  return;
1255  }
1256  for (const SCEV *BaseReg : F.BaseRegs) {
1257  if (VisitedRegs.count(BaseReg)) {
1258  Lose();
1259  return;
1260  }
1261  RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1262  if (isLoser())
1263  return;
1264  }
1265 
1266  // Determine how many (unfolded) adds we'll need inside the loop.
1267  size_t NumBaseParts = F.getNumRegs();
1268  if (NumBaseParts > 1)
1269  // Do not count the base and a possible second register if the target
1270  // allows to fold 2 registers.
1271  C.NumBaseAdds +=
1272  NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1273  C.NumBaseAdds += (F.UnfoldedOffset != 0);
1274 
1275  // Accumulate non-free scaling amounts.
1276  C.ScaleCost += getScalingFactorCost(TTI, LU, F, *L);
1277 
1278  // Tally up the non-zero immediates.
1279  for (const LSRFixup &Fixup : LU.Fixups) {
1280  int64_t O = Fixup.Offset;
1281  int64_t Offset = (uint64_t)O + F.BaseOffset;
1282  if (F.BaseGV)
1283  C.ImmCost += 64; // Handle symbolic values conservatively.
1284  // TODO: This should probably be the pointer size.
1285  else if (Offset != 0)
1286  C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1287 
1288  // Check with target if this offset with this instruction is
1289  // specifically not supported.
1290  if (LU.Kind == LSRUse::Address && Offset != 0 &&
1291  !TTI.isFoldableMemAccessOffset(Fixup.UserInst, Offset))
1292  C.NumBaseAdds++;
1293  }
1294 
1295  // If we don't count instruction cost exit here.
1296  if (!InsnsCost) {
1297  assert(isValid() && "invalid cost");
1298  return;
1299  }
1300 
1301  // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1302  // additional instruction (at least fill).
1303  unsigned TTIRegNum = TTI.getNumberOfRegisters(false) - 1;
1304  if (C.NumRegs > TTIRegNum) {
1305  // Cost already exceeded TTIRegNum, then only newly added register can add
1306  // new instructions.
1307  if (PrevNumRegs > TTIRegNum)
1308  C.Insns += (C.NumRegs - PrevNumRegs);
1309  else
1310  C.Insns += (C.NumRegs - TTIRegNum);
1311  }
1312 
1313  // If ICmpZero formula ends with not 0, it could not be replaced by
1314  // just add or sub. We'll need to compare final result of AddRec.
1315  // That means we'll need an additional instruction.
1316  // For -10 + {0, +, 1}:
1317  // i = i + 1;
1318  // cmp i, 10
1319  //
1320  // For {-10, +, 1}:
1321  // i = i + 1;
1322  if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd())
1323  C.Insns++;
1324  // Each new AddRec adds 1 instruction to calculation.
1325  C.Insns += (C.AddRecCost - PrevAddRecCost);
1326 
1327  // BaseAdds adds instructions for unfolded registers.
1328  if (LU.Kind != LSRUse::ICmpZero)
1329  C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1330  assert(isValid() && "invalid cost");
1331 }
1332 
1333 /// Set this cost to a losing value.
1334 void Cost::Lose() {
1335  C.Insns = ~0u;
1336  C.NumRegs = ~0u;
1337  C.AddRecCost = ~0u;
1338  C.NumIVMuls = ~0u;
1339  C.NumBaseAdds = ~0u;
1340  C.ImmCost = ~0u;
1341  C.SetupCost = ~0u;
1342  C.ScaleCost = ~0u;
1343 }
1344 
1345 /// Choose the lower cost.
1346 bool Cost::isLess(Cost &Other, const TargetTransformInfo &TTI) {
1347  if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1348  C.Insns != Other.C.Insns)
1349  return C.Insns < Other.C.Insns;
1350  return TTI.isLSRCostLess(C, Other.C);
1351 }
1352 
1353 void Cost::print(raw_ostream &OS) const {
1354  if (InsnsCost)
1355  OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1356  OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1357  if (C.AddRecCost != 0)
1358  OS << ", with addrec cost " << C.AddRecCost;
1359  if (C.NumIVMuls != 0)
1360  OS << ", plus " << C.NumIVMuls << " IV mul"
1361  << (C.NumIVMuls == 1 ? "" : "s");
1362  if (C.NumBaseAdds != 0)
1363  OS << ", plus " << C.NumBaseAdds << " base add"
1364  << (C.NumBaseAdds == 1 ? "" : "s");
1365  if (C.ScaleCost != 0)
1366  OS << ", plus " << C.ScaleCost << " scale cost";
1367  if (C.ImmCost != 0)
1368  OS << ", plus " << C.ImmCost << " imm cost";
1369  if (C.SetupCost != 0)
1370  OS << ", plus " << C.SetupCost << " setup cost";
1371 }
1372 
1373 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1374 LLVM_DUMP_METHOD void Cost::dump() const {
1375  print(errs()); errs() << '\n';
1376 }
1377 #endif
1378 
1379 LSRFixup::LSRFixup()
1380  : UserInst(nullptr), OperandValToReplace(nullptr),
1381  Offset(0) {}
1382 
1383 /// Test whether this fixup always uses its value outside of the given loop.
1384 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1385  // PHI nodes use their value in their incoming blocks.
1386  if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1387  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1388  if (PN->getIncomingValue(i) == OperandValToReplace &&
1389  L->contains(PN->getIncomingBlock(i)))
1390  return false;
1391  return true;
1392  }
1393 
1394  return !L->contains(UserInst);
1395 }
1396 
1397 void LSRFixup::print(raw_ostream &OS) const {
1398  OS << "UserInst=";
1399  // Store is common and interesting enough to be worth special-casing.
1400  if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1401  OS << "store ";
1402  Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1403  } else if (UserInst->getType()->isVoidTy())
1404  OS << UserInst->getOpcodeName();
1405  else
1406  UserInst->printAsOperand(OS, /*PrintType=*/false);
1407 
1408  OS << ", OperandValToReplace=";
1409  OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1410 
1411  for (const Loop *PIL : PostIncLoops) {
1412  OS << ", PostIncLoop=";
1413  PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1414  }
1415 
1416  if (Offset != 0)
1417  OS << ", Offset=" << Offset;
1418 }
1419 
1420 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1421 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1422  print(errs()); errs() << '\n';
1423 }
1424 #endif
1425 
1426 /// Test whether this use as a formula which has the same registers as the given
1427 /// formula.
1428 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1429  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1430  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1431  // Unstable sort by host order ok, because this is only used for uniquifying.
1432  std::sort(Key.begin(), Key.end());
1433  return Uniquifier.count(Key);
1434 }
1435 
1436 /// The function returns a probability of selecting formula without Reg.
1437 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1438  unsigned FNum = 0;
1439  for (const Formula &F : Formulae)
1440  if (F.referencesReg(Reg))
1441  FNum++;
1442  return ((float)(Formulae.size() - FNum)) / Formulae.size();
1443 }
1444 
1445 /// If the given formula has not yet been inserted, add it to the list, and
1446 /// return true. Return false otherwise. The formula must be in canonical form.
1447 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1448  assert(F.isCanonical(L) && "Invalid canonical representation");
1449 
1450  if (!Formulae.empty() && RigidFormula)
1451  return false;
1452 
1453  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1454  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1455  // Unstable sort by host order ok, because this is only used for uniquifying.
1456  std::sort(Key.begin(), Key.end());
1457 
1458  if (!Uniquifier.insert(Key).second)
1459  return false;
1460 
1461  // Using a register to hold the value of 0 is not profitable.
1462  assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1463  "Zero allocated in a scaled register!");
1464 #ifndef NDEBUG
1465  for (const SCEV *BaseReg : F.BaseRegs)
1466  assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1467 #endif
1468 
1469  // Add the formula to the list.
1470  Formulae.push_back(F);
1471 
1472  // Record registers now being used by this use.
1473  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1474  if (F.ScaledReg)
1475  Regs.insert(F.ScaledReg);
1476 
1477  return true;
1478 }
1479 
1480 /// Remove the given formula from this use's list.
1481 void LSRUse::DeleteFormula(Formula &F) {
1482  if (&F != &Formulae.back())
1483  std::swap(F, Formulae.back());
1484  Formulae.pop_back();
1485 }
1486 
1487 /// Recompute the Regs field, and update RegUses.
1488 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1489  // Now that we've filtered out some formulae, recompute the Regs set.
1490  SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1491  Regs.clear();
1492  for (const Formula &F : Formulae) {
1493  if (F.ScaledReg) Regs.insert(F.ScaledReg);
1494  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1495  }
1496 
1497  // Update the RegTracker.
1498  for (const SCEV *S : OldRegs)
1499  if (!Regs.count(S))
1500  RegUses.dropRegister(S, LUIdx);
1501 }
1502 
1503 void LSRUse::print(raw_ostream &OS) const {
1504  OS << "LSR Use: Kind=";
1505  switch (Kind) {
1506  case Basic: OS << "Basic"; break;
1507  case Special: OS << "Special"; break;
1508  case ICmpZero: OS << "ICmpZero"; break;
1509  case Address:
1510  OS << "Address of ";
1511  if (AccessTy.MemTy->isPointerTy())
1512  OS << "pointer"; // the full pointer type could be really verbose
1513  else {
1514  OS << *AccessTy.MemTy;
1515  }
1516 
1517  OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1518  }
1519 
1520  OS << ", Offsets={";
1521  bool NeedComma = false;
1522  for (const LSRFixup &Fixup : Fixups) {
1523  if (NeedComma) OS << ',';
1524  OS << Fixup.Offset;
1525  NeedComma = true;
1526  }
1527  OS << '}';
1528 
1529  if (AllFixupsOutsideLoop)
1530  OS << ", all-fixups-outside-loop";
1531 
1532  if (WidestFixupType)
1533  OS << ", widest fixup type: " << *WidestFixupType;
1534 }
1535 
1536 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1537 LLVM_DUMP_METHOD void LSRUse::dump() const {
1538  print(errs()); errs() << '\n';
1539 }
1540 #endif
1541 
1543  LSRUse::KindType Kind, MemAccessTy AccessTy,
1544  GlobalValue *BaseGV, int64_t BaseOffset,
1545  bool HasBaseReg, int64_t Scale,
1546  Instruction *Fixup = nullptr) {
1547  switch (Kind) {
1548  case LSRUse::Address:
1549  return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1550  HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1551 
1552  case LSRUse::ICmpZero:
1553  // There's not even a target hook for querying whether it would be legal to
1554  // fold a GV into an ICmp.
1555  if (BaseGV)
1556  return false;
1557 
1558  // ICmp only has two operands; don't allow more than two non-trivial parts.
1559  if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1560  return false;
1561 
1562  // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1563  // putting the scaled register in the other operand of the icmp.
1564  if (Scale != 0 && Scale != -1)
1565  return false;
1566 
1567  // If we have low-level target information, ask the target if it can fold an
1568  // integer immediate on an icmp.
1569  if (BaseOffset != 0) {
1570  // We have one of:
1571  // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1572  // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1573  // Offs is the ICmp immediate.
1574  if (Scale == 0)
1575  // The cast does the right thing with INT64_MIN.
1576  BaseOffset = -(uint64_t)BaseOffset;
1577  return TTI.isLegalICmpImmediate(BaseOffset);
1578  }
1579 
1580  // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1581  return true;
1582 
1583  case LSRUse::Basic:
1584  // Only handle single-register values.
1585  return !BaseGV && Scale == 0 && BaseOffset == 0;
1586 
1587  case LSRUse::Special:
1588  // Special case Basic to handle -1 scales.
1589  return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1590  }
1591 
1592  llvm_unreachable("Invalid LSRUse Kind!");
1593 }
1594 
1596  int64_t MinOffset, int64_t MaxOffset,
1597  LSRUse::KindType Kind, MemAccessTy AccessTy,
1598  GlobalValue *BaseGV, int64_t BaseOffset,
1599  bool HasBaseReg, int64_t Scale) {
1600  // Check for overflow.
1601  if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1602  (MinOffset > 0))
1603  return false;
1604  MinOffset = (uint64_t)BaseOffset + MinOffset;
1605  if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1606  (MaxOffset > 0))
1607  return false;
1608  MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1609 
1610  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1611  HasBaseReg, Scale) &&
1612  isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1613  HasBaseReg, Scale);
1614 }
1615 
1617  int64_t MinOffset, int64_t MaxOffset,
1618  LSRUse::KindType Kind, MemAccessTy AccessTy,
1619  const Formula &F, const Loop &L) {
1620  // For the purpose of isAMCompletelyFolded either having a canonical formula
1621  // or a scale not equal to zero is correct.
1622  // Problems may arise from non canonical formulae having a scale == 0.
1623  // Strictly speaking it would best to just rely on canonical formulae.
1624  // However, when we generate the scaled formulae, we first check that the
1625  // scaling factor is profitable before computing the actual ScaledReg for
1626  // compile time sake.
1627  assert((F.isCanonical(L) || F.Scale != 0));
1628  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1629  F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1630 }
1631 
1632 /// Test whether we know how to expand the current formula.
1633 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1634  int64_t MaxOffset, LSRUse::KindType Kind,
1635  MemAccessTy AccessTy, GlobalValue *BaseGV,
1636  int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1637  // We know how to expand completely foldable formulae.
1638  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1639  BaseOffset, HasBaseReg, Scale) ||
1640  // Or formulae that use a base register produced by a sum of base
1641  // registers.
1642  (Scale == 1 &&
1643  isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1644  BaseGV, BaseOffset, true, 0));
1645 }
1646 
1647 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1648  int64_t MaxOffset, LSRUse::KindType Kind,
1649  MemAccessTy AccessTy, const Formula &F) {
1650  return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1651  F.BaseOffset, F.HasBaseReg, F.Scale);
1652 }
1653 
1655  const LSRUse &LU, const Formula &F) {
1656  // Target may want to look at the user instructions.
1657  if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1658  for (const LSRFixup &Fixup : LU.Fixups)
1659  if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1660  F.BaseOffset, F.HasBaseReg, F.Scale,
1661  Fixup.UserInst))
1662  return false;
1663  return true;
1664  }
1665 
1666  return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1667  LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1668  F.Scale);
1669 }
1670 
1671 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1672  const LSRUse &LU, const Formula &F,
1673  const Loop &L) {
1674  if (!F.Scale)
1675  return 0;
1676 
1677  // If the use is not completely folded in that instruction, we will have to
1678  // pay an extra cost only for scale != 1.
1679  if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1680  LU.AccessTy, F, L))
1681  return F.Scale != 1;
1682 
1683  switch (LU.Kind) {
1684  case LSRUse::Address: {
1685  // Check the scaling factor cost with both the min and max offsets.
1686  int ScaleCostMinOffset = TTI.getScalingFactorCost(
1687  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1688  F.Scale, LU.AccessTy.AddrSpace);
1689  int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1690  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1691  F.Scale, LU.AccessTy.AddrSpace);
1692 
1693  assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1694  "Legal addressing mode has an illegal cost!");
1695  return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1696  }
1697  case LSRUse::ICmpZero:
1698  case LSRUse::Basic:
1699  case LSRUse::Special:
1700  // The use is completely folded, i.e., everything is folded into the
1701  // instruction.
1702  return 0;
1703  }
1704 
1705  llvm_unreachable("Invalid LSRUse Kind!");
1706 }
1707 
1708 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1709  LSRUse::KindType Kind, MemAccessTy AccessTy,
1710  GlobalValue *BaseGV, int64_t BaseOffset,
1711  bool HasBaseReg) {
1712  // Fast-path: zero is always foldable.
1713  if (BaseOffset == 0 && !BaseGV) return true;
1714 
1715  // Conservatively, create an address with an immediate and a
1716  // base and a scale.
1717  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1718 
1719  // Canonicalize a scale of 1 to a base register if the formula doesn't
1720  // already have a base register.
1721  if (!HasBaseReg && Scale == 1) {
1722  Scale = 0;
1723  HasBaseReg = true;
1724  }
1725 
1726  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1727  HasBaseReg, Scale);
1728 }
1729 
1730 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1731  ScalarEvolution &SE, int64_t MinOffset,
1732  int64_t MaxOffset, LSRUse::KindType Kind,
1733  MemAccessTy AccessTy, const SCEV *S,
1734  bool HasBaseReg) {
1735  // Fast-path: zero is always foldable.
1736  if (S->isZero()) return true;
1737 
1738  // Conservatively, create an address with an immediate and a
1739  // base and a scale.
1740  int64_t BaseOffset = ExtractImmediate(S, SE);
1741  GlobalValue *BaseGV = ExtractSymbol(S, SE);
1742 
1743  // If there's anything else involved, it's not foldable.
1744  if (!S->isZero()) return false;
1745 
1746  // Fast-path: zero is always foldable.
1747  if (BaseOffset == 0 && !BaseGV) return true;
1748 
1749  // Conservatively, create an address with an immediate and a
1750  // base and a scale.
1751  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1752 
1753  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1754  BaseOffset, HasBaseReg, Scale);
1755 }
1756 
1757 namespace {
1758 
1759 /// An individual increment in a Chain of IV increments. Relate an IV user to
1760 /// an expression that computes the IV it uses from the IV used by the previous
1761 /// link in the Chain.
1762 ///
1763 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1764 /// original IVOperand. The head of the chain's IVOperand is only valid during
1765 /// chain collection, before LSR replaces IV users. During chain generation,
1766 /// IncExpr can be used to find the new IVOperand that computes the same
1767 /// expression.
1768 struct IVInc {
1769  Instruction *UserInst;
1770  Value* IVOperand;
1771  const SCEV *IncExpr;
1772 
1773  IVInc(Instruction *U, Value *O, const SCEV *E):
1774  UserInst(U), IVOperand(O), IncExpr(E) {}
1775 };
1776 
1777 // The list of IV increments in program order. We typically add the head of a
1778 // chain without finding subsequent links.
1779 struct IVChain {
1780  SmallVector<IVInc,1> Incs;
1781  const SCEV *ExprBase;
1782 
1783  IVChain() : ExprBase(nullptr) {}
1784 
1785  IVChain(const IVInc &Head, const SCEV *Base)
1786  : Incs(1, Head), ExprBase(Base) {}
1787 
1788  typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1789 
1790  // Return the first increment in the chain.
1791  const_iterator begin() const {
1792  assert(!Incs.empty());
1793  return std::next(Incs.begin());
1794  }
1795  const_iterator end() const {
1796  return Incs.end();
1797  }
1798 
1799  // Returns true if this chain contains any increments.
1800  bool hasIncs() const { return Incs.size() >= 2; }
1801 
1802  // Add an IVInc to the end of this chain.
1803  void add(const IVInc &X) { Incs.push_back(X); }
1804 
1805  // Returns the last UserInst in the chain.
1806  Instruction *tailUserInst() const { return Incs.back().UserInst; }
1807 
1808  // Returns true if IncExpr can be profitably added to this chain.
1809  bool isProfitableIncrement(const SCEV *OperExpr,
1810  const SCEV *IncExpr,
1811  ScalarEvolution&);
1812 };
1813 
1814 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1815 /// between FarUsers that definitely cross IV increments and NearUsers that may
1816 /// be used between IV increments.
1817 struct ChainUsers {
1819  SmallPtrSet<Instruction*, 4> NearUsers;
1820 };
1821 
1822 /// This class holds state for the main loop strength reduction logic.
1823 class LSRInstance {
1824  IVUsers &IU;
1825  ScalarEvolution &SE;
1826  DominatorTree &DT;
1827  LoopInfo &LI;
1828  const TargetTransformInfo &TTI;
1829  Loop *const L;
1830  bool Changed;
1831 
1832  /// This is the insert position that the current loop's induction variable
1833  /// increment should be placed. In simple loops, this is the latch block's
1834  /// terminator. But in more complicated cases, this is a position which will
1835  /// dominate all the in-loop post-increment users.
1836  Instruction *IVIncInsertPos;
1837 
1838  /// Interesting factors between use strides.
1839  ///
1840  /// We explicitly use a SetVector which contains a SmallSet, instead of the
1841  /// default, a SmallDenseSet, because we need to use the full range of
1842  /// int64_ts, and there's currently no good way of doing that with
1843  /// SmallDenseSet.
1845 
1846  /// Interesting use types, to facilitate truncation reuse.
1848 
1849  /// The list of interesting uses.
1851 
1852  /// Track which uses use which register candidates.
1853  RegUseTracker RegUses;
1854 
1855  // Limit the number of chains to avoid quadratic behavior. We don't expect to
1856  // have more than a few IV increment chains in a loop. Missing a Chain falls
1857  // back to normal LSR behavior for those uses.
1858  static const unsigned MaxChains = 8;
1859 
1860  /// IV users can form a chain of IV increments.
1862 
1863  /// IV users that belong to profitable IVChains.
1865 
1866  void OptimizeShadowIV();
1867  bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1868  ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1869  void OptimizeLoopTermCond();
1870 
1871  void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1872  SmallVectorImpl<ChainUsers> &ChainUsersVec);
1873  void FinalizeChain(IVChain &Chain);
1874  void CollectChains();
1875  void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1876  SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1877 
1878  void CollectInterestingTypesAndFactors();
1879  void CollectFixupsAndInitialFormulae();
1880 
1881  // Support for sharing of LSRUses between LSRFixups.
1883  UseMapTy UseMap;
1884 
1885  bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1886  LSRUse::KindType Kind, MemAccessTy AccessTy);
1887 
1888  std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1889  MemAccessTy AccessTy);
1890 
1891  void DeleteUse(LSRUse &LU, size_t LUIdx);
1892 
1893  LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1894 
1895  void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1896  void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1897  void CountRegisters(const Formula &F, size_t LUIdx);
1898  bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1899 
1900  void CollectLoopInvariantFixupsAndFormulae();
1901 
1902  void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1903  unsigned Depth = 0);
1904 
1905  void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1906  const Formula &Base, unsigned Depth,
1907  size_t Idx, bool IsScaledReg = false);
1908  void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1909  void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1910  const Formula &Base, size_t Idx,
1911  bool IsScaledReg = false);
1912  void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1913  void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1914  const Formula &Base,
1915  const SmallVectorImpl<int64_t> &Worklist,
1916  size_t Idx, bool IsScaledReg = false);
1917  void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1918  void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1919  void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1920  void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1921  void GenerateCrossUseConstantOffsets();
1922  void GenerateAllReuseFormulae();
1923 
1924  void FilterOutUndesirableDedicatedRegisters();
1925 
1926  size_t EstimateSearchSpaceComplexity() const;
1927  void NarrowSearchSpaceByDetectingSupersets();
1928  void NarrowSearchSpaceByCollapsingUnrolledCode();
1929  void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1930  void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
1931  void NarrowSearchSpaceByDeletingCostlyFormulas();
1932  void NarrowSearchSpaceByPickingWinnerRegs();
1933  void NarrowSearchSpaceUsingHeuristics();
1934 
1935  void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1936  Cost &SolutionCost,
1938  const Cost &CurCost,
1939  const SmallPtrSet<const SCEV *, 16> &CurRegs,
1940  DenseSet<const SCEV *> &VisitedRegs) const;
1941  void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1942 
1944  HoistInsertPosition(BasicBlock::iterator IP,
1945  const SmallVectorImpl<Instruction *> &Inputs) const;
1947  AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1948  const LSRFixup &LF,
1949  const LSRUse &LU,
1950  SCEVExpander &Rewriter) const;
1951 
1952  Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1953  BasicBlock::iterator IP, SCEVExpander &Rewriter,
1954  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1955  void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
1956  const Formula &F, SCEVExpander &Rewriter,
1957  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1958  void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1959  SCEVExpander &Rewriter,
1960  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1961  void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
1962 
1963 public:
1964  LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
1965  LoopInfo &LI, const TargetTransformInfo &TTI);
1966 
1967  bool getChanged() const { return Changed; }
1968 
1969  void print_factors_and_types(raw_ostream &OS) const;
1970  void print_fixups(raw_ostream &OS) const;
1971  void print_uses(raw_ostream &OS) const;
1972  void print(raw_ostream &OS) const;
1973  void dump() const;
1974 };
1975 
1976 } // end anonymous namespace
1977 
1978 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
1979 /// the cast operation.
1980 void LSRInstance::OptimizeShadowIV() {
1981  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1982  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1983  return;
1984 
1985  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1986  UI != E; /* empty */) {
1987  IVUsers::const_iterator CandidateUI = UI;
1988  ++UI;
1989  Instruction *ShadowUse = CandidateUI->getUser();
1990  Type *DestTy = nullptr;
1991  bool IsSigned = false;
1992 
1993  /* If shadow use is a int->float cast then insert a second IV
1994  to eliminate this cast.
1995 
1996  for (unsigned i = 0; i < n; ++i)
1997  foo((double)i);
1998 
1999  is transformed into
2000 
2001  double d = 0.0;
2002  for (unsigned i = 0; i < n; ++i, ++d)
2003  foo(d);
2004  */
2005  if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2006  IsSigned = false;
2007  DestTy = UCast->getDestTy();
2008  }
2009  else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2010  IsSigned = true;
2011  DestTy = SCast->getDestTy();
2012  }
2013  if (!DestTy) continue;
2014 
2015  // If target does not support DestTy natively then do not apply
2016  // this transformation.
2017  if (!TTI.isTypeLegal(DestTy)) continue;
2018 
2019  PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2020  if (!PH) continue;
2021  if (PH->getNumIncomingValues() != 2) continue;
2022 
2023  Type *SrcTy = PH->getType();
2024  int Mantissa = DestTy->getFPMantissaWidth();
2025  if (Mantissa == -1) continue;
2026  if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2027  continue;
2028 
2029  unsigned Entry, Latch;
2030  if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2031  Entry = 0;
2032  Latch = 1;
2033  } else {
2034  Entry = 1;
2035  Latch = 0;
2036  }
2037 
2039  if (!Init) continue;
2040  Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2041  (double)Init->getSExtValue() :
2042  (double)Init->getZExtValue());
2043 
2044  BinaryOperator *Incr =
2046  if (!Incr) continue;
2047  if (Incr->getOpcode() != Instruction::Add
2048  && Incr->getOpcode() != Instruction::Sub)
2049  continue;
2050 
2051  /* Initialize new IV, double d = 0.0 in above example. */
2052  ConstantInt *C = nullptr;
2053  if (Incr->getOperand(0) == PH)
2054  C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2055  else if (Incr->getOperand(1) == PH)
2056  C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2057  else
2058  continue;
2059 
2060  if (!C) continue;
2061 
2062  // Ignore negative constants, as the code below doesn't handle them
2063  // correctly. TODO: Remove this restriction.
2064  if (!C->getValue().isStrictlyPositive()) continue;
2065 
2066  /* Add new PHINode. */
2067  PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2068 
2069  /* create new increment. '++d' in above example. */
2070  Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2071  BinaryOperator *NewIncr =
2072  BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2073  Instruction::FAdd : Instruction::FSub,
2074  NewPH, CFP, "IV.S.next.", Incr);
2075 
2076  NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2077  NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2078 
2079  /* Remove cast operation */
2080  ShadowUse->replaceAllUsesWith(NewPH);
2081  ShadowUse->eraseFromParent();
2082  Changed = true;
2083  break;
2084  }
2085 }
2086 
2087 /// If Cond has an operand that is an expression of an IV, set the IV user and
2088 /// stride information and return true, otherwise return false.
2089 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2090  for (IVStrideUse &U : IU)
2091  if (U.getUser() == Cond) {
2092  // NOTE: we could handle setcc instructions with multiple uses here, but
2093  // InstCombine does it as well for simple uses, it's not clear that it
2094  // occurs enough in real life to handle.
2095  CondUse = &U;
2096  return true;
2097  }
2098  return false;
2099 }
2100 
2101 /// Rewrite the loop's terminating condition if it uses a max computation.
2102 ///
2103 /// This is a narrow solution to a specific, but acute, problem. For loops
2104 /// like this:
2105 ///
2106 /// i = 0;
2107 /// do {
2108 /// p[i] = 0.0;
2109 /// } while (++i < n);
2110 ///
2111 /// the trip count isn't just 'n', because 'n' might not be positive. And
2112 /// unfortunately this can come up even for loops where the user didn't use
2113 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2114 /// will commonly be lowered like this:
2115 //
2116 /// if (n > 0) {
2117 /// i = 0;
2118 /// do {
2119 /// p[i] = 0.0;
2120 /// } while (++i < n);
2121 /// }
2122 ///
2123 /// and then it's possible for subsequent optimization to obscure the if
2124 /// test in such a way that indvars can't find it.
2125 ///
2126 /// When indvars can't find the if test in loops like this, it creates a
2127 /// max expression, which allows it to give the loop a canonical
2128 /// induction variable:
2129 ///
2130 /// i = 0;
2131 /// max = n < 1 ? 1 : n;
2132 /// do {
2133 /// p[i] = 0.0;
2134 /// } while (++i != max);
2135 ///
2136 /// Canonical induction variables are necessary because the loop passes
2137 /// are designed around them. The most obvious example of this is the
2138 /// LoopInfo analysis, which doesn't remember trip count values. It
2139 /// expects to be able to rediscover the trip count each time it is
2140 /// needed, and it does this using a simple analysis that only succeeds if
2141 /// the loop has a canonical induction variable.
2142 ///
2143 /// However, when it comes time to generate code, the maximum operation
2144 /// can be quite costly, especially if it's inside of an outer loop.
2145 ///
2146 /// This function solves this problem by detecting this type of loop and
2147 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2148 /// the instructions for the maximum computation.
2149 ///
2150 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2151  // Check that the loop matches the pattern we're looking for.
2152  if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2153  Cond->getPredicate() != CmpInst::ICMP_NE)
2154  return Cond;
2155 
2156  SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2157  if (!Sel || !Sel->hasOneUse()) return Cond;
2158 
2159  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2160  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2161  return Cond;
2162  const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2163 
2164  // Add one to the backedge-taken count to get the trip count.
2165  const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2166  if (IterationCount != SE.getSCEV(Sel)) return Cond;
2167 
2168  // Check for a max calculation that matches the pattern. There's no check
2169  // for ICMP_ULE here because the comparison would be with zero, which
2170  // isn't interesting.
2172  const SCEVNAryExpr *Max = nullptr;
2173  if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2174  Pred = ICmpInst::ICMP_SLE;
2175  Max = S;
2176  } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2177  Pred = ICmpInst::ICMP_SLT;
2178  Max = S;
2179  } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2180  Pred = ICmpInst::ICMP_ULT;
2181  Max = U;
2182  } else {
2183  // No match; bail.
2184  return Cond;
2185  }
2186 
2187  // To handle a max with more than two operands, this optimization would
2188  // require additional checking and setup.
2189  if (Max->getNumOperands() != 2)
2190  return Cond;
2191 
2192  const SCEV *MaxLHS = Max->getOperand(0);
2193  const SCEV *MaxRHS = Max->getOperand(1);
2194 
2195  // ScalarEvolution canonicalizes constants to the left. For < and >, look
2196  // for a comparison with 1. For <= and >=, a comparison with zero.
2197  if (!MaxLHS ||
2198  (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2199  return Cond;
2200 
2201  // Check the relevant induction variable for conformance to
2202  // the pattern.
2203  const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2204  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2205  if (!AR || !AR->isAffine() ||
2206  AR->getStart() != One ||
2207  AR->getStepRecurrence(SE) != One)
2208  return Cond;
2209 
2210  assert(AR->getLoop() == L &&
2211  "Loop condition operand is an addrec in a different loop!");
2212 
2213  // Check the right operand of the select, and remember it, as it will
2214  // be used in the new comparison instruction.
2215  Value *NewRHS = nullptr;
2216  if (ICmpInst::isTrueWhenEqual(Pred)) {
2217  // Look for n+1, and grab n.
2218  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2219  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2220  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2221  NewRHS = BO->getOperand(0);
2222  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2223  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2224  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2225  NewRHS = BO->getOperand(0);
2226  if (!NewRHS)
2227  return Cond;
2228  } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2229  NewRHS = Sel->getOperand(1);
2230  else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2231  NewRHS = Sel->getOperand(2);
2232  else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2233  NewRHS = SU->getValue();
2234  else
2235  // Max doesn't match expected pattern.
2236  return Cond;
2237 
2238  // Determine the new comparison opcode. It may be signed or unsigned,
2239  // and the original comparison may be either equality or inequality.
2240  if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2241  Pred = CmpInst::getInversePredicate(Pred);
2242 
2243  // Ok, everything looks ok to change the condition into an SLT or SGE and
2244  // delete the max calculation.
2245  ICmpInst *NewCond =
2246  new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2247 
2248  // Delete the max calculation instructions.
2249  Cond->replaceAllUsesWith(NewCond);
2250  CondUse->setUser(NewCond);
2251  Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2252  Cond->eraseFromParent();
2253  Sel->eraseFromParent();
2254  if (Cmp->use_empty())
2255  Cmp->eraseFromParent();
2256  return NewCond;
2257 }
2258 
2259 /// Change loop terminating condition to use the postinc iv when possible.
2260 void
2261 LSRInstance::OptimizeLoopTermCond() {
2263 
2264  // We need a different set of heuristics for rotated and non-rotated loops.
2265  // If a loop is rotated then the latch is also the backedge, so inserting
2266  // post-inc expressions just before the latch is ideal. To reduce live ranges
2267  // it also makes sense to rewrite terminating conditions to use post-inc
2268  // expressions.
2269  //
2270  // If the loop is not rotated then the latch is not a backedge; the latch
2271  // check is done in the loop head. Adding post-inc expressions before the
2272  // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2273  // in the loop body. In this case we do *not* want to use post-inc expressions
2274  // in the latch check, and we want to insert post-inc expressions before
2275  // the backedge.
2276  BasicBlock *LatchBlock = L->getLoopLatch();
2277  SmallVector<BasicBlock*, 8> ExitingBlocks;
2278  L->getExitingBlocks(ExitingBlocks);
2279  if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2280  return LatchBlock != BB;
2281  })) {
2282  // The backedge doesn't exit the loop; treat this as a head-tested loop.
2283  IVIncInsertPos = LatchBlock->getTerminator();
2284  return;
2285  }
2286 
2287  // Otherwise treat this as a rotated loop.
2288  for (BasicBlock *ExitingBlock : ExitingBlocks) {
2289 
2290  // Get the terminating condition for the loop if possible. If we
2291  // can, we want to change it to use a post-incremented version of its
2292  // induction variable, to allow coalescing the live ranges for the IV into
2293  // one register value.
2294 
2295  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2296  if (!TermBr)
2297  continue;
2298  // FIXME: Overly conservative, termination condition could be an 'or' etc..
2299  if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2300  continue;
2301 
2302  // Search IVUsesByStride to find Cond's IVUse if there is one.
2303  IVStrideUse *CondUse = nullptr;
2304  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2305  if (!FindIVUserForCond(Cond, CondUse))
2306  continue;
2307 
2308  // If the trip count is computed in terms of a max (due to ScalarEvolution
2309  // being unable to find a sufficient guard, for example), change the loop
2310  // comparison to use SLT or ULT instead of NE.
2311  // One consequence of doing this now is that it disrupts the count-down
2312  // optimization. That's not always a bad thing though, because in such
2313  // cases it may still be worthwhile to avoid a max.
2314  Cond = OptimizeMax(Cond, CondUse);
2315 
2316  // If this exiting block dominates the latch block, it may also use
2317  // the post-inc value if it won't be shared with other uses.
2318  // Check for dominance.
2319  if (!DT.dominates(ExitingBlock, LatchBlock))
2320  continue;
2321 
2322  // Conservatively avoid trying to use the post-inc value in non-latch
2323  // exits if there may be pre-inc users in intervening blocks.
2324  if (LatchBlock != ExitingBlock)
2325  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2326  // Test if the use is reachable from the exiting block. This dominator
2327  // query is a conservative approximation of reachability.
2328  if (&*UI != CondUse &&
2329  !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2330  // Conservatively assume there may be reuse if the quotient of their
2331  // strides could be a legal scale.
2332  const SCEV *A = IU.getStride(*CondUse, L);
2333  const SCEV *B = IU.getStride(*UI, L);
2334  if (!A || !B) continue;
2335  if (SE.getTypeSizeInBits(A->getType()) !=
2336  SE.getTypeSizeInBits(B->getType())) {
2337  if (SE.getTypeSizeInBits(A->getType()) >
2338  SE.getTypeSizeInBits(B->getType()))
2339  B = SE.getSignExtendExpr(B, A->getType());
2340  else
2341  A = SE.getSignExtendExpr(A, B->getType());
2342  }
2343  if (const SCEVConstant *D =
2344  dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2345  const ConstantInt *C = D->getValue();
2346  // Stride of one or negative one can have reuse with non-addresses.
2347  if (C->isOne() || C->isMinusOne())
2348  goto decline_post_inc;
2349  // Avoid weird situations.
2350  if (C->getValue().getMinSignedBits() >= 64 ||
2351  C->getValue().isMinSignedValue())
2352  goto decline_post_inc;
2353  // Check for possible scaled-address reuse.
2354  MemAccessTy AccessTy = getAccessType(UI->getUser());
2355  int64_t Scale = C->getSExtValue();
2356  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2357  /*BaseOffset=*/0,
2358  /*HasBaseReg=*/false, Scale,
2359  AccessTy.AddrSpace))
2360  goto decline_post_inc;
2361  Scale = -Scale;
2362  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2363  /*BaseOffset=*/0,
2364  /*HasBaseReg=*/false, Scale,
2365  AccessTy.AddrSpace))
2366  goto decline_post_inc;
2367  }
2368  }
2369 
2370  DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2371  << *Cond << '\n');
2372 
2373  // It's possible for the setcc instruction to be anywhere in the loop, and
2374  // possible for it to have multiple users. If it is not immediately before
2375  // the exiting block branch, move it.
2376  if (&*++BasicBlock::iterator(Cond) != TermBr) {
2377  if (Cond->hasOneUse()) {
2378  Cond->moveBefore(TermBr);
2379  } else {
2380  // Clone the terminating condition and insert into the loopend.
2381  ICmpInst *OldCond = Cond;
2382  Cond = cast<ICmpInst>(Cond->clone());
2383  Cond->setName(L->getHeader()->getName() + ".termcond");
2384  ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2385 
2386  // Clone the IVUse, as the old use still exists!
2387  CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2388  TermBr->replaceUsesOfWith(OldCond, Cond);
2389  }
2390  }
2391 
2392  // If we get to here, we know that we can transform the setcc instruction to
2393  // use the post-incremented version of the IV, allowing us to coalesce the
2394  // live ranges for the IV correctly.
2395  CondUse->transformToPostInc(L);
2396  Changed = true;
2397 
2398  PostIncs.insert(Cond);
2399  decline_post_inc:;
2400  }
2401 
2402  // Determine an insertion point for the loop induction variable increment. It
2403  // must dominate all the post-inc comparisons we just set up, and it must
2404  // dominate the loop latch edge.
2405  IVIncInsertPos = L->getLoopLatch()->getTerminator();
2406  for (Instruction *Inst : PostIncs) {
2407  BasicBlock *BB =
2408  DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2409  Inst->getParent());
2410  if (BB == Inst->getParent())
2411  IVIncInsertPos = Inst;
2412  else if (BB != IVIncInsertPos->getParent())
2413  IVIncInsertPos = BB->getTerminator();
2414  }
2415 }
2416 
2417 /// Determine if the given use can accommodate a fixup at the given offset and
2418 /// other details. If so, update the use and return true.
2419 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2420  bool HasBaseReg, LSRUse::KindType Kind,
2421  MemAccessTy AccessTy) {
2422  int64_t NewMinOffset = LU.MinOffset;
2423  int64_t NewMaxOffset = LU.MaxOffset;
2424  MemAccessTy NewAccessTy = AccessTy;
2425 
2426  // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2427  // something conservative, however this can pessimize in the case that one of
2428  // the uses will have all its uses outside the loop, for example.
2429  if (LU.Kind != Kind)
2430  return false;
2431 
2432  // Check for a mismatched access type, and fall back conservatively as needed.
2433  // TODO: Be less conservative when the type is similar and can use the same
2434  // addressing modes.
2435  if (Kind == LSRUse::Address) {
2436  if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2437  NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2438  AccessTy.AddrSpace);
2439  }
2440  }
2441 
2442  // Conservatively assume HasBaseReg is true for now.
2443  if (NewOffset < LU.MinOffset) {
2444  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2445  LU.MaxOffset - NewOffset, HasBaseReg))
2446  return false;
2447  NewMinOffset = NewOffset;
2448  } else if (NewOffset > LU.MaxOffset) {
2449  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2450  NewOffset - LU.MinOffset, HasBaseReg))
2451  return false;
2452  NewMaxOffset = NewOffset;
2453  }
2454 
2455  // Update the use.
2456  LU.MinOffset = NewMinOffset;
2457  LU.MaxOffset = NewMaxOffset;
2458  LU.AccessTy = NewAccessTy;
2459  return true;
2460 }
2461 
2462 /// Return an LSRUse index and an offset value for a fixup which needs the given
2463 /// expression, with the given kind and optional access type. Either reuse an
2464 /// existing use or create a new one, as needed.
2465 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2466  LSRUse::KindType Kind,
2467  MemAccessTy AccessTy) {
2468  const SCEV *Copy = Expr;
2469  int64_t Offset = ExtractImmediate(Expr, SE);
2470 
2471  // Basic uses can't accept any offset, for example.
2472  if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2473  Offset, /*HasBaseReg=*/ true)) {
2474  Expr = Copy;
2475  Offset = 0;
2476  }
2477 
2478  std::pair<UseMapTy::iterator, bool> P =
2479  UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2480  if (!P.second) {
2481  // A use already existed with this base.
2482  size_t LUIdx = P.first->second;
2483  LSRUse &LU = Uses[LUIdx];
2484  if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2485  // Reuse this use.
2486  return std::make_pair(LUIdx, Offset);
2487  }
2488 
2489  // Create a new use.
2490  size_t LUIdx = Uses.size();
2491  P.first->second = LUIdx;
2492  Uses.push_back(LSRUse(Kind, AccessTy));
2493  LSRUse &LU = Uses[LUIdx];
2494 
2495  LU.MinOffset = Offset;
2496  LU.MaxOffset = Offset;
2497  return std::make_pair(LUIdx, Offset);
2498 }
2499 
2500 /// Delete the given use from the Uses list.
2501 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2502  if (&LU != &Uses.back())
2503  std::swap(LU, Uses.back());
2504  Uses.pop_back();
2505 
2506  // Update RegUses.
2507  RegUses.swapAndDropUse(LUIdx, Uses.size());
2508 }
2509 
2510 /// Look for a use distinct from OrigLU which is has a formula that has the same
2511 /// registers as the given formula.
2512 LSRUse *
2513 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2514  const LSRUse &OrigLU) {
2515  // Search all uses for the formula. This could be more clever.
2516  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2517  LSRUse &LU = Uses[LUIdx];
2518  // Check whether this use is close enough to OrigLU, to see whether it's
2519  // worthwhile looking through its formulae.
2520  // Ignore ICmpZero uses because they may contain formulae generated by
2521  // GenerateICmpZeroScales, in which case adding fixup offsets may
2522  // be invalid.
2523  if (&LU != &OrigLU &&
2524  LU.Kind != LSRUse::ICmpZero &&
2525  LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2526  LU.WidestFixupType == OrigLU.WidestFixupType &&
2527  LU.HasFormulaWithSameRegs(OrigF)) {
2528  // Scan through this use's formulae.
2529  for (const Formula &F : LU.Formulae) {
2530  // Check to see if this formula has the same registers and symbols
2531  // as OrigF.
2532  if (F.BaseRegs == OrigF.BaseRegs &&
2533  F.ScaledReg == OrigF.ScaledReg &&
2534  F.BaseGV == OrigF.BaseGV &&
2535  F.Scale == OrigF.Scale &&
2536  F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2537  if (F.BaseOffset == 0)
2538  return &LU;
2539  // This is the formula where all the registers and symbols matched;
2540  // there aren't going to be any others. Since we declined it, we
2541  // can skip the rest of the formulae and proceed to the next LSRUse.
2542  break;
2543  }
2544  }
2545  }
2546  }
2547 
2548  // Nothing looked good.
2549  return nullptr;
2550 }
2551 
2552 void LSRInstance::CollectInterestingTypesAndFactors() {
2554 
2555  // Collect interesting types and strides.
2557  for (const IVStrideUse &U : IU) {
2558  const SCEV *Expr = IU.getExpr(U);
2559 
2560  // Collect interesting types.
2561  Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2562 
2563  // Add strides for mentioned loops.
2564  Worklist.push_back(Expr);
2565  do {
2566  const SCEV *S = Worklist.pop_back_val();
2567  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2568  if (AR->getLoop() == L)
2569  Strides.insert(AR->getStepRecurrence(SE));
2570  Worklist.push_back(AR->getStart());
2571  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2572  Worklist.append(Add->op_begin(), Add->op_end());
2573  }
2574  } while (!Worklist.empty());
2575  }
2576 
2577  // Compute interesting factors from the set of interesting strides.
2579  I = Strides.begin(), E = Strides.end(); I != E; ++I)
2581  std::next(I); NewStrideIter != E; ++NewStrideIter) {
2582  const SCEV *OldStride = *I;
2583  const SCEV *NewStride = *NewStrideIter;
2584 
2585  if (SE.getTypeSizeInBits(OldStride->getType()) !=
2586  SE.getTypeSizeInBits(NewStride->getType())) {
2587  if (SE.getTypeSizeInBits(OldStride->getType()) >
2588  SE.getTypeSizeInBits(NewStride->getType()))
2589  NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2590  else
2591  OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2592  }
2593  if (const SCEVConstant *Factor =
2594  dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2595  SE, true))) {
2596  if (Factor->getAPInt().getMinSignedBits() <= 64)
2597  Factors.insert(Factor->getAPInt().getSExtValue());
2598  } else if (const SCEVConstant *Factor =
2599  dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2600  NewStride,
2601  SE, true))) {
2602  if (Factor->getAPInt().getMinSignedBits() <= 64)
2603  Factors.insert(Factor->getAPInt().getSExtValue());
2604  }
2605  }
2606 
2607  // If all uses use the same type, don't bother looking for truncation-based
2608  // reuse.
2609  if (Types.size() == 1)
2610  Types.clear();
2611 
2612  DEBUG(print_factors_and_types(dbgs()));
2613 }
2614 
2615 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2616 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2617 /// IVStrideUses, we could partially skip this.
2618 static User::op_iterator
2620  Loop *L, ScalarEvolution &SE) {
2621  for(; OI != OE; ++OI) {
2622  if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2623  if (!SE.isSCEVable(Oper->getType()))
2624  continue;
2625 
2626  if (const SCEVAddRecExpr *AR =
2627  dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2628  if (AR->getLoop() == L)
2629  break;
2630  }
2631  }
2632  }
2633  return OI;
2634 }
2635 
2636 /// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2637 /// a convenient helper.
2638 static Value *getWideOperand(Value *Oper) {
2639  if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2640  return Trunc->getOperand(0);
2641  return Oper;
2642 }
2643 
2644 /// Return true if we allow an IV chain to include both types.
2645 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2646  Type *LType = LVal->getType();
2647  Type *RType = RVal->getType();
2648  return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2649  // Different address spaces means (possibly)
2650  // different types of the pointer implementation,
2651  // e.g. i16 vs i32 so disallow that.
2652  (LType->getPointerAddressSpace() ==
2653  RType->getPointerAddressSpace()));
2654 }
2655 
2656 /// Return an approximation of this SCEV expression's "base", or NULL for any
2657 /// constant. Returning the expression itself is conservative. Returning a
2658 /// deeper subexpression is more precise and valid as long as it isn't less
2659 /// complex than another subexpression. For expressions involving multiple
2660 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2661 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2662 /// IVInc==b-a.
2663 ///
2664 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2665 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2666 static const SCEV *getExprBase(const SCEV *S) {
2667  switch (S->getSCEVType()) {
2668  default: // uncluding scUnknown.
2669  return S;
2670  case scConstant:
2671  return nullptr;
2672  case scTruncate:
2673  return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2674  case scZeroExtend:
2675  return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2676  case scSignExtend:
2677  return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2678  case scAddExpr: {
2679  // Skip over scaled operands (scMulExpr) to follow add operands as long as
2680  // there's nothing more complex.
2681  // FIXME: not sure if we want to recognize negation.
2682  const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2683  for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2684  E(Add->op_begin()); I != E; ++I) {
2685  const SCEV *SubExpr = *I;
2686  if (SubExpr->getSCEVType() == scAddExpr)
2687  return getExprBase(SubExpr);
2688 
2689  if (SubExpr->getSCEVType() != scMulExpr)
2690  return SubExpr;
2691  }
2692  return S; // all operands are scaled, be conservative.
2693  }
2694  case scAddRecExpr:
2695  return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2696  }
2697 }
2698 
2699 /// Return true if the chain increment is profitable to expand into a loop
2700 /// invariant value, which may require its own register. A profitable chain
2701 /// increment will be an offset relative to the same base. We allow such offsets
2702 /// to potentially be used as chain increment as long as it's not obviously
2703 /// expensive to expand using real instructions.
2704 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2705  const SCEV *IncExpr,
2706  ScalarEvolution &SE) {
2707  // Aggressively form chains when -stress-ivchain.
2708  if (StressIVChain)
2709  return true;
2710 
2711  // Do not replace a constant offset from IV head with a nonconstant IV
2712  // increment.
2713  if (!isa<SCEVConstant>(IncExpr)) {
2714  const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2715  if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2716  return false;
2717  }
2718 
2719  SmallPtrSet<const SCEV*, 8> Processed;
2720  return !isHighCostExpansion(IncExpr, Processed, SE);
2721 }
2722 
2723 /// Return true if the number of registers needed for the chain is estimated to
2724 /// be less than the number required for the individual IV users. First prohibit
2725 /// any IV users that keep the IV live across increments (the Users set should
2726 /// be empty). Next count the number and type of increments in the chain.
2727 ///
2728 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2729 /// effectively use postinc addressing modes. Only consider it profitable it the
2730 /// increments can be computed in fewer registers when chained.
2731 ///
2732 /// TODO: Consider IVInc free if it's already used in another chains.
2733 static bool
2735  ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2736  if (StressIVChain)
2737  return true;
2738 
2739  if (!Chain.hasIncs())
2740  return false;
2741 
2742  if (!Users.empty()) {
2743  DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2744  for (Instruction *Inst : Users) {
2745  dbgs() << " " << *Inst << "\n";
2746  });
2747  return false;
2748  }
2749  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2750 
2751  // The chain itself may require a register, so intialize cost to 1.
2752  int cost = 1;
2753 
2754  // A complete chain likely eliminates the need for keeping the original IV in
2755  // a register. LSR does not currently know how to form a complete chain unless
2756  // the header phi already exists.
2757  if (isa<PHINode>(Chain.tailUserInst())
2758  && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2759  --cost;
2760  }
2761  const SCEV *LastIncExpr = nullptr;
2762  unsigned NumConstIncrements = 0;
2763  unsigned NumVarIncrements = 0;
2764  unsigned NumReusedIncrements = 0;
2765  for (const IVInc &Inc : Chain) {
2766  if (Inc.IncExpr->isZero())
2767  continue;
2768 
2769  // Incrementing by zero or some constant is neutral. We assume constants can
2770  // be folded into an addressing mode or an add's immediate operand.
2771  if (isa<SCEVConstant>(Inc.IncExpr)) {
2772  ++NumConstIncrements;
2773  continue;
2774  }
2775 
2776  if (Inc.IncExpr == LastIncExpr)
2777  ++NumReusedIncrements;
2778  else
2779  ++NumVarIncrements;
2780 
2781  LastIncExpr = Inc.IncExpr;
2782  }
2783  // An IV chain with a single increment is handled by LSR's postinc
2784  // uses. However, a chain with multiple increments requires keeping the IV's
2785  // value live longer than it needs to be if chained.
2786  if (NumConstIncrements > 1)
2787  --cost;
2788 
2789  // Materializing increment expressions in the preheader that didn't exist in
2790  // the original code may cost a register. For example, sign-extended array
2791  // indices can produce ridiculous increments like this:
2792  // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2793  cost += NumVarIncrements;
2794 
2795  // Reusing variable increments likely saves a register to hold the multiple of
2796  // the stride.
2797  cost -= NumReusedIncrements;
2798 
2799  DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2800  << "\n");
2801 
2802  return cost < 0;
2803 }
2804 
2805 /// Add this IV user to an existing chain or make it the head of a new chain.
2806 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2807  SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2808  // When IVs are used as types of varying widths, they are generally converted
2809  // to a wider type with some uses remaining narrow under a (free) trunc.
2810  Value *const NextIV = getWideOperand(IVOper);
2811  const SCEV *const OperExpr = SE.getSCEV(NextIV);
2812  const SCEV *const OperExprBase = getExprBase(OperExpr);
2813 
2814  // Visit all existing chains. Check if its IVOper can be computed as a
2815  // profitable loop invariant increment from the last link in the Chain.
2816  unsigned ChainIdx = 0, NChains = IVChainVec.size();
2817  const SCEV *LastIncExpr = nullptr;
2818  for (; ChainIdx < NChains; ++ChainIdx) {
2819  IVChain &Chain = IVChainVec[ChainIdx];
2820 
2821  // Prune the solution space aggressively by checking that both IV operands
2822  // are expressions that operate on the same unscaled SCEVUnknown. This
2823  // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2824  // first avoids creating extra SCEV expressions.
2825  if (!StressIVChain && Chain.ExprBase != OperExprBase)
2826  continue;
2827 
2828  Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2829  if (!isCompatibleIVType(PrevIV, NextIV))
2830  continue;
2831 
2832  // A phi node terminates a chain.
2833  if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2834  continue;
2835 
2836  // The increment must be loop-invariant so it can be kept in a register.
2837  const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2838  const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2839  if (!SE.isLoopInvariant(IncExpr, L))
2840  continue;
2841 
2842  if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2843  LastIncExpr = IncExpr;
2844  break;
2845  }
2846  }
2847  // If we haven't found a chain, create a new one, unless we hit the max. Don't
2848  // bother for phi nodes, because they must be last in the chain.
2849  if (ChainIdx == NChains) {
2850  if (isa<PHINode>(UserInst))
2851  return;
2852  if (NChains >= MaxChains && !StressIVChain) {
2853  DEBUG(dbgs() << "IV Chain Limit\n");
2854  return;
2855  }
2856  LastIncExpr = OperExpr;
2857  // IVUsers may have skipped over sign/zero extensions. We don't currently
2858  // attempt to form chains involving extensions unless they can be hoisted
2859  // into this loop's AddRec.
2860  if (!isa<SCEVAddRecExpr>(LastIncExpr))
2861  return;
2862  ++NChains;
2863  IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2864  OperExprBase));
2865  ChainUsersVec.resize(NChains);
2866  DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2867  << ") IV=" << *LastIncExpr << "\n");
2868  } else {
2869  DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2870  << ") IV+" << *LastIncExpr << "\n");
2871  // Add this IV user to the end of the chain.
2872  IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2873  }
2874  IVChain &Chain = IVChainVec[ChainIdx];
2875 
2876  SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2877  // This chain's NearUsers become FarUsers.
2878  if (!LastIncExpr->isZero()) {
2879  ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2880  NearUsers.end());
2881  NearUsers.clear();
2882  }
2883 
2884  // All other uses of IVOperand become near uses of the chain.
2885  // We currently ignore intermediate values within SCEV expressions, assuming
2886  // they will eventually be used be the current chain, or can be computed
2887  // from one of the chain increments. To be more precise we could
2888  // transitively follow its user and only add leaf IV users to the set.
2889  for (User *U : IVOper->users()) {
2890  Instruction *OtherUse = dyn_cast<Instruction>(U);
2891  if (!OtherUse)
2892  continue;
2893  // Uses in the chain will no longer be uses if the chain is formed.
2894  // Include the head of the chain in this iteration (not Chain.begin()).
2895  IVChain::const_iterator IncIter = Chain.Incs.begin();
2896  IVChain::const_iterator IncEnd = Chain.Incs.end();
2897  for( ; IncIter != IncEnd; ++IncIter) {
2898  if (IncIter->UserInst == OtherUse)
2899  break;
2900  }
2901  if (IncIter != IncEnd)
2902  continue;
2903 
2904  if (SE.isSCEVable(OtherUse->getType())
2905  && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2906  && IU.isIVUserOrOperand(OtherUse)) {
2907  continue;
2908  }
2909  NearUsers.insert(OtherUse);
2910  }
2911 
2912  // Since this user is part of the chain, it's no longer considered a use
2913  // of the chain.
2914  ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2915 }
2916 
2917 /// Populate the vector of Chains.
2918 ///
2919 /// This decreases ILP at the architecture level. Targets with ample registers,
2920 /// multiple memory ports, and no register renaming probably don't want
2921 /// this. However, such targets should probably disable LSR altogether.
2922 ///
2923 /// The job of LSR is to make a reasonable choice of induction variables across
2924 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2925 /// ILP *within the loop* if the target wants it.
2926 ///
2927 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2928 /// will not reorder memory operations, it will recognize this as a chain, but
2929 /// will generate redundant IV increments. Ideally this would be corrected later
2930 /// by a smart scheduler:
2931 /// = A[i]
2932 /// = A[i+x]
2933 /// A[i] =
2934 /// A[i+x] =
2935 ///
2936 /// TODO: Walk the entire domtree within this loop, not just the path to the
2937 /// loop latch. This will discover chains on side paths, but requires
2938 /// maintaining multiple copies of the Chains state.
2939 void LSRInstance::CollectChains() {
2940  DEBUG(dbgs() << "Collecting IV Chains.\n");
2941  SmallVector<ChainUsers, 8> ChainUsersVec;
2942 
2943  SmallVector<BasicBlock *,8> LatchPath;
2944  BasicBlock *LoopHeader = L->getHeader();
2945  for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2946  Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2947  LatchPath.push_back(Rung->getBlock());
2948  }
2949  LatchPath.push_back(LoopHeader);
2950 
2951  // Walk the instruction stream from the loop header to the loop latch.
2952  for (BasicBlock *BB : reverse(LatchPath)) {
2953  for (Instruction &I : *BB) {
2954  // Skip instructions that weren't seen by IVUsers analysis.
2955  if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
2956  continue;
2957 
2958  // Ignore users that are part of a SCEV expression. This way we only
2959  // consider leaf IV Users. This effectively rediscovers a portion of
2960  // IVUsers analysis but in program order this time.
2961  if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
2962  continue;
2963 
2964  // Remove this instruction from any NearUsers set it may be in.
2965  for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2966  ChainIdx < NChains; ++ChainIdx) {
2967  ChainUsersVec[ChainIdx].NearUsers.erase(&I);
2968  }
2969  // Search for operands that can be chained.
2970  SmallPtrSet<Instruction*, 4> UniqueOperands;
2971  User::op_iterator IVOpEnd = I.op_end();
2972  User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
2973  while (IVOpIter != IVOpEnd) {
2974  Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2975  if (UniqueOperands.insert(IVOpInst).second)
2976  ChainInstruction(&I, IVOpInst, ChainUsersVec);
2977  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2978  }
2979  } // Continue walking down the instructions.
2980  } // Continue walking down the domtree.
2981  // Visit phi backedges to determine if the chain can generate the IV postinc.
2982  for (BasicBlock::iterator I = L->getHeader()->begin();
2983  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2984  if (!SE.isSCEVable(PN->getType()))
2985  continue;
2986 
2987  Instruction *IncV =
2989  if (IncV)
2990  ChainInstruction(PN, IncV, ChainUsersVec);
2991  }
2992  // Remove any unprofitable chains.
2993  unsigned ChainIdx = 0;
2994  for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2995  UsersIdx < NChains; ++UsersIdx) {
2996  if (!isProfitableChain(IVChainVec[UsersIdx],
2997  ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2998  continue;
2999  // Preserve the chain at UsesIdx.
3000  if (ChainIdx != UsersIdx)
3001  IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3002  FinalizeChain(IVChainVec[ChainIdx]);
3003  ++ChainIdx;
3004  }
3005  IVChainVec.resize(ChainIdx);
3006 }
3007 
3008 void LSRInstance::FinalizeChain(IVChain &Chain) {
3009  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3010  DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3011 
3012  for (const IVInc &Inc : Chain) {
3013  DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
3014  auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3015  assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3016  IVIncSet.insert(UseI);
3017  }
3018 }
3019 
3020 /// Return true if the IVInc can be folded into an addressing mode.
3021 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3022  Value *Operand, const TargetTransformInfo &TTI) {
3023  const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3024  if (!IncConst || !isAddressUse(UserInst, Operand))
3025  return false;
3026 
3027  if (IncConst->getAPInt().getMinSignedBits() > 64)
3028  return false;
3029 
3030  MemAccessTy AccessTy = getAccessType(UserInst);
3031  int64_t IncOffset = IncConst->getValue()->getSExtValue();
3032  if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3033  IncOffset, /*HaseBaseReg=*/false))
3034  return false;
3035 
3036  return true;
3037 }
3038 
3039 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3040 /// user's operand from the previous IV user's operand.
3041 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3042  SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3043  // Find the new IVOperand for the head of the chain. It may have been replaced
3044  // by LSR.
3045  const IVInc &Head = Chain.Incs[0];
3046  User::op_iterator IVOpEnd = Head.UserInst->op_end();
3047  // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3048  User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3049  IVOpEnd, L, SE);
3050  Value *IVSrc = nullptr;
3051  while (IVOpIter != IVOpEnd) {
3052  IVSrc = getWideOperand(*IVOpIter);
3053 
3054  // If this operand computes the expression that the chain needs, we may use
3055  // it. (Check this after setting IVSrc which is used below.)
3056  //
3057  // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3058  // narrow for the chain, so we can no longer use it. We do allow using a
3059  // wider phi, assuming the LSR checked for free truncation. In that case we
3060  // should already have a truncate on this operand such that
3061  // getSCEV(IVSrc) == IncExpr.
3062  if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3063  || SE.getSCEV(IVSrc) == Head.IncExpr) {
3064  break;
3065  }
3066  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3067  }
3068  if (IVOpIter == IVOpEnd) {
3069  // Gracefully give up on this chain.
3070  DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3071  return;
3072  }
3073 
3074  DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3075  Type *IVTy = IVSrc->getType();
3076  Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3077  const SCEV *LeftOverExpr = nullptr;
3078  for (const IVInc &Inc : Chain) {
3079  Instruction *InsertPt = Inc.UserInst;
3080  if (isa<PHINode>(InsertPt))
3081  InsertPt = L->getLoopLatch()->getTerminator();
3082 
3083  // IVOper will replace the current IV User's operand. IVSrc is the IV
3084  // value currently held in a register.
3085  Value *IVOper = IVSrc;
3086  if (!Inc.IncExpr->isZero()) {
3087  // IncExpr was the result of subtraction of two narrow values, so must
3088  // be signed.
3089  const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3090  LeftOverExpr = LeftOverExpr ?
3091  SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3092  }
3093  if (LeftOverExpr && !LeftOverExpr->isZero()) {
3094  // Expand the IV increment.
3095  Rewriter.clearPostInc();
3096  Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3097  const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3098  SE.getUnknown(IncV));
3099  IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3100 
3101  // If an IV increment can't be folded, use it as the next IV value.
3102  if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3103  assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3104  IVSrc = IVOper;
3105  LeftOverExpr = nullptr;
3106  }
3107  }
3108  Type *OperTy = Inc.IVOperand->getType();
3109  if (IVTy != OperTy) {
3110  assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3111  "cannot extend a chained IV");
3112  IRBuilder<> Builder(InsertPt);
3113  IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3114  }
3115  Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3116  DeadInsts.emplace_back(Inc.IVOperand);
3117  }
3118  // If LSR created a new, wider phi, we may also replace its postinc. We only
3119  // do this if we also found a wide value for the head of the chain.
3120  if (isa<PHINode>(Chain.tailUserInst())) {
3121  for (BasicBlock::iterator I = L->getHeader()->begin();
3122  PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
3123  if (!isCompatibleIVType(Phi, IVSrc))
3124  continue;
3125  Instruction *PostIncV = dyn_cast<Instruction>(
3126  Phi->getIncomingValueForBlock(L->getLoopLatch()));
3127  if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3128  continue;
3129  Value *IVOper = IVSrc;
3130  Type *PostIncTy = PostIncV->getType();
3131  if (IVTy != PostIncTy) {
3132  assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3133  IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3134  Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3135  IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3136  }
3137  Phi->replaceUsesOfWith(PostIncV, IVOper);
3138  DeadInsts.emplace_back(PostIncV);
3139  }
3140  }
3141 }
3142 
3143 void LSRInstance::CollectFixupsAndInitialFormulae() {
3144  for (const IVStrideUse &U : IU) {
3145  Instruction *UserInst = U.getUser();
3146  // Skip IV users that are part of profitable IV Chains.
3147  User::op_iterator UseI =
3148  find(UserInst->operands(), U.getOperandValToReplace());
3149  assert(UseI != UserInst->op_end() && "cannot find IV operand");
3150  if (IVIncSet.count(UseI)) {
3151  DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3152  continue;
3153  }
3154 
3155  LSRUse::KindType Kind = LSRUse::Basic;
3156  MemAccessTy AccessTy;
3157  if (isAddressUse(UserInst, U.getOperandValToReplace())) {
3158  Kind = LSRUse::Address;
3159  AccessTy = getAccessType(UserInst);
3160  }
3161 
3162  const SCEV *S = IU.getExpr(U);
3163  PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3164 
3165  // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3166  // (N - i == 0), and this allows (N - i) to be the expression that we work
3167  // with rather than just N or i, so we can consider the register
3168  // requirements for both N and i at the same time. Limiting this code to
3169  // equality icmps is not a problem because all interesting loops use
3170  // equality icmps, thanks to IndVarSimplify.
3171  if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3172  if (CI->isEquality()) {
3173  // Swap the operands if needed to put the OperandValToReplace on the
3174  // left, for consistency.
3175  Value *NV = CI->getOperand(1);
3176  if (NV == U.getOperandValToReplace()) {
3177  CI->setOperand(1, CI->getOperand(0));
3178  CI->setOperand(0, NV);
3179  NV = CI->getOperand(1);
3180  Changed = true;
3181  }
3182 
3183  // x == y --> x - y == 0
3184  const SCEV *N = SE.getSCEV(NV);
3185  if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3186  // S is normalized, so normalize N before folding it into S
3187  // to keep the result normalized.
3188  N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3189  Kind = LSRUse::ICmpZero;
3190  S = SE.getMinusSCEV(N, S);
3191  }
3192 
3193  // -1 and the negations of all interesting strides (except the negation
3194  // of -1) are now also interesting.
3195  for (size_t i = 0, e = Factors.size(); i != e; ++i)
3196  if (Factors[i] != -1)
3197  Factors.insert(-(uint64_t)Factors[i]);
3198  Factors.insert(-1);
3199  }
3200 
3201  // Get or create an LSRUse.
3202  std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3203  size_t LUIdx = P.first;
3204  int64_t Offset = P.second;
3205  LSRUse &LU = Uses[LUIdx];
3206 
3207  // Record the fixup.
3208  LSRFixup &LF = LU.getNewFixup();
3209  LF.UserInst = UserInst;
3210  LF.OperandValToReplace = U.getOperandValToReplace();
3211  LF.PostIncLoops = TmpPostIncLoops;
3212  LF.Offset = Offset;
3213  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3214 
3215  if (!LU.WidestFixupType ||
3216  SE.getTypeSizeInBits(LU.WidestFixupType) <
3217  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3218  LU.WidestFixupType = LF.OperandValToReplace->getType();
3219 
3220  // If this is the first use of this LSRUse, give it a formula.
3221  if (LU.Formulae.empty()) {
3222  InsertInitialFormula(S, LU, LUIdx);
3223  CountRegisters(LU.Formulae.back(), LUIdx);
3224  }
3225  }
3226 
3227  DEBUG(print_fixups(dbgs()));
3228 }
3229 
3230 /// Insert a formula for the given expression into the given use, separating out
3231 /// loop-variant portions from loop-invariant and loop-computable portions.
3232 void
3233 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3234  // Mark uses whose expressions cannot be expanded.
3235  if (!isSafeToExpand(S, SE))
3236  LU.RigidFormula = true;
3237 
3238  Formula F;
3239  F.initialMatch(S, L, SE);
3240  bool Inserted = InsertFormula(LU, LUIdx, F);
3241  assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3242 }
3243 
3244 /// Insert a simple single-register formula for the given expression into the
3245 /// given use.
3246 void
3247 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3248  LSRUse &LU, size_t LUIdx) {
3249  Formula F;
3250  F.BaseRegs.push_back(S);
3251  F.HasBaseReg = true;
3252  bool Inserted = InsertFormula(LU, LUIdx, F);
3253  assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3254 }
3255 
3256 /// Note which registers are used by the given formula, updating RegUses.
3257 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3258  if (F.ScaledReg)
3259  RegUses.countRegister(F.ScaledReg, LUIdx);
3260  for (const SCEV *BaseReg : F.BaseRegs)
3261  RegUses.countRegister(BaseReg, LUIdx);
3262 }
3263 
3264 /// If the given formula has not yet been inserted, add it to the list, and
3265 /// return true. Return false otherwise.
3266 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3267  // Do not insert formula that we will not be able to expand.
3268  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3269  "Formula is illegal");
3270 
3271  if (!LU.InsertFormula(F, *L))
3272  return false;
3273 
3274  CountRegisters(F, LUIdx);
3275  return true;
3276 }
3277 
3278 /// Check for other uses of loop-invariant values which we're tracking. These
3279 /// other uses will pin these values in registers, making them less profitable
3280 /// for elimination.
3281 /// TODO: This currently misses non-constant addrec step registers.
3282 /// TODO: Should this give more weight to users inside the loop?
3283 void
3284 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3285  SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3287 
3288  while (!Worklist.empty()) {
3289  const SCEV *S = Worklist.pop_back_val();
3290 
3291  // Don't process the same SCEV twice
3292  if (!Visited.insert(S).second)
3293  continue;
3294 
3295  if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3296  Worklist.append(N->op_begin(), N->op_end());
3297  else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3298  Worklist.push_back(C->getOperand());
3299  else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3300  Worklist.push_back(D->getLHS());
3301  Worklist.push_back(D->getRHS());
3302  } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3303  const Value *V = US->getValue();
3304  if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3305  // Look for instructions defined outside the loop.
3306  if (L->contains(Inst)) continue;
3307  } else if (isa<UndefValue>(V))
3308  // Undef doesn't have a live range, so it doesn't matter.
3309  continue;
3310  for (const Use &U : V->uses()) {
3311  const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3312  // Ignore non-instructions.
3313  if (!UserInst)
3314  continue;
3315  // Ignore instructions in other functions (as can happen with
3316  // Constants).
3317  if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3318  continue;
3319  // Ignore instructions not dominated by the loop.
3320  const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3321  UserInst->getParent() :
3322  cast<PHINode>(UserInst)->getIncomingBlock(
3323  PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3324  if (!DT.dominates(L->getHeader(), UseBB))
3325  continue;
3326  // Don't bother if the instruction is in a BB which ends in an EHPad.
3327  if (UseBB->getTerminator()->isEHPad())
3328  continue;
3329  // Don't bother rewriting PHIs in catchswitch blocks.
3330  if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3331  continue;
3332  // Ignore uses which are part of other SCEV expressions, to avoid
3333  // analyzing them multiple times.
3334  if (SE.isSCEVable(UserInst->getType())) {
3335  const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3336  // If the user is a no-op, look through to its uses.
3337  if (!isa<SCEVUnknown>(UserS))
3338  continue;
3339  if (UserS == US) {
3340  Worklist.push_back(
3341  SE.getUnknown(const_cast<Instruction *>(UserInst)));
3342  continue;
3343  }
3344  }
3345  // Ignore icmp instructions which are already being analyzed.
3346  if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3347  unsigned OtherIdx = !U.getOperandNo();
3348  Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3349  if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3350  continue;
3351  }
3352 
3353  std::pair<size_t, int64_t> P = getUse(
3354  S, LSRUse::Basic, MemAccessTy());
3355  size_t LUIdx = P.first;
3356  int64_t Offset = P.second;
3357  LSRUse &LU = Uses[LUIdx];
3358  LSRFixup &LF = LU.getNewFixup();
3359  LF.UserInst = const_cast<Instruction *>(UserInst);
3360  LF.OperandValToReplace = U;
3361  LF.Offset = Offset;
3362  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3363  if (!LU.WidestFixupType ||
3364  SE.getTypeSizeInBits(LU.WidestFixupType) <
3365  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3366  LU.WidestFixupType = LF.OperandValToReplace->getType();
3367  InsertSupplementalFormula(US, LU, LUIdx);
3368  CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3369  break;
3370  }
3371  }
3372  }
3373 }
3374 
3375 /// Split S into subexpressions which can be pulled out into separate
3376 /// registers. If C is non-null, multiply each subexpression by C.
3377 ///
3378 /// Return remainder expression after factoring the subexpressions captured by
3379 /// Ops. If Ops is complete, return NULL.
3380 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3382  const Loop *L,
3383  ScalarEvolution &SE,
3384  unsigned Depth = 0) {
3385  // Arbitrarily cap recursion to protect compile time.
3386  if (Depth >= 3)
3387  return S;
3388 
3389  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3390  // Break out add operands.
3391  for (const SCEV *S : Add->operands()) {
3392  const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3393  if (Remainder)
3394  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3395  }
3396  return nullptr;
3397  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3398  // Split a non-zero base out of an addrec.
3399  if (AR->getStart()->isZero() || !AR->isAffine())
3400  return S;
3401 
3402  const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3403  C, Ops, L, SE, Depth+1);
3404  // Split the non-zero AddRec unless it is part of a nested recurrence that
3405  // does not pertain to this loop.
3406  if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3407  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3408  Remainder = nullptr;
3409  }
3410  if (Remainder != AR->getStart()) {
3411  if (!Remainder)
3412  Remainder = SE.getConstant(AR->getType(), 0);
3413  return SE.getAddRecExpr(Remainder,
3414  AR->getStepRecurrence(SE),
3415  AR->getLoop(),
3416  //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3418  }
3419  } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3420  // Break (C * (a + b + c)) into C*a + C*b + C*c.
3421  if (Mul->getNumOperands() != 2)
3422  return S;
3423  if (const SCEVConstant *Op0 =
3424  dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3425  C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3426  const SCEV *Remainder =
3427  CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3428  if (Remainder)
3429  Ops.push_back(SE.getMulExpr(C, Remainder));
3430  return nullptr;
3431  }
3432  }
3433  return S;
3434 }
3435 
3436 /// \brief Helper function for LSRInstance::GenerateReassociations.
3437 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3438  const Formula &Base,
3439  unsigned Depth, size_t Idx,
3440  bool IsScaledReg) {
3441  const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3443  const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3444  if (Remainder)
3445  AddOps.push_back(Remainder);
3446 
3447  if (AddOps.size() == 1)
3448  return;
3449 
3451  JE = AddOps.end();
3452  J != JE; ++J) {
3453 
3454  // Loop-variant "unknown" values are uninteresting; we won't be able to
3455  // do anything meaningful with them.
3456  if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3457  continue;
3458 
3459  // Don't pull a constant into a register if the constant could be folded
3460  // into an immediate field.
3461  if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3462  LU.AccessTy, *J, Base.getNumRegs() > 1))
3463  continue;
3464 
3465  // Collect all operands except *J.
3466  SmallVector<const SCEV *, 8> InnerAddOps(
3467  ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3468  InnerAddOps.append(std::next(J),
3469  ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3470 
3471  // Don't leave just a constant behind in a register if the constant could
3472  // be folded into an immediate field.
3473  if (InnerAddOps.size() == 1 &&
3474  isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3475  LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3476  continue;
3477 
3478  const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3479  if (InnerSum->isZero())
3480  continue;
3481  Formula F = Base;
3482 
3483  // Add the remaining pieces of the add back into the new formula.
3484  const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3485  if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3486  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3487  InnerSumSC->getValue()->getZExtValue())) {
3488  F.UnfoldedOffset =
3489  (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3490  if (IsScaledReg)
3491  F.ScaledReg = nullptr;
3492  else
3493  F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3494  } else if (IsScaledReg)
3495  F.ScaledReg = InnerSum;
3496  else
3497  F.BaseRegs[Idx] = InnerSum;
3498 
3499  // Add J as its own register, or an unfolded immediate.
3500  const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3501  if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3502  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3503  SC->getValue()->getZExtValue()))
3504  F.UnfoldedOffset =
3505  (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3506  else
3507  F.BaseRegs.push_back(*J);
3508  // We may have changed the number of register in base regs, adjust the
3509  // formula accordingly.
3510  F.canonicalize(*L);
3511 
3512  if (InsertFormula(LU, LUIdx, F))
3513  // If that formula hadn't been seen before, recurse to find more like
3514  // it.
3515  GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3516  }
3517 }
3518 
3519 /// Split out subexpressions from adds and the bases of addrecs.
3520 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3521  Formula Base, unsigned Depth) {
3522  assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3523  // Arbitrarily cap recursion to protect compile time.
3524  if (Depth >= 3)
3525  return;
3526 
3527  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3528  GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3529 
3530  if (Base.Scale == 1)
3531  GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3532  /* Idx */ -1, /* IsScaledReg */ true);
3533 }
3534 
3535 /// Generate a formula consisting of all of the loop-dominating registers added
3536 /// into a single register.
3537 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3538  Formula Base) {
3539  // This method is only interesting on a plurality of registers.
3540  if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3541  return;
3542 
3543  // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3544  // processing the formula.
3545  Base.unscale();
3546  Formula F = Base;
3547  F.BaseRegs.clear();
3549  for (const SCEV *BaseReg : Base.BaseRegs) {
3550  if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3551  !SE.hasComputableLoopEvolution(BaseReg, L))
3552  Ops.push_back(BaseReg);
3553  else
3554  F.BaseRegs.push_back(BaseReg);
3555  }
3556  if (Ops.size() > 1) {
3557  const SCEV *Sum = SE.getAddExpr(Ops);
3558  // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3559  // opportunity to fold something. For now, just ignore such cases
3560  // rather than proceed with zero in a register.
3561  if (!Sum->isZero()) {
3562  F.BaseRegs.push_back(Sum);
3563  F.canonicalize(*L);
3564  (void)InsertFormula(LU, LUIdx, F);
3565  }
3566  }
3567 }
3568 
3569 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3570 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3571  const Formula &Base, size_t Idx,
3572  bool IsScaledReg) {
3573  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3574  GlobalValue *GV = ExtractSymbol(G, SE);
3575  if (G->isZero() || !GV)
3576  return;
3577  Formula F = Base;
3578  F.BaseGV = GV;
3579  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3580  return;
3581  if (IsScaledReg)
3582  F.ScaledReg = G;
3583  else
3584  F.BaseRegs[Idx] = G;
3585  (void)InsertFormula(LU, LUIdx, F);
3586 }
3587 
3588 /// Generate reuse formulae using symbolic offsets.
3589 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3590  Formula Base) {
3591  // We can't add a symbolic offset if the address already contains one.
3592  if (Base.BaseGV) return;
3593 
3594  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3595  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3596  if (Base.Scale == 1)
3597  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3598  /* IsScaledReg */ true);
3599 }
3600 
3601 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3602 void LSRInstance::GenerateConstantOffsetsImpl(
3603  LSRUse &LU, unsigned LUIdx, const Formula &Base,
3604  const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3605  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3606  for (int64_t Offset : Worklist) {
3607  Formula F = Base;
3608  F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3609  if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3610  LU.AccessTy, F)) {
3611  // Add the offset to the base register.
3612  const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3613  // If it cancelled out, drop the base register, otherwise update it.
3614  if (NewG->isZero()) {
3615  if (IsScaledReg) {
3616  F.Scale = 0;
3617  F.ScaledReg = nullptr;
3618  } else
3619  F.deleteBaseReg(F.BaseRegs[Idx]);
3620  F.canonicalize(*L);
3621  } else if (IsScaledReg)
3622  F.ScaledReg = NewG;
3623  else
3624  F.BaseRegs[Idx] = NewG;
3625 
3626  (void)InsertFormula(LU, LUIdx, F);
3627  }
3628  }
3629 
3630  int64_t Imm = ExtractImmediate(G, SE);
3631  if (G->isZero() || Imm == 0)
3632  return;
3633  Formula F = Base;
3634  F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3635  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3636  return;
3637  if (IsScaledReg)
3638  F.ScaledReg = G;
3639  else
3640  F.BaseRegs[Idx] = G;
3641  (void)InsertFormula(LU, LUIdx, F);
3642 }
3643 
3644 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3645 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3646  Formula Base) {
3647  // TODO: For now, just add the min and max offset, because it usually isn't
3648  // worthwhile looking at everything inbetween.
3649  SmallVector<int64_t, 2> Worklist;
3650  Worklist.push_back(LU.MinOffset);
3651  if (LU.MaxOffset != LU.MinOffset)
3652  Worklist.push_back(LU.MaxOffset);
3653 
3654  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3655  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3656  if (Base.Scale == 1)
3657  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3658  /* IsScaledReg */ true);
3659 }
3660 
3661 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3662 /// == y -> x*c == y*c.
3663 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3664  Formula Base) {
3665  if (LU.Kind != LSRUse::ICmpZero) return;
3666 
3667  // Determine the integer type for the base formula.
3668  Type *IntTy = Base.getType();
3669  if (!IntTy) return;
3670  if (SE.getTypeSizeInBits(IntTy) > 64) return;
3671 
3672  // Don't do this if there is more than one offset.
3673  if (LU.MinOffset != LU.MaxOffset) return;
3674 
3675  assert(!Base.BaseGV && "ICmpZero use is not legal!");
3676 
3677  // Check each interesting stride.
3678  for (int64_t Factor : Factors) {
3679  // Check that the multiplication doesn't overflow.
3680  if (Base.BaseOffset == INT64_MIN && Factor == -1)
3681  continue;
3682  int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3683  if (NewBaseOffset / Factor != Base.BaseOffset)
3684  continue;
3685  // If the offset will be truncated at this use, check that it is in bounds.
3686  if (!IntTy->isPointerTy() &&
3687  !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3688  continue;
3689 
3690  // Check that multiplying with the use offset doesn't overflow.
3691  int64_t Offset = LU.MinOffset;
3692  if (Offset == INT64_MIN && Factor == -1)
3693  continue;
3694  Offset = (uint64_t)Offset * Factor;
3695  if (Offset / Factor != LU.MinOffset)
3696  continue;
3697  // If the offset will be truncated at this use, check that it is in bounds.
3698  if (!IntTy->isPointerTy() &&
3699  !ConstantInt::isValueValidForType(IntTy, Offset))
3700  continue;
3701 
3702  Formula F = Base;
3703  F.BaseOffset = NewBaseOffset;
3704 
3705  // Check that this scale is legal.
3706  if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3707  continue;
3708 
3709  // Compensate for the use having MinOffset built into it.
3710  F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3711 
3712  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3713 
3714  // Check that multiplying with each base register doesn't overflow.
3715  for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3716  F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3717  if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3718  goto next;
3719  }
3720 
3721  // Check that multiplying with the scaled register doesn't overflow.
3722  if (F.ScaledReg) {
3723  F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3724  if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3725  continue;
3726  }
3727 
3728  // Check that multiplying with the unfolded offset doesn't overflow.
3729  if (F.UnfoldedOffset != 0) {
3730  if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3731  continue;
3732  F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3733  if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3734  continue;
3735  // If the offset will be truncated, check that it is in bounds.
3736  if (!IntTy->isPointerTy() &&
3737  !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3738  continue;
3739  }
3740 
3741  // If we make it here and it's legal, add it.
3742  (void)InsertFormula(LU, LUIdx, F);
3743  next:;
3744  }
3745 }
3746 
3747 /// Generate stride factor reuse formulae by making use of scaled-offset address
3748 /// modes, for example.
3749 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3750  // Determine the integer type for the base formula.
3751  Type *IntTy = Base.getType();
3752  if (!IntTy) return;
3753 
3754  // If this Formula already has a scaled register, we can't add another one.
3755  // Try to unscale the formula to generate a better scale.
3756  if (Base.Scale != 0 && !Base.unscale())
3757  return;
3758 
3759  assert(Base.Scale == 0 && "unscale did not did its job!");
3760 
3761  // Check each interesting stride.
3762  for (int64_t Factor : Factors) {
3763  Base.Scale = Factor;
3764  Base.HasBaseReg = Base.BaseRegs.size() > 1;
3765  // Check whether this scale is going to be legal.
3766  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3767  Base)) {
3768  // As a special-case, handle special out-of-loop Basic users specially.
3769  // TODO: Reconsider this special case.
3770  if (LU.Kind == LSRUse::Basic &&
3771  isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3772  LU.AccessTy, Base) &&
3773  LU.AllFixupsOutsideLoop)
3774  LU.Kind = LSRUse::Special;
3775  else
3776  continue;
3777  }
3778  // For an ICmpZero, negating a solitary base register won't lead to
3779  // new solutions.
3780  if (LU.Kind == LSRUse::ICmpZero &&
3781  !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3782  continue;
3783  // For each addrec base reg, if its loop is current loop, apply the scale.
3784  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3785  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3786  if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3787  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3788  if (FactorS->isZero())
3789  continue;
3790  // Divide out the factor, ignoring high bits, since we'll be
3791  // scaling the value back up in the end.
3792  if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3793  // TODO: This could be optimized to avoid all the copying.
3794  Formula F = Base;
3795  F.ScaledReg = Quotient;
3796  F.deleteBaseReg(F.BaseRegs[i]);
3797  // The canonical representation of 1*reg is reg, which is already in
3798  // Base. In that case, do not try to insert the formula, it will be
3799  // rejected anyway.
3800  if (F.Scale == 1 && (F.BaseRegs.empty() ||
3801  (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3802  continue;
3803  // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3804  // non canonical Formula with ScaledReg's loop not being L.
3805  if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
3806  F.canonicalize(*L);
3807  (void)InsertFormula(LU, LUIdx, F);
3808  }
3809  }
3810  }
3811  }
3812 }
3813 
3814 /// Generate reuse formulae from different IV types.
3815 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3816  // Don't bother truncating symbolic values.
3817  if (Base.BaseGV) return;
3818 
3819  // Determine the integer type for the base formula.
3820  Type *DstTy = Base.getType();
3821  if (!DstTy) return;
3822  DstTy = SE.getEffectiveSCEVType(DstTy);
3823 
3824  for (Type *SrcTy : Types) {
3825  if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3826  Formula F = Base;
3827 
3828  if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3829  for (const SCEV *&BaseReg : F.BaseRegs)
3830  BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3831 
3832  // TODO: This assumes we've done basic processing on all uses and
3833  // have an idea what the register usage is.
3834  if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3835  continue;
3836 
3837  F.canonicalize(*L);
3838  (void)InsertFormula(LU, LUIdx, F);
3839  }
3840  }
3841 }
3842 
3843 namespace {
3844 
3845 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3846 /// modifications so that the search phase doesn't have to worry about the data
3847 /// structures moving underneath it.
3848 struct WorkItem {
3849  size_t LUIdx;
3850  int64_t Imm;
3851  const SCEV *OrigReg;
3852 
3853  WorkItem(size_t LI, int64_t I, const SCEV *R)
3854  : LUIdx(LI), Imm(I), OrigReg(R) {}
3855 
3856  void print(raw_ostream &OS) const;
3857  void dump() const;
3858 };
3859 
3860 } // end anonymous namespace
3861 
3862 void WorkItem::print(raw_ostream &OS) const {
3863  OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3864  << " , add offset " << Imm;
3865 }
3866 
3867 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3868 LLVM_DUMP_METHOD void WorkItem::dump() const {
3869  print(errs()); errs() << '\n';
3870 }
3871 #endif
3872 
3873 /// Look for registers which are a constant distance apart and try to form reuse
3874 /// opportunities between them.
3875 void LSRInstance::GenerateCrossUseConstantOffsets() {
3876  // Group the registers by their value without any added constant offset.
3877  typedef std::map<int64_t, const SCEV *> ImmMapTy;
3879  DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3881  for (const SCEV *Use : RegUses) {
3882  const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
3883  int64_t Imm = ExtractImmediate(Reg, SE);
3884  auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
3885  if (Pair.second)
3886  Sequence.push_back(Reg);
3887  Pair.first->second.insert(std::make_pair(Imm, Use));
3888  UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
3889  }
3890 
3891  // Now examine each set of registers with the same base value. Build up
3892  // a list of work to do and do the work in a separate step so that we're
3893  // not adding formulae and register counts while we're searching.
3894  SmallVector<WorkItem, 32> WorkItems;
3895  SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3896  for (const SCEV *Reg : Sequence) {
3897  const ImmMapTy &Imms = Map.find(Reg)->second;
3898 
3899  // It's not worthwhile looking for reuse if there's only one offset.
3900  if (Imms.size() == 1)
3901  continue;
3902 
3903  DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3904  for (const auto &Entry : Imms)
3905  dbgs() << ' ' << Entry.first;
3906  dbgs() << '\n');
3907 
3908  // Examine each offset.
3909  for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3910  J != JE; ++J) {
3911  const SCEV *OrigReg = J->second;
3912 
3913  int64_t JImm = J->first;
3914  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3915 
3916  if (!isa<SCEVConstant>(OrigReg) &&
3917  UsedByIndicesMap[Reg].count() == 1) {
3918  DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3919  continue;
3920  }
3921 
3922  // Conservatively examine offsets between this orig reg a few selected
3923  // other orig regs.
3924  ImmMapTy::const_iterator OtherImms[] = {
3925  Imms.begin(), std::prev(Imms.end()),
3926  Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3927  2)
3928  };
3929  for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3930  ImmMapTy::const_iterator M = OtherImms[i];
3931  if (M == J || M == JE) continue;
3932 
3933  // Compute the difference between the two.
3934  int64_t Imm = (uint64_t)JImm - M->first;
3935  for (unsigned LUIdx : UsedByIndices.set_bits())
3936  // Make a memo of this use, offset, and register tuple.
3937  if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3938  WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3939  }
3940  }
3941  }
3942 
3943  Map.clear();
3944  Sequence.clear();
3945  UsedByIndicesMap.clear();
3946  UniqueItems.clear();
3947 
3948  // Now iterate through the worklist and add new formulae.
3949  for (const WorkItem &WI : WorkItems) {
3950  size_t LUIdx = WI.LUIdx;
3951  LSRUse &LU = Uses[LUIdx];
3952  int64_t Imm = WI.Imm;
3953  const SCEV *OrigReg = WI.OrigReg;
3954 
3955  Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3956  const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3957  unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3958 
3959  // TODO: Use a more targeted data structure.
3960  for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3961  Formula F = LU.Formulae[L];
3962  // FIXME: The code for the scaled and unscaled registers looks
3963  // very similar but slightly different. Investigate if they
3964  // could be merged. That way, we would not have to unscale the
3965  // Formula.
3966  F.unscale();
3967  // Use the immediate in the scaled register.
3968  if (F.ScaledReg == OrigReg) {
3969  int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3970  // Don't create 50 + reg(-50).
3971  if (F.referencesReg(SE.getSCEV(
3972  ConstantInt::get(IntTy, -(uint64_t)Offset))))
3973  continue;
3974  Formula NewF = F;
3975  NewF.BaseOffset = Offset;
3976  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3977  NewF))
3978  continue;
3979  NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3980 
3981  // If the new scale is a constant in a register, and adding the constant
3982  // value to the immediate would produce a value closer to zero than the
3983  // immediate itself, then the formula isn't worthwhile.
3984  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3985  if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
3986  (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
3987  .ule(std::abs(NewF.BaseOffset)))
3988  continue;
3989 
3990  // OK, looks good.
3991  NewF.canonicalize(*this->L);
3992  (void)InsertFormula(LU, LUIdx, NewF);
3993  } else {
3994  // Use the immediate in a base register.
3995  for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3996  const SCEV *BaseReg = F.BaseRegs[N];
3997  if (BaseReg != OrigReg)
3998  continue;
3999  Formula NewF = F;
4000  NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4001  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4002  LU.Kind, LU.AccessTy, NewF)) {
4003  if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4004  continue;
4005  NewF = F;
4006  NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4007  }
4008  NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4009 
4010  // If the new formula has a constant in a register, and adding the
4011  // constant value to the immediate would produce a value closer to
4012  // zero than the immediate itself, then the formula isn't worthwhile.
4013  for (const SCEV *NewReg : NewF.BaseRegs)
4014  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4015  if ((C->getAPInt() + NewF.BaseOffset)
4016  .abs()
4017  .slt(std::abs(NewF.BaseOffset)) &&
4018  (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4019  countTrailingZeros<uint64_t>(NewF.BaseOffset))
4020  goto skip_formula;
4021 
4022  // Ok, looks good.
4023  NewF.canonicalize(*this->L);
4024  (void)InsertFormula(LU, LUIdx, NewF);
4025  break;
4026  skip_formula:;
4027  }
4028  }
4029  }
4030  }
4031 }
4032 
4033 /// Generate formulae for each use.
4034 void
4035 LSRInstance::GenerateAllReuseFormulae() {
4036  // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4037  // queries are more precise.
4038  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4039  LSRUse &LU = Uses[LUIdx];
4040  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4041  GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4042  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4043  GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4044  }
4045  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4046  LSRUse &LU = Uses[LUIdx];
4047  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4048  GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4049  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4050  GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4051  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4052  GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4053  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4054  GenerateScales(LU, LUIdx, LU.Formulae[i]);
4055  }
4056  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4057  LSRUse &LU = Uses[LUIdx];
4058  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4059  GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4060  }
4061 
4062  GenerateCrossUseConstantOffsets();
4063 
4064  DEBUG(dbgs() << "\n"
4065  "After generating reuse formulae:\n";
4066  print_uses(dbgs()));
4067 }
4068 
4069 /// If there are multiple formulae with the same set of registers used
4070 /// by other uses, pick the best one and delete the others.
4071 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4072  DenseSet<const SCEV *> VisitedRegs;
4075 #ifndef NDEBUG
4076  bool ChangedFormulae = false;
4077 #endif
4078 
4079  // Collect the best formula for each unique set of shared registers. This
4080  // is reset for each use.
4081  typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
4082  BestFormulaeTy;
4083  BestFormulaeTy BestFormulae;
4084 
4085  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4086  LSRUse &LU = Uses[LUIdx];
4087  DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4088 
4089  bool Any = false;
4090  for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4091  FIdx != NumForms; ++FIdx) {
4092  Formula &F = LU.Formulae[FIdx];
4093 
4094  // Some formulas are instant losers. For example, they may depend on
4095  // nonexistent AddRecs from other loops. These need to be filtered
4096  // immediately, otherwise heuristics could choose them over others leading
4097  // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4098  // avoids the need to recompute this information across formulae using the
4099  // same bad AddRec. Passing LoserRegs is also essential unless we remove
4100  // the corresponding bad register from the Regs set.
4101  Cost CostF;
4102  Regs.clear();
4103  CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, SE, DT, LU, &LoserRegs);
4104  if (CostF.isLoser()) {
4105  // During initial formula generation, undesirable formulae are generated
4106  // by uses within other loops that have some non-trivial address mode or
4107  // use the postinc form of the IV. LSR needs to provide these formulae
4108  // as the basis of rediscovering the desired formula that uses an AddRec
4109  // corresponding to the existing phi. Once all formulae have been
4110  // generated, these initial losers may be pruned.
4111  DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4112  dbgs() << "\n");
4113  }
4114  else {
4116  for (const SCEV *Reg : F.BaseRegs) {
4117  if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4118  Key.push_back(Reg);
4119  }
4120  if (F.ScaledReg &&
4121  RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4122  Key.push_back(F.ScaledReg);
4123  // Unstable sort by host order ok, because this is only used for
4124  // uniquifying.
4125  std::sort(Key.begin(), Key.end());
4126 
4127  std::pair<BestFormulaeTy::const_iterator, bool> P =
4128  BestFormulae.insert(std::make_pair(Key, FIdx));
4129  if (P.second)
4130  continue;
4131 
4132  Formula &Best = LU.Formulae[P.first->second];
4133 
4134  Cost CostBest;
4135  Regs.clear();
4136  CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, SE, DT, LU);
4137  if (CostF.isLess(CostBest, TTI))
4138  std::swap(F, Best);
4139  DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4140  dbgs() << "\n"
4141  " in favor of formula "; Best.print(dbgs());
4142  dbgs() << '\n');
4143  }
4144 #ifndef NDEBUG
4145  ChangedFormulae = true;
4146 #endif
4147  LU.DeleteFormula(F);
4148  --FIdx;
4149  --NumForms;
4150  Any = true;
4151  }
4152 
4153  // Now that we've filtered out some formulae, recompute the Regs set.
4154  if (Any)
4155  LU.RecomputeRegs(LUIdx, RegUses);
4156 
4157  // Reset this to prepare for the next use.
4158  BestFormulae.clear();
4159  }
4160 
4161  DEBUG(if (ChangedFormulae) {
4162  dbgs() << "\n"
4163  "After filtering out undesirable candidates:\n";
4164  print_uses(dbgs());
4165  });
4166 }
4167 
4168 // This is a rough guess that seems to work fairly well.
4169 static const size_t ComplexityLimit = UINT16_MAX;
4170 
4171 /// Estimate the worst-case number of solutions the solver might have to
4172 /// consider. It almost never considers this many solutions because it prune the
4173 /// search space, but the pruning isn't always sufficient.
4174 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4175  size_t Power = 1;
4176  for (const LSRUse &LU : Uses) {
4177  size_t FSize = LU.Formulae.size();
4178  if (FSize >= ComplexityLimit) {
4179  Power = ComplexityLimit;
4180  break;
4181  }
4182  Power *= FSize;
4183  if (Power >= ComplexityLimit)
4184  break;
4185  }
4186  return Power;
4187 }
4188 
4189 /// When one formula uses a superset of the registers of another formula, it
4190 /// won't help reduce register pressure (though it may not necessarily hurt
4191 /// register pressure); remove it to simplify the system.
4192 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4193  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4194  DEBUG(dbgs() << "The search space is too complex.\n");
4195 
4196  DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4197  "which use a superset of registers used by other "
4198  "formulae.\n");
4199 
4200  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4201  LSRUse &LU = Uses[LUIdx];
4202  bool Any = false;
4203  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4204  Formula &F = LU.Formulae[i];
4205  // Look for a formula with a constant or GV in a register. If the use
4206  // also has a formula with that same value in an immediate field,
4207  // delete the one that uses a register.
4209  I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4210  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4211  Formula NewF = F;
4212  NewF.BaseOffset += C->getValue()->getSExtValue();
4213  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4214  (I - F.BaseRegs.begin()));
4215  if (LU.HasFormulaWithSameRegs(NewF)) {
4216  DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4217  LU.DeleteFormula(F);
4218  --i;
4219  --e;
4220  Any = true;
4221  break;
4222  }
4223  } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4224  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4225  if (!F.BaseGV) {
4226  Formula NewF = F;
4227  NewF.BaseGV = GV;
4228  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4229  (I - F.BaseRegs.begin()));
4230  if (LU.HasFormulaWithSameRegs(NewF)) {
4231  DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4232  dbgs() << '\n');
4233  LU.DeleteFormula(F);
4234  --i;
4235  --e;
4236  Any = true;
4237  break;
4238  }
4239  }
4240  }
4241  }
4242  }
4243  if (Any)
4244  LU.RecomputeRegs(LUIdx, RegUses);
4245  }
4246 
4247  DEBUG(dbgs() << "After pre-selection:\n";
4248  print_uses(dbgs()));
4249  }
4250 }
4251 
4252 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4253 /// allocate a single register for them.
4254 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4255  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4256  return;
4257 
4258  DEBUG(dbgs() << "The search space is too complex.\n"
4259  "Narrowing the search space by assuming that uses separated "
4260  "by a constant offset will use the same registers.\n");
4261 
4262  // This is especially useful for unrolled loops.
4263 
4264  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4265  LSRUse &LU = Uses[LUIdx];
4266  for (const Formula &F : LU.Formulae) {
4267  if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4268  continue;
4269 
4270  LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4271  if (!LUThatHas)
4272  continue;
4273 
4274  if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4275  LU.Kind, LU.AccessTy))
4276  continue;
4277 
4278  DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4279 
4280  LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4281 
4282  // Transfer the fixups of LU to LUThatHas.
4283  for (LSRFixup &Fixup : LU.Fixups) {
4284  Fixup.Offset += F.BaseOffset;
4285  LUThatHas->pushFixup(Fixup);
4286  DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4287  }
4288 
4289  // Delete formulae from the new use which are no longer legal.
4290  bool Any = false;
4291  for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4292  Formula &F = LUThatHas->Formulae[i];
4293  if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4294  LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4295  DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4296  dbgs() << '\n');
4297  LUThatHas->DeleteFormula(F);
4298  --i;
4299  --e;
4300  Any = true;
4301  }
4302  }
4303 
4304  if (Any)
4305  LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4306 
4307  // Delete the old use.
4308  DeleteUse(LU, LUIdx);
4309  --LUIdx;
4310  --NumUses;
4311  break;
4312  }
4313  }
4314 
4315  DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4316 }
4317 
4318 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4319 /// we've done more filtering, as it may be able to find more formulae to
4320 /// eliminate.
4321 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4322  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4323  DEBUG(dbgs() << "The search space is too complex.\n");
4324 
4325  DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4326  "undesirable dedicated registers.\n");
4327 
4328  FilterOutUndesirableDedicatedRegisters();
4329 
4330  DEBUG(dbgs() << "After pre-selection:\n";
4331  print_uses(dbgs()));
4332  }
4333 }
4334 
4335 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4336 /// Pick the best one and delete the others.
4337 /// This narrowing heuristic is to keep as many formulae with different
4338 /// Scale and ScaledReg pair as possible while narrowing the search space.
4339 /// The benefit is that it is more likely to find out a better solution
4340 /// from a formulae set with more Scale and ScaledReg variations than
4341 /// a formulae set with the same Scale and ScaledReg. The picking winner
4342 /// reg heurstic will often keep the formulae with the same Scale and
4343 /// ScaledReg and filter others, and we want to avoid that if possible.
4344 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4345  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4346  return;
4347 
4348  DEBUG(dbgs() << "The search space is too complex.\n"
4349  "Narrowing the search space by choosing the best Formula "
4350  "from the Formulae with the same Scale and ScaledReg.\n");
4351 
4352  // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4353  typedef DenseMap<std::pair<const SCEV *, int64_t>, size_t> BestFormulaeTy;
4354  BestFormulaeTy BestFormulae;
4355 #ifndef NDEBUG
4356  bool ChangedFormulae = false;
4357 #endif
4358  DenseSet<const SCEV *> VisitedRegs;
4360 
4361  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4362  LSRUse &LU = Uses[LUIdx];
4363  DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4364 
4365  // Return true if Formula FA is better than Formula FB.
4366  auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4367  // First we will try to choose the Formula with fewer new registers.
4368  // For a register used by current Formula, the more the register is
4369  // shared among LSRUses, the less we increase the register number
4370  // counter of the formula.
4371  size_t FARegNum = 0;
4372  for (const SCEV *Reg : FA.BaseRegs) {
4373  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4374  FARegNum += (NumUses - UsedByIndices.count() + 1);
4375  }
4376  size_t FBRegNum = 0;
4377  for (const SCEV *Reg : FB.BaseRegs) {
4378  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4379  FBRegNum += (NumUses - UsedByIndices.count() + 1);
4380  }
4381  if (FARegNum != FBRegNum)
4382  return FARegNum < FBRegNum;
4383 
4384  // If the new register numbers are the same, choose the Formula with
4385  // less Cost.
4386  Cost CostFA, CostFB;
4387  Regs.clear();
4388  CostFA.RateFormula(TTI, FA, Regs, VisitedRegs, L, SE, DT, LU);
4389  Regs.clear();
4390  CostFB.RateFormula(TTI, FB, Regs, VisitedRegs, L, SE, DT, LU);
4391  return CostFA.isLess(CostFB, TTI);
4392  };
4393 
4394  bool Any = false;
4395  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4396  ++FIdx) {
4397  Formula &F = LU.Formulae[FIdx];
4398  if (!F.ScaledReg)
4399  continue;
4400  auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4401  if (P.second)
4402  continue;
4403 
4404  Formula &Best = LU.Formulae[P.first->second];
4405  if (IsBetterThan(F, Best))
4406  std::swap(F, Best);
4407  DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4408  dbgs() << "\n"
4409  " in favor of formula ";
4410  Best.print(dbgs()); dbgs() << '\n');
4411 #ifndef NDEBUG
4412  ChangedFormulae = true;
4413 #endif
4414  LU.DeleteFormula(F);
4415  --FIdx;
4416  --NumForms;
4417  Any = true;
4418  }
4419  if (Any)
4420  LU.RecomputeRegs(LUIdx, RegUses);
4421 
4422  // Reset this to prepare for the next use.
4423  BestFormulae.clear();
4424  }
4425 
4426  DEBUG(if (ChangedFormulae) {
4427  dbgs() << "\n"
4428  "After filtering out undesirable candidates:\n";
4429  print_uses(dbgs());
4430  });
4431 }
4432 
4433 /// The function delete formulas with high registers number expectation.
4434 /// Assuming we don't know the value of each formula (already delete
4435 /// all inefficient), generate probability of not selecting for each
4436 /// register.
4437 /// For example,
4438 /// Use1:
4439 /// reg(a) + reg({0,+,1})
4440 /// reg(a) + reg({-1,+,1}) + 1
4441 /// reg({a,+,1})
4442 /// Use2:
4443 /// reg(b) + reg({0,+,1})
4444 /// reg(b) + reg({-1,+,1}) + 1
4445 /// reg({b,+,1})
4446 /// Use3:
4447 /// reg(c) + reg(b) + reg({0,+,1})
4448 /// reg(c) + reg({b,+,1})
4449 ///
4450 /// Probability of not selecting
4451 /// Use1 Use2 Use3
4452 /// reg(a) (1/3) * 1 * 1
4453 /// reg(b) 1 * (1/3) * (1/2)
4454 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4455 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4456 /// reg({a,+,1}) (2/3) * 1 * 1
4457 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4458 /// reg(c) 1 * 1 * 0
4459 ///
4460 /// Now count registers number mathematical expectation for each formula:
4461 /// Note that for each use we exclude probability if not selecting for the use.
4462 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4463 /// probabilty 1/3 of not selecting for Use1).
4464 /// Use1:
4465 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4466 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4467 /// reg({a,+,1}) 1
4468 /// Use2:
4469 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4470 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4471 /// reg({b,+,1}) 2/3
4472 /// Use3:
4473 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4474 /// reg(c) + reg({b,+,1}) 1 + 2/3
4475 
4476 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4477  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4478  return;
4479  // Ok, we have too many of formulae on our hands to conveniently handle.
4480  // Use a rough heuristic to thin out the list.
4481 
4482  // Set of Regs wich will be 100% used in final solution.
4483  // Used in each formula of a solution (in example above this is reg(c)).
4484  // We can skip them in calculations.
4486  DEBUG(dbgs() << "The search space is too complex.\n");
4487 
4488  // Map each register to probability of not selecting
4490  for (const SCEV *Reg : RegUses) {
4491  if (UniqRegs.count(Reg))
4492  continue;
4493  float PNotSel = 1;
4494  for (const LSRUse &LU : Uses) {
4495  if (!LU.Regs.count(Reg))
4496  continue;
4497  float P = LU.getNotSelectedProbability(Reg);
4498  if (P != 0.0)
4499  PNotSel *= P;
4500  else
4501  UniqRegs.insert(Reg);
4502  }
4503  RegNumMap.insert(std::make_pair(Reg, PNotSel));
4504  }
4505 
4506  DEBUG(dbgs() << "Narrowing the search space by deleting costly formulas\n");
4507 
4508  // Delete formulas where registers number expectation is high.
4509  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4510  LSRUse &LU = Uses[LUIdx];
4511  // If nothing to delete - continue.
4512  if (LU.Formulae.size() < 2)
4513  continue;
4514  // This is temporary solution to test performance. Float should be
4515  // replaced with round independent type (based on integers) to avoid
4516  // different results for different target builds.
4517  float FMinRegNum = LU.Formulae[0].getNumRegs();
4518  float FMinARegNum = LU.Formulae[0].getNumRegs();
4519  size_t MinIdx = 0;
4520  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4521  Formula &F = LU.Formulae[i];
4522  float FRegNum = 0;
4523  float FARegNum = 0;
4524  for (const SCEV *BaseReg : F.BaseRegs) {
4525  if (UniqRegs.count(BaseReg))
4526  continue;
4527  FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4528  if (isa<SCEVAddRecExpr>(BaseReg))
4529  FARegNum +=
4530  RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4531  }
4532  if (const SCEV *ScaledReg = F.ScaledReg) {
4533  if (!UniqRegs.count(ScaledReg)) {
4534  FRegNum +=
4535  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4536  if (isa<SCEVAddRecExpr>(ScaledReg))
4537  FARegNum +=
4538  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4539  }
4540  }
4541  if (FMinRegNum > FRegNum ||
4542  (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4543  FMinRegNum = FRegNum;
4544  FMinARegNum = FARegNum;
4545  MinIdx = i;
4546  }
4547  }
4548  DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4549  dbgs() << " with min reg num " << FMinRegNum << '\n');
4550  if (MinIdx != 0)
4551  std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4552  while (LU.Formulae.size() != 1) {
4553  DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4554  dbgs() << '\n');
4555  LU.Formulae.pop_back();
4556  }
4557  LU.RecomputeRegs(LUIdx, RegUses);
4558  assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4559  Formula &F = LU.Formulae[0];
4560  DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4561  // When we choose the formula, the regs become unique.
4562  UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4563  if (F.ScaledReg)
4564  UniqRegs.insert(F.ScaledReg);
4565  }
4566  DEBUG(dbgs() << "After pre-selection:\n";
4567  print_uses(dbgs()));
4568 }
4569 
4570 
4571 /// Pick a register which seems likely to be profitable, and then in any use
4572 /// which has any reference to that register, delete all formulae which do not
4573 /// reference that register.
4574 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4575  // With all other options exhausted, loop until the system is simple
4576  // enough to handle.
4578  while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4579  // Ok, we have too many of formulae on our hands to conveniently handle.
4580  // Use a rough heuristic to thin out the list.
4581  DEBUG(dbgs() << "The search space is too complex.\n");
4582 
4583  // Pick the register which is used by the most LSRUses, which is likely
4584  // to be a good reuse register candidate.
4585  const SCEV *Best = nullptr;
4586  unsigned BestNum = 0;
4587  for (const SCEV *Reg : RegUses) {
4588  if (Taken.count(Reg))
4589  continue;
4590  if (!Best) {
4591  Best = Reg;
4592  BestNum = RegUses.getUsedByIndices(Reg).count();
4593  } else {
4594  unsigned Count = RegUses.getUsedByIndices(Reg).count();
4595  if (Count > BestNum) {
4596  Best = Reg;
4597  BestNum = Count;
4598  }
4599  }
4600  }
4601 
4602  DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4603  << " will yield profitable reuse.\n");
4604  Taken.insert(Best);
4605 
4606  // In any use with formulae which references this register, delete formulae
4607  // which don't reference it.
4608  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4609  LSRUse &LU = Uses[LUIdx];
4610  if (!LU.Regs.count(Best)) continue;
4611 
4612  bool Any = false;
4613  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4614  Formula &F = LU.Formulae[i];
4615  if (!F.referencesReg(Best)) {
4616  DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4617  LU.DeleteFormula(F);
4618  --e;
4619  --i;
4620  Any = true;
4621  assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4622  continue;
4623  }
4624  }
4625 
4626  if (Any)
4627  LU.RecomputeRegs(LUIdx, RegUses);
4628  }
4629 
4630  DEBUG(dbgs() << "After pre-selection:\n";
4631  print_uses(dbgs()));
4632  }
4633 }
4634 
4635 /// If there are an extraordinary number of formulae to choose from, use some
4636 /// rough heuristics to prune down the number of formulae. This keeps the main
4637 /// solver from taking an extraordinary amount of time in some worst-case
4638 /// scenarios.
4639 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4640  NarrowSearchSpaceByDetectingSupersets();
4641  NarrowSearchSpaceByCollapsingUnrolledCode();
4642  NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4643  if (FilterSameScaledReg)
4644  NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4645  if (LSRExpNarrow)
4646  NarrowSearchSpaceByDeletingCostlyFormulas();
4647  else
4648  NarrowSearchSpaceByPickingWinnerRegs();
4649 }
4650 
4651 /// This is the recursive solver.
4652 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4653  Cost &SolutionCost,
4655  const Cost &CurCost,
4656  const SmallPtrSet<const SCEV *, 16> &CurRegs,
4657  DenseSet<const SCEV *> &VisitedRegs) const {
4658  // Some ideas:
4659  // - prune more:
4660  // - use more aggressive filtering
4661  // - sort the formula so that the most profitable solutions are found first
4662  // - sort the uses too
4663  // - search faster:
4664  // - don't compute a cost, and then compare. compare while computing a cost
4665  // and bail early.
4666  // - track register sets with SmallBitVector
4667 
4668  const LSRUse &LU = Uses[Workspace.size()];
4669 
4670  // If this use references any register that's already a part of the
4671  // in-progress solution, consider it a requirement that a formula must
4672  // reference that register in order to be considered. This prunes out
4673  // unprofitable searching.
4675  for (const SCEV *S : CurRegs)
4676  if (LU.Regs.count(S))
4677  ReqRegs.insert(S);
4678 
4680  Cost NewCost;
4681  for (const Formula &F : LU.Formulae) {
4682  // Ignore formulae which may not be ideal in terms of register reuse of
4683  // ReqRegs. The formula should use all required registers before
4684  // introducing new ones.
4685  int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4686  for (const SCEV *Reg : ReqRegs) {
4687  if ((F.ScaledReg && F.ScaledReg == Reg) ||
4688  is_contained(F.BaseRegs, Reg)) {
4689  --NumReqRegsToFind;
4690  if (NumReqRegsToFind == 0)
4691  break;
4692  }
4693  }
4694  if (NumReqRegsToFind != 0) {
4695  // If none of the formulae satisfied the required registers, then we could
4696  // clear ReqRegs and try again. Currently, we simply give up in this case.
4697  continue;
4698  }
4699 
4700  // Evaluate the cost of the current formula. If it's already worse than
4701  // the current best, prune the search at that point.
4702  NewCost = CurCost;
4703  NewRegs = CurRegs;
4704  NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, SE, DT, LU);
4705  if (NewCost.isLess(SolutionCost, TTI)) {
4706  Workspace.push_back(&F);
4707  if (Workspace.size() != Uses.size()) {
4708  SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4709  NewRegs, VisitedRegs);
4710  if (F.getNumRegs() == 1 && Workspace.size() == 1)
4711  VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4712  } else {
4713  DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4714  dbgs() << ".\n Regs:";
4715  for (const SCEV *S : NewRegs)
4716  dbgs() << ' ' << *S;
4717  dbgs() << '\n');
4718 
4719  SolutionCost = NewCost;
4720  Solution = Workspace;
4721  }
4722  Workspace.pop_back();
4723  }
4724  }
4725 }
4726 
4727 /// Choose one formula from each use. Return the results in the given Solution
4728 /// vector.
4729 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4731  Cost SolutionCost;
4732  SolutionCost.Lose();
4733  Cost CurCost;
4735  DenseSet<const SCEV *> VisitedRegs;
4736  Workspace.reserve(Uses.size());
4737 
4738  // SolveRecurse does all the work.
4739  SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4740  CurRegs, VisitedRegs);
4741  if (Solution.empty()) {
4742  DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4743  return;
4744  }
4745 
4746  // Ok, we've now made all our decisions.
4747  DEBUG(dbgs() << "\n"
4748  "The chosen solution requires "; SolutionCost.print(dbgs());
4749  dbgs() << ":\n";
4750  for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4751  dbgs() << " ";
4752  Uses[i].print(dbgs());
4753  dbgs() << "\n"
4754  " ";
4755  Solution[i]->print(dbgs());
4756  dbgs() << '\n';
4757  });
4758 
4759  assert(Solution.size() == Uses.size() && "Malformed solution!");
4760 }
4761 
4762 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4763 /// we can go while still being dominated by the input positions. This helps
4764 /// canonicalize the insert position, which encourages sharing.
4766 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4768  const {
4769  Instruction *Tentative = &*IP;
4770  while (true) {
4771  bool AllDominate = true;
4772  Instruction *BetterPos = nullptr;
4773  // Don't bother attempting to insert before a catchswitch, their basic block
4774  // cannot have other non-PHI instructions.
4775  if (isa<CatchSwitchInst>(Tentative))
4776  return IP;
4777 
4778  for (Instruction *Inst : Inputs) {
4779  if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4780  AllDominate = false;
4781  break;
4782  }
4783  // Attempt to find an insert position in the middle of the block,
4784  // instead of at the end, so that it can be used for other expansions.
4785  if (Tentative->getParent() == Inst->getParent() &&
4786  (!BetterPos || !DT.dominates(Inst, BetterPos)))
4787  BetterPos = &*std::next(BasicBlock::iterator(Inst));
4788  }
4789  if (!AllDominate)
4790  break;
4791  if (BetterPos)
4792  IP = BetterPos->getIterator();
4793  else
4794  IP = Tentative->getIterator();
4795 
4796  const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4797  unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4798 
4799  BasicBlock *IDom;
4800  for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4801  if (!Rung) return IP;
4802  Rung = Rung->getIDom();
4803  if (!Rung) return IP;
4804  IDom = Rung->getBlock();
4805 
4806  // Don't climb into a loop though.
4807  const Loop *IDomLoop = LI.getLoopFor(IDom);
4808  unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4809  if (IDomDepth <= IPLoopDepth &&
4810  (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4811  break;
4812  }
4813 
4814  Tentative = IDom->getTerminator();
4815  }
4816 
4817  return IP;
4818 }
4819 
4820 /// Determine an input position which will be dominated by the operands and
4821 /// which will dominate the result.
4823 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4824  const LSRFixup &LF,
4825  const LSRUse &LU,
4826  SCEVExpander &Rewriter) const {
4827  // Collect some instructions which must be dominated by the
4828  // expanding replacement. These must be dominated by any operands that
4829  // will be required in the expansion.
4831  if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4832  Inputs.push_back(I);
4833  if (LU.Kind == LSRUse::ICmpZero)
4834  if (Instruction *I =
4835  dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4836  Inputs.push_back(I);
4837  if (LF.PostIncLoops.count(L)) {
4838  if (LF.isUseFullyOutsideLoop(L))
4839  Inputs.push_back(L->getLoopLatch()->getTerminator());
4840  else
4841  Inputs.push_back(IVIncInsertPos);
4842  }
4843  // The expansion must also be dominated by the increment positions of any
4844  // loops it for which it is using post-inc mode.
4845  for (const Loop *PIL : LF.PostIncLoops) {
4846  if (PIL == L) continue;
4847 
4848  // Be dominated by the loop exit.
4849  SmallVector<BasicBlock *, 4> ExitingBlocks;
4850  PIL->getExitingBlocks(ExitingBlocks);
4851  if (!ExitingBlocks.empty()) {
4852  BasicBlock *BB = ExitingBlocks[0];
4853  for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4854  BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4855  Inputs.push_back(BB->getTerminator());
4856  }
4857  }
4858 
4859  assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
4860  && !isa<DbgInfoIntrinsic>(LowestIP) &&
4861  "Insertion point must be a normal instruction");
4862 
4863  // Then, climb up the immediate dominator tree as far as we can go while
4864  // still being dominated by the input positions.
4865  BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4866 
4867  // Don't insert instructions before PHI nodes.
4868  while (isa<PHINode>(IP)) ++IP;
4869 
4870  // Ignore landingpad instructions.
4871  while (IP->isEHPad()) ++IP;
4872 
4873  // Ignore debug intrinsics.
4874  while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4875 
4876  // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4877  // IP consistent across expansions and allows the previously inserted
4878  // instructions to be reused by subsequent expansion.
4879  while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
4880  ++IP;
4881 
4882  return IP;
4883 }
4884 
4885 /// Emit instructions for the leading candidate expression for this LSRUse (this
4886 /// is called "expanding").
4887 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
4888  const Formula &F, BasicBlock::iterator IP,
4889  SCEVExpander &Rewriter,
4890  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
4891  if (LU.RigidFormula)
4892  return LF.OperandValToReplace;
4893 
4894  // Determine an input position which will be dominated by the operands and
4895  // which will dominate the result.
4896  IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4897  Rewriter.setInsertPoint(&*IP);
4898 
4899  // Inform the Rewriter if we have a post-increment use, so that it can
4900  // perform an advantageous expansion.
4901  Rewriter.setPostInc(LF.PostIncLoops);
4902 
4903  // This is the type that the user actually needs.
4904  Type *OpTy = LF.OperandValToReplace->getType();
4905  // This will be the type that we'll initially expand to.
4906  Type *Ty = F.getType();
4907  if (!Ty)
4908  // No type known; just expand directly to the ultimate type.
4909  Ty = OpTy;
4910  else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4911  // Expand directly to the ultimate type if it's the right size.
4912  Ty = OpTy;
4913  // This is the type to do integer arithmetic in.
4914  Type *IntTy = SE.getEffectiveSCEVType(Ty);
4915 
4916  // Build up a list of operands to add together to form the full base.
4918 
4919  // Expand the BaseRegs portion.
4920  for (const SCEV *Reg : F.BaseRegs) {
4921  assert(!Reg->isZero() && "Zero allocated in a base register!");
4922 
4923  // If we're expanding for a post-inc user, make the post-inc adjustment.
4924  Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
4925  Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
4926  }
4927 
4928  // Expand the ScaledReg portion.
4929  Value *ICmpScaledV = nullptr;
4930  if (F.Scale != 0) {
4931  const SCEV *ScaledS = F.ScaledReg;
4932 
4933  // If we're expanding for a post-inc user, make the post-inc adjustment.
4934  PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4935  ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
4936 
4937  if (LU.Kind == LSRUse::ICmpZero) {
4938  // Expand ScaleReg as if it was part of the base regs.
4939  if (F.Scale == 1)
4940  Ops.push_back(
4941  SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
4942  else {
4943  // An interesting way of "folding" with an icmp is to use a negated
4944  // scale, which we'll implement by inserting it into the other operand
4945  // of the icmp.
4946  assert(F.Scale == -1 &&
4947  "The only scale supported by ICmpZero uses is -1!");
4948  ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
4949  }
4950  } else {
4951  // Otherwise just expand the scaled register and an explicit scale,
4952  // which is expected to be matched as part of the address.
4953 
4954  // Flush the operand list to suppress SCEVExpander hoisting address modes.
4955  // Unless the addressing mode will not be folded.
4956  if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4957  isAMCompletelyFolded(TTI, LU, F)) {
4958  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4959  Ops.clear();
4960  Ops.push_back(SE.getUnknown(FullV));
4961  }
4962  ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
4963  if (F.Scale != 1)
4964  ScaledS =
4965  SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4966  Ops.push_back(ScaledS);
4967  }
4968  }
4969 
4970  // Expand the GV portion.
4971  if (F.BaseGV) {
4972  // Flush the operand list to suppress SCEVExpander hoisting.
4973  if (!Ops.empty()) {
4974  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4975  Ops.clear();
4976  Ops.push_back(SE.getUnknown(FullV));
4977  }
4978  Ops.push_back(SE.getUnknown(F.BaseGV));
4979  }
4980 
4981  // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4982  // unfolded offsets. LSR assumes they both live next to their uses.
4983  if (!Ops.empty()) {
4984  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4985  Ops.clear();
4986  Ops.push_back(SE.getUnknown(FullV));
4987  }
4988 
4989  // Expand the immediate portion.
4990  int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4991  if (Offset != 0) {
4992  if (LU.Kind == LSRUse::ICmpZero) {
4993  // The other interesting way of "folding" with an ICmpZero is to use a
4994  // negated immediate.
4995  if (!ICmpScaledV)
4996  ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4997  else {
4998  Ops.push_back(SE.getUnknown(ICmpScaledV));
4999  ICmpScaledV = ConstantInt::get(IntTy, Offset);
5000  }
5001  } else {
5002  // Just add the immediate values. These again are expected to be matched
5003  // as part of the address.
5004  Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5005  }
5006  }
5007 
5008  // Expand the unfolded offset portion.
5009  int64_t UnfoldedOffset = F.UnfoldedOffset;
5010  if (UnfoldedOffset != 0) {
5011  // Just add the immediate values.
5013  UnfoldedOffset)));
5014  }
5015 
5016  // Emit instructions summing all the operands.
5017  const SCEV *FullS = Ops.empty() ?
5018  SE.getConstant(IntTy, 0) :
5019  SE.getAddExpr(Ops);
5020  Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5021 
5022  // We're done expanding now, so reset the rewriter.
5023  Rewriter.clearPostInc();
5024 
5025  // An ICmpZero Formula represents an ICmp which we're handling as a
5026  // comparison against zero. Now that we've expanded an expression for that
5027  // form, update the ICmp's other operand.
5028  if (LU.Kind == LSRUse::ICmpZero) {
5029  ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5030  DeadInsts.emplace_back(CI->getOperand(1));
5031  assert(!F.BaseGV && "ICmp does not support folding a global value and "
5032  "a scale at the same time!");
5033  if (F.Scale == -1) {
5034  if (ICmpScaledV->getType() != OpTy) {
5035  Instruction *Cast =
5036  CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5037  OpTy, false),
5038  ICmpScaledV, OpTy, "tmp", CI);
5039  ICmpScaledV = Cast;
5040  }
5041  CI->setOperand(1, ICmpScaledV);
5042  } else {
5043  // A scale of 1 means that the scale has been expanded as part of the
5044  // base regs.
5045  assert((F.Scale == 0 || F.Scale == 1) &&
5046  "ICmp does not support folding a global value and "
5047  "a scale at the same time!");
5049  -(uint64_t)Offset);
5050  if (C->getType() != OpTy)
5052  OpTy, false),
5053  C, OpTy);
5054 
5055  CI->setOperand(1, C);
5056  }
5057  }
5058 
5059  return FullV;
5060 }
5061 
5062 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5063 /// effectively happens in their predecessor blocks, so the expression may need
5064 /// to be expanded in multiple places.
5065 void LSRInstance::RewriteForPHI(
5066  PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5067  SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5069  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5070  if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5071  BasicBlock *BB = PN->getIncomingBlock(i);
5072 
5073  // If this is a critical edge, split the edge so that we do not insert
5074  // the code on all predecessor/successor paths. We do this unless this
5075  // is the canonical backedge for this loop, which complicates post-inc
5076  // users.
5077  if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5078  !isa<IndirectBrInst>(BB->getTerminator()) &&
5079  !isa<CatchSwitchInst>(BB->getTerminator())) {
5080  BasicBlock *Parent = PN->getParent();
5081  Loop *PNLoop = LI.getLoopFor(Parent);
5082  if (!PNLoop || Parent != PNLoop->getHeader()) {
5083  // Split the critical edge.
5084  BasicBlock *NewBB = nullptr;
5085  if (!Parent->isLandingPad()) {
5086  NewBB = SplitCriticalEdge(BB, Parent,
5088  .setMergeIdenticalEdges()
5089  .setDontDeleteUselessPHIs());
5090  } else {
5092  SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5093  NewBB = NewBBs[0];
5094  }
5095  // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5096  // phi predecessors are identical. The simple thing to do is skip
5097  // splitting in this case rather than complicate the API.
5098  if (NewBB) {
5099  // If PN is outside of the loop and BB is in the loop, we want to
5100  // move the block to be immediately before the PHI block, not
5101  // immediately after BB.
5102  if (L->contains(BB) && !L->contains(PN))
5103  NewBB->moveBefore(PN->getParent());
5104 
5105  // Splitting the edge can reduce the number of PHI entries we have.
5106  e = PN->getNumIncomingValues();
5107  BB = NewBB;
5108  i = PN->getBasicBlockIndex(BB);
5109  }
5110  }
5111  }
5112 
5113  std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5114  Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5115  if (!Pair.second)
5116  PN->setIncomingValue(i, Pair.first->second);
5117  else {
5118  Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5119  Rewriter, DeadInsts);
5120 
5121  // If this is reuse-by-noop-cast, insert the noop cast.
5122  Type *OpTy = LF.OperandValToReplace->getType();
5123  if (FullV->getType() != OpTy)
5124  FullV =
5126  OpTy, false),
5127  FullV, LF.OperandValToReplace->getType(),
5128  "tmp", BB->getTerminator());
5129 
5130  PN->setIncomingValue(i, FullV);
5131  Pair.first->second = FullV;
5132  }
5133  }
5134 }
5135 
5136 /// Emit instructions for the leading candidate expression for this LSRUse (this
5137 /// is called "expanding"), and update the UserInst to reference the newly
5138 /// expanded value.
5139 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5140  const Formula &F, SCEVExpander &Rewriter,
5141  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5142  // First, find an insertion point that dominates UserInst. For PHI nodes,
5143  // find the nearest block which dominates all the relevant uses.
5144  if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5145  RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5146  } else {
5147  Value *FullV =
5148  Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5149 
5150  // If this is reuse-by-noop-cast, insert the noop cast.
5151  Type *OpTy = LF.OperandValToReplace->getType();
5152  if (FullV->getType() != OpTy) {
5153  Instruction *Cast =
5154  CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5155  FullV, OpTy, "tmp", LF.UserInst);
5156  FullV = Cast;
5157  }
5158 
5159  // Update the user. ICmpZero is handled specially here (for now) because
5160  // Expand may have updated one of the operands of the icmp already, and
5161  // its new value may happen to be equal to LF.OperandValToReplace, in
5162  // which case doing replaceUsesOfWith leads to replacing both operands
5163  // with the same value. TODO: Reorganize this.
5164  if (LU.Kind == LSRUse::ICmpZero)
5165  LF.UserInst->setOperand(0, FullV);
5166  else
5167  LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5168  }
5169 
5170  DeadInsts.emplace_back(LF.OperandValToReplace);
5171 }
5172 
5173 /// Rewrite all the fixup locations with new values, following the chosen
5174 /// solution.
5175 void LSRInstance::ImplementSolution(
5176  const SmallVectorImpl<const Formula *> &Solution) {
5177  // Keep track of instructions we may have made dead, so that
5178  // we can remove them after we are done working.
5180 
5182  "lsr");
5183 #ifndef NDEBUG
5184  Rewriter.setDebugType(DEBUG_TYPE);
5185 #endif
5186  Rewriter.disableCanonicalMode();
5187  Rewriter.enableLSRMode();
5188  Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5189 
5190  // Mark phi nodes that terminate chains so the expander tries to reuse them.
5191  for (const IVChain &Chain : IVChainVec) {
5192  if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5193  Rewriter.setChainedPhi(PN);
5194  }
5195 
5196  // Expand the new value definitions and update the users.
5197  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5198  for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5199  Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5200  Changed = true;
5201  }
5202 
5203  for (const IVChain &Chain : IVChainVec) {
5204  GenerateIVChain(Chain, Rewriter, DeadInsts);
5205  Changed = true;
5206  }
5207  // Clean up after ourselves. This must be done before deleting any
5208  // instructions.
5209  Rewriter.clear();
5210 
5211  Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5212 }
5213 
5214 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5215  DominatorTree &DT, LoopInfo &LI,
5216  const TargetTransformInfo &TTI)
5217  : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L), Changed(false),
5218  IVIncInsertPos(nullptr) {
5219  // If LoopSimplify form is not available, stay out of trouble.
5220  if (!L->isLoopSimplifyForm())
5221  return;
5222 
5223  // If there's no interesting work to be done, bail early.
5224  if (IU.empty()) return;
5225 
5226  // If there's too much analysis to be done, bail early. We won't be able to
5227  // model the problem anyway.
5228  unsigned NumUsers = 0;
5229  for (const IVStrideUse &U : IU) {
5230  if (++NumUsers > MaxIVUsers) {
5231  (void)U;
5232  DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n");
5233  return;
5234  }
5235  // Bail out if we have a PHI on an EHPad that gets a value from a
5236  // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5237  // no good place to stick any instructions.
5238  if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5239  auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5240  if (isa<FuncletPadInst>(FirstNonPHI) ||
5241  isa<CatchSwitchInst>(FirstNonPHI))
5242  for (BasicBlock *PredBB : PN->blocks())
5243  if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5244  return;
5245  }
5246  }
5247 
5248 #ifndef NDEBUG
5249  // All dominating loops must have preheaders, or SCEVExpander may not be able
5250  // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5251  //
5252  // IVUsers analysis should only create users that are dominated by simple loop
5253  // headers. Since this loop should dominate all of its users, its user list
5254  // should be empty if this loop itself is not within a simple loop nest.
5255  for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5256  Rung; Rung = Rung->getIDom()) {
5257  BasicBlock *BB = Rung->getBlock();
5258  const Loop *DomLoop = LI.getLoopFor(BB);
5259  if (DomLoop && DomLoop->getHeader() == BB) {
5260  assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
5261  }
5262  }
5263 #endif // DEBUG
5264 
5265  DEBUG(dbgs() << "\nLSR on loop ";
5266  L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
5267  dbgs() << ":\n");
5268 
5269  // First, perform some low-level loop optimizations.
5270  OptimizeShadowIV();
5271  OptimizeLoopTermCond();
5272 
5273  // If loop preparation eliminates all interesting IV users, bail.
5274  if (IU.empty()) return;
5275 
5276  // Skip nested loops until we can model them better with formulae.
5277  if (!L->empty()) {
5278  DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
5279  return;
5280  }
5281 
5282  // Start collecting data and preparing for the solver.
5283  CollectChains();
5284  CollectInterestingTypesAndFactors();
5285  CollectFixupsAndInitialFormulae();
5286  CollectLoopInvariantFixupsAndFormulae();
5287 
5288  assert(!Uses.empty() && "IVUsers reported at least one use");
5289  DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
5290  print_uses(dbgs()));
5291 
5292  // Now use the reuse data to generate a bunch of interesting ways
5293  // to formulate the values needed for the uses.
5294  GenerateAllReuseFormulae();
5295 
5296  FilterOutUndesirableDedicatedRegisters();
5297  NarrowSearchSpaceUsingHeuristics();
5298 
5300  Solve(Solution);
5301 
5302  // Release memory that is no longer needed.
5303  Factors.clear();
5304  Types.clear();
5305  RegUses.clear();
5306 
5307  if (Solution.empty())
5308  return;
5309 
5310 #ifndef NDEBUG
5311  // Formulae should be legal.
5312  for (const LSRUse &LU : Uses) {
5313  for (const Formula &F : LU.Formulae)
5314  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
5315  F) && "Illegal formula generated!");
5316  };
5317 #endif
5318 
5319  // Now that we've decided what we want, make it so.
5320  ImplementSolution(Solution);
5321 }
5322 
5323 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5324  if (Factors.empty() && Types.empty()) return;
5325 
5326  OS << "LSR has identified the following interesting factors and types: ";
5327  bool First = true;
5328 
5329  for (int64_t Factor : Factors) {
5330  if (!First) OS << ", ";
5331  First = false;
5332  OS << '*' << Factor;
5333  }
5334 
5335  for (Type *Ty : Types) {
5336  if (!First) OS << ", ";
5337  First = false;
5338  OS << '(' << *Ty << ')';
5339  }
5340  OS << '\n';
5341 }
5342 
5343 void LSRInstance::print_fixups(raw_ostream &OS) const {
5344  OS << "LSR is examining the following fixup sites:\n";
5345  for (const LSRUse &LU : Uses)
5346  for (const LSRFixup &LF : LU.Fixups) {
5347  dbgs() << " ";
5348  LF.print(OS);
5349  OS << '\n';
5350  }
5351 }
5352 
5353 void LSRInstance::print_uses(raw_ostream &OS) const {
5354  OS << "LSR is examining the following uses:\n";
5355  for (const LSRUse &LU : Uses) {
5356  dbgs() << " ";
5357  LU.print(OS);
5358  OS << '\n';
5359  for (const Formula &F : LU.Formulae) {
5360  OS << " ";
5361  F.print(OS);
5362  OS << '\n';
5363  }
5364  }
5365 }
5366 
5367 void LSRInstance::print(raw_ostream &OS) const {
5368  print_factors_and_types(OS);
5369  print_fixups(OS);
5370  print_uses(OS);
5371 }
5372 
5373 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5374 LLVM_DUMP_METHOD void LSRInstance::dump() const {
5375  print(errs()); errs() << '\n';
5376 }
5377 #endif
5378 
5379 namespace {
5380 
5381 class LoopStrengthReduce : public LoopPass {
5382 public:
5383  static char ID; // Pass ID, replacement for typeid
5384 
5385  LoopStrengthReduce();
5386 
5387 private:
5388  bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5389  void getAnalysisUsage(AnalysisUsage &AU) const override;
5390 };
5391 
5392 } // end anonymous namespace
5393 
5394 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5396 }
5397 
5398 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5399  // We split critical edges, so we change the CFG. However, we do update
5400  // many analyses if they are around.
5402 
5410  // Requiring LoopSimplify a second time here prevents IVUsers from running
5411  // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5416 }
5417 
5419  DominatorTree &DT, LoopInfo &LI,
5420  const TargetTransformInfo &TTI) {
5421  bool Changed = false;
5422 
5423  // Run the main LSR transformation.
5424  Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5425 
5426  // Remove any extra phis created by processing inner loops.
5427  Changed |= DeleteDeadPHIs(L->getHeader());
5428  if (EnablePhiElim && L->isLoopSimplifyForm()) {
5430  const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5431  SCEVExpander Rewriter(SE, DL, "lsr");
5432 #ifndef NDEBUG
5433  Rewriter.setDebugType(DEBUG_TYPE);
5434 #endif
5435  unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5436  if (numFolded) {
5437  Changed = true;
5439  DeleteDeadPHIs(L->getHeader());
5440  }
5441  }
5442  return Changed;
5443 }
5444 
5445 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5446  if (skipLoop(L))
5447  return false;
5448 
5449  auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5450  auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5451  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5452  auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5453  const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5454  *L->getHeader()->getParent());
5455  return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5456 }
5457 
5460  LPMUpdater &) {
5461  if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5462  AR.DT, AR.LI, AR.TTI))
5463  return PreservedAnalyses::all();
5464 
5466 }
5467 
5468 char LoopStrengthReduce::ID = 0;
5469 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5470  "Loop Strength Reduction", false, false)
5476 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5477 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5478  "Loop Strength Reduction", false, false)
5479 
5480 Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }
Pass interface - Implemented by all &#39;passes&#39;.
Definition: Pass.h:81
static bool isValueValidForType(Type *Ty, uint64_t V)
This static method returns true if the type Ty is big enough to represent the value V...
Definition: Constants.cpp:1172
uint64_t CallInst * C
DomTreeNodeBase< NodeT > * getNode(NodeT *BB) const
getNode - return the (Post)DominatorTree node for the specified basic block.
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:243
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
iterator_range< use_iterator > uses()
Definition: Value.h:350
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
raw_ostream & errs()
This returns a reference to a raw_ostream for standard error.
BlockT * getLoopLatch() const
If there is a single latch block for this loop, return it.
Definition: LoopInfoImpl.h:149
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
PreservedAnalyses getLoopPassPreservedAnalyses()
Returns the minimum set of Analyses that all loop passes must preserve.
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:234
DiagnosticInfoOptimizationBase::Argument NV
typename SuperClass::const_iterator const_iterator
Definition: SmallVector.h:329
static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE)
Return true if this AddRec is already a phi in its loop.
Pass * createLoopStrengthReducePass()
This is a &#39;bitvector&#39; (really, a variable-sized bit array), optimized for the case when the array is ...
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
const SCEV * getConstant(ConstantInt *V)
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
bool isLegalAddImmediate(int64_t Imm) const
Return true if the specified immediate is legal add immediate, that is the target has add instruction...
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:449
size_type size() const
Determine the number of elements in the SetVector.
Definition: SetVector.h:78
This header provides classes for managing a pipeline of passes over loops in LLVM IR...
Type * getEffectiveSCEVType(Type *Ty) const
Return a type with the same bitwidth as the given type and which represents how SCEV will treat the g...
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
APInt sdiv(const APInt &RHS) const
Signed division function for APInt.
Definition: APInt.cpp:1590
an instruction that atomically checks whether a specified value is in a memory location, and, if it is, stores a new value there.
Definition: Instructions.h:514
const SCEV * normalizeForPostIncUse(const SCEV *S, const PostIncLoopSet &Loops, ScalarEvolution &SE)
Normalize S to be post-increment for all loops present in Loops.
static const size_t ComplexityLimit
Implements a dense probed hash-table based set.
Definition: DenseSet.h:221
unsigned getLoopDepth() const
Return the nesting level of this loop.
Definition: LoopInfo.h:96
The main scalar evolution driver.
bool isZero() const
Return true if the expression is a constant zero.
unsigned getNumberOfRegisters(bool Vector) const
BlockT * getLoopPreheader() const
If there is a preheader for this loop, return it.
Definition: LoopInfoImpl.h:101
static cl::opt< bool > InsnsCost("lsr-insns-cost", cl::Hidden, cl::init(false), cl::desc("Add instruction count to a LSR cost model"))
static PointerType * get(Type *ElementType, unsigned AddressSpace)
This constructs a pointer to an object of the specified type in a numbered address space...
Definition: Type.cpp:617
unsigned less than
Definition: InstrTypes.h:885
static const unsigned MaxIVUsers
MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for bail out...
bool properlyDominates(const SCEV *S, const BasicBlock *BB)
Return true if elements that makes up the given SCEV properly dominate the specified basic block...
void initializeLoopStrengthReducePass(PassRegistry &)
uint64_t getTypeSizeInBits(Type *Ty) const
Return the size in bits of the specified type, for which isSCEVable must return true.
void setDebugType(const char *s)
NodeT * findNearestCommonDominator(NodeT *A, NodeT *B) const
findNearestCommonDominator - Find nearest common dominator basic block for basic block A and B...
bool isLoopInvariant(const SCEV *S, const Loop *L)
Return true if the value of the given SCEV is unchanging in the specified loop.
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:818
void resize(unsigned N, bool t=false)
Grow or shrink the bitvector.
static void dump(StringRef Title, SpillInfo const &Spills)
Definition: CoroFrame.cpp:285
static bool isCanonical(const MDString *S)
bool properlyDominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
properlyDominates - Returns true iff A dominates B and A != B.
The adaptor from a function pass to a loop pass computes these analyses and makes them available to t...
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:503
An instruction for reading from memory.
Definition: Instructions.h:164
an instruction that atomically reads a memory location, combines it with another value, and then stores the result back.
Definition: Instructions.h:677
Value * getCondition() const
iv Induction Variable Users
Definition: IVUsers.cpp:51
void reserve(size_type N)
Definition: SmallVector.h:380
loop data prefetch
iterator end()
Get an iterator to the end of the SetVector.
Definition: SetVector.h:93
INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce", "Loop Strength Reduction", false, false) INITIALIZE_PASS_END(LoopStrengthReduce
This is the base class for unary cast operator classes.
A templated base class for SmallPtrSet which provides the typesafe interface that is common across al...
Definition: SmallPtrSet.h:345
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:176
#define DEBUG_TYPE
AnalysisUsage & addRequired()
const Module * getModule() const
Return the module owning the function this basic block belongs to, or nullptr it the function does no...
Definition: BasicBlock.cpp:116
bool isLegalICmpImmediate(int64_t Imm) const
Return true if the specified immediate is legal icmp immediate, that is the target has icmp instructi...
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
Hexagon Hardware Loops
This class represents the LLVM &#39;select&#39; instruction.
Predicate getInversePredicate() const
For example, EQ -> NE, UGT -> ULE, SLT -> SGE, OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc.
Definition: InstrTypes.h:958
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
Option class for critical edge splitting.
int getBasicBlockIndex(const BasicBlock *BB) const
Return the first index of the specified basic block in the value list for this PHI.
LoopT * getLoopFor(const BlockT *BB) const
Return the inner most loop that BB lives in.
Definition: LoopInfo.h:585
bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg, int64_t Scale, unsigned AddrSpace=0, Instruction *I=nullptr) const
Return true if the addressing mode represented by AM is legal for this target, for a load/store of th...
void clearPostInc()
Disable all post-inc expansion.
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
static User::op_iterator findIVOperand(User::op_iterator OI, User::op_iterator OE, Loop *L, ScalarEvolution &SE)
Helper for CollectChains that finds an IV operand (computed by an AddRec in this loop) within [OI...
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
LLVMContext & getContext() const
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:664
Reg
All possible values of the reg field in the ModR/M byte.
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:284
This node represents multiplication of some number of SCEVs.
ELFYAML::ELF_STO Other
Definition: ELFYAML.cpp:695
This file implements a class to represent arbitrary precision integral constant values and operations...
const APInt & getAPInt() const
BlockT * getHeader() const
Definition: LoopInfo.h:103
ConstantInt * getValue() const
auto reverse(ContainerTy &&C, typename std::enable_if< has_rbegin< ContainerTy >::value >::type *=nullptr) -> decltype(make_range(C.rbegin(), C.rend()))
Definition: STLExtras.h:250
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:201
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
#define F(x, y, z)
Definition: MD5.cpp:55
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
bool isTruncateFree(Type *Ty1, Type *Ty2) const
Return true if it&#39;s free to truncate a value of type Ty1 to type Ty2.